Policy Research Working Paper 10507 Violent Discipline and Parental Behavior Short- and Medium-term Effects of Virtual Parenting Support to Caregivers Lelys Dinarte-Diaz Saravana Ravindran Manisha Shah Shawn Powers Helen Baker-Henningham Development Economics Development Research Group June 2023 Policy Research Working Paper 10507 Abstract Approximately 75% of children aged 2 to 4 worldwide against children. Treatment children also experience fewer are regularly subjected to violent discipline across the emotional problems (0.17 SD). Medium-term results (nine globe. This paper studies the impact of a virtually-deliv- months later) show reductions in caregiver depression (0.12 ered intervention on positive parenting practices in Jamaica. SD), anxiety (0.16 SD), and parental stress (0.16 SD) for Short-term results indicate that the intervention improves treatment caregivers. The virtual delivery has important caregiver knowledge (0.52 SD) and attitudes around vio- scalable policy implications which could help decrease vio- lence (0.2 SD) and leads to meaningful changes in caregiver lence against children across the globe. disciplining behaviors, with a 0.12 SD reduction in violence This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at ldinartediaz@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Violent Discipline and Parental Behavior: Short- and Medium-term Effects of Virtual Parenting Support to Caregivers* Lelys Dinarte-Diaz† Saravana Ravindran‡ Manisha Shah§ Shawn Powers¶ Helen Baker-Henningham|| Keywords: Child maltreatment, violence against children, positive parenting, digital de- livery JEL Classification: I24, J12, J13, J16 * We are thankful to Anna Aizer, S Anukriti, Kathleen Beegle, Janet Currie, Andrew Dustan, Bilge Erten, Emanuela Galasso, Melanie Guldi, and Jevay Grooms for their useful comments and feedback. We also appreciate comments from participants at the 2023 NBER Childrens Meeting, 2023 EDW, 2023 WECJr, PacDev 2023, 1st WAP Workshop, 2023 LACEA IEN Meeting, Atlanta Workshop on Public Policy and Child Well-Being 2023, Development Research Group (World Bank), IFPRI, Exeter, University of Western Australia, NUS, PUC Chile, Universidad Los Andes, and Universidad Andres Bello. Carlos Guzman, Gabriela Lecaro, and Nathy Andrade provided superb research assistance. We are truly grateful for the support of the Early Childhood Commission (ECC) for implementing the intervention. We thank Taja Francis for supporting the development of the virtual Irie Homes Toolbox and the training of officers. This work was supported by the World Bank SIEF and ELP Trust Funds and the Research Support Budget. Ravindran is funded by a startup grant at the Lee Kuan Yew School of Public Policy, National University of Singapore. This research project’s protocol was reviewed and approved by the UWI Institutional Review Board in Jamaica (approval ID No. CREC-MN.86, 20-21) and by the IPA Research Ethics Committee (approval ID No. 15586). The trial was registered at the American Economic Association RCT registry (AEARCTR- 0008266). The authors have no conflicts of interest to report. The findings, interpretations, and conclusions expressed in this report are entirely those of the authors. They do not necessarily represent the views of the World Bank and its affiliated organizations, its Executive Directors, or the governments they represent. † Development Research Group. The World Bank. Email: ldinartediaz@worldbank.org ‡ Lee Kuan Yew School of Public Policy, National University of Singapore. Email: saravana@nus.edu.sg § Luskin School of Public Affairs, UCLA, and NBER. Email: manishashah@ucla.edu ¶ The World Bank. Email: spowers@worldbank.org || University of West Indies and Bangor University. Email: h.henningham@bangor.ac.uk 1 Introduction Approximately 75% of children aged 2 to 4 worldwide—close to 300 million—are regu- larly subjected to violent discipline (physical punishment and/or psychological aggres- sion) by parents and/or caregivers at home (UNICEF, 2017).1 Such exposure to violence can hinder children’s development and undermine their sense of self-worth (Boden et al., 2007; Fry et al., 2018; Mersky and Topitzes, 2010). Moreover, research shows children who are victims of abuse and neglect are more likely to exhibit risky behaviors as teenagers (Hamby et al., 2011); be absent from school, and have higher levels of aggression, mental distress, and social problems (Lansford et al., 2002). These children can negatively affect peers’ test scores and behavior in the classroom (Carrell and Hoekstra, 2010). Early expo- sure to violence also negatively affects outcomes during adulthood (Doyle Jr and Aizer, 2018), resulting in worse labor market outcomes (Currie and Spatz Widom, 2010) and/or increased involvement in crime (Currie and Tekin, 2012; Sviatschi, 2022). In this paper we test whether a virtually delivered, scalable intervention on positive parenting practices can improve caregivers attitudes and behaviors related to violence against children in Jamaica, where 82% of children aged 2-14 are regularly subjected to violent discipline at home (UNICEF, 2022). Within Latin America, Jamaica ranks third highest in terms of violence against children (Haiti and Suriname rank first and second) (UNICEF, 2022). We adapt an in-person Jamaican behavior change violence-prevention parenting program—the Irie Homes Toolbox (IHT)—for virtual delivery. The intervention is targeted at parents of 2-to-6-year-old children (we refer to the eldest child aged 2 to 6 in each household as the “target child”). The virtual IHT is a 10-week program that includes content related to four core concepts: i) building positive relationships between parent and child, ii) preventing misbehavior, iii) managing misbehavior, and iv) supporting chil- dren’s emotional self-regulation (Francis and Baker-Henningham, 2020). The content is delivered through three components: First, participants receive three SMS messages per 1 This exposure to violence can start even earlier. According to data from 30 countries, nearly half of children aged 12 to 23 months are subjected to corporal punishment and a similar proportion are exposed to verbal abuse at home (UNICEF, 2017). 1 week (for a total of 30 SMS messages). Each SMS message briefly describes the techniques caregivers should practice from each of the four core elements of the intervention. Second, caregivers receive access to a data-free app with weekly content (videos and other related materials). Third, participants were offered the opportunity to join weekly, one-hour, vir- tual group parenting sessions led by field officers of the Early Childhood Commission in Jamaica (ECC). To measure the impacts of the virtual program, we collected three rounds of data: be- fore the start of the intervention (baseline), right after its completion (short-term follow- up), and 9 months later (medium-term follow-up). We collected data on caregivers’ atti- tudes and perpetration of violence against children, caregivers’ mental distress, parental self-efficacy, and changes in caregiver support networks. We also collected data on con- duct and emotional problems of their children as well as demographic characteristics of other household members. We complement this household survey data with administra- tive data from TrendMedia on SMS messages sent and App usage, reports on attendance at the virtual sessions provided by ECC field officers, and focus group discussions with caregivers. Intent-to-treat estimates show that the intervention led to significant improvements in caregivers’ attitudes toward violence against children (VAC). Caregivers in the treatment group improved attitudes by 0.2 SD in the VAC index (p < 0.01) in the short run. This is driven by a 0.19 SD improvement in the physical VAC sub-index (p < 0.01) and a 0.12 SD improvement in the psychological VAC sub-index (p < 0.05). These effects are persistent, with caregivers in the treatment group improving attitudes by 0.14 SD on the VAC index (p < 0.05) in the medium-term follow-up. Consistent with improvements in attitudes toward VAC, caregivers in the treatment group also changed their child disciplining behaviors. Caregivers in the treatment group reduced violence against the target child (VATC) by 0.12 SD (p < 0.05). We construct sub-indices for the physical and psychological VATC and find 0.14 SD (p < 0.05) and 0.1 SD (p < 0.1) reductions in these indices in the short term. These effects persist in the medium-term, where we estimate a 0.13 SD reduction in VATC (p < 0.05). Furthermore, 2 while we do not observe statistically significant impacts on caregivers’ mental well-being at the first follow-up, we show reductions in caregiver depression (0.12 SD, p < 0.1), anxiety (0.16 SD, p < 0.05), and parental stress (0.16 SD, p < 0.05) at the second follow- up. Consistent with improvements in caregiver attitudes and disciplining practices, we find that target children in the treatment group scored 0.17 SD lower on the index of the emotional problem in the short term (p < 0.01). We introduce a simple conceptual framework to understand the results and poten- tial mechanisms. We hypothesize that the intervention may have improved caregivers (i) knowledge about positive parenting practices, (ii) self-efficacy, and/or (iii) support net- works. We show that caregivers in the treatment group learned from the material and scored 0.53 SD higher (p < 0.01) on the information module relative to caregivers in the control group at the first follow-up. The treatment effects on caregiver knowledge are persistent: the treatment group scored 0.39 SD higher (p < 0.01) on the information index relative to caregivers in the control group in the second follow-up. Administrative and survey data show that 92% of weekly messages were sent and successfully received by treated caregivers. Ninety-seven percent of caregivers reported they read the messages, and 98% of the caregivers who read the messages found them to be useful. We also show positive impacts of treatment on caregivers’ self-efficacy at the second follow-up (0.21 SD, p < 0.01). However, it appears that caregiver parenting support and/or borrowing networks are not relevant in explaining the improvements. The posi- tive treatment effects and analysis of mechanisms suggest that the intervention provided caregivers with the necessary parenting tools and boosted their confidence in their own parenting skills. We examine the robustness of the results in the following ways. First, we address po- tential concerns related to experimenter demand effects since many of the outcomes are self-reports. Following Asadullah et al. (2021) and Dhar et al. (2022), we test whether the treatment had an effect on the caregivers’ social desirability index.2 We do not find 2 This index captures the study participant’s individual-level propensity to misreport sensitive items, which indicates whether or not the respondent is driven by the need for social approval. 3 evidence that desirability bias changed among treated caregivers relative to those in the comparison group. In additional robustness checks, we include the social desirability in- dex as a control variable and this does not impact the ITT estimates or standard errors. Second, we examine whether the intervention displaced violence from the target child to other children in the household. We find the opposite. Including data on the eldest child aged 7-12 in the household, we find 0.14 SD and 0.15 SD reductions in the index for vio- lence against these children in the short- and medium-term (p < 0.01). Third, we explore issues related to selective attrition and show that our results are robust to attrition. Fourth, we verify the robustness of our results to the inclusion of additional control variables as selected by a double LASSO algorithm. Our results remain similar in terms of magnitude and inference. Lastly, we take a more agnostic approach to the structure of the standard errors and estimate them using a randomization inference approach. We find that the magnitude of the p-values is similar to the magnitudes of p-values obtained by estimating heteroskedasticity-robust standard errors. Our paper makes several important contributions to the literature. First, as far as we know, this is the first causal study of a virtually-delivered intervention to caregivers with the primary goal of reducing violence against children. The majority of previous early childhood interventions are more costly in-person interventions that focus on nutrition as well as cognitive and socio-emotional development of children (Carneiro et al., 2023; Bos et al., 2022; Mehrin et al., 2022; Chandra et al., 2021; Heckman et al., 2020; Attanasio et al., 2020; Hamadani et al., 2019; Levere et al., 2016; Attanasio et al., 2014; Macours et al., 2012; Paxson and Schady, 2010) and few in-person studies target violence against children as the primary outcome (de Simone et al., 2022; Francis and Baker-Henningham, 2021; Lachman et al., 2020; Cluver et al., 2018; Altafim and Linhares, 2016; Knerr et al., 2013). Second, we contribute to a small and nascent literature on digital and low-cost parent- ing programs, since most parenting programs occur in person. The few interventions that have been virtually-delivered (see, for example, Smith et al. (2023); Amaral et al. (2021); Barrera et al. (2020); Widen et al. (2020); York et al. (2019); Cortes et al. (2019); Doss et al. (2019); Hurwitz et al. (2015)) do not target violence against children as the main outcome. In low-income settings, the widespread availability of mobile phones, the high incidence 4 of violence against children, and the general social acceptance of this issue make our inter- vention potentially attractive, but the effectiveness of this type of digital programming is not yet well understood. Lastly, this study contributes to questions of scalability by testing an alternative and low-cost delivery mode of a parenting intervention. Overall, the high prevalence of child maltreatment and its negative long-term impacts call for innovative and effective strategies to address this global problem. 2 Intervention and Conceptual Framework 2.1 Intervention The Irie Homes Toolbox (IHT) consists of a behavior change violence prevention pro- gram targeting parents of 2-to-6-year-old children. The content of the program is based on evidenced-informed parenting practices that improve parenting behavior and reduce child behavior problems (Chorpita and Daleiden, 2009; Garland et al., 2008; Michie et al., 2011). We generated a digitally adapted version of the Irie Homes Toolbox (vIHT) for this study. The vIHT intervention includes content related to four key concepts: i) building posi- tive relationships between parent and child (e.g. praise, child-led play, involving the child in everyday activities), ii) preventing misbehavior (e.g. understanding why children mis- behave, giving children independence and autonomy, giving clear instructions, setting rules and expectations, modeling appropriate behavior), iii) managing misbehavior (e.g. redirecting child, withdrawing attention, setting limits, giving appropriate consequences) and iv) child emotional regulation and stress reduction techniques (Francis and Baker- Henningham, 2020).3 The vIHT content was delivered over a 10-week period via smartphones. First, ben- 3 The in person version of the program consists of an eight-week, group-based, parenting intervention, delivered through community preschools. Francis and Baker-Henningham (2021) show that the in person program reduced parents’ use of harsh punishment by 0.29 SD, increased caregiver involvement with their child by 0.30 SD, and reduced child behavior difficulties among children at or above the 50th percentile on initial behavior difficulties by 0.36 SD. 5 eficiaries received three SMS messages per week (30 SMS messages in total) relating to content from the Irie Homes Toolbox.4 Each SMS briefly describes the techniques par- ents should practice during the week. It also includes a link to the content embedded in the program App and information on the Irie Challenge for the week. The Irie Chal- lenge consists of suggestions on how to put the strategies learned from the program into practice with children. Second, caregivers received access to a data-free App. This pro- gram App included videos of parents utilizing the strategies with their children, the Irie Tower (which depicts the list of all strategies thought during the program), and the Irie Challenge. The content was uploaded and updated every week during the 10 weeks.5 Lastly, caregivers were offered the opportunity to join a virtual parent group that met once per week for ten weeks through GoogleMeet video calls to discuss the specific topic of the week with an early childhood education specialist. We partnered with the Early Childhood Commission (ECC) in Jamaica for the implementation of this third component. ECC officers were trained in the curriculum and its implementation by the Irie Toolbox Team based in the Caribbean Institute for Health Research. The groups included 8 to 9 participants and were formed randomly (see more details in the research design sec- tion). Participant caregivers received data packages that allowed them to join the virtual groups. During these sessions, the messages received by text and video were reinforced through discussion and practice. Moreover, ECC officers sent e-summaries of the session via WhatsApp at the end of the session. 2.2 Conceptual Framework Our analysis is organized around the conceptual framework presented in Figure 1 which was also pre-registered at the AEA. This framework approaches the incipient problem of high levels of violence against children in households in low- and middle-income coun- tries. We expect the intervention to potentially impact at least three potential primary out- 4 See the content delivered during each weekly session in Table A1 in the appendix. The complete list of SMS messages can be found in Table A2 in the Appendix. 5 See Figure A1 in the Appendix with examples of materials corresponding to Week 4. 6 comes. First, caregivers could learn about positive parenting practices and violence pre- vention. This could then impact their attitudes and behaviors related to violence against children. If we observe changes on these margins, then secondary outcomes of interest will be related to caregiver mental health and/or children’s emotional/behavioral prob- lems (Chorpita and Daleiden, 2009; Garland et al., 2008; Francis and Baker-Henningham, 2021). In terms of potential change pathways or mechanisms besides increased parental knowl- edge on violence prevention parenting, parental self-efficacy, and/or a parent’s belief in their ability to perform the role of parent successfully could be important mechanisms (Wittkowski et al., 2017). Evidence shows that caregiver’s perceived support is associated with parental self-efficacy (Fang et al., 2021). Since the intervention aims to help care- givers through the provision of information and skills training on non-violent practices, this sense of support could assist them in increasing their parental self-efficacy. More- over, considering the strong association between parental stress, maternal depression, and parental self-efficacy (Fang et al., 2021), the intervention could indirectly improve caregiver’s mental health. Lastly, social networks and the support that network members provide could be es- sential resources for caregivers in sustaining their caregiving role. By joining the virtual meetings, caregivers could expand their support networks through their interactions with other caregivers. For example, they could share information, knowledge, and experience about improving attitudes and violent behaviors toward their children. Information trans- mission has been shown to be a mechanism operating in programs that exploit variation in peers and networks (Dahl et al., 2014). Moreover, an improvement in caregivers’ net- works can also indirectly improve their mental health. As evidence shows, caregivers’ perceptions of support networks, including family and friends, have been linked to their health status (Owen and Anderson, 2017; Balaji et al., 2007). 7 3 Experimental Design and Data 3.1 Participant Recruitment and Enrollment We mainly recruited participants via SMS messages to Digicel’s customers. Digicel is a mobile phone network provider headquartered in Jamaica, and is the main provider of mobile phone services in the country. As shown in Table A3, 93.4% of participants were recruited via this channel. We recruited the rest of the participants through the ECC and preschool principals and/or via social media. We partnered with TrendMedia,6 a subsidiary of Digicel Group, to send every participant a link to an enrollment survey through these three channels. The enrollment survey included the following eligibility criteria: caregivers had to 1) live in the same house with at least one child aged 2 to 6 years old, 2) have access to a smartphone or tablet, and 3) provide consent to participate in the intervention and study. We also asked about gender and parish of residence7 in the enrollment survey, but these two variables were not part of the eligibility criteria.8 A potential concern with access to a smartphone or tablet as eligibility criteria is that it might shrink the potential sample of participants. However, mobile subscription is large in Jamaica. Data from the Office of Utilities Regulation indicates that mobile coverage in Jamaica was close to 100% in 2021 and 70.3% of the population had access to smartphones (Operator Watch, 2021).9 Digicel’s SMS messages did not target particular parishes. Figure A5 in the Appendix shows a parish-wise comparison of the distribution of our sample relative to the distribu- tion of the population of caregivers in Jamaica. Despite a few exceptions in Kingston and Saint Andrew and Saint Catherine, the proportions are comparable. As Figure 2 shows, we identified a total of 1,993 eligible caregivers based on the criteria 6 TrendMedia offers business solutions based on information and communication technologies, to corpo- rate clients, SMEs, and governments. 7 Parishes are the primary unit of local government in Jamaica. 8 See SMS messages Snapshot in Figure A2 and snapshots of the full survey in Figures A3 and A4. 9 See more details in https://our.org.jm/sectors/telecommunications/telecommunications-market- information-data/ 8 mentioned above. In August 2021, we contacted all the 1,993 caregivers and collected baseline data from 1,113 individuals distributed across 14 parishes. The remaining 880 individuals did not complete the baseline survey for several reasons, including they did not provide a correct phone number (14%); we were directed to voicemail (32%); they changed their mind and decided not to participate in the study (24%); etc. 3.2 Randomization We randomly assigned the sample of 1,113 enrolled caregivers to either the treatment or the control group with equal probability. Our sample includes 557 caregivers in the treat- ment group and 556 caregivers in the control group. Caregivers in the control group also received three SMS messages per week (30 SMS messages in total) with content related to good practices to avoid COVID-19.10 Randomization was stratified by gender of the caregiver (male or female) and the mode of recruitment into the study (SMS messages campaign or ECC/Principal referral and social media campaign). Appendix Table A3 indicates the size of each stratum. 3.3 Data Collection We collected data at baseline (before the intervention) in August 2021 and conducted two follow-up rounds (one right after the intervention ended and another one 9 months later). We administered phone-based surveys in all data collection rounds. We piloted the survey instrument before conducting the baseline (between May and June 2021). The intervention was completed at the end of November 2021, and we collected the first round of follow-up data in December 2021 from 985 caregivers (an 88.5% response rate). We timed the first follow-up survey to test the short-term effects of the program and to minimize attrition. To measure the medium-term effects of the vIHT, we collected a second round of follow-up data in August-September 2022. We were able to collect data from 705 caregivers (63.3% response rate). All data collected during the three rounds are self-reported. Participants 10 See the complete list of SMS messages sent to the control group in Table A4 in the Appendix. 9 received a small monetary incentive (US$2.50) to complete each of the three surveys. We also conducted focus group discussions with 43 study participants in April and May 2023. A detailed description of this qualitative study is presented in Appendix A2. A. SMS viewership, attendance, App use, and learning We collected administrative data on SMS delivery and whether caregivers logged in to the App and the time (in minutes) they used the App from TrendMedia. ECC officers collected attendance data for the virtual groups for each caregiver. In addition, we collected self- reports from individuals on whether they read the text messages, how useful they found them, etc. We also asked questions to measure whether caregivers learned some concepts and practices that were taught in the program in the two follow-up rounds to measure learning from the information provided. B. Main outcomes11 Attitudes toward violence against children: We use an adapted version of the UNICEF Multi- ple Indicator Cluster Survey (MICS) questionnaire (UNICEF, 2011) to measure parental attitudes toward physical and psychological violence against children at baseline and follow-up rounds. The adapted instrument includes 5 items asking about some attitudes such as if they agree that a good parent can slap the child if he misbehaves, if shouting and yelling would make the child more obedient, etc. Due to time constraints, we selected 5 out of 13 items with the greatest variation based on the results from the survey instrument pilot that was conducted prior to baseline data collection. Violence against children: We use the UNICEF MICS questionnaire to measure caregivers’ perpetration of physical or psychological violence against children. The adapted instru- ment includes 5 items asking about physical violence (hitting the child with a bare hand or with an object) and/or psychological violence (shouting, yelling, or screaming at the child; saying to send the child away; threatening to hit the child). We asked each caregiver about perpetrating these violent acts to the “target child” or to another older child within 11 All outcomes were registered at the American Economic Association RCT registry - AEARCTR-0008266. A detailed description of the items included in the estimation of the outcomes and other variables is pre- sented in Appendix A1. 10 the household (eldest child between 7 to 12 years old). Using these reports, we created two indices: violence against the target child and violence against any child within the household. C. Secondary Outcomes Caregiver’s mental health: We administered the Patient Health Questionnaire (PHQ-2 sur- vey, Kroenke et al. (2003)) and a question on having difficulty sleeping at night. We mea- sure anxiety using the Generalized Anxiety Disorder (GAD-2, Donker et al. (2011)) instru- ment. We included these instruments during all three rounds of data collection. More- over, during the second follow-up, we also included questions from the 18-items Parental Stress Scale (PSS-18, Berry and Jones (1995)). Our main outcomes of interest are indices of depression, anxiety, and stress. Child conduct and emotional problems: We use 10 items related to children’s conduct and emotional problems (5 items each) from the Strengths and Difficulties Questionnaire (SDQ) instrument to measure a child’s behavior. We collected this information in each of the sur- vey rounds. Each question is answered on a 0-2 scale (Not true, somewhat true, certainly true). D. Potential Mechanisms Parental self-efficacy: We measure parental self-efficacy at baseline and during the first follow-up round using the 5 items from the Brief Parental Self Efficacy Scale (BPSES) in- strument. The scale asks parents about their agreement with statements that can describe their ability to improve a child’s behavior. For the second follow-up, we adapted the Tool to Measure Parenting Self-Efficacy (TOPSE) for more detailed questions relating to disci- pline and self-acceptance. Caregiver’s support networks: The effectiveness of positive parenting programs can be driven by support networks for participant caregivers. To test this potential mechanism, we col- lect information on whether caregivers obtained support from friends, family, or profes- sionals to solve parenting or financial issues. We asked how many people they could reach out to in case they need to talk about issues related to parenting and child rearing or borrowing money. 11 3.4 Baseline Summary Statistics Table 1 reports summary statistics of the sample. On average, control group caregivers are 33 years old, have 14 years of education, and 85% of them are female. Furthermore, 38% of the caregivers are married, 79% reported being employed and the average income in the past month was USD 910. With respect to household characteristics, the average household has 4.6 members, with approximately 2 of them being children under 17. Target children are, on average, 4.1 years old and are gender-balanced (48% are female). Caregivers exhibit relatively high support for violence against children, with 37% agree- ing to the statement: “Shouting, yelling, and threatening to slap will not harm the child.” Furthermore, the average caregiver draws on harsh conduct to discipline their children 2 to 6 years old approximately 1.5 days per week (“Shouted, yelled, or screamed at him/her”). For child behavior-related outcomes, we observe that 41.7% of children exhibit conduct problems, while 24.7% display emotional problems. Average prevalence rates of depres- sion and generalized anxiety disorder are 20% and 13%, respectively. How do caregivers in our sample compare to the representative caregiver of a child aged 2-6 in Jamaica? We compare our sample of 1,113 caregivers with the Jamaica Survey of Living Conditions (JSLC) 2019. The JLSC is a living standards measurement survey that is representative of the Jamaican population. We restrict the JSLC sample to only caregivers with a child aged 2-6 and compare key demographic variables for which data is available across surveys in Appendix Table A6. The last column of the table provides p-values for the comparison of means across the two samples. We find that children across the samples are quite similar in terms of age and gender. However caregivers in our sam- ple are slightly younger, have one more year of education and are more likely to be mar- ried and employed. Caregivers in our sample seem to have slightly better socioeconomic status than the average caregiver in Jamaica. We also address the potential external validity concern that our results may be spe- cific to the COVID-19 pandemic period during which the intervention was delivered in Jamaica. To better understand counterfactual rates of violent disciplining by caregivers and mental health absent the intervention, we present summary statistics of these vari- 12 ables for control group caregivers for each of the three rounds of surveys in Table A7. The rates of violent disciplining and responses to questions regarding depression and anxiety are very similar across baseline, first, and second follow-up. Importantly, this time pe- riod spans August 2021 to September 2022, thereby covering a significant period after the relaxation of all COVID restrictions in Jamaica in March 2022. Table A7 shows that the counterfactual rates of violent discipline and mental health are stable across the various time periods, and we do not observe any meaningful changes in counterfactual trends post-COVID restrictions. 3.5 Baseline Balance We test for balance in pre-intervention outcomes and socio-demographics between treat- ment and control groups. These results are presented in Table 1. Columns 2 and 5 present sample means for each variable (control and treatment) and columns 3 and 6 present the standard deviations (SD). Column 7 provides p-values for t-tests for equality of means between the treatment and control groups. With the exception of two out of 27 variables tested, we do not find significant differences in these variables across treatment and con- trol groups at p-values less than 0.1. The only two differences we find are that caregivers in the treatment group told their children they would send them away less in the past seven days (0.14 days in the treatment group versus 0.23 days in the control group; p = 0.023), and were more likely to suffer from generalized anxiety disorder at baseline (17.1% in the treatment group versus 13.3% in the control group; p = 0.084). The p-value for the overall F-test of joint orthogonality is 0.968, highlighting that jointly, the means of the variables are not statistically distinguishable across the treatment and control groups. The random- ization produced comparable treatment and control groups. 13 4 Estimation Framework 4.1 Empirical methods To study intent-to-treat (ITT) impacts, we estimate Ordinary Least Squares (OLS) regres- sions using the following Analysis of Covariance (ANCOVA) specification for caregiver i in period t and stratum s: Yi,t = β0 + β1 Ti + β2 Yi,t−1 + γs + εi,t (1) where Yi,t refers to the outcome variable of interest of caregiver i as measured at first or second follow-up, defined in Section 3.3. Ti is an indicator variable capturing the assign- ment of caregiver i to the treatment group, and Yi,t−1 refers to the outcome variable of interest measured at baseline. γs captures stratum fixed effects for the four strata; the interaction between the gender of the caregiver (female or male) and the two modes of re- cruitment (SMS messages or social media). The estimate of β1 captures the ITT effect of the treatment. We estimate and report heteroskedasticity-robust standard errors. As a robust- ness check, we follow a more agnostic approach to the structure of the standard errors (or a potential fuzzy clustering) and estimate randomization inference standard errors. Ran- domization inference gives us precise p-values based on the empirical distribution of all estimated treatment effects that could arise under our design and data (after randomly reassigning the treatment status 1,000 times) under the null hypothesis of no effect for any unit. To address potential concerns relating to multiple hypothesis testing, we construct in- dices for broad families of outcomes using Anderson (2008). Summary index tests offer three advantages: (i) they are robust to over-testing because each index represents a single test, (ii) they provide a statistical test for whether a program has a "general effect" on a set of outcomes, and (iii) they are potentially more powerful than individual-level tests by reducing random error in each outcome measure (Anderson, 2008). Each summary index is a weighted mean of several standardized outcomes, where the weights are calculated 14 to maximize the amount of information captured in the index using an efficient general- ized least squares (GLS) estimator.12 We orient caregivers and child outcomes such that a reduction in the index is always an improvement in the outcome of interest. We also conduct Least Absolute Shrinkage and Selection Operator (LASSO) analysis to identify variables with strong relationships with Yi,t , to assess their suitability for in- clusion as controls in Equation (1) following Bruhn and McKenzie (2009). Since LASSO consistently only selected the outcome variable of interest measured at baseline (Yi,t−1 ) for inclusion across all outcomes, we do not include other control variables in Equation (1).13 5 Short- and Medium-term Results We use the Conceptual Framework and pre-analysis plan to guide the empirical work. We first test whether treatment was effective in improving parents knowledge of positive parenting practices related to violence against children. If treatment is effective, we might expect changes in caregiver attitudes and behaviors related to VAC. We also investigate impacts on secondary outcomes related to caregiver mental health and child emotional problems (in Sections 5.4 and 5.5, respectively). Later we discuss the role of potential mechanisms. 5.1 ITT Impact on Violence Prevention Knowledge To assess whether the vIHT content increased caregiver knowledge on positive parenting practices, we administered an information module “test” to all caregivers at endline. Table 2 presents ITT impacts of the treatment on caregiver knowledge. Columns 1 to 8 present the treatment impacts on eight statements relating to parenting practices. The eight state- ments were designed to assess understanding of the four key concepts of the intervention 12 We show robustness of our results to the use of unweighted summary indices following Kling et al. (2007) in Section 5. 13 Our main specification only includes the corresponding outcome variable of interest measured at base- line as a control, since it is selected consistently across all outcome variables. We assess robustness to the inclusion of other controls selected only for some outcomes in Section 7. 15 as outlined in Section 2 (two statements were asked for each of the four concepts). All statements are true; thus higher values represent greater knowledge. Panel A presents the short-term results from the first follow-up conducted immediately after the intervention ended. We observe that caregivers in the treatment group are significantly more likely to state that the statements are true relative to caregivers in the control group for six out of the eight statements (four statements significant at the 1% level, one statement significant at the 5% level, and one significant at the 10% level). To address potential concerns re- lating to multiple hypothesis testing, we aggregate the responses to the eight statements into one information index as outlined in Section 4. Column 9 of Table 2 shows that care- givers in the treatment group score 0.53 SD higher on the information module relative to caregivers in the control group. The impact is large in magnitude and statistically signifi- cant at the 1% level. Panel B presents the medium-term results from the second follow-up conducted nine months after the intervention ended. We see persistent treatment impacts on caregiver knowledge: the treatment group scored 0.39 SD higher on the information index relative to caregivers in the control group (p < 0.01). Table A8 (panel D) shows that the medium-term impact on the information index is not statistically different from the short-term impact. 5.2 ITT Impacts on Attitudes Toward Violence Against Children Next we investigate the impact of treatment on caregiver attitudes toward violence against children (VAC). Figure 3 and column (1) of Table 3 (Panel A) show that caregivers in the treatment group improved their attitudes on VAC by 0.2 SD in the short-run (VAC index, p < 0.01) in the short-run. Panel A of Table A9 presents results from the first follow-up and breaks down the attitudes toward VAC into five individual components. The first three columns constitute the attitudes toward physical VAC, while columns 4 and 5 comprise attitudes toward psychological VAC. In the short term, caregivers in the treatment group are 3.9 percentage points less likely to agree with the statement that children need to be physically punished in order to bring up, raise, or educate a child properly (this is 32% reduction from the control mean). They are 6.2 percentage points (27%) less likely to agree 16 that a good parent slaps their child when they misbehave and 5.1 percentage points (53%) less likely to agree that when a child is beaten, he/she will stop doing the unwanted behavior (p < 0.01). The sub-index for attitudes toward physical VAC is shown in Figure 3 and column (2) of Table 3 (Panel A), and we estimate a 0.19 SD improvement in this sub-index (p < 0.01). In terms of short-run changes in attitudes toward psychological VAC, caregivers in the treatment group were 2.3 percentage points less likely to agree that shouting and yelling makes the child more obedient and 2.7 percentage points more likely to disagree with the statement that "shouting, yelling, and threatening to slap will harm the child," although these results were not statistically significant at conventional levels. Figure 3 also presents the treatment impact on the sub-index for attitudes toward psychological VAC. As shown in column (3) of Table 3 (Panel A), at first follow-up, caregivers in the treatment group improve attitudes toward psychological VAC by 0.12 SD (p < 0.05). These changes in attitudes persist into the medium-term. Figure 3 and column (1) of Table 3 (Panel B) show that caregivers in the treatment group improved attitudes toward violence against children by 0.14 SD (p < 0.05) in the medium-run. Analyzing the sub- indices for attitudes toward physical and psychological VAC, columns (2) and (3) of Table 3 (Panel B) show persistent impacts on attitudes toward physical VAC (0.15 SD reduction, p < 0.05), although the changes in attitudes toward psychological VAC are no longer statistically significant. Panels A and B of Table A8 shows that the medium-term impacts are not statistically different from the short-term impacts. Columns (1)-(3) of Table A10 show that the results are very similar when we use indices constructed following Kling et al. (2007). 5.3 ITT Impacts on Violent Behaviors Against Children The results in Figure 4 and column (4) of Table 3 (Panel A) highlight that caregivers in the treatment group reduced violence against the target child (VATC) by 0.12 SD (p < 0.05). The figure also highlights the treatment impacts on the sub-indices for physical and psychological VATC. We estimate a 0.14 SD (p < 0.05) and 0.10 SD (p < 0.1) reduction in 17 these indices (columns 5-6 of Table 3, Panel A).14 The larger treatment impact on physical versus psychological VATC is in line with the larger treatment effects for the attitudes toward physical versus psychological VAC. These results persist nine months after the end of the intervention: column (4) of Table 3 (Panel B) highlight that caregivers in the treatment group scored 0.13 SD lower at the second follow-up on the violence against target child (VATC) index (p < 0.05). Columns (5)-(6) of Table 3 (Panel B) show that we estimate statistically significant reductions in physical VATC (0.12 SD reduction, p < 0.1), but do not find statistically significant impacts on psychological VATC. Panels A and B of Table A8 show that the medium-term impacts are not statistically different from the short-term impacts. Columns (4)-(6) of Table A10 show that the results from the short and medium term are very similar when we use indices constructed following Kling et al. (2007). Panel B of Table A9 presents the short-term impacts for the five components that com- prise the VATC index. The first two columns are components of the physical VATC index, while columns 3, 4, and 5 comprise components of the psychological VATC index. Care- givers in the treatment group are 9.7 percentage points (25%) less likely to hit their child on the bottom, hand, arm, or leg with their bare hand (p < 0.01). In terms of psychological VATC, caregivers in the treatment group are 10.1 percentage points (14%) less likely to shout, yell, or scream at their child (p < 0.01) and 5.5 percentage points (8%) less likely to threaten to hit their child without doing so (p < 0.1). However, they are 4 percentage points (44%) more likely to report they tell their child they will send them away when misbehaving (p < 0.05). This is suggestive of caregivers in the treatment group substituting away from other more violent forms of discipline. Overall, we find that the treatment led to positive and persistent impacts on caregivers’ attitudes toward VAC, which in turn led to persistent reductions in VATC. 14 The magnitudes of these effects are similar to the impacts from a digital stress management and positive parenting intervention in El Salvador on physical violence perpetrated by female caregivers (Amaral et al., 2021). 18 5.4 ITT Impacts on Caregiver Depression, Anxiety, & Stress Harsh behaviors toward children may be explained by stress, anxiety, and frustration (Persson and Rossin-Slater, 2018; Bendini and Dinarte, 2020), which affect parental func- tioning through psychological well-being (Belsky, 1984; Belsky and Jaffee, 2006; Taraban and Shaw, 2018). We estimate ITT impacts on caregivers’ well-being as measured by de- pression, anxiety, and parental stress. Depression and anxiety were measured at both follow-ups using the Patient Health Questionnaire (PHQ-2) and Generalized Anxiety Dis- order (GAD-2) instruments, respectively. Parental stress was measured using the Parental Stress Scale (PSS-18) developed by Berry and Jones (1995) at the second follow-up only. While we do not observe statistically significant impacts on caregivers’ well-being at the first follow-up, Figure 5 and columns (1)-(3) of Table 4 (Panel B) show reductions in caregiver depression (0.12 SD, p < 0.1), anxiety (0.16 SD, p < 0.05), and parental stress (0.16 SD, p < 0.05) at the second follow-up. The lag in improvements in caregivers’ mental health suggests that mental health improved only after they applied new tools learned from the intervention. Alternatively, caregivers may have needed time to apply their learning from the intervention. 5.5 ITT Impacts on Child Behaviors The conceptual framework suggests that positive impacts on caregiver attitudes and be- haviors can lead to positive impacts on child behaviors. Figure 6 and columns (4)-(5) of Table 4 present the ITT impacts of the intervention on the target child’s conduct and emo- tional problems. As outlined in Section 3.3, conduct and emotional problems were mea- sured using caregivers’ responses to the Strengths and Difficulties Questionnaire (SDQ) instrument. The estimates show that children in the treatment group scored 0.17 SD lower on the emotional problems index in the short term (p < 0.01). These children also scored 0.03 SD lower on the conduct problems index in the short term; however, this result is not statistically significant at conventional levels. In the medium term, though coefficients 19 are negative, we do not find statistically significant reductions in conduct and emotional problems. Panel C of Table A8 shows that the medium-term impacts are not statistically different from the short-term impacts. 5.6 Intervention Take-up As highlighted in Section 2, the intervention consisted of three components: three SMS messages per week, access to a data-free App with vIHT content, and weekly virtual ses- sions with ECC officers. We cannot causally unpack the relative contribution of each com- ponent. However, in this section we explore take-up of the various components using both administrative data and self-reports in Table 5. Panel A of Table 5 shows that 92% of the 30 SMS messages were sent by TrendMedia to caregivers. Panel B highlights that 91% of caregivers in the treatment group reported receiving SMS and/or WhatsApp messages as part of the intervention. Of the 499 care- givers who reported receiving the SMS and/or WhatsApp messages, 97% reported having read them. Furthermore, 98% of the 444 caregivers who read the messages found them to be useful. Using streaming data from the App, we also tracked the total duration of time spent accessing content on the App in Panel C. The mean duration spent on the App (across the 10-week intervention) was 6.9 minutes across all treatment groups. Moreover, the mean number of sessions accessed on the App was 1 out of 10. Results from the qual- itative study show that the App was not as good a resource to deliver this intervention and that some adjustments will be required to make it more attractive/accessible to this population (Szekely, 2023). In Panel D, we show that across all caregivers assigned to the treatment group, the mean number of sessions attended was 4.6 (out of 10), and 79% of caregivers attended at least one session.15 From the focus group discussions, we document that there was some substitution in take-up across the virtual meetings and SMS messages. When we asked participants with low virtual meeting attendance their reasons for not at- tending many sessions, one of the main reasons given was that it was because they knew 15 This attendance rate is 23 percentage points lower than the rate for the in-person intervention evaluated in Francis and Baker-Henningham (2021). 20 they were going to receive similar materials via SMS/WhatsApp (Szekely, 2023). Does attendance at a greater number of sessions lead to more learning? Acknowledg- ing that selection into attendance at sessions is endogenous, we explore this dose-response relationship in Figure A6, where we present coefficient plots of the impact of session at- tendance on the caregiver information index. To improve precision, we group attendance into 5 pairs of possible combinations, i.e. attendance at one or two sessions, three or four sessions, and so on (where attendance at none of the sessions represents the base refer- ence group). The regression is estimated over caregivers in the treatment group only. We observe an increasing dose-response relationship: while caregivers attending one to four sessions did not score significantly higher than caregivers who did not attend any sessions, caregivers attending five to ten sessions scored significantly higher on the infor- mation index relative to caregivers who did not attend any sessions. With the exception of attendance at seven or eight sessions, caregiver knowledge monotonically increases in the short term in the number of virtual sessions attended. Results from the second follow-up show a very similar pattern in the medium term, with caregivers attending seven to ten sessions scoring significantly higher on the information index relative to caregivers who did not attend any sessions. These dose-response results suggest that our results are likely driven by learning new information rather than the SMS messages being just nudges or reminders of practices that caregivers already knew. In sum, although this intervention was delivered through three components, the take- up rate of the App was relatively low and attendance at the weekly groups was low, suggesting the SMS and/or WhatsApp messages might have played an important role. 6 Potential Mechanisms As highlighted in Figure 1, our conceptual framework hypothesizes at least three poten- tial change pathways: (i) information, (ii) self-efficacy, and (iii) support networks. These mechanisms are based on previous research. First, research has highlighted that a lack of knowledge and skills on parenting practices may lead to harsh and unhealthy parent- ing behaviors (Baker-Henningham and Francis, 2018). Second, responsiveness-oriented 21 parenting approaches have shown the importance of self-efficacy beliefs in parenting functioning strategies (Michl-Petzing et al., 2019). Responsiveness-oriented behaviors are based on acceptance and warm responses toward child actions (Landry et al., 2012) and parents’ perceptions about their self-efficacy (i.e. the extent to which they consider themselves capable and prepared to raise a child and deal with the associated parenting tasks) has become a cornerstone for good parenting practices (Izzo et al., 2000; Hoover- Dempsey et al., 2005; Jones and Prinz, 2005). Third, prior work has highlighted that par- enting beliefs, attitudes, and behaviors may also be influenced by parents’ social networks (Cochran, 2019). The results in Section 5.1 documented the impacts of the intervention on violence pre- vention knowledge (the first potential pathway). We find large increases in knowledge related to positive parenting practices. In Figure 7, we present ITT impacts on the sec- ond and third potential pathways - self-efficacy and support networks. For self-efficacy, we used the Brief Parental Self-Efficacy Scale (BPSES) at the first follow-up but adapted to the Tool to Measure Parenting Self-Efficacy (TOPSE) at the second follow-up for more detailed questions relating to discipline and self-acceptance. For networks, we collected day on parenting support and borrowing support networks. Figure 7 shows positive impacts on caregivers’ self-efficacy relating to self-acceptance at the second follow-up (0.21 SD, p < 0.01). We do not find statistically significant impacts on caregivers’ parenting support or borrowing networks. The lack of significant impacts on support networks highlights that peer interactions during the weekly virtual sessions were likely limited. These findings are confirmed by the qualitative study. Participants across the different focus groups reported that they did not make any friends from the program and that they are not in touch with anyone from a virtual group. Some of the reasons they mentioned included the vast geographic spread of other participants (the sessions were not organized to take into consideration the proximity across participants) and the virtual nature of the meetings (Szekely, 2023). The positive treatment effects and analysis of mechanisms suggest that the interven- tion provided caregivers with the necessary parenting tools and boosted their confidence 22 in their own parenting skills. Taken together, our results suggest that the harsh behav- iors of the caregivers could be explained, to a large extent, by a lack of knowledge and skills, self-efficacy, and emotional self-regulation. As Baker-Henningham and Francis (2018) have suggested, integrating new interventions that aim to train parents in alter- native discipline strategies could help to improve the quality of parenting and reduce violent behaviors against the children. 7 Robustness of the Results In this section, we address potential concerns regarding the results presented in Section 5, such as experimenter demand effects, displacement of violence toward other children in the household, and differential attrition across treatment and control groups in the first follow-up. We also present the results of additional robustness checks. 7.1 Assessing Potential Bias Due To Experimenter Demand Effects Self-reported measures are susceptible to potential experimenter demand effects when as- sessing sensitive information such as attitudes and perpetration of violence because par- ticipants’ responses regarding sensitive topics might be influenced by social desirability bias (Aguero and Frisancho, 2021; Amaral et al., 2021). Recent empirical evidence sug- gests a limited quantitative importance of experimenter demand effects in some domains (Haaland et al., 2023; de Quidt et al., 2018). The concern is whether participants report statements on violent attitudes and behaviors differently from their attitudes and behav- iors outside of the study environment and whether this is differential across the treatment and control groups. We take the possibility of experimenter demand effects seriously and conduct several exercises to address this potential concern. First, we test the intervention’s direct effect on a social desirability index (SDI) estimated using the Marlowe-Crowne Social Desirability Scale (Crowne and Marlowe, 1960) which we collected during the second follow-up. The Marlowe-Crowne Social Desirability Scale has been shown in different settings to be in- 23 formative about one’s propensity to report in socially desirable ways when asked about physical and psychological violence perpetration and victimization. For example, Bell and Naugle (2007) and Fernández-González et al. (2013) show sizeable correlations between SDI and physical and psychological aggression. First, we test whether treatment and control caregivers differ by SDI. Column (1) of Table 6 (Panel A) shows there are no statistically significant differences between treated and control caregivers on the SDI. Second, we include the SDI as an additional control variable in the main regressions. These results in columns (2)-(7) of Table 6 (Panel A) show that the estimated effects remain similar in magnitude and statistical significance. Third, we are concerned that caregivers in the treatment group might have a higher propensity to give socially desirable answers, thereby biasing the estimated treatment ef- fects. Our results in Panel B of Table 6 show that there are no heterogeneous treatment impacts by SDI; the interaction terms in the third row are not statistically significant. The only instance where the interaction term is statistically significant is in column (2) (overall attitudes toward VAC), driven by attitudes toward psychological VAC (column 4). How- ever, the interaction term is in fact positive, and in the opposite direction of what we would expect from treated individuals who are inclined to report in socially desirable ways. Lastly, we investigate the potential concern that in addition to behaviors relating to violence, the SDI may also capture social desirability relating to other non-violence re- lated behaviors. To address this concern, we select 5 items out of the 13 items included in the original SDI instrument that are most closely related to violence and aggression (i.e., “I have deliberately said something that hurt someone’s feelings”) and create a new violence-focused SDI. We repeat the analysis on social desirability with this new scale, and the results are presented in Appendix Table A11. Overall, the main conclusions relating to social desirability are sustained. Together, these findings suggest that it is unlikely that experimenter demand effects have biased respondents’ likelihood to misreport their attitudes and behaviors about the use of violence with their children across the treatment and control groups. 24 7.2 Potential Displacement of Violence Within the Household A potential concern might be that the intervention succeeded in reducing caregiver violent discipline against the target child, but this may have displaced violence toward other children in the household. To address this concern, we surveyed caregivers about violence against their eldest child aged 7-12 in addition to the target child. We stack the responses on caregiver attitudes and behaviors to violence against children to form a child-level dataset that allows us to study violence against the target child, as well as the eldest child aged 7-12.16 Figure 8 shows 0.14 SD and 0.15 SD ITT reductions in the index for violence against these children in the short- and medium-term, respectively (p < 0.01). We also find 0.1 SD (p < 0.1) and 0.13 SD (p < 0.05) reductions in the indices of physical violence in the short- and medium-term, respectively. Moreover, we also find that caregivers in the treatment group scored 0.13 SD (p < 0.01) and 0.14 SD (p < 0.05) lower in the indices of psychological violence against any of these children at the first and second follow-ups. Overall, these results suggest that the intervention did not displace violence from the target child to other children in the household. Another potential concern from the intervention is the displacement of violent disci- plining from the caregiver participating in the program to the other caregivers within their household that did not join the program. To address this concern, we asked participants of the focus groups if they shared information with other caregivers within the household and if they observed changes in the dynamics with their partners. Results from the qual- itative study suggest there were positive spillovers within the household. Participants reported that they shared information from the intervention with their partners. For ex- ample, they encouraged their partner to praise their children for good actions. Moreover, they reported that family dynamics had improved after the program. They reported try- ing to be more gentle with their children instead of immediately being aggressive and also 16 Panel A of Table 1 shows that households in our sample have an average of 1.9 children aged 17 or below. Given the constraints of the phone survey, we collected information on one elder child in the household. 25 played games as a family (Szekely, 2023). 7.3 Assessing Potential Bias Due To Differential Attrition Between the baseline and first follow up survey, we lost 128 individuals, so approximately 11.5% of the sample. We then lost another 280 individuals, or 25.2% of the sample between the short and medium term. We test for differential attrition between treatment and con- trol groups as we are concerned attrition might bias the estimates. In particular, if the higher proportion of treatment group caregivers who attrited are more likely to exhibit violent disciplining behaviors, our estimates may overestimate the true treatment effects. We assess differential attrition between treatment and control groups for each follow- up round and present the results of this analysis in Table 7. Column (1) shows that care- givers in the treatment group were 4.7 percentage points more likely to complete the first follow-up survey relative to caregivers in the control group. In column (2), we study whether the differential attrition in the first follow-up was correlated with any demo- graphic characteristics or outcome variables measured at baseline. None of the interac- tion terms in this regression are statistically significant at the 10% level, suggesting no evidence of relationships between demographic characteristics or outcome variables and the differential attrition. In line with this result, we cannot reject the null hypothesis that all interaction terms between the treatment indicator and the relevant variables are not statistically significant (p-value of 0.528). In the second follow-up, we find no significant differences in the probability of not completing the survey between treatment and control groups (Column 3). Moreover, we find that none of the relevant variables explain any differential attrition between treatment and control, we do not reject the null hypothesis that the interaction terms between the treatment indicator and all the relevant variables are not different from zero (p-value = 0.916). To address potential differential attrition in the first follow-up, we estimate Lee bounds to account for sample selection (Lee, 2009) and present these results in Table A12. This procedure is a conservative estimate of the treatment effect, as it corresponds to extreme 26 assumptions about the missing information. We find that all upper and lower bounds significantly differ from zero except the upper bounds of the index of attitudes toward psychological VAC (p-value = 0.177) and violence against the target child (p-value = 0.177) (driven by the psychological violence dimension), suggesting that our results are overall robust to differential attrition. As an additional check, we re-estimate the ITT impacts on caregiver attitudes and be- haviors, as well as child outcomes using a balanced panel of caregivers that were present at both the first and second follow-ups. This addresses concerns of potential selection in who might be present at either follow-up round. The results are shown in Figures A7 - A10. Overall, the results are very similar to the estimates presented in Section 5. 7.4 Assessing Sensitivity From Selection of Control Variables As we discuss in Section 4, we use a double LASSO approach to identify the variables that should be included in our estimations as controls. We find that LASSO consistently selects, across all outcomes, the measure of the outcome at baseline. LASSO also selects other variables that can be used as controls, but they vary across outcomes. For instance, for some outcomes age and gender are selected, whereas for others LASSO selects edu- cation level and household composition. Considering this, we test for the stability of our estimated coefficients after including the control variables selected by LASSO for each of our main outcomes. As we show in Table A13, the estimated coefficients and their statis- tical significance do not change after including these additional control variables selected by LASSO. 7.5 Estimating the exact p-values using randomization inference To take a more agnostic approach to the structure of the standard errors, we estimate standard errors using the randomization inference (RI) approach. As discussed in Section 4, what RI allows us to do is to assign a p-value for a given treatment effect by observing where that treatment effect falls in the distribution of all possible estimated effects from 27 the 1,000 randomizations we simulate under the assumption of no effects (Blattman et al., 2021). As we show in the row “RI p-value” in Table A14 in the Appendix, the magnitude of p-value is similar to the magnitudes of p-values obtained by estimating heteroskedasticity- robust standard errors. 8 Heterogeneity Recent work has highlighted that the treatment impacts of parenting interventions may differ for sub-groups of the sample. For example, Amaral et al. (2021) highlight differ- ential effects of a digital stress management and positive parenting intervention by the gender of the caregiver, with male caregivers experiencing increased stress and anxiety while female caregivers saw no impacts on mental health. Baranov et al. (2020) study a psychotherapy intervention for prenatally depressed mothers in Pakistan, and show significant treatment heterogeneity along wealth for outcomes including parenting style and children’s socio-emotional development. Francis and Baker-Henningham (2021) also show, in an evaluation of the in-person Irie Homes Toolbox in Jamaica, that there were sig- nificant reductions in behavior difficulties for children with above-median baseline levels of behavior difficulties. Studying such treatment heterogeneity is important for assessing the scale-up and policy applications of our study. We investigate treatment impact heterogeneity in two stages. First, we use machine learning techniques to understand if there is any evidence of significant treatment hetero- geneity for our primary outcomes of interest. Second, conditional on evidence of hetero- geneity, we use traditional interaction term analysis to understand the dimensions along which this heterogeneity arises. In the first stage, to understand if there is any evidence of heterogeneity, we use casual forest algorithms to estimate Conditional Average Treatment Effects (CATEs) following Athey et al. (2019), Athey and Wager (2019), and Chernozhukov et al. (2023). Given subject characteristics, we use a subject-specific treatment prioritization rule that assigns scores to subjects, with higher scores assigned to caregivers with the largest benefit from the treatment as given by the CATE. To quantify treatment benefits, we use the Targeting 28 Operator Characteristic (TOC), which compares the ATE in smaller groups defined by the prioritization rule to the overall ATE from treating everyone in the treatment group. We present results investigating evidence of treatment heterogeneity for our primary outcome variables (information index, attitudes toward VAC index, and violence against the target child index) in Table 8. To do so, we categorize caregivers into high and low ATE groups by creating below and above median CATE sub-groups, and estimate the ATE in each group. We observe a statistically significant 0.56 SD difference in above versus be- low median CATE for attitudes toward VAC. This shows evidence of significant treatment heterogeneity for attitudes toward VAC. While the information index and violence against target child variables show large differences – 0.33 SD and 0.12 SD, respectively – these differences are not statistically significant. Therefore, we conclude that we do not have sufficient evidence suggesting significant treatment heterogeneity for the information in- dex and the violence against target child index. Figure 9 presents the TOC corresponding to attitudes toward VAC, where the x-axis denotes the top q-th fraction of individuals with the largest prioritization score. We pool our first and second rounds of follow-up data to estimate the CATEs and plot the TOC. We see significant treatment heterogeneity for attitudes toward VAC. For example, the top 20% of caregivers with the largest prioritization score show treatment effects that are approximately 0.4 SD larger than the ATE. The heterogeneity is statistically significant at the 5% level for most of the distribution, as indicated by the 95% confidence interval bars. Given the evidence of significant heterogeneity for attitudes toward VAC, in the sec- ond stage of analysis, we investigate the specific dimensions along which this heterogene- ity arises using a traditional interaction term analysis. These results are presented in Table 9. Motivated by the recent literature outlined at the start of this section, we consider het- erogeneity along the following dimensions: (i) gender of the caregiver, (ii) income, (iii) baseline attitudes toward VAC, (iv) baseline conduct problems of the target child, and (v) baseline emotional problems of the target child. Panels A and B present the analysis for the short- and medium-term follow-ups, respectively. We observe significant heterogeneity by baseline attitudes toward VAC: the treatment 29 effects are 0.19 SD and 0.21 SD larger in the short- and medium-term for caregivers with worse than median attitudes toward VAC at baseline (column 3). Columns (4) and (5) highlight that treatment effects are 0.2 and 0.18 SD larger for target children with above- median conduct and emotional problems, respectively. However, these results are only statistically significant in the short-run. We do not find any significant heterogeneity by gender of the caregiver or income. Taken together, the results highlight significant heterogeneity in attitudes toward VAC, shown using machine learning techniques. Traditional interaction term analysis high- lights the role of baseline attitudes toward VAC, as well as emotional and conduct prob- lems of the target child in the short-run, in explaining some of this heterogeneity. 9 Cost Effectiveness Table 10 shows the cost-effectiveness of our intervention relative to the face-to-face IHT in Jamaica (Francis and Baker-Henningham, 2021) and a cash transfer intervention in Mali (Heath et al., 2020). We show these comparisons because these interventions also reduced physical or psychological violence toward children or the female partner. Excluding the cost of developing the App, which is a fixed cost for the intervention component that was least used by caregivers, the vIHT cost USD 62.4 per person and reduced violent behaviors against children by 0.11 to 0.13 SD in the medium-term. The cost per 0.13 SD effect is between USD 62.4 and USD 73.75. In comparison, the face-to-face IHT cost USD 123.95 per person and reduced physical violence against children by 0.29 SD, yielding a USD 55.56 cost per 0.13 SD effect. The per person cost of the vIHT is a similar magnitude to the in person IHT. However, we note that the vIHT has greater potential to be scaled- up given the virtual delivery via smart phones. It is less demanding on staff capacity to conduct the virtual meetings relative to the number of professionals required to conduct the face-to-face activities. In contexts with a limited number of local professionals for early childhood development and parenting interventions, the implementation of face-to-face parenting interventions are more difficult to implement at scale. In a cash transfer intervention in Mali, Heath et al. (2020) show that cash decreased 30 physical and psychological violence against the female partner (0.13 and 0.12 SD, respec- tively) and physical violence against children (0.17 SD). This translates to a per person cost between USD 496 and USD 702.8 per 0.13 SD effect size. While we note that cash transfer programs may aim to improve several other outcomes, this result highlights the strong cost-effectiveness of our intervention relative to cash transfers to reduce violent behaviors toward children.17 10 Conclusion Violence against children is a global problem that will require new and creative solutions. We evaluate the impact of the virtual Irie Homes Toolbox, a positive parenting program for Jamaican parents of children aged two to six years. Our results show that the inter- vention improved caregivers’ attitudes and violent behaviors against children in the short term. Furthermore, the program reduced children’s emotional problems. Results in the medium term from a second follow-up indicate that these effects on caregivers’ attitudes and violent behaviors against children persist nine months after the intervention ended. We also document important improvements in caregivers’ mental health in the medium term. In terms of potential mechanisms driving these effects, we find evidence consistent with improvements in parental self-efficacy most likely due to the knowledge and skills they gained from the virtual Irie Homes Toolbox. From a policy perspective, the positive effects of the intervention on attitudes and be- haviors related to violent discipline against children provide evidence of the importance of developing and implementing positive training for caregivers. Digital behavioral change information campaigns that aim to build positive relationships between parent and child, prevent misbehavior, manage misbehavior, and improve emotional self-regulation can help to increase the quality of parenting. Moreover, this learning can help reduce harmful parenting practices during the early ages of the children, diminishing exposure to violence at home (Francis and Baker-Henningham, 2021). Altogether, a more positive and healthy 17 For further detail, Table A15 shows detailed cost information (fixed and variable) associated with our intervention as well as the face-to-face program. 31 rearing environment has the potential to increase the long-term well-being of children and their families. 32 References Aguero, J. and Frisancho, V. (2021). Measuring violence against women with experimental meth- ods. Economic Development and Cultural Change. Altafim, E. R. P. and Linhares, M. B. M. (2016). Universal violence and child maltreatment preven- tion programs for parents: A systematic review. Psychosocial Intervention, 25(1):27–38. Amaral, S., Dinarte Diaz, L., Dominguez, P., and Perez-Vincent, S. M. (2021). Helping families help themselves? Heterogeneous effects of a digital parenting program. Anderson, M. L. (2008). Multiple inference and gender differences in the effects of early interven- tion: A reevaluation of the Abecedarian, Perry Preschool, and Early Training projects. Journal of the American Statistical Association, 103(484):1481–1495. Asadullah, M. N., De Cao, E., Khatoon, F. Z., and Siddique, Z. (2021). Measuring gender attitudes using list experiments. Journal of Population Economics, 34(2):367–400. Athey, S., Tibshirani, J., and Wager, S. (2019). Generalized random forests. The Annals of Statistics, 47(2):1148 – 1178. Athey, S. and Wager, S. (2019). Estimating treatment effects with causal forests: An application. Attanasio, O., Cattan, S., Fitzsimons, E., Meghir, C., and Rubio-Codina, M. (2020). Estimating the production function for human capital: Results from a randomized controlled trial in Colombia. American Economic Review, 110(1):48–85. Attanasio, O. P., Fernández, C., Fitzsimons, E. O., Grantham-McGregor, S. M., Meghir, C., and Rubio-Codina, M. (2014). Using the infrastructure of a conditional cash transfer program to deliver a scalable integrated early child development program in colombia: cluster randomized controlled trial. Bmj, 349. Baker-Henningham, H. and Francis, T. (2018). Parents’ use of harsh punishment and young chil- dren’s behaviour and achievement: a longitudinal study of jamaican children with conduct problems. Global Mental Health, 5. Balaji, A. B., Claussen, A. H., Smith, D. C., Visser, S. N., Morales, M. J., and Perou, R. (2007). Social support networks and maternal mental health and well-being. Journal of women’s health, 16(10):1386–1396. Baranov, V., Bhalotra, S., Biroli, P., and Maselko, J. (2020). Maternal depression, women’s em- powerment, and parental investment: Evidence from a randomized controlled trial. American Economic Review, 110(3):824–59. Barrera, O., Macours, K., Premand, P., and Vakis, R. (2020). Texting parents about early child development. World Bank Policy Research Working Paper, (9492). Bell, K. M. and Naugle, A. E. (2007). Effects of social desirability on students’ self-reporting of partner abuse perpetration and victimization. Violence and victims, 22(2):243–256. Belsky, J. (1984). The determinants of parenting: A process model. Child development, pages 83–96. 33 Belsky, J. and Jaffee, S. R. (2006). The multiple determinants of parenting. Bendini, M. and Dinarte, L. (2020). Does maternal depression undermine childhood cognitive de- velopment? Evidence from the Young Lives Survey in Peru. International Journal of Environmental Research and Public Health, 17(19):7248. Berry, J. O. and Jones, W. H. (1995). The parental stress scale: Initial psychometric evidence. Journal of social and personal relationships, 12(3):463–472. Blattman, C., Green, D. P., Ortega, D., and Tobón, S. (2021). Place-based interventions at scale: The direct and spillover effects of policing and city services on crime. Journal of the European Economic Association, 19(4):2022–2051. Boden, J. M., Horwood, L. J., and Fergusson, D. M. (2007). Exposure to childhood sexual and physi- cal abuse and subsequent educational achievement outcomes. Child abuse & neglect, 31(10):1101– 1114. Bos, J., Shonchoy, A., Ravindran, S., and Khan, A. (2022). Early childhood human capital formation at scale. Available at SSRN 3906697. Bruhn, M. and McKenzie, D. (2009). In pursuit of balance: Randomization in practice in develop- ment field experiments. American Economic Journal: Applied Economics, 1(4):200–232. Carneiro, P., Galasso, E., García, I. L., Bedregal, P., and Cordero, M. (2023). Parental beliefs, in- vestments, and child development: Evidence from a large-scale experiment. Journal of Political Economy, forthcoming. Carrell, S. E. and Hoekstra, M. L. (2010). Externalities in the classroom: How children exposed to domestic violence affect everyone’s kids. American Economic Journal: Applied Economics, 2(1):211– 28. Chandra, A., Mani, S., Dolphin, H., Dyson, M., and Marah, Y. (2021). Experimental evidence from an integrated early childhood parenting program in sierra leone. Chernozhukov, V., Demirer, M., Duflo, E., and Fernández-Val, I. (2023). Generic Machine Learning Inference on Heterogeneous Treatment Effects in Randomized Experiments, with an Application to Immunization in India. NBER Working Paper, (w24678). Chorpita, B. F. and Daleiden, E. L. (2009). Mapping evidence-based treatments for children and adolescents: application of the distillation and matching model to 615 treatments from 322 ran- domized trials. Journal of consulting and clinical psychology, 77(3):566. Cluver, L. D., Meinck, F., Steinert, J. I., Shenderovich, Y., Doubt, J., Romero, R. H., Lombard, C. J., Redfern, A., Ward, C. L., Tsoanyane, S., Nzima, D., Sibanda, N., Wittesaele, C., Stone, S. D., Boyes, M. E., Catanho, R., Lachman, J. M., Salah, N., Nocuza, M., and Gardner, F. (2018). Parent- ing for Lifelong Health: a pragmatic cluster randomised controlled trial of a non-commercialised parenting programme for adolescents and their families in South Africa. BMJ Global Health, 3(1). Cochran, M. (2019). Chapter 6: Parenting and Social Networks. Handbook of Parenting: Volume 4 Social Conditions and Applied Parenting. 34 Cortes, K. E., Fricke, H. D., Loeb, S., Song, D. S., and York, B. N. (2019). When behavioral barriers are too high or low–how timing matters for parenting interventions. Technical report, National Bureau of Economic Research. Crowne, D. P. and Marlowe, D. (1960). A new scale of social desirability independent of psy- chopathology. Journal of Consulting Psychology, 24(4):349. Currie, J. and Spatz Widom, C. (2010). Long-term consequences of child abuse and neglect on adult economic well-being. Child maltreatment, 15(2):111–120. Currie, J. and Tekin, E. (2012). Understanding the cycle childhood maltreatment and future crime. Journal of Human Resources, 47(2):509–549. Dahl, G. B., Løken, K. V., and Mogstad, M. (2014). Peer effects in program participation. American Economic Review, 104(7):2049–74. de Quidt, J., Haushofer, J., and Roth, C. (2018). Measuring and bounding experimenter demand. American Economic Review, 108(11):3266–3302. de Simone, F., Mejia, C., Martinez-Carrasco, J., Perez-Vincent, S. M., and Villalba, H. (2022). Miti- gating Coercive Parenting Through Home Visitations: Impacts of a Parenting Program Targeted at Vulnerable Communities in Jamaica. IDB Working Paper. Dhar, D., Jain, T., and Jayachandran, S. (2022). Reshaping adolescents’ gender attitudes: Evidence from a school-based experiment in india. American Economic Review, 112(3):899–927. Donker, T., van Straten, A., Marks, I., and Cuijpers, P. (2011). Quick and easy self-rating of gen- eralized anxiety disorder: validity of the dutch web-based gad-7, gad-2 and gad-si. Psychiatry research, 188(1):58–64. Doss, C., Fahle, E. M., Loeb, S., and York, B. N. (2019). More than just a nudge supporting kinder- garten parents with differentiated and personalized text messages. Journal of Human Resources, 54(3):567–603. Doyle Jr, J. J. and Aizer, A. (2018). Economics of child protection: Maltreatment, foster care, and intimate partner violence. Annual review of economics, 10:87–108. Fang, Y., Boelens, M., Windhorst, D. A., Raat, H., and van Grieken, A. (2021). Factors associated with parenting self-efficacy: A systematic review. Journal of Advanced Nursing, 77(6):2641–2661. Fernández-González, L., O’Leary, K. D., and Muñoz-Rivas, M. J. (2013). We are not joking: Need for controls in reports of dating violence. Journal of Interpersonal violence, 28(3):602–620. Francis, T. and Baker-Henningham, H. (2020). Design and implementation of the irie homes tool- box: a violence prevention, early childhood, parenting program. Frontiers in public health, 8:702. Francis, T. and Baker-Henningham, H. (2021). The Irie Homes Toolbox: A cluster randomized controlled trial of an early childhood parenting program to prevent violence against children in Jamaica. Children and Youth Services Review, 126:106060. Fry, D., Fang, X., Elliott, S., Casey, T., Zheng, X., Li, J., Florian, L., and McCluskey, G. (2018). The relationships between violence in childhood and educational outcomes: A global systematic review and meta-analysis. Child abuse & neglect, 75:6–28. 35 Garland, A. F., Hawley, K. M., Brookman-Frazee, L., and Hurlburt, M. S. (2008). Identifying common elements of evidence-based psychosocial treatments for children’s disruptive behavior problems. Journal of the American Academy of Child & Adolescent Psychiatry, 47(5):505–514. Haaland, I., Roth, C., and Wohlfart, J. (2023). Designing information provision experiments. Journal of Economic Literature, 61(1):3–40. Hamadani, J. D., Mehrin, S. F., Tofail, F., Hasan, M. I., Huda, S. N., Baker-Henningham, H., Rid- out, D., and Grantham-McGregor, S. (2019). Integrating an early childhood development pro- gramme into bangladeshi primary health-care services: an open-label, cluster-randomised con- trolled trial. The Lancet Global Health, 7(3):e366–e375. Hamby, S., Finkelhor, D., Turner, H., and Ormrod, R. (2011). Children’s Exposure to Intimate Partner Violence and Other Family Violence. Washington, DC: U.S. Department of Justice. Heath, R., Hidrobo, M., and Roy, S. (2020). Cash transfers, polygamy, and intimate partner vio- lence: Experimental evidence from mali. Journal of Development Economics, 143(102410). Heckman, J. J., Liu, B., Lu, M., and Zhou, J. (2020). The impacts of a prototypical home visiting program on child skills. NBER Working Paper, (w27356). Hoover-Dempsey, K. V., Walker, J. M., Sandler, H. M., Whetsel, D., Green, C. L., Wilkins, A. S., and Closson, K. (2005). Why do parents become involved? research findings and implications. The elementary school journal, 106(2):105–130. Hurwitz, L. B., Lauricella, A. R., Hanson, A., Raden, A., and Wartella, E. (2015). Supporting head start parents: impact of a text message intervention on parent–child activity engagement. Early Child Development and Care, 185(9):1373–1389. Izzo, C., Weiss, L., Shanahan, T., and Rodriguez-Brown, F. (2000). Parental self-efficacy and social support as predictors of parenting practices and children’s socioemotional adjustment in mexi- can immigrant families. Journal of Prevention & Intervention in the Community, 20(1-2):197–213. Jones, T. L. and Prinz, R. J. (2005). Potential roles of parental self-efficacy in parent and child adjustment: A review. Clinical psychology review, 25(3):341–363. Kling, J. R., Liebman, J. B., and Katz, L. F. (2007). Experimental analysis of neighborhood effects. Econometrica, 75(1):83–119. Knerr, W., Gardner, F., and Cluver, L. (2013). Improving Positive Parenting Skills and Reducing Harsh and Abusive Parenting in Low- and Middle-Income Countries: A Systematic Review . Prevention Science, 14:352–363. Kroenke, K., Spitzer, R. L., and Williams, J. B. (2003). The patient health questionnaire-2: validity of a two-item depression screener. Medical care, pages 1284–1292. Lachman, J., Wamoyi, J., Spreckelsen, T., Wight, D., Maganga, J., and Gardner, F. (2020). Combin- ing parenting and economic strengthening programmes to reduce violence against children: a cluster randomised controlled trial with predominantly male caregivers in rural Tanzania. BMJ Global Health, 5(7). 36 Landry, S. H., Smith, K. E., Swank, P. R., Zucker, T., Crawford, A. D., and Solari, E. F. (2012). The effects of a responsive parenting intervention on parent–child interactions during shared book reading. Developmental psychology, 48(4):969. Lansford, J. E., Dodge, K. A., Pettit, G. S., Bates, J. E., Crozier, J., and Kaplow, J. (2002). A 12-year prospective study of the long-term effects of early child physical maltreatment on psychological, behavioral, and academic problems in adolescence. Archives of pediatrics & adolescent medicine, 156(8):824–830. Lee, D. S. (2009). Training, wages, and sample selection: Estimating sharp bounds on treatment effects. The Review of Economic Studies, 76(3):1071–1102. Levere, M., Acharya, G., and Bharadwaj, P. (2016). The role of information and cash transfers on early childhood development: evidence from nepal. Technical report, National Bureau of Economic Research. Macours, K., Schady, N., and Vakis, R. (2012). Cash transfers, behavioral changes, and cognitive development in early childhood: evidence from a randomized experiment. American Economic Journal: Applied Economics, 4(2):247–73. Mehrin, S. F., Hasan, M. I., Tofail, F., Shiraji, S., Ridout, D., Grantham-McGregor, S., Hamadani, J. D., and Baker-Henningham, H. (2022). Integrating a group-based, early childhood parent- ing intervention into primary health care services in rural bangladesh: A cluster-randomised controlled trial. Frontiers in Pediatrics. Mersky, J. P. and Topitzes, J. (2010). Comparing early adult outcomes of maltreated and non- maltreated children: A prospective longitudinal investigation. Children and Youth Services Review, 32(8):1086–1096. Michie, S., Van Stralen, M. M., and West, R. (2011). The behaviour change wheel: a new method for characterising and designing behaviour change interventions. Implementation science, 6(1):1–12. Michl-Petzing, L. C., Handley, E. D., Sturge-Apple, M., Cicchetti, D., and Toth, S. L. (2019). Re- examining the “cycle of abuse”: Parenting determinants among previously maltreated, low- income mothers. Journal of family psychology, 33(6):742. Operator Watch (2021). Jamaica has good 4G coverage but no certainty on 5G. Available at https: //www.operatorwatch.com/2021/04/jamaica-has-good-4g-coverage-but-no. html. Owen, A. and Anderson, B. (2017). Informal community support for parents of pre-school children: A comparative study investigating the subjective experience of parents attending community- based toddler groups in different socio-economic situations. Journal of Early Childhood Research, 15(2):212–224. Paxson, C. and Schady, N. (2010). Does money matter? the effects of cash transfers on child development in rural ecuador. Economic development and cultural change, 59(1):187–229. Persson, P. and Rossin-Slater, M. (2018). Family ruptures, stress, and the mental health of the next generation. American Economic Review, 108(4-5):1214–52. 37 Smith, J. A., Chang, S. M., Brentani, A., Fink, G., Lopez-Boo, F., Torino, B. M., Codina, M. R., and Walker, S. P. (2023). A Remote Parenting Program and Parent and Staff Perspectives: A Randomized Trial. Pediatrics, 151(Supplement 2). e2023060221F. Sviatschi, M. M. (2022). Making a narco: Childhood exposure to illegal labor markets and criminal life paths. Econometrica, 90(4):1835–1878. Szekely, M. (2023). Results from focus group discussions with participants of the program Virtual Irie Homes in Jamaica. Technical report. Taraban, L. and Shaw, D. S. (2018). Parenting in context: Revisiting belsky’s classic process of parenting model in early childhood. Developmental Review, 48:55–81. UNICEF (2011). 2011 Multiple Indicator Cluster Survey. Technical report, UNICEF. UNICEF (2017). A familiar face: Violence in the lives of children and adolescents. New York, NY: UNICEF. UNICEF (2022). A statistical profile of violence against children in latin america and the caribbean. Technical report, United Nations Children’s Fund (UNICEF). Widen, S. C., Orozco, M., Horng, E. L., and Loeb, S. (2020). Reaching unconnected caregivers: Us- ing a text-message education program to better understand how to support informal caregivers roles in child development. Journal of Early Childhood Research, 18(1):29–43. Wittkowski, A., Garrett, C., Calam, R., and Weisberg, D. (2017). Self-report measures of parental self-efficacy: A systematic review of the current literature. Journal of child and family studies, 26:2960–2978. York, B. N., Loeb, S., and Doss, C. (2019). One step at a time the effects of an early literacy text- messaging program for parents of preschoolers. Journal of Human Resources, 54(3):537–566. 38 Tables and Figures Figures Figure 1: Conceptual Framework Notes: This diagram shows the conceptual framework for our analysis. For more details about this framework, see the text. 39 Figure 2: Experimental Design Notes: This figure summarizes the experimental design of the study. 880 enrolled individuals did not complete the baseline survey for several reasons, including they did not provide a correct phone number; we were unable to reach them after the maximum number of attempts determined in the ethics protocol; they changed their mind and decided not to participate in the study, among others. 40 Figure 3: ITT Impacts on Caregiver Attitudes Toward Violence Against Children Attitudes toward p = 0.000/ b = -0.198 violence against children (index) p = 0.018/ b = -0.144 Attitudes toward p = 0.000/ b = -0.192 physical violence Short-term against children Medium-term (index) p = 0.017/ b = -0.150 Attitudes toward p = 0.040/ b = -0.120 psychological violence against children (index) p = 0.273/ b = -0.075 -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate, b) from Equation 1 (and the respective p-value) on caregivers’ attitudes toward violence against children. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey (short-term). The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow- up survey (medium-term). Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. All specifications include controls for strata fixed effects. Standard deviation units are used for the x-axis. 41 Figure 4: ITT Impacts on Caregiver Behaviors Related to Violence Against the Target Child p = 0.030/ b = -0.121 Violence against target child (index) p = 0.041/ b = -0.127 Physical violence p = 0.024/ b = -0.136 Short-term against target child (index) p = 0.074/ b = -0.124 Medium-term Psychological p = 0.065/ b = -0.101 violence against target child (index) p = 0.106/ b = -0.105 -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate, b) from Equation 1 (and the respective p-value) on caregivers’ behaviors relating to violence against the target child. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey (short-term). The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow-up survey (medium-term). Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. All specifications include controls for strata fixed effects. Standard deviation units are used for the x-axis. 42 Figure 5: ITT Impacts on Caregivers’ Mental Health p = 0.574/ b = -0.033 Depression (index) p = 0.065/ b = -0.124 p = 0.238/ b = -0.070 Short-term Anxiety (index) p = 0.021/ b = -0.157 Medium-term Parental stress scale (index) p = 0.038/ b = -0.156 -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate, b) from Equation 1 (and the respective p-value) on caregivers’ depression, anxiety, and parental stress. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey (short-term). The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow- up survey (medium-term). Caregivers’ depression and anxiety were measured using PHQ-2 and GAD-2 at both follow-ups, respectively. Parental stress was measured using PSS-18 only at the second follow-up (medium-term). Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. All specifications include controls for strata fixed effects. Standard deviation units are used for the x-axis. 43 Figure 6: ITT Impacts on Child Conduct and Emotional Problems p = 0.547/ b = -0.030 Conduct problems (index) p = 0.367/ b = -0.053 Short-term Medium-term p = 0.002/ b = -0.166 Emotional problems (index) p = 0.434/ b = -0.051 -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate, b) from Equation 1 (and the respective p-value) on the target child’s conduct and emotional problems. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey (short-term). The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow- up survey (medium-term). Conduct and emotional problems were measured using caregivers’ responses to the Strengths and Difficulties Questionnaire (SDQ) instrument. Each outcome consists of a standardized index esti- mated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. All specifications include controls for strata fixed effects. Standard deviation units are used for the x-axis. 44 Figure 7: Potential Mechanisms Parental p = 0.250/ b = 0.071 self-efficacy (index) Parental self-efficacy discipline (index) p = 0.583/ b = 0.041 Parental Short-term self-efficacy acceptance (index) p = 0.007/ b = 0.205 Medium-term p = 0.675/ b = 0.025 Parenting support networks (index) p = 0.525/ b = 0.035 Borrowing support networks (index) -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate, b) from Equation 1 (and the respective p-value) on five variables that explore potential mechanisms through which the intervention improved the caregivers’ self-efficacy and support networks. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey (short-term). The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow-up survey (medium-term). To measure self-efficacy, we used the Brief Parental Self-Efficacy Scale (BPSES) at the first follow-up but adapted to the Tool to Measure Parenting Self-Efficacy (TOPSE) at the second follow-up for more detailed questions relating to discipline and self-acceptance. For networks, we examined parenting support networks and borrowing support networks. Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. All specifications include controls for strata fixed effects. Standard deviation units are used for the x-axis. 45 Figure 8: ITT Impacts on Caregiver Behaviors Related to Violence Against Other Children in Household Violence against p = 0.003/ b = -0.145 children in the household (index) p = 0.006/ b = -0.149 Physical violence p = 0.067/ b = -0.100 against children in Short-term the household Medium-term (index) p = 0.026/ b = -0.131 Psychological p = 0.004/ b = -0.135 violence against children in the household (index) p = 0.015/ b = -0.136 -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate, b) from Equation 1 (and the respective p-value) on caregivers’ behaviors relating to violence against the target child and the eldest child (if any, aged 7-12) in the household. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey (short-term). The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow-up survey (medium-term). Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. All specifications include controls for strata fixed effects. Standard deviation units are used for the x-axis. 46 Figure 9: Heterogeneity in Treatment Impacts: Machine Learning Approach Notes: This figure presents the Targeting Operator Characteristic for the outcome attitudes toward violence against children (index). Data from the short- and medium-term follow-ups have been pooled for statistical power. The x-axis denotes the top q-th fraction of individuals with the largest prioritization score, i.e. caregivers believed to have the largest benefit from the treatment as given by the Conditional Average Treatment Effect (CATE). The TOC compares the ATE in smaller groups defined by the prioritization rule to the overall ATE from treat- ing everyone in the treatment group. The CATEs are estimated non-parametrically using the causal_f orest and rank _average_treatment_ef f ect functions in R. 95% confidence intervals are shown. Standard deviation units are used for the y-axis. 47 Tables Table 1: Summary Statistics by Group and Balance Tests Control Treatment p-value (1) (2) (3) (4) (5) Variable Mean SD Mean SD (1) - (3) Panel A. Caregiver’s characteristics Age (years) 33.405 7.727 33.070 7.247 0.443 Female (%) 0.856 0.351 0.853 0.355 0.634 Education level completed (years) 14.220 2.761 14.445 2.733 0.166 Married (%) 0.384 0.487 0.356 0.479 0.450 Employed (%) 0.784 0.412 0.792 0.407 0.700 Income in the past month (USD) 910.306 1,166.991 855.201 1,062.471 0.489 Household size (N) 4.559 1.873 4.598 2.025 0.734 Children 17 years or younger (N) 1.950 1.021 1.873 1.044 0.213 Panel B. Target child’s characteristics Age (years) 4.171 1.429 4.070 1.425 0.236 Female (%) 0.480 0.500 0.496 0.500 0.623 Panel C. Primary Outcomes Attitudes toward violence against children (% in agreement) Shouting and yelling makes the child more obedient 0.096 0.295 0.115 0.320 0.311 Shouting, yelling, and threatening to slap will not harm the child 0.368 0.483 0.362 0.481 0.830 To raise a child properly, the child needs to be physically punished 0.115 0.319 0.131 0.338 0.391 A good parent slaps their child when they misbehave 0.203 0.403 0.204 0.404 0.941 When a child is beaten, he/she will stop doing the unwanted behavior 0.097 0.296 0.098 0.297 0.948 Violence against target child (Number of days in the past 7 days) Shouted, yelled, or screamed at him/her? 1.541 1.394 1.632 1.352 0.270 Said you would send him/her away? 0.228 0.719 0.142 0.506 0.023 Threatened to hit him/her but not actually done it? 1.479 1.498 1.540 1.485 0.511 Hit him/her on the bottom, hand, arm, or leg with your bare hand? 0.519 0.790 0.525 0.728 0.926 Hit him/her on the bottom, hand, arm, or leg with a hard object? 0.081 0.356 0.097 0.393 0.482 Panel D. Secondary Outcomes Conduct problems (%) 41.655 27.063 41.939 26.966 0.849 Emotional problems (%) 24.676 23.305 25.458 22.925 0.566 Depression (%) 19.964 40.009 18.133 38.564 0.440 Generalized Anxiety Disorder (%) 13.309 33.998 17.056 37.646 0.084 Panel E. Mechanisms Borrowing money support networks (N) 2.302 2.555 2.336 2.490 0.825 Parenting issues support networks (N) 2.558 2.691 2.743 2.680 0.247 Belong to parent support group (%) 0.155 0.362 0.162 0.368 0.761 F-test of joint significance (p-value) 0.968 Notes: This table shows average characteristics at baseline for the study participants assigned to the treatment and control groups. Columns (1) and (2) present the mean and standard errors of the variables for the control group, while columns (3) and (4) present the mean and standard errors of the variables for the treatment group, respectively. Column (5) shows the p-value associated with the hypothesis of the mean values across both groups being the same. We imputed the mean to have consistent sample sizes in the following variables: Panel E: Borrowing money support networks (N), Parenting issues support networks (N), Belong to parent support group (%). 48 Table 2: First Stage: ITT Impacts on Learning Information module (1) (2) (3) (4) (5) (6) (7) (8) (9) Praising Important for Clear Understand Calm down Withdraw Redirect Consequences children parents to instructions why child before attention from rather than and timeout Information helps play with child help misbehaves disciplining child’s whining reprimand appropriate Index Panel A: Short-term Treatment 0.189∗∗∗ 0.059 0.224∗∗∗ 0.063∗ -0.011 0.624∗∗∗ 0.278∗∗∗ 0.080∗∗ 0.525∗∗∗ (0.05) (0.04) (0.05) (0.04) (0.04) (0.07) (0.05) (0.04) (0.07) Observations 978 979 978 974 974 971 973 971 979 Control mean 4.13 4.27 3.98 4.29 4.38 2.73 3.74 4.16 -0.00 Panel B: Medium-term Treatment 0.104∗∗ 0.039 0.248∗∗∗ -0.058 -0.026 0.480∗∗∗ 0.165∗∗∗ 0.071 0.394∗∗∗ (0.05) (0.04) (0.06) (0.04) (0.04) (0.08) (0.06) (0.04) (0.08) Observations 698 699 699 699 698 691 691 697 699 Control mean 4.196 4.290 3.989 4.408 4.450 2.743 3.810 4.164 0.000 49 Notes: This table presents estimates of β1 (the ITT estimate) from Equation 1. Each column is a separate dependent variable. Panel A displays the short-term effects and Panel B the medium-term effects. Columns 1 to 8 show the impact of treatment over eight statements relating to parenting practices. Each statement is designed to evaluate the understanding of the four key concepts of the intervention. All outcome variables are Likert-scale variables ranging from 1 (Strongly disagree) to 5 (Strongly agree). Column 9 presents the ITT impacts of the treatment on the information index, which aggregates the eight outcome statements as described in Section 4. All specifications include controls for strata fixed effects. The control mean in Panels A and B refer to the mean of the control group from the first and second round of data, respectively. Number of observations vary across variables due to differences in response rate. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficient estimates. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table 3: ITT Impacts on Caregiver Attitudes & Behaviors Primary hypotheses (1) (2) (3) (4) (5) (6) Attitudes toward Attitudes toward Violence Psychological Attitudes toward violence against physical violence pychological violence against Physical violence violence against children against children against children target child against target child target child (index) (index) (index) (index) (index) (index) Panel A: Short-term Treatment -0.198∗∗∗ -0.192∗∗∗ -0.120∗∗ -0.121∗∗ -0.136∗∗ -0.101∗ (0.05) (0.05) (0.06) (0.06) (0.06) (0.05) Observations 977 974 961 943 920 942 50 Control mean -0.000 -0.000 0.000 -0.000 0.000 -0.000 Panel B: Medium-term Treatment -0.144∗∗ -0.150∗∗ -0.075 -0.127∗∗ -0.124∗ -0.105 (0.06) (0.06) (0.07) (0.06) (0.07) (0.06) Observations 696 694 685 681 676 681 Control mean -0.000 -0.000 -0.014 0.000 0.017 0.008 Notes: This table presents estimates of β1 (the ITT estimate) from Equation 1. Each column is a separate dependent variable. Panel A displays the short-term effects and Panel B the medium-term effects. Columns (1)-(3) present treatment impacts on the caregiver attitudes while columns (4)-(6) present treatment impacts on caregiver behaviors. All dependent variables are index variables, constructed as described in Section 4. All specifications include controls for strata fixed effects and the baseline dependent variable as a control. The control mean in Panels A and B refer to the mean of the control group from the first and second round of data, respectively. Number of observations vary across variables due to differences in response rate. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficient estimates. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table 4: ITT Impacts on Caregivers’ Mental Health & Child Outcomes Caregivers’ Mental Health Child Outcomes (1) (2) (3) (4) (5) Parental Conduct Emotional Depression Anxiety stress scale problems problems (index) (index) (index) (index) (index) Panel A: Short-term Treatment -0.033 -0.070 -0.030 -0.166∗∗∗ (0.06) (0.06) (0.05) (0.05) Observations 982 982 961 961 Control mean 0.000 -0.000 0.000 0.000 Panel B: Medium-term 51 Treatment -0.124∗ -0.157∗∗ -0.156∗∗ -0.053 -0.051 (0.07) (0.07) (0.08) (0.06) (0.07) Observations 699 699 699 685 685 Control mean 0.000 -0.000 0.000 0.000 0.000 Notes: This table presents estimates of β1 (the ITT estimate) from Equation 1. Each col- umn is a separate dependent variable. Panel A displays the short-term effects and Panel B the medium-term effects. Columns (1)-(3) present treatment impacts on the caregivers’ mental health while columns (4)-(5) present treatment impacts on child conduct and emo- tional problems. The parental stress scale was only measured at second follow-up. All dependent variables are index variables, constructed as described in Section 4. All specifi- cations include controls for strata fixed effects and columns (1), (2), (4), and (5) additionally include the baseline dependent variable as a control. The control mean in Panels A and B refer to the mean of the control group from the first and second round of data, respec- tively. Number of observations vary across variables due to differences in response rate. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficient estimates. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table 5: First Stage: Take-up of Intervention Variable Mean Std. Dev. Min. Max. N Panel A. SMS Delivery (Admin Data) Sent SMS (%) 92.28 6.42 69 98 30 Panel B. SMS/WhatsApp Receipt (Survey Data) Received any SMS/WhatsApp (%) 91.38 28.09 0 100 499 Read SMS/WhatsApp if received (%) 96.94 17.23 0 100 458 Found the SMS/WhatsApp useful if read (%) 98.20 13.32 0 100 444 Panel C. App usage (Admin Data) Number of sessions accessed 1.04 1.85 0 10 557 Total time in sessions (mins) 6.94 15.58 0 75 557 Panel D. Virtual sessions (Admin Data) Number of sessions attended 4.55 3.44 0 10 557 Notes: This table shows descriptive statistics for the take-up of relevant outcomes for each of the three components of the intervention. The table uses survey data on the reception of the SMS messages, App usage data from the phone company, and ECC officer reports of attendance at the virtual sessions. The unit of observation is a message in “Sent SMS (%)” and a treated caregiver in all other variables. 52 Table 6: Social Desirability Bias Analysis Primary hypotheses (1) (2) (3) (4) (5) (6) (7) Attitudes toward Attitudes toward Violence Physical violence Psychological Attitudes toward violence against physical violence psychological violence against against target violence against SDB children against children against children target child child target child (index) (index) (index) (index) (index) (index) (index) Panel A: Controlling for SDB Treatment -0.016 -0.145∗∗ -0.151∗∗ -0.076 -0.132∗∗ -0.120∗ -0.115∗ (0.08) (0.06) (0.06) (0.07) (0.06) (0.07) (0.06) SDB (index) -0.047 -0.050 -0.050 -0.019 -0.027 -0.033 (0.03) (0.03) (0.04) (0.03) (0.03) (0.03) Observations 700 696 694 685 677 672 677 Control Mean 0.000 -0.000 -0.000 -0.014 0.001 0.016 0.009 Panel B: Heterogeneity by SDB Treatment -0.243∗∗∗ -0.158∗ -0.204∗∗ -0.128 -0.103 -0.079 53 (0.09) (0.09) (0.10) (0.09) (0.10) (0.10) High SDB Score -0.223∗∗ -0.128 -0.271∗∗∗ -0.138 -0.119 -0.112 (0.09) (0.09) (0.10) (0.10) (0.11) (0.10) Treatment × High SDB Score 0.198∗ 0.015 0.256∗ -0.009 -0.036 -0.073 (0.12) (0.13) (0.14) (0.13) (0.14) (0.13) Treat + Treat × High SDB score -0.045 -0.143∗ 0.053 -0.138∗ -0.138 -0.152∗ (0.08) (0.08) (0.09) (0.08) (0.09) (0.09) Observations 696 694 685 677 672 677 Control Mean -0.000 -0.000 -0.014 0.001 0.016 0.009 Notes: Panel A presents estimates of β1 (the ITT estimate) from Equation 1, additionally controlling for SDB (index). Panel B presents heterogeneity by SDB and reports estimates of β1 , β2 , and β3 (the ITT estimates) from the following equation: Yi,t = β0 + β1 Ti + β2 HighSDBi + β3 Ti ∗ HighSDBi + β4 Yi,t−1 + γs + εi,t . “High SDB score” is a dummy variable that takes the value 1 if the SDB score was above the median SDB score for the sample, and 0 otherwise. Each column is a separate dependent variable. Medium-term effects are shown, estimated using data from the second follow-up and baseline. Column (1) presents treatment impacts on the Social Desirability Bias (index). Columns (2)-(7) of this table show our results on caregiver attitudes and behaviors controlling for SDB (index). The SDB score was only measured at second follow-up. The term “Treat + Treat x High SDB score” denotes the total effect of the treatment for those in the treatment group with above- median SDB scores. All dependent variables are index variables, constructed as described in Section 4. All specifications include controls for strata fixed effects, while specifications (2) - (7) additionally include the baseline dependent variable as a control. The estimation sample for columns (5)-(7) uses 4 fewer observations relative to the estimation sample for columns (4)-(6) in Table 3 as SDB measures were not collected for the 4 caregivers. The control mean refers to the mean of the control group from the second follow-up round of data. The sample size in each specification varies according to the number of observations available for each outcome and to the SDB index. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficient estimates. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table 7: Attrition Analysis First follow-up Second follow-up (1) (2) (3) (4) In endline In endline In endline In endline Treatment 0.047∗∗ -0.090 0.008 -0.273 (0.02) (0.18) (0.03) (0.24) Treatment × Age (years) 0.003 0.004 (0.00) (0.00) Treatment × Female (%) -0.033 0.050 (0.07) (0.09) Treatment × Education level completed (years) 0.009 0.001 (0.01) (0.01) Treatment × Married (%) -0.024 -0.038 (0.05) (0.07) Treatment × Employed (%) 0.021 0.057 (0.05) (0.07) Treatment × Income in the past month (USD) -0.000 0.000 (0.00) (0.00) Treatment × Household size (N) -0.009 -0.016 (0.02) (0.02) Treatment × Children 17 years or younger (N) 0.004 0.039 (0.03) (0.04) Treatment × Violence against target child (index) 0.007 -0.027 (0.02) (0.03) Treatment × Depression (index) -0.025 -0.023 (0.02) (0.03) Treatment × Anxiety (index) -0.010 0.009 (0.02) (0.03) Treatment × Conduct problems (index) -0.016 -0.017 (0.02) (0.03) Treatment × Emotional problems (index) 0.023 0.014 (0.02) (0.03) Treatment × Attitudes toward violence against children (index) -0.001 0.015 (0.02) (0.03) Constant 0.843∗∗∗ 0.684∗∗∗ 0.633∗∗∗ 0.377∗ (0.02) (0.15) (0.02) (0.20) Observations 1113 1113 1113 1113 Q test pvalue 0.897 0.916 Notes: This table uses data from the first and second follow-up rounds to show the differences in attrition between treatment and control groups for each follow-up. We present estimates of β1 (the ITT estimate) from Equation 1. The dependent variable “In Endline” in all columns is a dummy indicating if a caregiver responded to the follow-up surveys. Models 1 and 3 measure the impact of the treatment on the follow-up survey respondent. Models 2 and 4 measure the impact of any demographic characteristics or outcome variables measured at the baseline on the probability of complet- ing the follow-ups (these variables were also included on their own but their output has been suppressed for brevity). All regressions include strata fixed-effects. Heteroskedasticity-robust standard errors are reported in parenthesis below the coefficient estimates. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. 54 Table 8: Conditional Average Treatment Effect Regressions Information Attitude toward violence Violence against target index against children (index) child (index) (1) (2) (3) (4) (5) (6) (7) (8) (9) Above Below Above Below Above Below Median Median Difference Median Median Difference Median Median Difference CATE CATE CATE CATE CATE CATE Treatment 0.822*** 0.490*** 0.333 0.05 -0.514*** 0.564*** -0.092 -0.212 0.12 (0.169) (0.170) (0.240) (0.087) (0.163) (0.185) (0.135) (0.154) (0.205) Observations 1080 1080 1080 1079 1079 1079 1044 1044 1044 Control Mean 0.002 0.002 0.002 0.042 0.042 0.042 0.007 0.007 0.007 Notes: This table uses data from the first and second follow-up rounds and presents results of average treatment effects (ATEs) for caregivers based on groups defined by caregivers with high and low conditional average treatment effects (CATEs, estimated using grf model in R) for the three primary outcomes. All dependent variables are index variables, constructed as described in Section 4. Data from the short- and medium-term follow-ups have been pooled for statistical 55 power. Above (below) median groups represent caregivers whose CATEs are greater than (less than) the median esti- mated CATEs. The sample size in each specification varies according to the number of observations available for each outcome. Heteroskedasticity-robust standard errors are reported in parenthesis. * p < 0.10, ** p < 0.05, *** p < 0.01. Table 9: Heterogeneity in Treatment Impacts: Interaction Terms Approach Attitudes toward violence against children (index) (1) (2) (3) (4) (5) Panel A: Short-term Treatment -0.241* -0.213** -0.099* -0.094 -0.107 (0.13) (0.09) (0.06) (0.07) (0.07) Treatment × Female caregiver 0.051 (0.15) Treatment × High income 0.027 (0.11) Treatment × Worse attitudes toward violence against children -0.185* (0.10) Treatment × More conduct problems -0.203** (0.10) Treatment × More emotional problems -0.175* (0.10) Observations 977 774 977 977 977 Panel B: Medium-term 56 Treatment -0.058 -0.181** -0.032 -0.213** -0.198** (0.17) (0.11) (0.07) (0.09) (0.08) Treatment × Female caregiver -0.100 (0.18) Treatment × High income 0.102 (0.14) Treatment × Worse attitudes toward violence against children -0.206* (0.12) Treatment × More conduct problems 0.136 (0.12) Treatment × More emotional problems 0.111 (0.12) Observations 696 555 696 696 696 Notes: This table presents treatment heterogeneity of our results for caregiver attitudes toward violence against children (VAC). We present estimates of β1 and β3 (the ITT estimates) from the following equation: Yi,t = β0 + β1 Ti + β2 V ari + β3 Ti ∗ V ari + β4 Yi,t−1 + γs + εi,t . The variables (V ari ) “high income”, “worse attitudes toward VAC”, “more conduct problems”, and “more emotional problems” are indicator variables equal to one for above-median values of the underlying variables. The variables that were interacted with “treatment” were also included in the regressions, but their output has been suppressed for brevity. Panel A displays the short-term effects and Panel B the medium-term effects. The dependent variable is an index variable, constructed as described in Section 4. All specifications include controls for strata fixed effects and the baseline dependent variable. The control mean in Panels A and B refer to the mean of the control group from the first and second round of data, respectively. The sample size in each specification varies according to the number of observations available for each outcome. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficient estimates. * p < 0.10, ** p < 0.05, *** p < 0.01. Table 10: Cost-Effectiveness Comparison (1) (2) (3) (4) Cost Treatment Type of Cost per per person effect violence 0.13 SD effect (USD) (USD) vIHT 62.4 -0.13 SD Violence against (target) child 62.40 -0.12 SD Physical violence against (target) child 67.60 -0.11 SD Psychological violence against (target) child 73.75 IHT 123.95 -0.29 SD Physical violence against children 55.56 Cash transfer intervention 648.72 -0.13 SD Physical violence against female partner 648.72 -0.12 SD Psychological violence against female partner 702.78 -0.17 SD Physical violence against children 496.08 57 Notes: This table presents a cost effectiveness comparison between the vIHT and other two interventions: face-to-face IHT in Jamaica (Francis and Baker-Henningham, 2021) and a cash transfer intervention in Mali (Heath et al., 2020). Costs for the vIHT excludes the cost of setting up the App (USD 54.21). Column (1) presents the cost of the intervention (in USD) per person. Columns (2) and (3) present the estimated treatment effect on different violence-related outcomes, respectively. Column (4) presents the cost (in USD) per an average effect of 0.13 SD. Appendix Figures & Tables For Online Publication Only Figure A1: Features Available in the App for Week 4 58 Figure A2: Enrollment of Participants - SMS Snapshot Figure A3: Enrollment of Participants 59 Figure A4: Enrollment Survey Figure A5: Distribution of Participants Compared to Population 6.1 Clarendon 9.1 1.1 Hanover 2.6 31.5 Kingston & Saint Andrew 24.6 3.6 Manchester 7.0 2.2 Portland 3.0 4.9 Saint Ann 6.4 29.6 Saint Catherine 19.1 2.7 Saint Elizabeth 5.6 8.6 Saint James 6.8 2.9 Saint Mary 4.2 2.2 Saint Thomas 3.5 2.1 Trelawny 2.8 2.5 Baseline Westmoreland 5.3 Population 0 10 20 30 Percentage (%) Notes: This figure presents the distribution of study participants compared to the population of caregivers in Jamaica. We report Kingston and St. Andrew as a combined parish because the Kingston urban area includes both parishes. 60 Figure A6: Dose-response Regressions on Information Index by Number of Sessions Attended 1.5 1 p = 0.000/ b = 0.814 p = 0.001/ b = 0.625 p = 0.005/ b = 0.507 .5 p = 0.002/ b = 0.536 p = 0.001/ b = 0.544 p = 0.423/ b = 0.173 p = 0.358/ b = 0.186 p = 0.613/ b = 0.097 p = 0.506/ b = 0.144 Short-term 0 Medium-term p = 0.362/ b = -0.170 -.5 -1 -1.5 1 or 2 3 or 4 5 or 6 7 or 8 9 or 10 sessions sessions sessions sessions sessions Notes: This figure plots point estimates and corresponding 95% confidence intervals from OLS regressions of the information index (dependent variable) on the number of sessions attended by caregivers (the independent variable). These dose-response regressions are only run with caregivers in the treatment group. To improve precision, we group the number of sessions attended into five categories (with attendance at zero sessions being the omitted base category). The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey. The gray triangles and corresponding dotted lines represent the point estimates and 95% confidence intervals from the second follow-up survey. Standard deviation units are used for the y-axis. 61 Figure A7: ITT Impacts on Caregiver Attitudes Toward Violence Against Children (Balanced Panel) Attitudes toward p = 0.000/ b = -0.325 violence against children (index) p = 0.023/ b = -0.141 Attitudes toward p = 0.000/ b = -0.294 physical violence Short-term against children Medium-term (index) p = 0.031/ b = -0.138 Attitudes toward p = 0.001/ b = -0.232 psychological violence against children (index) p = 0.322/ b = -0.069 -.5 -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 .5 Notes: This figure presents estimates of β1 (the ITT estimate) on caregivers’ attitudes toward violence against children. This figure uses a balanced panel of caregivers who were present at both the first and second follow-ups. Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey. The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow-up survey. Standard deviation units are used for the x-axis. 62 Figure A8: ITT Impacts on Caregiver Behaviors Related to Violence Against Target Child (Balanced Panel) p = 0.023/ b = -0.163 Violence against target child (index) p = 0.096/ b = -0.107 Physical violence p = 0.023/ b = -0.175 Short-term against target child (index) p = 0.121/ b = -0.115 Medium-term Psychological p = 0.025/ b = -0.154 violence against target child (index) p = 0.214/ b = -0.083 -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 Notes: This figure presents estimates of β1 (the ITT estimate) on caregivers’ behaviors relating to violence against target child. This figure uses a balanced panel of caregivers present at both first and second follow-ups. Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description of the indices, see Section 3.3. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey. The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow-up survey. Standard deviation units are used for the x-axis. 63 Figure A9: ITT Impacts on Caregivers’ Mental Health (Balanced panel) p = 0.306/ b = -0.073 Depression (index) p = 0.094/ b = -0.114 p = 0.206/ b = -0.090 Short-term Anxiety (index) p = 0.014/ b = -0.168 Medium-term Parental stress scale (index) p = 0.038/ b = -0.156 -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate) on caregivers’ depression (measured using PHQ-2 at both follow-ups), anxiety (measured using GAD-2 at both follow-ups), and parental stress (measured using PSS-18 only at second follow-up). This figure uses a balanced panel of caregivers present at both first and second follow-ups. Each outcome consists of a standardized index estimated following Anderson (2008) and standardized relative to the control group. For a detailed description, see Section 3.3. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey. The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow-up survey. Standard deviation units are used for the x-axis. 64 Figure A10: ITT Impacts on Child Conduct and Emotional Problems (Balanced Panel) p = 0.535/ b = -0.038 Conduct problems (index) p = 0.270/ b = -0.067 Short-term Medium-term p = 0.001/ b = -0.209 Emotional problems (index) p = 0.278/ b = -0.073 -.3 -.2 -.1 0 .1 .2 .3 Notes: This figure presents estimates of β1 (the ITT estimate) on the target child’s conduct and emotional problems. This figure uses a balanced panel of caregivers present at both first and second follow-ups. Conduct and emotional problems were measured using caregivers’ responses to the Strengths and Difficulties Questionnaire (SDQ) instru- ment. Each outcome consists of a standardized index estimated following Anderson (2008) and standardized rela- tive to the control group. For a detailed description, see Section 3.3. The black circles and corresponding solid lines represent the point estimates and 95% confidence intervals from the first follow-up survey. The gray triangles and corresponding dashed lines represent the point estimates and 95% confidence intervals from the second follow-up survey. Standard deviation units are used for the x-axis. 65 Table A1: Structure of Virtual Parenting Programme Session TOPICS COVERED EACH SESSION PRAISING YOUR CHILD - Importance of praising your child Session 1 - How to praise your child - What to praise your child for PRAISING YOURSELF: Importance of praising yourself for being a good parent INTRODUCING IRIE TIME - The importance of Irie Time - How to follow your child’s lead in play Session 2 - How to talk about what your child is doing - Using Respond, Describe, and Praise when playing with your child - Ideas for Irie Time activities GIVING YOUR CHILD POSITIVE ATTENTION THROUGHOUT THE DAY -Paying attention to positive behaviour during daily routines -Getting children involved in chores Session 3 - Using Describe, Respond, and Praise during daily activities MODELLING: Modelling the behaviour you want / Being a good role model IRIE TIME: Playing with toys with your child GIVING CLEAR INSTRUCTIONS - How to give clear instructions (clear, specific, short, positively phrased, realistic, get child attention first) Session 4 - Using labelled praise to praise child when they follow an instruction KNOW YOUR CHILD: Understand what your child likes/dislikes and factors that affect his/her behaviour IRIE TIME: looking at books with your child (using ‘Going to School’ book) TEACHING YOUR CHILD NEW SKILLS: - teach children how to follow the rules and expectations in the household. INDEPENDENCE: giving children independence Session 5 CHOICES: -giving children choices. IRIE TIME: colouring with your child REASONS WHY CHILDREN MISBEHAVE: Identify why children misbehave Session 6 ME-TIME: the importance of taking time to do something that you like to do. IRIE TIME: playing outside (ball, skipping, chasing) – how to do ‘outside play activities’ during Irie Time. MANAGING YOUR EMOTIONS -How we feel affects the way we behave. -Link back to the importance of trying to understand why children misbehave -How to recognize our own emotions. Session 7 -How to calm down when feeling angry. HELPING CHILDREN UNDERSTAND THEIR OWN EMOTIONS -Labelling children’s emotions IRIE TIME: Looking at books with your child (Use Emotions Book) HOW TO MANAGE YOUR CHILD’S MISBEHAVIOUR USING WITHDRAW ATTENTION AND REDIRECT -How to redirect children Session 8 -How to withdraw attention from attention seeking behaviours. -How to use redirect and withdraw attention together. IRIE TIME: singing & dancing during Irie Time HOW TO MANAGE YOUR CHILD’S BEHAVIOUR USING CONSEQUENCES AND CHILLAX -How to use chillax Session 9 -Giving children consequences -Problem solving IRIE TIME: pretend play (how to do pretend play activities during Irie Time) REVIEW OF THE YELLLOW AND GREEN BLOCKS OF THE IRIE TOWER Session 10 I AM AN IRIE PARENT GOAL SETTING 66 Table A2: Structure of Virtual Parenting Programme Week # Date Message Message 1 Children love to be praised. When you praise your child for doing something good, they will want to do it again and again. 1 Sept 20 – 24 Learn more at: LINK HERE FOR SESSION 1 Message 2 When praising, describe exactly what your child did, praise him/her and use your child’s name. Add a clap or hug to make praise extra special. LINK HERE FOR SESSION 1 Message 3 Irie Challenge for the week: Praise your child every day for all the good things they do. Praise yourself for being an Irie Parent. LINK HERE FOR SESSION 1 Message 1 Irie time is when we play and have fun with our child, doing what they want. Irie Time makes our child feel special and loved. 2 Sept 27 - Oct 1 LINK HERE FOR SESSION 2 Message 2 During Irie Time, we can play with toys, look at books and play games. We let our child choose what they want to do and follow their lead. LINK HERE FOR SESSION 2 Message 3 Irie Challenge for the week: Have Irie Time with your child for at least 10 minutes every day. Have fun! Great job for being an Irie Parent. LINK HERE FOR SESSION 2 Message 1 Give your child positive attention and praise throughout the day. This will help your child learn to behave well and to learn new things. 3 Oct 4 - 8 LINK HERE FOR SESSION 3 67 Message 2 We are role models for our children. Children copy our behaviour. We need to speak and act in ways that we want our child to speak and act. LINK HERE FOR SESSION 3 Message 3 Irie Challenge this week: Give your child positive attention throughout the day. Have Irie Time every day. Good job for being an Irie Parent. LINK HERE FOR SESSION 3 Message 1 Try to give your child clear instructions and praise them whenever they follow your instruction. Praise encourages positive behaviour. 4 Oct 11 – 15 LINK HERE FOR SESSION 4 Message 2 As parents we know our child best. We know when they are most likely to misbehave. This can help us to prevent bad behaviour. LINK HERE FOR SESSION 4 Message 3 Irie Challenge this week: Give your child clear instructions and praise them when they do what you say. Have Irie Time every day. Awesome Job. LINK HERE FOR SESSION 4 Message 1 We can teach our child the little rules we have in our house and teach them important daily skills. This makes our life easier. 5 Oct 18 – 22 LINK HERE FOR SESSION 5 Message 2 We can help our child to behave well and learn well by giving them simple choices and allowing them some independence. This makes them feel good. LINK HERE FOR SESSION 5 Message 3 Irie Challenge: Teach your child one skill this week. Give your child choices and some independence. Praise yourself for being an Irie Parent. LINK HERE FOR SESSION 5 Message 1 To prevent our child from misbehaving, we need to understand the reason why our child is behaving in a certain way. 6 Oct 25 - 29 LINK HERE FOR SESSION 6 Message 2 As parents we are very busy. We need to take some time out for ourselves. In Me Time we do something that relaxes us and makes us feel happy. LINK HERE FOR SESSION 6 Message 3 Irie Challenge: If your child misbehaves, try to understand why so you can prevent the behavior next time. Have some Me Time every day. LINK HERE FOR SESSION 6 Message 1 When we feel angry with our child, we can stop, think and calm down before dealing with the situation. We can be a good role model for our child. 7 Nov 1 – 5 LINK HERE FOR SESSION 7 Message 2 We can help our child understand and manage their feelings by naming the emotion and describing why our child is feeling that particular emotion. LINK HERE FOR SESSION 7 Message 3 Irie Challenge: Find ways to calm down when you are angry. Name your child’s emotions and explain why they feel that way. You are an Irie Parent. LINK HERE FOR SESSION 7 Message 1 Sometimes our child misbehaves when they are having fun, exploring & copying us. We can redirect our child’s attention away from these behaviors. 8 Nov 8 - 12 LINK HERE FOR SESSION 8 Message 2 Children may try to get attention by crying for things, complaining, and nagging. We can withdraw attention from these behaviors. LINK HERE FOR SESSION 8 68 Message 3 Irie Challenge: Use redirect and withdraw attention to deal with misbehavior. Praise your child when they behave well. Great job Irie Parent. LINK HERE FOR SESSION 8 Message 1 We can give consequences for more serious misbehaviour. Consequences work best when they are short and not too harsh. We can also use Chillax. 9 Nov 15 - 19 LINK HERE FOR SESSION 9 Message 2 Chillax & consequences work best at managing our child’s behaviour when we give them praise throughout the day for the good things they do. LINK HERE FOR SESSION 9 Message 3 Irie Challenge: Give your child positive attention & praise for all the good things they do throughout the day. Awesome! You are an Irie Parent. LINK HERE FOR SESSION 9 Message 1 Irie parents make Irie Homes for their children. Children feel safe, secure and loved and they have lots of opportunities to learn and play. 10 Nov 22 - 26 LINK HERE FOR SESSION 10 Message 2 Go back through the previous sessions. Set yourself parenting challenges each week to help you to continue making an Irie Home. Great job! LINK HERE FOR SESSION 10 Message 3 CONGRATULATIONS. You have completed the Irie Homes Toolbox sessions. Praise yourself for being an Irie Parent and for making an Irie Home. LINK HERE FOR SESSION 10 Table A3: Strata Composition Strata Frequency Percentage SMS - female 875 78.62 SMS - male 163 14.65 Principals or social media - female 70 6.29 Principals or social media - male 5 0.45 Total 1,113 100.00 Notes: This table shows the gender decomposition of the differ- ent recruitment sources for all participants in the intervention. The first two rows correspond to the participants recruited via SMS messages. Rows three and four correspond to participants recruited via school principals or social media. The columns show the absolute and relative frequencies associated with those variables. 69 Table A4: Text Messages for Control Group No. Message 1 Thank you for participating in our parenting survey. You will now receive 3 SMS/week with tips on how to keep you and your child safe from COVID. Stay safe 2 COVID-19 Tip: Remind children to avoid sharing food, toys, pencils, books with friends. 3 COVID-19 Tip: Encourage children to wash their hands often with soap and water. 4 COVID-19 Tip: Remind children to avoid touching their face during COVID-19. 5 COVID-19 Tip: Wash your hands regularly when interacting with children 6 COVID-19 Tip: Avoid crowded spaces and close contact with others when traveling with children. 7 COVID-19 Tip: Keep children a safe distance from anyone with a cold or flu symptoms. 8 COVID-19 Tip: Remind children two years and over to avoid touching masks to reduce risks of contamination. 9 COVID-19 Tip: Remember to have children two years or older wear a mask outdoors and supervise mask wearing. 10 COVID-19 Tip: Show children 2 years and older, proper way to wear a mask and supervise them during mask wearing. 11 COVID-19 Tip: For children over two years, discard single use masks after each use and throw mask away without delay 12 COVID-19 Tip: Teach children proper ways to cover nose and mouth, when sneezing or coughing. 13 COVID-19 Tip: Remember to follow the COVID-19 protocols of the Ministries of Health and Education 14 COVID-19 Tip: Ensure children two years and older wear mask properly over nose, mouth, chin, it is secure, and they are supervised. 15 COVID-19 Tip: Remember to regularly wash reusable masks used by children over two years old. 70 16 COVID-19 Tip: Remove face masks of children in the right way, by removing from the ties and not touching the front. 17 COVID-19 Tip: For children over two years, remind them to wash their hands with soap and water after touching a used mask. 18 COVID-19 Tip: Ensure children have their own resources, to limit sharing with others 19 COVID-19 Tip: Remind children to avoid sharing food, toys, pencils, books with friends. 20 COVID-19 Tip: Dispose of single use face masks right after removal in a closed bin. 21 COVID-19 Tip: Remember to keep child at home or see a doctor if they are unwell with fever and cough. 22 COVID-19 Tip: Avoid crowded spaces and close contact with others when traveling with children. 23 COVID-19 Tip: Teach children to throw tissues used for sneezing or coughing into closed bin right after use, and wash hands. 24 COVID-19 Tip: Remember to teach your children over two years the proper way to wash their hands. 25 COVID-19 Tip: Remind children to keep safe distance from non-family members. 26 COVID-19 Tip: Remind children two years and over to wash their hands after coughing or sneezing. 27 COVID-19 Tip: Remember to clean your phones before giving to children to play. 28 COVID-19 Tip: Clean and disinfect high touch areas around the home used by children 29 COVID-19 Tip: Teach children how to properly wash hands with soap and water. 30 COVID-19 Tip: Regularly disinfect or wash toys and resources of children. 31 COVID-19 Tip: Wash your hands regularly when interacting with children. Notes: This table enumerates the different COVID-19-related weekly SMS tips received by the caregivers in the control group. Table A5: Survey Modules Survey Modules Baseline First Follow up Second Follow up Caregivers Outcomes Attitudes towards violence against children X X X Violence against target child X X X Depression and anxiety X X X Parental Stress Scale X Child Outcomes Conduct and emotional problems (SDQ) X X X Mechanisms Brief Parental Self-Efficacy Scale (BPSES) X Parental Self-Efficacy [From TOPSE - Discipline & Self-Acceptance] X Support networks X X Caregiver and target child socio-demographic characteristics Household Roster X X X Social Desirability Bias X Intervention take up and learning Information module X X Receipt of parenting support (+ take-up) X X Notes: This table shows if the data for each of the survey modules was collected at baseline and/or during the first or second follow up. 71 Table A6: Comparison of Study Sample with Representative Survey Mean JLCS 2019 Mean Study Sample p-value Panel A. Caregiver’s characteristics Age (years) 36.9 33.2 0.000 (11.4) (7.49) Female (%) 0.89 0.85 0.076 (0.31) (0.35) Education level completed (years) 13.3 14.3 0.000 (2.64) (2.75) Married (%) 0.21 0.37 0.000 (0.41) (0.48) Employed (%) 0.62 0.79 0.000 (0.48) (0.41) Household size (N) 4.61 4.58 0.793 72 (1.86) (1.95) Children 17 years or younger (N) 2.23 1.91 0.000 (1.19) (1.03) Panel B. Target child’s characteristics Age (years) 4.16 4.12 0.651 (1.37) (1.43) Female (%) 0.47 0.49 0.518 (0.50) (0.50) Notes: This table compares relevant descriptive statistics between our main sample and Ja- maica Survey of Living Conditions restricted to caregivers with a child aged between two and six years. Panel A contains demographic variables associated with the caregiver. Panel B shows information about the eldest child aged two to six years old for each caregiver in the JLCS (2019) sample and information about the target child in the study sample. Standard deviations are presented in parentheses below the means. The last column shows the p-value associated with the null hypothesis of the mean values across both groups being the same. Table A7: Summary Statistics of Violence Against Target Child and Caregivers’ Mental Health in the Control Group First Second Variable Baseline Follow-up Follow-up Violence against target child Shouted, yelled, or screamed at him/her? 1.54 1.49 1.49 (1.39) (1.32) (1.42) Said you would send him/her away? 0.23 0.18 0.20 (0.72) (0.66) (0.69) Threatened to hit him/her but not actually done it? 1.48 1.54 1.58 (1.50) (1.49) (1.53) Hit him/her on the bottom, hand, arm, or leg with your bare hand? 0.52 0.52 0.51 (0.79) (0.75) (0.86) Hit him/her on the bottom, hand, arm, or leg with a hard object? 0.08 0.09 0.11 (0.36) (0.35) (0.38) Caregivers’ mental health Have you been feeling down, depressed, or hopeless? 0.62 0.58 0.62 (0.91) (0.87) (0.95) Have you been feeling nervous, anxious or on edge? 0.46 0.39 0.51 (0.83) (0.75) (0.84) Have you been feeling little interest or pleasure in doing things? 0.69 0.65 0.76 (0.99) (0.92) (1.00) Have you not been able to stop or control worrying? 0.55 0.55 0.68 (0.90) (0.91) (1.04) Notes: This table presents mean and SD values for violence against target child and caregivers’ mental health at baseline, first, and second follow-up for the control group only. Standard deviations are presented in paren- theses below the means. 73 Table A8: Comparison of Treatment Impacts Between First & Second Follow-ups First follow-up Second follow-up Variable βf 1 SEf 1 βf 2 SEf 2 p-value βf 1 − βf 2 Panel A. First Stage: Learning Praising children helps 0.189 0.045 0.104 0.049 0.203 Important for parents to play with child 0.059 0.038 0.039 0.044 0.729 Clear instructions help 0.224 0.046 0.248 0.060 0.753 Understand why child misbehaves 0.063 0.037 -0.058 0.038 0.023 Calm down before disciplining -0.011 0.039 -0.026 0.040 0.788 Withdraw attention from child’s whining 0.624 0.068 0.480 0.079 0.168 Redirect rather than reprimand 0.278 0.053 0.165 0.058 0.150 Consequences and timeout appropriate 0.080 0.040 0.071 0.044 0.870 Information index 0.525 0.070 0.394 0.080 0.217 Panel B. Primary Outcomes Attitudes towards violence against children (index) -0.198 0.050 -0.144 0.061 0.497 Attitudes towards physical violence against children (index) -0.191 0.051 -0.150 0.063 0.615 Attitudes towards physchological violence against children (index) -0.120 0.058 -0.075 0.068 0.612 Violence against target child (index) -0.121 0.056 -0.127 0.062 0.950 Physical violence against target child (index) -0.136 0.060 -0.124 0.069 0.897 Psychological violence against target child (index) -0.101 0.055 -0.105 0.065 0.961 Panel C. Secondary Outcomes Depression (index) -0.033 0.059 -0.124 0.067 0.309 Anxiety (index) -0.070 0.059 -0.157 0.068 0.330 Conduct problems (index) -0.030 0.049 -0.053 0.059 0.760 Emotional problems (index) -0.166 0.053 -0.051 0.065 0.170 Notes: This table presents a comparison of treatment impacts between the first and second follow-ups as estimated using our main specification (1). β refers to the estimated coefficients and SE refers to the estimated heteroskedasticity-robust standard errors. The last column presents the p-value for a test of the difference in means between the estimated coefficients for the first and second follow-ups. 74 Table A9: ITT Impacts on Caregiver Attitudes and Violent Behaviors Against Children (First Follow-up) Caregivers survey (1) (2) (3) (4) (5) Shout, yell, threaten Physical punishment Slap when Beat to stop Shout and yell to slap needed misbehave unwanted behavior for obedience not harmful Panel A: Attitudes towards violence against children Treatment -0.039∗∗ -0.062∗∗ -0.051∗∗∗ -0.023 -0.027 (0.02) (0.02) (0.02) (0.02) (0.03) Observations 916 847 882 886 870 Control Mean 0.121 0.226 0.097 0.077 0.329 Hit with Hit with Threaten to send Threaten to bare hand an object Yelled child away hit Panel B: Violence against target child Treatment -0.097∗∗∗ -0.026 -0.101∗∗∗ 0.040∗∗ -0.055∗ (0.03) (0.02) (0.03) (0.02) (0.03) Observations 916 885 934 884 928 Control Mean 0.395 0.074 0.745 0.090 0.675 Notes: This table shows the estimated short-term impacts for each of the components in the caregivers’ attitude and behavior indices described in Figure 3. We present the estimated coefficient β1 from the specification (1) in all the panels. "Treatment" is a dummy variable taking a value of one if the observed caregiver is in the treatment group, while "Control Mean" is the mean of the outcome for the control group. The sample size in each specification varies according to the number of observations available for each outcome. All results correspond to the first follow-up of the intervention. Since all dependent variables are measured as dummy variables whose descriptions are available in Appendix A2, these results are the extensive margin impacts. Panel A describes the results of caregivers’ attitudes toward violence against children. The first three columns correspond to attitudes toward physical violence, while the last two columns correspond to psychological violence. These components are framed as Yes/No questions regarding the following statements: (1) In order to bring up, raise, or educate a child properly, the child needs to be physically punished. (2) good parent slaps their child when they misbehave. (3) When a child is beaten, he/she will stop doing the unwanted behavior. (4) Shouting and yelling make the child more obedient. (5) Shouting, yelling, and threatening to slap will harm the child. Panel B shows the extensive margin short- term impacts for the five components that comprise the violence against the target child (index). These are framed as Agree/Disagree responses to the questions regarding the following statements: (1) hit the child on the bottom, hand, arm, or legs with their bare hands (2) hit the child with something like a belt, hairbrush, stick, or some other hard object, (3) shout, yell, or scream at their child, (4) threaten to send the child away in response to behavior perceived as inappropriate, (5) threaten to hit the child in response to behavior perceived as inappropriate. All specifications include strata-fixed effects for the four strata. Strata were defined as the cross between the gender of the caregiver (male or female) and the mode of recruitment into the study (SMS messages campaign or ECC/Principal referral and social media campaign). The sample size in each specification varies according to the number of observations available for each outcome. Heteroskedasticity-robust standard errors are reported in parenthesis. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. 75 Table A10: ITT Impacts on Caregiver Attitudes & Behaviors - Kling et al. (2007) Indices Primary hypotheses (1) (2) (3) (4) (5) (6) Attitudes toward Attitudes toward Violence Psychological Attitudes toward violence against physical violence pychological violence against Physical violence violence against children against children against children target child against target child target child (index) (index) (index) (index) (index) (index) Panel A: Short-term Treatment -0.201∗∗∗ -0.186∗∗∗ -0.110∗ -0.145∗∗∗ -0.142∗∗ -0.122∗∗ (0.05) (0.05) (0.06) (0.05) (0.06) (0.05) Observations 977 974 961 943 920 942 76 Control mean -0.002 -0.005 0.006 0.000 0.002 -0.000 Panel B: Medium-term Treatment -0.149∗∗ -0.148∗∗ -0.076 -0.140∗∗ -0.123∗ -0.110∗ (0.06) (0.06) (0.07) (0.06) (0.07) (0.06) Observations 696 694 685 681 676 681 Control mean -0.002 -0.002 -0.014 -0.000 -0.007 -0.000 Notes: This table presents estimates of β1 (the ITT estimate) from Equation 1. Each column is a separate dependent variable. Panel A displays the short-term effects and Panel B the medium-term effects. Columns (1)-(3) present treatment impacts on the caregiver attitudes while columns (4)-(6) present treatment impacts on caregiver behaviors. All dependent variables are index variables, constructed following Kling et al. (2007). All specifications include controls for strata fixed effects and the baseline dependent variable as a control. The control mean in Panels A and B refer to the mean of the control group from the first and second round of data, respectively. The sample size in each specification varies according to the number of observations available for each outcome. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficient estimates. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table A11: Social Desirability Bias Analysis - Violence-Specific SDB Items Primary hypotheses (1) (2) (3) (4) (5) (6) (7) Attitudes toward Attitudes toward Attitudes toward Violence Physical violence Psychological violence against physical violence psychological violence against against target violence against target SDB children against children against children target child child child (Violence index) (index) (index) (index) (index) (index) (index) Panel A: Controlling for SDB Treatment 0.027 -0.142∗∗ -0.149∗∗ -0.073 -0.131∗∗ -0.120∗ -0.113∗ (0.08) (0.06) (0.06) (0.07) (0.06) (0.07) (0.06) SDB (Violence index) -0.061∗ -0.050 -0.073∗∗ 0.010 0.003 -0.009 (0.03) (0.03) (0.04) (0.03) (0.03) (0.03) Observations 700 696 694 685 677 672 677 Control Mean 0.000 -0.000 -0.000 -0.014 0.001 0.016 0.009 Panel B: Heterogeneity by SDB Treatment -0.272∗∗∗ -0.189∗∗ -0.224∗∗ -0.124 -0.056 -0.145 77 (0.08) (0.08) (0.09) (0.08) (0.09) (0.09) High SDB (Violence) -0.283∗∗∗ -0.123 -0.351∗∗∗ 0.028 0.066 -0.003 (0.09) (0.10) (0.10) (0.10) (0.11) (0.10) Treatment × High SDB (Violence) 0.295∗∗ 0.089 0.340∗∗ -0.016 -0.149 0.075 (0.12) (0.12) (0.14) (0.13) (0.14) (0.13) Treat + Treat × High SDB (Violence) 0.023 -0.100 0.116 -0.140 -0.205∗ -0.070 (0.09) (0.09) (0.10) (0.10) (0.11) (0.10) Observations 696 694 685 677 672 677 Control Mean -0.000 -0.000 -0.014 0.001 0.016 0.009 Notes: Panel A presents estimates of β1 (the ITT estimate) from Equation 1, additionally controlling for SDB (violence-specific index). Panel B presents heterogeneity by SDB (violence- specific index) and reports estimates of β1 , β2 , and β3 (the ITT estimates) from the following equation: Yi,t = β0 + β1 Ti + β2 HighSDBi + β3 Ti ∗ HighSDBi + β4 Yi,t−1 + γs + εi,t . “High SDB (Violence)” is a dummy variable that takes the value 1 if the SDB score was above the median SDB score for the sample, and 0 otherwise. Each column is a separate dependent variable. Medium-term effects are shown, estimated using data from the second follow-up and baseline. Column (1) presents treatment impacts on the Social Desirability Bias (violence-specific index). Columns (2)-(7) of this table show our results on caregiver attitudes and behaviors controlling for SDB (violence-specific index). The SDB score was only measured at second follow-up. The term “Treat + Treat x High SDB (Violence)” denotes the total effect of the treatment for those in the treatment group with above-median SDB scores (violence-specific index). All dependent variables are index variables, constructed as described in Section 4. All specifications include controls for strata fixed effects, while specifications (2) - (7) additionally include the baseline dependent variable as a control. The estimation sample for columns (5)-(7) uses 4 fewer observations relative to the estimation sample for columns (4)-(6) in Table 3 as SDB measures were not collected for the 4 caregivers. The control mean refers to the mean of the control group from the second follow-up round of data. The sample size in each specification varies according to the number of observations available for each outcome and the SDB index. Heteroskedasticity-robust standard errors are reported in parentheses below the coefficient estimates. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table A12: Robustness Check: Lee Bounds for Attrition Analysis (First Follow-up) Primary hypotheses (1) (2) (3) (4) (5) (6) Attitudes toward Attitudes toward Violence Physical violence Psychological Attitudes toward violence against physical violence psychological violence against against target violence against children against children against children target child child target child (index) (index) (index) (index) (index) (index) Treatment Lower Bound -0.301∗∗∗ -0.289∗∗∗ -0.231∗∗∗ -0.291∗∗∗ -0.277∗∗∗ -0.249∗∗∗ (0.07) (0.07) (0.07) (0.08) (0.10) (0.09) Upper Bound -0.168∗∗∗ -0.184∗∗∗ -0.088 -0.110 -0.120∗ -0.081 (0.06) (0.06) (0.06) (0.08) (0.07) (0.08) Observations 1111 1109 1105 1095 1079 1094 Control mean 0.001 0.004 -0.005 -0.000 -0.002 -0.000 78 Notes: This table shows the Lee bounds associated with the estimates for treatment effects. P-values for non-significant bounds: (i) Attitudes toward psychological violence against children (index) upper bound: 0.177; (ii) Violence against target child (index) upper bound: 0.177; (iii) Psychological violence against target child (index) upper bound: 0.302. Bounds tightened by quartiles of the dependent variable, in spirit of the ANCOVA specification (since bounds cannot be tightened using continuous variables). Sample size in each specification varies according to the number of observations available for each outcome. The sample size in each specification varies according to the number of observations available for each outcome. Heteroskedasticity-robust standard errors are reported in parenthesis. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table A13: Double LASSO for Selection of Controls Primary hypotheses (1) (2) (3) (4) (5) (6) Attitudes toward Attitudes toward Violence Physical violence Psychological Attitudes toward violence against physical violence psychological violence against against target violence against children against children against children target child child target child (index) (index) (index) (index) (index) (index) Panel A: Short-term Treatment -0.186∗∗∗ -0.182∗∗∗ -0.109∗ -0.112∗∗ -0.133∗∗ -0.091∗ (0.05) (0.05) (0.06) (0.06) (0.06) (0.05) Observations 977 974 961 943 920 942 # of controls selected 4 4 4 5 4 6 Control mean -0.000 -0.000 0.000 -0.000 0.000 -0.000 Panel B: Medium-term -0.142∗∗ -0.149∗∗ -0.107∗ -0.120∗ 79 Treatment -0.072 -0.087 (0.06) (0.06) (0.07) (0.06) (0.07) (0.06) Observations 696 694 685 681 676 681 # of controls selected 3 3 3 4 4 6 Control mean -0.000 -0.000 -0.014 0.000 0.017 0.008 Notes: This table presents estimates of β1 (the ITT estimate) from Equation 1. Each column is a separate dependent variable. Each specification includes the Double Lasso suggested controls for each of the main outcomes. Panel A describes the short-term effects corresponding to the first follow-up, while Panel B presents the analog results for the second follow-up. ”Number of Controls Selected” refers to the total number of controls included in the specification. ”Control Mean” describes the mean of the outcome for the control group. The sample size in each specification varies according to the number of observations available for each outcome. The sample size in each specification varies according to the number of observations available for each outcome. Heteroskedasticity-robust standard errors are reported in parenthesis. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table A14: Randomization Inference Adjusted p-values Primary hypotheses (1) (2) (3) (4) (5) (6) Attitudes toward Attitudes toward Violence Physical violence Psychological Attitudes toward violence against physical violence psychological violence against against target violence against children against children against children target child child target child (index) (index) (index) (index) (index) (index) Panel A: Short-term Treatment -0.198∗∗∗ -0.192∗∗∗ -0.120∗∗ -0.121∗∗ -0.136∗∗ -0.101∗ (0.05) (0.05) (0.06) (0.06) (0.06) (0.05) Observations 977 974 961 943 920 942 Control mean -0.000 -0.000 0.000 -0.000 0.000 -0.000 RI p-value 0.000 0.000 0.050 0.025 0.023 0.043 Panel B: Medium-term -0.144∗∗ -0.150∗∗ -0.127∗∗ -0.124∗ 80 Treatment -0.075 -0.105 (0.06) (0.06) (0.07) (0.06) (0.07) (0.06) Observations 696 694 685 681 676 681 Control mean -0.000 -0.000 -0.014 0.000 0.017 0.008 RI p-value 0.015 0.017 0.269 0.040 0.071 0.106 Notes: This table presents estimates of β1 (the ITT estimate) from Equation 1. Each column is a separate dependent variable. Panel A describes the short-term effects corresponding to the first follow-up, while Panel B presents the analog results for the second follow-up. ”Control Mean” describes the mean of the outcome for the control group. The sample size in each specification varies according to the number of observations available for each outcome. All regressions include strata-fixed effects. ”RI p-value” presents p-values estimated using randomization inference.The sample size in each specification varies according to the number of observations available for each outcome. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table A15: Comparison of Intervention Costs with Face-to-Face Intervention Costs Cost per Caregiver Targeted (USD) vIHT Costs IHT Costs Cost Category Fixed Variable Total Fixed Variable Total Costs Costs Costs Costs Costs Costs SMS 0.00 1.43 1.43 Costs to send SMS messages 0.00 1.43 1.43 App 54.21 0.00 54.21 Consultant and staff costs 6.40 0.00 6.40 Costs associated with filming videos for the App 33.93 0.00 33.93 KnowHub App development costs 13.87 0.00 13.87 Sessions 26.45 34.53 60.98 86.61 37.34 123.95 Staff salaries 19.19 0.00 19.19 60.07 0.00 60.07 Staff training and materials 0.73 0.00 0.73 19.17 0.00 19.17 Indirect costs to administer sessions 6.53 0.00 6.53 7.37 0.00 7.37 81 Data plans to caregivers to participate in sessions 0.00 34.53 34.53 0.00 12.28 12.28 Intervention materials for parents/children 0.00 25.06 25.06 Total per Caregiver Targeted (USD) 80.66 35.96 116.61 86.61 37.34 123.95 Notes: This table presents fixed and variable costs (in 2021 USD) associated with each component of the vIHT and the comparison with the face-to-face IHT costs. Costs for the vIHT were divided across 557 treatment group caregivers to compute the costs per caregiver targeted. IHT intervention costs were divided across 115 treatment group caregivers (Francis and Baker-Henningham, 2021). For costs incurred in Jamaican Dollars, the following exchange rate was used: 1 USD = 154.94 JMD. Appendix For Online Publication Only A1 Survey Instruments and Outcomes This appendix provides additional details on the survey instruments used to collect data and details on the components and questions included in the estimation of each outcome. Following information protection protocols, we stored collected data on the survey firm’s private server. Access to the data was restricted to project staff and researchers. In the three rounds, we fol- lowed the same best practices on survey protocols. For example, we trained enumerators in the content and structure of the baseline instrument and protocol and contacted each participant up to 10 times. To reduce the risk of respondent fatigue, we limited the instrument length to approximately 35 minutes in each round. The baseline survey includes modules on the caregiver’s characteristics and outcomes, in- cluding employment status and well-being; attitudes and perpetration of violence against chil- dren; parental self-efficacy; social networks; and economic anxiety. We also include questions to measure the child’s behavior and other socio-demographic characteristics. Finally, we add a module that collects information on household socioeconomic conditions. The structure of the first follow-up survey was similar to that of the baseline survey. In the former, however, we included a module on content assimilation to measure if caregivers self-report comprehending concepts taught in the vIHT. The structures of the first and second follow-up surveys were similar in most of the mod- ules, except that in the second follow up we excluded questions to measure caregivers’ social networks and economic anxiety; and included alternative measures of parental self-efficacy and mental health distress, vignettes to measure parental discipline, and a module on social desirability bias. A. SMS viewership, attendance, App use, and learning We collected information to measure the take-up or use of each program component from dif- ferent sources. First, TrendMedia shared information on weekly SMS delivery to participants. Also, we asked caregivers in the first follow-up survey if they received SMS messages with in- 82 formation on positive parenting and, conditional on receiving them, how relevant the content was for them. Second, TrendMedia shared individual-level information on whether caregivers logged in to the App and the time (in minutes) they were connected to the App. Finally, ECC officers collected attendance data for the virtual groups for each caregiver. We also collected information to measure if caregivers learned some concepts and practices that were taught in the program in the two follow-up rounds. We included two statements for each of the four key concepts taught in the intervention. For example, to measure the learning of the “building positive relationships between parent and child” concept, we asked caregivers to what extent they agreed with the following statements “Praising children helps them learn to behave well” and “It is important that parents take some time every day to play with their child doing what their child wants.” The response options for each statement were on a 1- to 5-point Likert scale (1–Strongly disagree, and 5–Strongly agree). We estimate a learning index—the greater the index, the more caregivers knew about the program content. B. Main outcomes Attitudes towards violence against children: We use an adapted version of the UNICEF MICS ques- tionnaire to measure parental attitudes towards physical and psychological violence against children at baseline and follow-up rounds. The adapted instrument includes 5 items asking about some attitudes such as if they agree that a good parent can slap the child if he misbe- haves, and if shouting and yelling would make the child more obedient, among others. The response options were “Yes” or “No.” Since the survey was over the phone and our goal was to conduct a 35-minute length survey, we only selected the 5 items (out of 13 items) with the great- est variation according to the results from the instrument piloting that was conducted before baseline data collection. The greater the index, the more pro-violence the caregivers’ attitudes. The items on attitudes to VAC included in the survey instrument were the following: • In order to bring up, raise, or educate a child properly, the child needs to be physically punished. • A good parent slaps their child when they misbehave. • When a child is beaten, he/she will stop doing the unwanted behavior. • Shouting and yelling makes the child more obedient. 83 • Shouting, yelling, and threatening to slap will harm the child Violence against children (self-reported): In the three rounds of data collection, we used a shortened version of the UNICEF MICS questionnaire to measure caregivers’ perpetration of physical or psychological violence against children. The adapted instrument includes 5 items asking about some violent behaviors. These can be grouped into physical violence (hitting the child with a bare hand or with an object) or psychological violence (shouting, yelling, or screaming at the child; saying to send the child away; threatening to hit the child). We asked each caregiver about perpetrating these violent acts separately to the “target child” (eldest child 2 to 6 years old) or to another older child within the household (eldest child between 7 to 12 years old). Using these reports, we created two indexes: violence against the target child and violence against any child within the household. The latter is a pooled measure for both the target child and any other older child within the household. Moreover, using the items for each of the types of violence against the target child, we also created separate indexes for physical and psycho- logical violence. The greater any of the indexes, the more violent acts were perpetrated by the caregiver against the target child or another child in the household. We asked the caregiver to report how many days in the past week they did the action stated in the following list: • Shouted, yelled, or screamed at him/her? • Said you would send him/her away? • Hit him/her on the bottom, hand, arm, or leg with your bare hand? • Threatened to hit him/her but not actually done it? • Hit him/her on the bottom, hand, arm, or leg with something like a belt, hairbrush, stick, or some other hard object. C. Secondary Outcomes Caregiver’s mental health: We collected data to measure caregiver’s depression using the Patient Health Questionnaire (PHQ-2 survey, Kroenke et al. (2003)) and a question on having difficulty sleeping at night. We also measure anxiety using the Generalized Anxiety Disorder (GAD-2, Donker et al. (2011)) instrument. The PHQ-2 and GAD-2 questionnaires include two items each 84 asking how often the caregiver had been bothered by any of the problems over the last two weeks. Specifically, we asked caregivers to report how often they have been bothered by any of the two following issues during the last two weeks. The items included for depression were: • Have you been feeling little interest or pleasure in doing things? • Have you been feeling down, depressed, or hopeless? The two items included for anxiety were: • Have you been feeling nervous, anxious or on edge? • Have you not been able to stop or control worrying? Moreover, during the second follow-up, we also included the questions from the 18-items Parental Stress Scale (PSS-18, Berry and Jones (1995)). Our main outcomes of interest are the aggregate indexes of depression, anxiety, and stress separately. The greater the index, the higher the levels of each mental distress. Child conduct and emotional problems: We use the 10 items related to children’s conduct and emotional problems (5 items each) from the Strengths and Difficulties Questionnaire (SDQ) instrument to measure a child’s behavior index. We collected this information in each of the survey rounds. Each question is answered on a 0-2 scale (Not true, somewhat true, certainly true). The items ask if some behaviors related to conduct and emotional problems occurred for the child during the previous 3 months. The greater the index, the more conduct or emotional problems the child has. The questions included in emotional strengths and difficulties are as follows: • He/she often complains of headaches, stomach-aches, or sickness • He/she has many worries, often seems worried • He/she is often unhappy, downhearted, or tearful • He/she is nervous or clingy in new situations, easily loses confidence 85 • He/she has many fears, is easily scared And the questions included in conduct strengths and difficulties are the following: • He/she often has temper tantrums or hot tempers • He/she is generally obedient, usually does what adults ask/request • He/she often fights with other children or bullies them • He/she is often argumentative with adults • He/she is often spiteful to others D. Mechanisms Parental self-efficacy: We measure parental self-efficacy at baseline and during the first follow-up round using the 5 items from the Brief Parental Self Efficacy Scale (BPSES) instrument. The scale asks parents about their agreement with statements that can describe their ability to improve a child’s behavior. For the second follow-up, we adapted and used the Tool to Measure Parenting Self-Efficacy (TOPSE) for more detailed questions relating to discipline and self-acceptance. In each survey wave, the main outcome of interest is the aggregate index of parental self-efficacy. The greater the index, the higher the self-assessment of efficacy. Caregiver’s support networks: As existing evidence shows, the effectiveness of positive parenting programs can be driven by the creation of support networks for participant caregivers. To test this potential mechanism in our context, we collected information on whether caregivers obtained support from friends, family, or professionals to solve parenting or financial issues. We asked how many people they could reach out to in case they need to talk about issues related to parenting and child rearing or borrowing money. The support network variables were divided in the following groups: Borrowing Money Support Instruction was: “How many people could you go to if you needed to borrow JMD 5,000? Please indicate separately for friends, family, and professionals. Profes- sionals here could include bank officers and other moneylenders.” Parenting Issues Support Instruction was: “How many people could you go to if you wanted to talk about issues relating to parenting and child rearing? Please indicate separately for 86 friends, family, and professionals. Professionals here could include ECC officers and other such individuals.” Using this information, we created two indexes: one for parenting support and another for financial support. The greater the index, the larger the network the caregivers have. E. Sociodemographic characteristics and other controls Socio-economic and demographics. We collected the following socio-demographic data on the main caregivers: age, gender, education, marital status, employment status and occupation, income, household composition, and recent changes in lifestyle as a result of COVID-19 ex- periences. We also collected the age and gender of the target child. We collected information from all children aged 17 and below and their caregivers who regularly live under the same roof in the household. For both children and caregivers, we asked about their age and gender. In addition, we also asked about the education, marital status, employment, and occupation of all household members for each caregiver living in the household. All this information was provided by the caregiver enrolled in the study. Social Desirability Bias: A potential concern with using self-reported data to measure sensitive outcomes such as violence is the experimenter demand effect. To account for this potential bias in our estimations, we included the short form of the Marlowe-Crowne Social Desirability Scale (Crowne and Marlowe, 1960). It consists of a survey module developed by social psychologists to measure a person’s propensity to give socially desirable answers. The module, which we included in the second follow-up survey, asks respondents if they have several too-good-to- be-true traits such as never being jealous of another person’s good fortune and always being a good listener. Using this information, we created an index of social desirability bias for each participant. The greater the index, the more socially desirable the responses of the participant. 87 A2 Qualitative Study This Appendix provides further information on the methods used in the qualitative component of the study and their main results. A2.1 Objective and selection of participants The qualitative study aims to complement the quantitative results by gathering information on (i) perceptions about the different components of the intervention (SMS messages, App, and Virtual meetings with ECC specialists); (ii) dynamics within the households and perceived changes due to the intervention; (iii) transmission of information to other caregivers; and (iv) general perceptions about the program. To this end, we selected caregivers from the group of treated participants from three of the largest parishes (Saint Catherine, Kingston, and Saint Andrew) based on (i) the change between baseline and medium-term follow-up on their violence against children (VAC) index and (ii) attendance at the virtual meetings. Based on these characteristics, we assigned them to one of the following groups: a) high change in VAC and high attendance, b) high change in VAC and low attendance, c) low change in VAC and high attendance, and d) low change in VAC and low attendance. We conducted eight focus group discussions (between 5-6 participants per group) in May 2023 with a total of 43 participants. A2.2 Approach The interviewers used a narrative technique that employed a semi-structured approach of open-ended questions to permit more variation in responses. These interviews and focus groups create a natural in-depth discussion that yields specific details on the different components in- cluded in the instruments. Focus group discussions lasted up to one hour. A local consultant with expertise in quali- tative research conducted the discussions. She was responsible for recruiting participants who met the eligibility criteria, obtaining their informed consent, conducting the focus group dis- cussion, and producing their transcripts. Special care was taken to preserve participant anonymity and freedom to consent. Indeed, 88 the strategy for maintaining trust and safety was to be extremely clear to all participants that the purpose of the survey was purely academic. The focus group discussions were conducted virtually to increase participation. Participants were invited to turn their cameras on but it was not mandatory. The sessions were held via Zoom and were recorded after the participants consented. These recordings are stored in secure servers that only the research team can access through an encrypted password. A2.3 Focus Group Questions We developed a semi-structured guide to lead the focus group discussions. This guide in- cluded three main components. First, there were questions that allowed participants to intro- duce themselves. Second, we included questions on the four main topics mentioned above. To gather information about participants’ perceptions of the different components of the interven- tion, we included specific questions for each of the components. For example, we asked them 1. Did you find that the SMS messages gave you new information? 2. How useful was the information you received through SMS messages? In the case of the App, we included some questions such as 1. Do you remember what the App consisted of? 2. Did you use the App? If you did, how many times per week? Why did you or did not? 3. What features of the App should have been included to facilitate your use of it? 4. Can the App be improved or are alternative methods necessary? Similar questions were included to gather perceptions about the virtual meetings. For example, 1. Do you remember what the virtual meetings with ECC officers consisted of? 2. Did you join the meetings? 3. Why did/did not you join these meetings To collect information on the dynamics within the households and perceived changes due to the intervention, we asked questions such as 1. Who was in charge of disciplining your child before the intervention? 2. During the intervention or after it ended, who was in charge of disciplining your child? If the person in charge changed, why did it change? among other questions. Similarly, to understand potential spillovers to other non-relatives, we asked 1. Have you shared the informa- tion you learned from the intervention with other adult friends? With whom? What is your relationship with them? Are they caregivers of children 2-6 years of age? Why did you share this information with them? Lastly, we also asked for information on general perceptions about the program. Specifically, we asked 1. Do you remember what the program was about? What activities or components were 89 included? 2. Overall, what is your assessment of the program? 3. What are the areas or elements of the program that you liked the most? What areas or elements could be improved?1 A2.4 Main results This subsection summarizes the main results presented in Szekely (2023). Use of three different components of the intervention. We asked participants about their use and assessment of each component of the program (SMS messages, App, and virtual meetings). In terms of the SMS messages, we observe that participants with low attendance to the vir- tual meetings were more likely to report that the SMS messages were very useful and helpful, some even reported that these were an even better resource than the meetings because the SMS messages felt more one-on-one. In terms of the virtual meetings, all participants reported that they found the virtual meetings with ECC officers very relevant, engaging, and interactive. For instance, the officers gave participants practical activities to practice at home. Like the SMS messages, participants reported substitutability between the SMS messages and virtual meet- ings. When we asked participants with low virtual meeting attendance their reasons for not attending many sessions, they explained it was because they knew they were going to receive similar materials via SMS/WhatsApp. The second most relevant reason was that they had to work at the same time as the meetings. Across the different groups, most of them recalled they had access to the App. Yet, they explained that they did not use it because they thought it was not useful or they felt that it was not user-friendly. The only group reporting that the App was useful was the group with high attendance to virtual sessions but a low improvement on VAC. In sum, these results con- firm that the App was not as good a resource to deliver this information-based program and that some adjustments will be required to make it more attractive/accessible to this popula- tion. Moreover, these results also suggest there was some substitutability across the other two components of the intervention. Networks and spillovers to other non-related caregivers. We included a section to explore the mechanism of the network and whether the participants shared information with other care- givers. Similar to our findings using the survey data, participants across the different focus 1 The complete interview guide is available upon request. 90 groups reported that they did not make any friends from the program and that they are not in touch with anyone from a virtual group. Some of the reasons they mentioned include the loca- tion (the groups were randomly formed, that is, they were not formed considering proximity across participants) and the virtual nature of the meetings. In terms of sharing the informa- tion they acquired through the intervention, the participants mentioned that they only shared it with close relatives (family members, such as partners and spouses). Since we excluded other caregivers within the same household from the study, we are fairly confident there was no con- tamination from treatment to control caregivers. Dynamics within the household. A potential concern from our study was the displacement of violent disciplining from the caregiver participating in the program to the other that did not join. As we discussed above, the participants reported sharing the information they learned from the program with their partner/spouse. This could have caused two potential outcomes. On the one hand, the partner/spouse may have learned the new practices from the participant caregiver. On the other hand, after the parent/spouse observes that the participant caregiver has learned new positive parenting practices that differ from the social norm (violent disci- pline), then the parent/spouse may feel responsible to compensate for this and discipline the child using violence. Results from the focus group discussions suggest there were positive spillovers within the household. Participants report they encouraged their partners to praise their children for good actions. In fact, they report that family dynamics have improved after the program. They now try to be more gentle with their children instead of immediately being aggressive, and now they play games as a family as well. 91