Policy Research Working Paper 9270 The Important Role of Equivalence Scales Household Size, Composition, and Poverty Dynamics in the Russian Federation Kseniya Abanokova Hai-Anh H. Dang Michael M. Lokshin Development Economics Development Data Group & Europe and Central Asia Region Office of the Chief Economist June 2020 Policy Research Working Paper 9270 Abstract Hardly any literature exists on the relationship between result in lower estimates of poverty lines. The study decom- equivalence scales and poverty dynamics for transitional poses poverty into chronic and transient components and countries. This paper offers a new study on the impacts finds that chronic poverty is positively related to the adult of equivalence scale adjustments on poverty dynamics scale parameter. However, chronic poverty is less sensitive in the Russian Federation, using equivalence scales con- to the child scale factor compared with the adult scale structed from subjective wealth and more than 20 waves of factor. Interestingly, the direction of income mobility might household panel survey data from the Russia Longitudinal change depending on the specific scale parameters that are Monitoring Survey. The analysis suggests that the equiva- employed. The results are robust to different measures of lence scale elasticity is sensitive to household demographic chronic poverty, income expectations, reference groups, composition. The adjustments for the equivalence of scales functional forms, and various other specifications. This paper is a product of the Development Data Group, Development Economics and the Office of the Chief Economist, Europe and Central Asia Region. 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 hdang@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 The Important Role of Equivalence Scales: Household Size, Composition, and Poverty Dynamics in the Russian Federation Kseniya Abanokova, Hai-Anh H. Dang and Michael M. Lokshin * Key words: poverty, poverty dynamics, equivalence scale, Russia, panel survey JEL: I30, J10, O15 * Abanokova (kabanokova@hse.ru) is junior research fellow with Higher School of Economics, National Research University, Russia; Dang (hdang@worldbank.org; corresponding author) is economist with the Analytics and Tools Unit, Development Data Group, World Bank and is also affiliated with IZA, Indiana University, and Vietam Academy of Social Sciences; Lokshin (mlokshin@worldbank.org) is lead economist with ECA Chief Economist’s Office, World Bank. We would like to thank Conchita D’Ambrosio, Sam Freije-Rodriguez, Jacques Silber, Rostislav Kapeliushnikov, Ambar Narayan, Sergey Roshchin, Ruslan Yemtsov, and participants at IARIW conference (Dresden), IARIW-HSE conference (Moscow), and seminars at the Laboratory for Labour Market Studies and the Centre for Labour Market Studies (Moscow) and the World Bank for useful feedback on earlier versions. We are grateful to Martin Biewen for sharing his Stata code. We would also like to thank the UK Department of International Development for additional funding assistance through its Strategic Research Program (SRP) and Knowledge for Change (KCP) programs. We also acknowledge support from the NRU-HSE Basic Research Program. 1. Introduction Obtaining comparable measures of household incomes across households of different sizes and composition—or converting these incomes on a common (equivalence) scale—is a crucial task for welfare measurement. Indeed, a large body of literature has demonstrated that there are substantial effects of scale adjustments on poverty and profiles of the poor for various countries at different income levels (Lanjouw and Ravallion, 1995; Peichl et al., 2012; Bishop et al., 2014). Equivalence scales are often estimated based on expenditure data; one major disadvantage of this method is that it requires strong identifying assumptions (Deaton and Paxson, 1998). In this paper, we make several contributions to the literature on equivalence scales and poverty measurement. First, we estimate equivalence scales using an alternative source of data, subjective well-being data. While a growing literature has followed this approach using panel data, these studies mostly investigate data on life satisfaction and income satisfaction. 1 We analyze instead a subjective well-being question where individuals are asked to evaluate their own level of material welfare on a nine-point scale from "poor" to "rich". This question arguably better captures the multidimensional nature of welfare and is more related to household welfare than satisfaction variables (Ravallion and Lokshin, 2001 and 2002). Second, we offer new and interesting findings regarding the dynamics of poverty given equivalence scale adjustments (scaling) on long-run household panel data from the Russian Longitudinal Monitoring Surveys (RLMS). It is well-known that policies to address short-term 1 Two main types of subjective well-being data have been analyzed in the economic literature. The first type asks respondents about a hypothetical minimum income level that is required to reach a specified level of well-being (e.g., Garner and Short, 2004). Since this method assumes that people know what their true minimum income level is, the hypothetical assessment of the situation may lead to interpretation issues of minimum income questions (Steiger et al., 1997). The second type asks respondents to evaluate their level of satisfaction with life or income, and does not have such disadvantage (e.g., Biewen and Juhasz, 2017; Borah et al., 2018). Our paper is more related to the second approach. But we also offer robustness checks using life satisfaction data that are collected in the same household surveys. 2 static poverty are quite different from those for long-term chronic poverty. Yet, while these dynamics, by definition, requires analysis that must be based on panel data, the data used in the existing literature to investigate the effects of scaling on poverty measurement typically come from cross-sectional surveys (e.g., Newhouse et al., 2017). 2 Such data do not provide a good understanding of how household demographics impact transient or chronic poverty, or to put it differently, how employing different scaling parameters affects household poverty dynamic patterns. To our knowledge, we are among the first to investigate the impacts of scale adjustments on poverty dynamics. As discussed later, we employ several different definitions of poverty dynamics for more robust analysis. Furthermore, the RLMS offers panel data with longer time intervals than most existing studies. Longer-run panel data allow us to extend our analysis to broader definitions of households— including multigenerational households—and to better capture demographic changes caused by the formation of complex extended families. 3 Finally, the richer countries examined in existing studies, such as Germany, Switzerland or the United Kingdom, have a smaller household size on average than that of the Russian Federation. This different demographic structure implies that findings on the former countries may not necessarily apply to Russia. Furthermore, our study is especially relevant for Russia for two other reasons. Firstly, the equivalence scale currently embedded in the official poverty lines allows for unequal consumption needs but completely ignores the economies of scale in household size. A direct policy implication of no scale adjustment is that the official poverty lines would oftentimes 2 But see Dang, Jolliffe, and Carletto (2019) for a review of alternative poverty measurement methods in contexts where no panel data exist. 3 Only Borah et al. (2018) used longer panel data to analyze equivalence scales but their analysis was restricted to “classical households”, which consist of either a single adult or two partnered adults, with or without children for Germany. 3 identify large families with children as those most in need of financial support, regardless of their actual living standards. Secondly, in his recent address to the federal assembly, the Russian president discussed the falling incomes of the country and the need to create favorable conditions to raise real incomes significantly. 4 But recent evidence also points to more upward mobility than downward mobility for the population over the past two decades (Dang et al., 2019). Consequently, it would be useful for policy makers to monitor income trends correctly, and to understand whether, and to what extent, results can be affected by scale adjustments. To our knowledge, Ravallion and Lokshin (2002) and Takeda (2010) are the only two other papers that estimate the relationship between household size and composition and subjective well- being in Russia using panel data. However, besides analyzing older data, these two papers use shorter panels and cross-sectional data, respectively. Consequently, their findings are likely biased by insufficient variation in household size and unobserved heterogeneity issues. We better control for unobservable characteristics by using a recently developed econometric technique, the fixed- effect-ordered-logit-type “blow-up and cluster” (BUC) estimator (Baetschmann et al., 2015) that respects the ordinal nature of subjective well-being data. We also tested our results using more flexible econometric models. Our results suggest that the elasticity is higher for adding another adult to a two-adult household than a child, and scaling results in lower estimates of poverty lines. We decompose poverty into chronic and transient components and find that chronic poverty as a share of total poverty, defined against an absolute poverty line, is positively related to the adult scale parameter. But chronic poverty is less sensitive to the child scale factor than the adult scale factor. Interestingly, income mobility can be classified as either upward or downward depending on the 4 See http://en.kremlin.ru/events/president/news/62582. 4 specific scale parameters that are employed. Our results are robust to different measures of poverty, income expectations, reference groups, functional forms, and various other specifications. This paper consists of seven sections. We briefly review the literature in the next section, before discussing our empirical strategy in Section 3. We subsequently describe the data in section 4, and present estimation results in section 5. We offer a wide range of robustness checks and further extensions in Section 6 before finally concluding in Section 7. 2. Brief Literature Review A number of studies estimate equivalence scales using panel subjective well-being data, but these studies mostly investigate data on life and income satisfaction and focus on richer countries that have panel data such as Germany or the United Kingdom (Charlier, 2002; Schwarze, 2003; Falter, 2006; Bollinger et al., 2012; Biewen and Juhasz, 2017; Borah et al., 2018). Our brief overview of these studies, shown in Table A.1, Appendix A, offers several findings. First, although the magnitude of the estimated equivalence parameters differs considerably across studies, all the four studies for Germany find a lower weight for children than that of an additional adult. Only one study by Bollinger et al. (2012) finds that children in the United Kingdom are associated with diseconomies, but this result mostly applies to the first child. Second, although most studies suggest larger returns to scale than the (old or modified) OECD equivalence scales, non-parametric scales recently estimated for Germany by Biewen and Juhasz (2017) are fairly close to “square- root” equivalence scales. 5 Third, equivalence parameters depend on the types of subjective data/questions used for analysis. For example, analyzing life satisfaction or minimum income data 5 The old OECD scale assigns a value of 1 to the first household member, 0.7 to each additional adult, and 0.5 to each child. The corresponding figures for the modified OECD scale are 1, 0.5, and 0.3. We discuss the definitions of the square root and other scales in Section 3. 5 leads to lower estimates of equivalence scales than using income satisfaction data (Charlier, 2002; Falter, 2006). Yet, these findings may not necessarily apply to Russia, given the latter’s different demographic structures. We show four such indicators in Figure 1: the average household size (Panel A), single-person households as a percentage of the total population (Panel B), three-or- more-adults households as a percentage of all households (Panel C), and three-or-more-adults households with children as a percentage of all households (Panel D). Russia has the largest household size, which averaged at least 2.6 persons per household for the last 10 years, which is followed by the United Kingdom (2.3 persons), Switzerland (2.2 persons), and Germany (2 persons) (Panel A). Single-person households are also least common in Russia, accounting for less than 10 percent of the total population on average, while the corresponding figure for Germany is roughly twice higher at 20 percent (Panel B). The corresponding figures for the United Kingdom and Switzerland fall somewhere in between, with Switzerland catching up quickly with Germany. Figure 1, Panels C and D also display a clear cross-country difference in the proportion of extended households (i.e., households where multiple adults are present). While less than 10 percent of households in the other three European countries consist of three or more adults (with or without children) on average, the corresponding figure is at least three times higher for Russia. 3. Empirical Strategy 3.1. Measuring Scale Elasticity We assume the following equation that determines an individual’s satisfaction ∗ ′ = + �ℎ � + + , = 1. . , t = 1 … T (1) where W∗ is the total household income. is a vector of is individual i’s latent utility and personal and household characteristics, is an individual-level unobserved component, is the 6 error term. It was expected that satisfaction positively depends on income and negatively depends on household size. 6 Importantly, ℎ is the household’s equivalence weight that depends on the number of adults ( ) and children ( ), such that ℎ = + ; [0,1] is the scale elasticity parameter to be estimated that also depends on the numbers of adults and children in the household. In particular, when equals 1, we have the usual per capita household income variable (without any scale adjustment), and when equals 0.5, we have the square root scale. Equation (1) was first proposed by Schwarze (2003), which assumes that individuals evaluate their welfare level based on equivalent income rather than total household income when answering the satisfaction question. 7 Following Schwarze (2003), we also define as the equivalence scale elasticity of a household consisting of adults only, and as the scale parameter when there are children in the household, such that = − . Both these parameters capture the effects of household size and composition. Parameter is a “baseline elasticity” that will be lowered b times for each child in the household. The smaller is, the greater is the effect of household size. If b is positive, children cost less than adults, and the opposite result holds vice versa. High values of intensify the effect of household composition when the household has many children. Plugging these values for ℎ and into Equation (1), we can rewrite it as ∗ ′ − ( + ) + ( + ) + + = + (2) 6 These results are supported by empirical evidence from both richer and developing countries such as Germany and Great Britain (Van Praag and Ferrer-i-Carbonell, 2004) and Mexico (Rojas, 2007). 7 Compared to other models, the advantages of Equation (1) are that it is easy to implement, it differentiates between adults and children, and it permits estimates of a wide range of possible values of elasticity. This equation assumes a logarithmic relationship between equivalent income and subjective welfare (with decreasing marginal utility from equivalent income). We reexamine this relationship using the non-parametric approach of Biewen and Juhasz (2017) in the sensitivity analysis. 7 Clearly, the equivalence scale elasticity can be directly derived from the parameters in Equation (2). In particular, dividing the absolute value of the coefficient on ( + ) by that on , we have (= ). Similarly, (= ) is the scale parameter when there are children in the household. Equation (2) can be stated in the latent continuous utility function when we can observe having a limited J number of outcomes, which is related to ∗ as follows = < ∗ ≤ +1 , = 1, … , (3) where the individual-specific thresholds ’s are increasing, < +1 , 1 = −∞, and +1 = ∞. The probability of observing outcome j for individual i at time t is then Pr( = | , (. ), ) = Λ �+1 − ′ − �ℎ � − � − Λ � − ′ − �ℎ � − � (4) If we assume that Λ(.) has a cumulative logistic distribution and unobserved individual heterogeneity does not exist (i.e., = 0), Equation (4) can be estimated as an ordered logit model using pooled cross-sectional data. Indeed, this model is usually employed as the starting point for analysis in most existing studies (Table 1). However, since unobserved individual heterogeneity such as personality traits and preferences likely exist (i.e., ≠ 0) and it can be correlated with household income or serially correlated over time, it can result in inconsistent estimates (Ravallion and Lokshin, 2002; Ferrer-i-Carbonell and Frijters, 2004; Ravallion, 2012). 8 The individual fixed- effects model is an appropriate model to deal with these issues. 8 Van Praag and Ferrer-i-Carbonell (2004) also observe that there will always be omitted variables in satisfaction equations. 8 We apply the most recent statistical model, the BUC fixed-effects model that is developed by Baetschmann et al. (2015). 9 Consistent estimations of parameters (, ) are performed by collapsing ordered variables (J levels of ) into binary outcomes for each choice (0, .., J-1). The conditional maximum likelihood estimator by Chamberlain (1980) can be subsequently applied to each of these binary choice models. By copying each observation J-1 times in the data set (i.e., “blowing-up” the sample size), so that for every J−1 copy of the observation, it is possible to dichotomize the dependent variable at each different threshold. This procedure helps avoid the (severe) loss of information as with the binary (Chamberlain) logit model with fixed effects. We use two-way clustering and cluster the standard errors at both the individual and household-wave levels. The BUC approach was found to outperform other existing estimators (e.g., Riedl and Geishecker, 2014), but for robustness checks, we also estimate other models such as the pooled ordered logit (POL) model and the linear fixed effects model (FE OLS). While both these models likely yield biased results, they can provide some comparison estimates. 10 For example, empirical evidence for Germany suggests that the equivalence scale parameters in FE models are significantly reduced compared to the pooled regressions (Schwarze, 2003; Borah et al., 2018), but the opposite result holds for Switzerland (Falter, 2006). 3.2. Chronic Poverty and Income Mobility 9 Subsequently, Das and Van Soest (1999), Ferrer-i-Carbonell and Frijters (2004) and Baetschmann et al. (2015) introduced new estimators for the fixed effects ordered logit model using the extensions of existing binary choice panel data models. Baetschmann et al.’s model is observed to outperform Das and van Soest’s estimator if some categories on the ordered scale have small sample size and Ferrer-i-Carbonell and Frijters’ estimator if the number of categories on the ordered scale is large (Riedl and Geishecker, 2012). 10 The POL provides biased estimates if the fixed effects are statistically significant, while the FE OLS does not model well the categorical dependent variable. 9 A common approach to measuring chronic poverty is to identify individuals’ permanent incomes, and then define these individuals as chronically poor if their permanent incomes are below a specified poverty line (Jalan and Ravallion, 2000). In this approach, the intertemporal mean of poverty for each individual is defined as 1 = ∑ =1 ( < ) �1 − � (5) where α is a measure of the sensitivity of poverty to inequality among the poor (i.e. poverty aversion indicator), I(.) is the indicator function which is one if the condition is satisfied and zero 1 otherwise. Total poverty is obtained by averaging across all individuals = ( … … ) = ∑ . The aggregate chronic poverty index is defined as 1 � = ∑ =1 � � < � �1 − � (6) � is obtained by averaging all income of over the period for each individual, In Equation (6), irrespective of the poverty status of the individual at any time. To provide robustness checks on estimation results, we also follow alternative approaches in measuring poverty. These include the spell approach, which defines individuals as chronically poor if they are poor in a certain number of periods and the equally distributed equivalent approach by Duclo et al. (2010). 11 Let yt and ztk respectively represent individuals’ income (consumption) and the income threshold k in year t, where t= 1 or 2, and k= 0, 1,…, K, and a higher number for k indicating a higher income threshold. The minimal and maximal thresholds 0 and correspond to -∞ and +∞, respectively. Let be the population’s relative mobility measure of interest, where l= u 11 For the spell approach, we employ Foster’s (2009) measure of chronic poverty, which considers an individual to be chronically poor if the percentage of time he spends below the poverty line (z) is at least the duration cutoff () as 1 follows = ∑ =1 [∑=1( < ) ≥ ] �1 − � , where is the minimum percentage of time a person must be in poverty in order to be chronically poor, α is a sensitivity of poverty measure to inequality among poor (i.e. poverty aversion indicator), I(.) is the indicator function which is one if the condition is satisfied and zero if not. 10 (upward mobility) or d (downward mobility), and o= n (unconditional mobility) or c (conditional mobility). We define the unconditional (probability of) upward mobility for the whole population as follows = ∑ =0 ( ≤ 1 ≤ +1 2 ≥ +1 ) (7) Note that this higher income category k+1 is not just the next higher income category, but can generally include any higher income category. The corresponding probabilities of unconditional downward mobility can be obtained by reversing the inequality signs in Equation (7) for individuals’ income level in the second year. Focusing on the income category k in year 1, we define the measure of conditional upward mobility for the whole population as follows 12 = ∑ =+1 (2 ≥ | ≤ 1 ≤ −1 ) (8) 4. Data We analyze the Russian Longitudinal Monitoring Survey (RLMS), which is an annual and nationally representative panel household survey. Our analysis covers 24 years (22 survey waves) from 1994 to 2017. We restrict the estimation sample to working-age adults, who are 16 years old or older, and exclude households where all members are younger. We also exclude households with an unusually large number of members (e.g., having more than five adults and three children). 13 12 See Dang et al. (2019) for more discussion on these measures of mobility. 13 Such households represent less than 3% of the data. See Appendix A, Table A.3 for the distribution of household types. But we offer estimates using the whole unrestricted sample in Table 5. The results suggest that the scale parameters for children are lower when using a pooled model and even negative (but insignificant) when using fixed effect ordered logit. At the same time, the adult scale parameter is robust to using an unrestricted sample. 11 Our outcome variable of interest, subjective wealth, captures individual responses to the following question on a scale ranging from one to nine: “Please imagine a nine-step ladder where on the bottom, the first step, stand the poorest people, and on the highest step, the ninth, stand the rich. On which step of the nine steps are you personally standing today?” We plot the distribution of this variable in Figure A.1 in Appendix A, which resembles a somewhat bell-shaped distribution. 14 There is also a reasonable degree of churning over time, with only about 40% of those who score in the range 3 to 5 keeping the same score for the next period. There are in total 44,010 individuals with 254,822 observations. We also offer robustness checks on our estimates by analyzing two other questions in the RLMS asking about satisfaction with life and personal economic conditions. Our measure of income is the household’s total monetary income, which is temporarily deflated and adjusted for regional differences. To reduce the effects of outliers, we trim one-quarter of a percent of the data at both the top and the bottom of the income distribution and only keep individuals with a positive income level. For the other control variables, we include in all models: individual’s age (in groups), education level, marital status, employment status, health status, dummy variables indicating whether there are other household members with poor health, and per capita living space. 15 To estimate the pooled regressions, we additionally include individuals’ gender, nationality, and an extended set of regional variables. Table A.2 in Appendix A provides the summary statistics for the control variables. 14 Since responses with a score of eight or nine account for less than 1% of the sample, we combine these in one group. But we also estimate scale parameters without this aggregation and obtain similar results (results available upon request). 15 Frijters and Beatton (2012) show that age effects are better captured with more flexible forms (such as using 5-year age groups) rather than with age and age squared. But we also implement robustness checks with age and age squared and obtain similar results. Since unemployment and health variables may be considered endogenous variables (e.g., Oswald and Powdthavee, 2008; Kassenboehmer and Haisken‐De New, 2019), we re-estimate our scale parameters without these variables and obtain similar results. 12 5. Estimation Results 5.1. Scale Parameters We provide in Table 1 estimates for the equivalence weights of adults and children, using three models: the pooled ordered logit (POL), the linear fixed-effects (FE OLS), and our preferred BUC model. Compared to the FE OLS model, the number of individuals in the BUC model decreases by almost 13,000, since these individuals were observed only once in the RLMS, or their subjective welfare levels did not change during the period of study. In all three model specifications, the � have the expected signs and are both statistically significant, � and estimated parameters � is slightly weaker at the 6 percent level for the BUC although the statistical significance for � is positive, indicating that households with a higher number of children need more model. resources and bear higher costs. Using the estimates from Table 1, Table 2 calculates the equivalence scales. Estimates based on the POL model yield 0.6 for the adult parameter and 0.08 for the child parameter b, suggesting that the overall elasticity is higher for adding another adult than a child to a two-adult household. Controlling for unobserved individual heterogeneity in the panel models reduces by about one-third both the estimated equivalence scale parameter for adults (from 0.6 to 0.4) and for children (from 0.08 to 0.05). The scale elasticities estimated from the FE OLS and the BUC models are nearly the same, but we switch to presenting results using the BUC model in the subsequent discussion. 16 Our estimates suggest a larger scale impact for children on household income in Russia than in Germany and Switzerland (Schwarze, 2003; Falter 2006; Borah et al., 2018), which can be 16 This result is consistent with that of Ferrer-i-Carbonell and Frijters (2004), who find little difference in estimates for the determinants of happiness in the FE Ordered Logit and FE OLS models, and with that of Riedl and Geishecker (2014) who show that linear and ordered fixed effect models offer similar estimates for the relative size of parameters. 13 explained by more generous transfers to households with children in Russia. At the same time, our results are consistent with those for Germany and Switzerland in terms of the smaller effect of additional children compared to additional adults. Figure 2 compares our preferred BUC estimated scales with some other common scales, including the simple per-capita adjustment, the square-root adjustment, the OECD scales, and the poverty line scale, each normalized to a single adult. 17 For each additional adult (or child), while the per capita and OECD scales display a constant marginal cost, our estimated scales, as well as the square-root scale, have a decreasing marginal cost. Compared to our estimated scales, all the other scales overestimate the weights for either an additional adult or an additional child. Interestingly, our estimated scales also provide lower elasticities than the equivalence scale embedded in the official poverty line for Russia, particularly for large-size households. 5.2. Adjusted Poverty Lines A natural question arises. What are the implications of these decreasing marginal costs for both adults and children for poverty measurement? We present in Table 3 our proposed population- weighted poverty lines for different family types, based on the estimated parameters of equivalence scales, and compare them with the official poverty thresholds employed by Rosstat. Our absolute poverty lines are derived from Rosstat’s official poverty thresholds for different age groups. Our relative poverty lines are computed as two-thirds of the median income per adult equivalent for each household type, using the parameters that account for differences in economies of scale and composition in the household. To make Table 3 easier to read, we leave out the standard errors (see Appendix A, Table A.5 for the full results). We offer a comparison of our results with those in studies for Germany and Switzerland that use similar estimation 17 methods in Appendix A, Table A.4. 14 Table 3 suggests that our proposed poverty lines for households, in both absolute and relative terms, are generally much lower than the official poverty thresholds for the country. In particular, for our absolute poverty lines, the official poverty threshold ranges from 50 percent (for a two- adult household) to 160 percent higher (for a five-adult-no-children household) for households without any children. It ranges from 160 percent (for a one-adult-one-child household) to more than 200 percent (for a five-adult-one-child household) higher for households with children. The corresponding differences for our relative poverty lines are less, but are still considerable. The official poverty thresholds are from about 20 percent to 90 percent higher and 40 percent to 100 percent higher respectively for households without children and households with children. We provide in Appendix A, Figure A.2 the poverty rates that corresponds to the official poverty line adjusted with the estimated equivalence scales in Table 2. Consistent with our previous discussion, the revised poverty rates based on the estimated equivalence scales are lower than the official poverty rates. 5.3. Poverty and Income Dynamics We start by examining in Figure 3 the extent to which the (headcount) poverty rate for Russia can be affected by the scale parameters. Again, the values of 1 and 0.5 for respectively correspond to the per capita scale and square root scale. The value of 0.1 for indicates an extremely large effect of household size. When b increases from 0 to 0.1, it is a situation where for the same household size, households with children have a lower equivalence scale elasticity (i.e., a higher economy of size) than households without children. We also examine poverty using either the absolute poverty line (Panel A) or the relative poverty line (Panel B). 15 Since the relative poverty line is adjusted to scaling by construction, it unsurprisingly provides the opposite scaling effects compared to the absolute poverty line. 18 Yet, Figure 3, Panel A shows that the poverty rate using the absolute poverty line can decrease by 9 to 15 percentage points (from 12 or 18 percent to 3 percent) if decreases from 1 to 0.5, depending on the child parameter values. The poverty rate subsequently remains almost the same, and decreases by one to two percentage points if decreases from 0.5 to 0.1. Figure 3, Panel B displays the opposite results where the poverty rate using the relative poverty line increases slightly by at most four percentage points if decreases from 1 to 0.5, again depending on the child parameter values. It then increases faster by four percentage points if decreases from 0.5 to 0.1. On the other hand, poverty is less sensitive to the choice of the child discount factor. It varies by at most 6 and 2 percentage points respectively for the absolute poverty line and the relative poverty line, when the child scale factor is varied from 0 to 0.1 and keeping fixed. 19 A natural question then arises. Are households with children always richer than those without children? The answer turns out to depend on the specific equivalence scales that are employed. For example, Table A.6 in Appendix A shows that for the same child scale parameter of 0.1, households with children are less poor if the adult scale parameter falls in the range [0.1, 0.5], but are poorer if the adult scale parameter falls in the range [0.6, 1]. Clearly, selecting a larger child discount factor, say at 0.1, will result in households with children being less poor than households with children for most values of the adult scale parameter. But still, if we set the latter at 1, households with children are poorer. 18 When we make scale adjustments for income, this results in changes to the population distribution of income and the relative poverty line. For example, most European countries set their relative poverty line at 60% of the national median equivalized disposable income. 19 We employ the range of [0, 0.1] for the child scale parameter since it is observed to be less than 0.1 in previous studies. For example, the scale elasticity for each additional child aged between 15 and 17 years was estimated to be 0.086 for Switzerland (Falter, 2006). 16 We turn next to examining poverty duration, which is defined as the average number of consecutive survey years (rounds) an individual spends in poverty. Figure 3, Panels C and D produce qualitatively similar results. Poverty duration is sensitive to changes in , and ranges from 1.8 to 2.6 years and from 2 to 2.7 years respectively with the absolute poverty line and the relative poverty line. But poverty duration is less sensitive to child scaling and varies by less than 0.2 year for both the absolute and the relative poverty lines. We provide in Table 4 transient and chronic poverty estimates using Jalan and Ravallion’s (2000) method for three common poverty measures: the headcount poverty rate, the poverty gap index, and the squared poverty gap index. Table 4 shows that the shares of chronic poverty of total poverty are positively related to the adult scale parameter, regardless of the poverty measures we use. For example, for headcount poverty, the share of chronic poverty decreases by almost 10 percentage points when increases from 0.3 to 0.7. For the poverty gap and squared poverty gap, the corresponding figures are a-7-percentage-point and a 5-percentage-point increase. We plot the alternative chronic poverty measures (Foster, 2009; Duclos et al., 2010) against the scale factors in Appendix A, Figure A.3, which also shows that these measures are more sensitive to scale adjustments for adults than for children. Figure 4 examines the relationship between scale parameters and unconditional income mobility (Panel A) and conditional income mobility (Panel B) (see Appendix A, Figure A.4 for the corresponding three-dimensional graphs). Three possible scenarios can happen with income mobility: more upward mobility (as represented by the area in orange), more downward mobility (as represented by the area in purple), and a mixed situation where neither upward mobility or downward mobility dominates (as represented by the gray area in between the two colors above). Interestingly, the selection of specific scale parameters can even change estimation results for 17 mobility. In particular, when income is measured on a per capita basis ( =1), there is always more upward unconditional mobility, regardless of the (different values for the) child parameter (Panel A). There is also mostly more upward conditional mobility, except for when the child parameter falls in the interval [0.09, 0.1] (Panel B). Yet when income is measured on a square-root scale, we have more upward unconditional mobility when the child parameter ranges from 0 to 0.03, a mixed situation when the child parameter ranges from 0.03 to 0.08, and even more downward mobility for the rest of the child parameter values. For conditional mobility, the square-root scale results in more downward mobility, for all values of the child parameter (Panel B). These results further emphasize the important role that equivalence scales have in determining estimation results with income dynamics. 6. Robustness Checks and Further Extensions 6.1. Robustness Checks In addition, we examine a number of other robustness checks and extensions, which include income expectations, different reference groups, other satisfaction variables as dependent variables, measurement error in incomes, and no sample restrictions. We briefly summarize the results below. Changes in household size or structure are typically expected and may affect subjective well- being well before their actual realization. We control for income expectations in the (t-1) period and find that this does not affect the estimates of baseline elasticity but slightly increases the child scale parameter up to 0.08 in the pooled model (Table 5, row 1). 20 20 We analyze the answer to the following question in the RLMS “Do you think that in the next 12 months you and your family will live better than today or worse?” The regression results are shown in Appendix A, Table A.7. 18 Relative income rather than total income may affect satisfaction, and if ignored, may result in biased estimates (Borah et al., 2018). We include dummy variables to indicate the relative position of the household in the reference group`s distribution of household income quartiles (Appendix A, Table A.8). The reference group is determined for each year and consists of individuals living in households with a similar size in the same primary sampling units. To ensure stability, we only consider the number of households in the reference group as having 10 or more households. We report estimates of the scale parameters for the POL model only, since the variable used to define the reference groups is largely time-invariant, especially at the primary sampling unit level. Controlling for the reference group decreases the child scale parameter to 0.05 in the POL model but does not change the baseline elasticity. More importantly, we still obtain the earlier result that an additional child has a smaller effect compared to an additional adult (Table 5, row 2). 21 We also analyze the other satisfaction variables in the RLMS as alternative dependent variables for the subjective wealth variables, which are satisfaction with one’s life and satisfaction with one’s economic conditions. The estimated coefficients on household income and household size are still statistically significant as expected (Appendix A, Table A.9). To save space, we only report the scale parameters derived from the regressions for life satisfaction (Table 5, row 3). The estimation results of the BUC model are robust with the adult scale parameter of about 0.6 and child scale parameter about 0.04. 22 As a check on the total household income variable, we generate a new total household income by summing all the net incomes reported by household members (Appendix A, Table A.10). Yet, the estimated scale parameters of 0.3 for adults and 0.06 for children obtained from the BUC 21 The full regression results are shown in Table 1. 22 The adult scale parameter is still high when using satisfaction with economic conditions (0.8), but the child scale parameter decreases to 0.02. The POL model also similarly provides a higher elasticity for adults (0.8), as well as for children (0.1) (Table 5, row 3). 19 models are close to our estimation results (Table 5, row 4). We use the unrestricted sample containing households with more than five adults and three children and estimate our main regressions. Estimation results for children are no longer statistically significant for the BUC model, and are only statistically significant in the POL model (Appendix A, Table A.11). At the same time, the estimates for the adult scale parameter remain similar at about 0.6 (Table 5, row 5). 6.2. Role of Pensioners Our earlier analysis has focused on household sizes and children, but has not discussed the impacts of elderly pensioners on the total household income. Pensioners may have disability or health issues and thus can impose significant costs on the household. On the other side, pensioners often consume less than a working-age adult and can contribute their pension salary to the household income. Our estimates from the RLMS suggest that the share of individuals (in total population) who received any pension in the past month hovers around 30 percent over the period 1994-2017. The majority of these pensioners (more than 70%) receive retirement or old-age pensions. We assume that the presence of a pensioner has an effect on subjective well-being through the cost channel only. The inclusion of the number of registered pensioners is additional: a pensioner enters the regression twice as a family member in his age group and as a pensioner. We can then modify Equation (1) as follows ∗ ′ = + 1 � � + 2 + + (ℎ) −− = ′ + 1 − 1 ℎ + 1 ℎ + 1 ℎ + 2 + + (9) where is the number of pensioners in the household. In this specification, the total effect of pensioners is then (1 ℎ + 2 ). 20 Although the interaction term for the household size and the number of pensioners is not statistically significant in both the pooled and BUC regressions (Appendix A, Table A.12), the total effect of pensioners is statistically significant and positive in the BUC model (Appendix A, Table A.13). But the inclusion of pensioners does not change the estimated scale parameters significantly: the adult scale parameter still varies between 0.4-0.6 and the child scale parameter is about 0.05-0.06 (Table 5, row 6). 6.3. Alternative Functional Form As an alternative to Equation (1), we can estimate a non-parametrical function recently proposed by Biewen and Juhasz (2017) as follows �ℎ � = (10) ( + ) where ( + ) = 1 ∗ 1 0 + 2 ∗ 2 0 + 21 ∗ 2 1 + ⋯ + 51 ∗ 5 1, and 2 1 indicates a household with two adults and one child. The estimated parameters for this scale are given in Appendix A, Table A.14. The table shows that the “non-parametric” scales for household types are smaller than those estimated using the parametric functional form as in Equation (1). The estimated equivalence weight of a second adult is 24 percent of the first adult, and the estimated equivalence weight of a child is 13 percent, or about half of the second adult. The child scale parameter is similar to our BUC estimates, and also to those obtained by Biewen and Juhasz (2017) for Germany. 7. Conclusion We estimate equivalence scales using unique subjective wealth data from Russia, and apply these scale adjustments to examine new poverty lines as well as the sensitivity of poverty dynamics. Our findings suggest that the country’s official poverty threshold ranges from 50 21 percent (for a two-adult household) to more than 200 percent (for a five-adult-one-child household) higher than our estimated poverty lines. The poverty rate varies for different adult scale parameters, but less so for children. The shares of chronic poverty of total poverty, defined against an absolute poverty line, are positively related to the adult scale parameter, regardless of the poverty measure. More interestingly, income mobility could be classified as either upward or downward depending on the specific scale parameters that are employed. Our results are robust to different measures of poverty, income expectations, reference groups, functional forms, and various other specifications. There is significant heterogeneity in terms of economic growth and demographic composition among the regions of Russia, which is caused by geographical differences in relative prices and consumption preferences. Since the RLMS data are, unfortunately, not representive at the regional level, we are unable to offer this analysis. But given a larger panel survey that is representative at the regional level, a promising direction for further research is to apply our scale adjustments to better understand the effects of equivalence of scale on the composition of poverty and poverty dynamics. Such knowledge is essential for regional poverty comparisons and the development of well-targeted policy interventions. 22 References Baetschmann, Gregori, Kevin E. Staub, and Rainer Winkelmann. (2015). “Consistent estimation of the fixed effects ordered logit model”. Journal of the Royal Statistical Society: Series A, 178(3): 685–703. Biewen, Martin, and Andos Juhasz. (2017). “Direct estimation of equivalence scales and more evidence on independence of base.” Oxford Bulletin of Economics and Statistics, 79(5): 875- 905. Bishop, John A., Andrew Grodner, Haiyong Liu, and Ismael Ahamdanech-Zarco. (2014). “Subjective poverty equivalence scales for Euro Zone countries.” Journal of Economic Inequality, 12(2): 265-278. Bollinger, Christopher R., Cheti Nicoletti, and Stephen Pudney. (2012) “Two can live as cheaply as one... but three's a crowd.” Institute for Social and Economic Research Working Papers, No. 2012-10. Borah, Melanie, Carina Keldenich, and Andreas Knabe. (2018). “Reference income effects in the determination of equivalence scales using income satisfaction data.” Review of Income and Wealth, 0(0): 1-35 Charlier, Erwin. (2002). “Equivalence scales in an intertemporal setting with an application to the former West Germany.” Review of Income and Wealth, 48(1): 99-126. Dang, Hai‐Anh, Dean Jolliffe, and Calogero Carletto. (2019). “Data Gaps, Data Incomparability, and Data Imputation: A Review of Poverty Measurement Methods for Data‐Scarce Environments.” Journal of Economic Surveys, 33(3): 757-797. Dang, Hai‐Anh, Michael M. Lokshin, Kseniya Abanokova, and Maurizio Bussolo. (2019). “Welfare Dynamics and Inequality in the Russian Federation during 1994-2015.” European Journal of Development Research. doi: https://doi.org/10.1057/s41287-019-00241-3. Das, Marcel, and Arthur Van Soest. (1999) “A panel data model for subjective information on household income growth.” Journal of Economic Behavior & Organization, 40(4): 409-426. Deaton, Angus, and Christina Paxson. (1998) “Economies of scale, household size, and the demand for food.” Journal of Political Economy, 106(5): 897-930. 23 Duclos, Jean-Yves, Abdelkrim Araar, and John Giles. (2010). "Chronic and transient poverty: Measurement and estimation, with evidence from China." Journal of Development Economics, 91(2): 266-277. Falter, Jean-Marc. (2006). “Equivalence Scales, Poverty Lines and Subjective Data in Switzerland.” Swiss Journal of Economics and Statistics, 142(2). Ferrer-i-Carbonell, Ada, and Bernard MS Van Praag. (2001). “Poverty in Russia.” Journal of Happiness Studies, 2(2): 147-172. Ferrer‐i‐Carbonell, Ada, and Paul Frijters. (2004). “How important is methodology for the estimates of the determinants of happiness?” Economic Journal 114: 641-659. Foster, James E. (2009). “A class of chronic poverty measures.” Poverty dynamics: interdisciplinary perspectives, Part 2, Chapter 3: 59-76 Frijters, Paul, and Tony Beatton. (2012) “The mystery of the U-shaped relationship between happiness and age.” Journal of Economic Behavior & Organization, 82(2-3): 525-542. Garner, Thesia I., and Kathleen S. Short. (2004). “Economic Well-Being Based on Income, Consumer Expenditures and Personal Assessments of Minimum Needs.” In Studies on Economic Well-being: Essays in the Honor of John P. Formby, pp. 319-361. Emerald Group Publishing Limited. Hulme, David, and Andrew Shepherd. (2003). “Conceptualizing chronic poverty.” World Development 31(3): 403-423. Jalan, Jyotsna, and Martin Ravallion. (2000). “Is transient poverty different? Evidence for rural China.” Journal of Development Studies, 36(6): 82-99. Kassenboehmer, Sonja C., and John P. Haisken‐De New. (2009). “You’re fired! The causal negative effect of entry unemployment on life satisfaction.” Economic Journal, 119(536): 448- 462. Lanjouw, Jean O., Peter Lanjouw, Branko Milanovic, and Stefano Paternostro. (2004). “Relative price shifts, economies of scale and poverty during economic transition.” Economics of Transition, 12(3): 509-536. Lanjouw, Peter, and Martin Ravallion. (1995). “Poverty and household size.” The Economic Journal, 105(433): 1415-1434. 24 Newhouse, David, Pablo Suárez Becerra, and Martin Evans. (2017). “New global estimates of child poverty and their sensitivity to alternative equivalence scales.” Economics Letters, 157: 125-128. Oswald, Andrew J., and Nattavudh Powdthavee. (2008). “Does happiness adapt? A longitudinal study of disability with implications for economists and judges.” Journal of Public Economics, 92(5-6): 1061-1077 Peichl, Andreas, Nico Pestel, and Hilmar Schneider. (2012). “Does size matter? The impact of changes in household structure on income distribution in Germany.” Review of Income and Wealth, 58(1): 118-141. Ravallion, Martin, and Michael Lokshin. (2001). “Identifying welfare effects from subjective questions.” Economica, 68(271): 335-357. Ravallion, Martin, and Michael Lokshin. (2002). “Self-rated economic welfare in Russia." European Economic Review, 46(8): 1453-1473. Ravallion, Martin. (2012). “Poor, or just feeling poor? On using subjective data in measuring poverty”. World Bank Policy Research Working Papers, No.5968 Riedl, Maximilian and Ingo Geishecker (2014) “Keep it simple: estimation strategies for ordered response models with fixed effects”. Journal of Applied Statistics, 41(11): 2358-2374 Riedl, Maximilian and Ingo Geishecker (2010) “Ordered response models and non-random personality traits: Monte Carlo simulations and a practical guide.” Center for European Governance and Economic Development Research Discussion Paper 116 (revised edition 2012). Rojas, Mariano. (2007). “A subjective well-being equivalence scale for Mexico: Estimation and poverty and income-distribution implications.” Oxford Development Studies 35(3): 273-293. Schwarze, Johannes. (2003) “Using panel data on income satisfaction to estimate equivalence scale elasticity.” Review of Income and Wealth 49(3): 359-372. Steiger, Darby Miller, Tina Mainieri, and Linda Stinson. (1997). “Subjective assessments of economic well-being: Understanding the minimum income question.” In annual American Association for Public Opinion Research Conference, Norfolk, Virginia, in the Proceedings of 25 the Joint Statistical Meetings Section on Survey Research Methods, Alexandria, VA: American Statistical Association, pp. 899-903. Takeda, Yuka. (2010) “Equivalence scales for measuring poverty in transitional Russia: Engel's food share method and the subjective economic well-being method.” Applied Economics Letters 17(4): 351-355. Van Praag, Bernard MS, Bernard Van Praag, and Ada Ferrer-i-Carbonell. (2004). Happiness quantified: A satisfaction calculus approach. Oxford University Press (revised edition 2008). Yang, Yang. (2008). “Social inequalities in happiness in the United States, 1972 to 2004: An age- period-cohort analysis.” American Sociological Review 73(2): 204-226. 26 Table 1. Detailed regression results, RLMS 1994-2017 Variables Pooled OL FE OLS BUC 0.655*** 0.249*** 0.412*** Ln household income ( ) (0.011) (0.007) (0.079) -0.417*** -0.100*** -0.167*** Ln household size (− ) (0.023) (0.016) (0.034) 0.051*** 0.012** 0.020* Children# Ln household size () (0.008) (0.005) (0.010) 1.027*** 0.404*** 0.668*** Age 16-20 (0.031) (0.029) (0.077) 0.386*** 0.178*** 0.296*** Age 21-30 (0.021) (0.021) (0.057) 0.188*** 0.062*** 0.101*** Age 31-40 (0.019) (0.014) (0.036) -0.092*** -0.004 -0.006 Age 51-60 (0.019) (0.013) (0.039) -0.009 0.057*** 0.097 Age 61-70 (0.025) (0.019) (0.065) 0.103*** 0.094*** 0.159* Age 71-80 (0.028) (0.023) (0.083) 0.399*** 0.330*** 0.542*** Age 80+ (0.038) (0.030) (0.086) Female -0.027* (0.014) Russian nationality -0.262*** (0.024) 0.193*** -0.010 -0.014 Complete secondary (0.019) (0.012) (0.041) 0.298*** -0.043*** -0.071 Secondary + vocational (0.021) (0.016) (0.056) 0.446*** -0.004 0.003 University and higher (0.023) (0.022) (0.082) -0.260*** -0.029* -0.048 Single (0.023) (0.017) (0.037) -0.333*** -0.156*** -0.263*** Divorced/widowed/separated (0.019) (0.013) (0.040) -0.275*** -0.162*** -0.269*** Unemployed/out of labor force (0.015) (0.009) (0.030) -0.136*** -0.047*** -0.080*** Bad health (0.010) (0.006) (0.012) -0.075*** -0.013* -0.022* Other members with bad health (0.011) (0.007) (0.012) 0.000 0.001 0.002 Log of per capita living space (0.002) (0.001) (0.002) Number of observations 237,395 240,640 712,448 Log pseudolikelihood -403,224 -346,509 -263,848 Number of individuals 42,326 42,894 30,058 Pseudo-R squared 0.043 0.036 0.0285 Note: Robust standard errors are in parentheses, controlling for two-way clustering (i.e., at the individual for the POL model and at the household-wave level for the FE OLS and BUC models). All regressions include year fixed effects, pooled model includes regional fixed effects (not reported). *** p<0.01, ** p<0.05, * p<0.1 27 Table 2. Scale elasticity parameters, RLMS 1994-2017 Dependent variable: subjective wealth Scale parameters Pooled Ordered Logit FE OLS BUC Baseline elasticity 0.636*** 0.399*** 0.407*** = / (0.032) (0.060) (0.088) Additional child 0.078*** 0.050** 0.048* = / (0.012) (0.021) (0.026) Overall elasticity 0.636-0.078*k 0.399-0.050*k 0.407-0.048*k Note: Standard errors in parentheses are calculated using delta-method. All regressions include age groups, education level, marital status, employment status, respondent`s poor health, dummy whether there are other household members in poor health, dummy indicating whether the person was employed at survey time and per capita living space and time effects as additional variables. Pooled model additionally includes gender, nationality and regional state effects. 28 Table 3. Alternative poverty thresholds by household size in 2017 (in rubles per month) Estimated with absolute line Estimated with relative line Household Type Official Pooled OL BUC Pooled OL BUC Households without children One adult, no children 9,607 9,607 10,800 10,800 9,607 Two adults, no children 14,891 12,777 13,913 16,306 19,214 Three adults, no children 19,310 14,987 15,931 20,488 28,821 Four adults, no children 23,153 16,908 17,520 24,066 38,428 Five adults, no children 26,707 18,542 17,397 25,150 48,035 Household with children One adult, one child 14,122 12,297 12,226 14,035 19,532 Two adults, one child 17,773 14,218 14,973 18,632 29,139 Two adults, two children 18,734 14,795 15,422 19,493 39,064 Three adults, one child 20,847 15,755 15,225 20,062 38,746 Three adults, two children 20,847 15,852 18,885 24,788 48,671 Four adults, one child 23,537 17,100 20,098 27,685 48,353 Four adults, two children 22,673 16,812 20,311 27,494 58,278 Five adults, one child 26,131 18,253 20,201 28,855 57,960 Note: Population weights are applied. Standard errors for poverty rates are adjusted for complex survey design. Poverty line for reference “one adult” is defined as an average of minimum subsistence levels for working-age individual and pensioners in 2017. The level of absolute poverty line is 10899 rubles per month for working-age individual, is 8315 rubles per month for pensioner and is 9925 rubles per month for child in 2017. Relative poverty line is set on 60% of household size-weighted median equivalized income for each household type using RLMS data in 2017. Poverty lines of reference adult are adjusted with weights in Table 2 using BUC model (where baseline elasticity equals 0.407 and every child has a weight 0.048) and using Pooled Ordered Logit model (where baseline elasticity equals 0.636 and every child has a weight 0.078). 29 Table 4. Chronic and transient poverty by adult scale factors, Jalan-Ravallion decomposition, RLMS 1994-2017 Equivalent income is computed using =0.3 =0.4 =0.5 =0.6 =0.7 Headcount Poverty Total Poverty 0.085 0.1 0.119 0.142 0.17 Transient Poverty 0.036 0.041 0.045 0.05 0.054 Chronic Poverty 0.049 0.059 0.074 0.092 0.115 Share of chronic poverty (%) 57.3 59.4 62.1 64.9 67.9 Poverty Gap Total Poverty 0.03 0.035 0.042 0.05 0.06 Transient Poverty 0.015 0.017 0.02 0.023 0.026 Chronic Poverty 0.015 0.018 0.022 0.027 0.035 Share of chronic poverty (%) 49.6 50.7 52.3 54.5 57.2 Squared Poverty Gap Total Poverty 0.016 0.019 0.022 0.026 0.032 Transient Poverty 0.009 0.01 0.012 0.014 0.016 Chronic Poverty 0.007 0.009 0.01 0.013 0.016 Share of chronic poverty (%) 45.6 46.1 47.1 48.5 50.5 Note: Absolute poverty line is defined as a minimum regional subsistence level per person for each year. Both the poverty thresholds and household income are converted to constant 2011 rubles using regional CPI indices provided by the Rosstat. The child scale parameter is set at 0.04 30 Table 5. The effect of alternative specifications on scale parameters estimates, RLMS 1994- 2017 Pooled OL BUC Sensitivity scenarios Baseline Additional Additional Baseline elasticity elasticity child child 0.649*** 0.080*** 0.410*** 0.050* 1 Expectations (0.03) (0.01) (0.09) (0.03) 0.497*** 0.057* 2 Reference group (0.06) (0.02) 0.762*** 0.117*** 0.659*** 0.043* 3 Life satisfaction (0.03) (0.01) (0.11) (0.03) 0.571*** 0.089*** 0.342*** 0.056* 4 Measurement error (0.04) (0.01) (0.10) (0.03) 0.577*** 0.020* 0.306*** -0.013 5 Unrestricted sample (0.03) (0.01) (0.09) (0.02) 6 Pensioners 0.560*** 0.064*** 0.352* 0.046* (0.04) (0.01) (0.15) (0.03) Note: Standard errors in parentheses are calculated using delta-method. All regressions include the same controls as in Table 1. 31 Figure 1. Distribution of household types in Germany, Russia, Switzerland, and the UK Panel A: Average household size Panel B: Single person 3.0 25 % of total population 2.5 20 2.0 15 1.5 1.0 10 0.5 5 0.0 0 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2008200920102011201220132014201520162017 Germany United Kingdom Russia Switzerland Germany United Kingdom Russia Switzerland Panel C: Three or more adults 40.0 35.0 Percentage of all hhs 30.0 25.0 20.0 15.0 10.0 5.0 0.0 2008200920102011201220132014201520162017 Germany United Kingdom Russia Switzerland Source: European Union Statistics on Income and Living Conditions (EU-SILC) and RLMS-HSE 32 Figure 2. Comparison of different equivalence scales Adult equivalent weights Child equivalence weight (first child) Child equivalence weights (second child) 100 90 Weight given to additional member (%) 80 70 60 50 40 30 20 10 0 2 3 4 5 1 2 3 4 5 2 3 4 Number of adults Number of adults Number of adults Per capita Square-root Per capita Square-root RUS Poverty Line Per capita Square-root RUS Poverty Line RUS Poverty Line OECD modified BUC OECD modified BUC OECD modified BUC 33 Figure 3. Scale Factors and Headcount Poverty Rate and Poverty Duration, RLMS 1994- 2017 Note: Absolute poverty line is defined as a minimum regional subsistence level per person for each year (for cross-sectional poverty in 2017). Relative poverty line is set on 60% of household size-weighted median equivalized income for each year (for cross-sectional poverty in 2017). Both the poverty thresholds and household income are converted to constant 2011 rubles using regional CPI indices provided by the Rosstat. 34 Figure 4. Scale Factors and Income Mobility, RLMS 1994-2017 1 1 .9 .9 .8 .8 .7 .7 Adult parameter e Adult parameter e .6 .6 .5 .5 .4 .4 .3 .3 .2 More upward mobility .2 Mixed .1 More downward mobility .1 0 .01 .02 .03 .04 .05 .06 .07 .08 .09 .1 0 .01 .02 .03 .04 .05 .06 .07 .08 .09 .1 Child parameter b Child parameter b 35 Appendix A: Additional Tables and Figures Table A.1: Overview of subjective scales estimated from panel data Weight given to the additional household member Author Data Welfare indicator Subsample Specification 2nd adult (1st 1st child (in 2nd child (in adult = 1) 2/1 adult hh) 2/1 adult hh) Pooled ordered logit 0,5 0.23/0.34 0.15/0.23 Satisfaction with life/ Charlier (2002)* GSOEP, 1984–91 Sample of household heads FE ordered logit (Das and van Soest (1999) approach) 0,43 0.20/0.28 0.10/0.16 Satisfaction with hh income RE ordered logit 0,5 0.26/0.39 0.19/0.28 Respondents who answered Pooled ordered logit 0,34 0.17/0.30 0.08/0.14 Schwarze (2003) GSOEP, 1992–99 Satisfaction with hh income at least twice are included FE binary logit 0,28 0.13/0.24 0.06/0.11 OLS 0,43 0.25/0.39 0.14/0.22 Satisfaction with hh income Households who answered at Orderel probit 0,43 0.26/0.4 0.15/0.23 Falter (2006)** SHP, 1999-2002 least twice are included FE linear model 0,48 0.28/0.43 0.14/0.24 Minimum Income Question FE linear model 0,07 0.09/0.1 0.11/0.11 Individuals 18-80 yo FE ordered logit (Baetschmann et al. (2015) approach), men 0,15 1.12/1.64 0.68/1.22 Self-assessed financial who are household heads or Bollinger et al.(2012) BHPS, 1991-2008 situation the partner of the head with FE ordered logit (Baetschmann et al. (2015) approach), 0,31 1.17/1.52 0.41/0.78 or without minor children women Individuals≥ 17 yo living in Biewen and Juhasz Nonlinear FE ordered logit (based on Baetschmann et al. GSOEP, 1999–2009 Satisfaction with hh income households with less than 6 0,35 0,13 0,13 (2017)*** (2015)) members 0.18-0.29/ 0.10-0.27/ Pooled ordered logit 0.31-0.36 Individuals≥18 yo, one- or 0.29-0.37 0.16-0.30 Borah et al. (2018)**** GSOEP, 1984-2013 Satisfaction with hh income two-adult-households with or Nonlinear least squares 0.31-0.25 0.12-0.31 0.12-0.31 without minor children 0.16-0.13/ 0.10-0.12/ FE ordered logit (Baetschmann et al. (2015) approach) 0.26-0.15 0.24-0.17 0.15-0.14 Note: GSOEP - German Socio-Economic Panel; SHP - Swiss Household Panel; BHPS - British Household Panel Survey *Equivalence weights reflect period-specific equivalence scales based on satisfaction with income in case when the first child is 12 years old and the second child is 6 years old. **Equivalence weights refer to the model with control variables ***Equivalence weights refer to OECD-type scale ****Equivalence weights reflect the cases without and with reference effect measured with household Mincer equation 36 Table A.2: Summary statistics, RLMS 1994-2017 (237 395 obs) Mean SD Subjective welfare 3.83 1.5 Log of household income 9.93 0.9 Household size 3.14 1.3 Number of adults 2.59 1.0 Number of children 0.56 0.8 age16_20 0.07 0.3 age21_30 0.19 0.4 age31_40 0.19 0.4 age41_50 0.17 0.4 age51_60 0.16 0.4 age61_70 0.12 0.3 age71_80 0.08 0.3 age80plus 0.03 0.2 Female 0.58 0.5 Russian nationality 0.87 0.3 Incomplete secondary 0.21 0.4 Complete secondary 0.32 0.5 Secondary + vocational 0.25 0.4 University and higher 0.22 0.4 Single 0.16 0.4 Married 0.63 0.5 Divorced/widowed/separated 0.21 0.4 Have poor health 0.40 0.5 Other household members in poor health 0.53 0.5 Employed 0.61 0.5 Unemployed/out of labour force 0.39 0.5 Log of per capita living space (sqm) 3.10 3.1 Note: data are unweighted 37 Table A.3: Distribution of household types, RLMS 1994-2017 No of adults No of children Share of sample (%) 1 0 9.36 1 1 1.18 1 2 0.29 1 3 0.03 2 0 23.99 2 1 12.07 2 2 6.01 2 3 0.81 3 0 15.24 3 1 8.54 3 2 2.47 3 3 0.42 4 0 7.60 4 1 4.52 4 2 1.64 4 3 0.34 5 0 2.57 5 1 1.94 5 2 0.75 5 3 0.22 Total 100 Number of observations 260,133 Note: data are unweighted 38 Table A.4: Comparison of different equivalence scales Schwarze Falter Borah et al. RUS Estimated scales Per Square- Modified (2003) (2006) (2018) Weights Poverty capita root OECD* FE FE FE Line** POL BUC POL POL POL BUC OLS BL*** OLS 1 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 2 2.00 1.41 1.50 2.00 1.55 1.32 1.33 1.34 1.28 1.43 1.48 1.31 1.26 Adults 3 3.00 1.73 2.00 3.00 2.01 1.55 1.56 1.59 1.47 1.77 1.86 1.53 1.44 4 4.00 2.00 2.50 4.00 2.41 1.74 1.77 1.78 1.63 2.06 2.18 1.71 1.58 5 5.00 2.24 3.00 5.00 2.78 1.90 1.93 1.98 1.76 2.31 2.47 1.87 1.70 1 Adult 1 Child 2.00 1.41 1.30 1.90 1.47 1.27 1.28 1.30 1.24 1.40 1.43 1.29 1.24 1 Child 3.00 1.73 1.80 2.90 1.85 1.47 1.48 1.52 1.41 1.70 1.78 1.49 1.41 2 Adults 2 Children 4.00 2.00 2.10 3.80 1.95 1.51 1.54 1.59 1.47 1.85 1.90 1.59 1.51 1 Child 4.00 2.00 2.30 3.90 2.17 1.62 1.64 1.69 1.55 1.95 2.04 1.65 1.55 3 Adults 2 Children 5.00 2.24 2.60 4.80 2.17 1.62 1.65 1.71 1.56 2.04 2.11 1.71 1.62 1 Child 5.00 2.24 2.80 4.90 2.45 1.75 1.78 1.84 1.66 2.17 2.28 1.79 1.66 4 Adults 2 Children 6.00 2.45 3.10 5.80 2.36 1.71 1.75 1.82 1.65 2.21 2.29 1.82 1.71 5 Adults 1 Child 6.00 2.45 3.30 5.90 2.72 1.87 1.90 1.97 1.76 2.37 2.51 1.91 1.76 Note: Household types whose population share is at least 1%. Children are defined as individuals aged below 16 years (OECD and Rosstat def.) *First adult has weight 1.0, every further adult 0.5, children 0.3. **Working-age adult has weight 1.0, pensioner 0.8, children 0.9 *** Fixed Effects Binary Logit 39 Table A.5. Alternative poverty thresholds and corresponding poverty rates for Russia by household size in 2017 Poverty Line (in rubles) Poverty Headcount (in percent) Household Type Pooled OL BUC Official Pooled OL BUC Official Households without children 6.6 6.6 6.6 One adult, no children 9,607 9,607 9,607 (1.41) (1.41) (1.41) 14,891 12,777 3.7 2.0 9.1 Two adults, no children 19,214 (343) (784) (0.76) (0.46) (1.58) 19,310 14,987 1.8 0.8 12.7 Three adults, no children 28,821 (735) (1,618) (0.61) (0.40) (2.93) 23,153 16,908 2.7 0.7 10.0 Four adults, no children 38,428 (1,127) (2,304) (1.25) (0.70) (2.56) 26,707 18,542 7.0 Five adults, no children 48,035 0.0 0.0 (1,470) (2,990) (3.70) Household with children 14,122 12,297 8.7 7.2 27.5 One adult, one child 19,532 (196) (539) (3.21) (2.97) (6.45) 17,773 14,218 2.7 0.9 16.9 Two adults, one child 29,139 (392) (1,029) (1.14) (0.52) (2.91) 18,734 14,795 4.1 0.4 30.8 Two adults, two children 39,064 (196) (784) (1.10) (0.39) (4.47) 20,847 15,755 4.1 0.9 21.0 Three adults, one child 38,746 (588) (1,519) (1.47) (0.60) (3.49) 20,847 15,852 5.6 1.5 32.0 Three adults, two children 48,671 (245) (980) (2.61) (1.48) (6.41) 23,537 17,100 1.0 1.0 14.7 Four adults, one child 48,353 (784) (1,912) (1.02) (1.02) (3.74) 22,673 16,812 20.0 Four adults, two children 58,278 0.0 0.0 (343) (1,127) (7.61) 26,131 18,253 11.0 Five adults, one child 57,960 0.0 0.0 (931) (2,255) (5.31) Note: Population weights are applied. Standard errors for poverty rates are adjusted for complex survey design. Poverty line for reference “one adult” is defined as an average of minimum subsistence levels for working-age individual and pensioners in 2017. The level of absolute poverty line is 10899 rubles per month for working-age individual, is 8315 rubles per month for pensioner and is 9925 rubles per month for child in 2017. Absolute poverty line of reference adult is adjusted with weights in Table 2 using BUC model (where baseline elasticity equals 0.407 and every child has a weight 0.048) and using Pooled Ordered Logit model (where baseline elasticity equals 0.636 and every child has a weight 0.078). 40 Table A.6. Differences in Poverty Rates for Households with and without Children for Different Equivalence Scales, RLMS 1994-2017 Panel A. Headcount poverty in households with children (percent) Child scale Adult scale parameter e parameter 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 b 0.01 0.4 0.8 1.2 1.7 2.8 5.0 7.5 10.9 15.9 23.1 0.02 0.3 0.7 1.1 1.5 2.5 4.8 6.6 9.9 15.1 21.5 0.03 0.2 0.6 1.0 1.4 2.2 4.1 6.1 9.4 13.9 19.7 0.04 0.2 0.6 1.0 1.4 2.1 3.6 5.5 8.7 13.0 18.6 0.05 0.2 0.6 1.0 1.4 2.0 3.1 5.2 8.1 11.9 17.3 0.06 0.2 0.4 0.9 1.2 1.8 3.0 5.0 7.5 10.9 16.6 0.07 0.2 0.3 0.8 1.1 1.6 2.7 4.6 6.8 10.2 15.5 0.08 0.2 0.3 0.8 1.1 1.6 2.6 4.3 6.2 9.8 14.1 0.09 0.2 0.3 0.8 1.1 1.6 2.4 4.1 6.1 9.4 13.0 0.1 0.2 0.3 0.8 1.0 1.4 2.2 3.7 5.6 8.7 12.5 Panel B. Headcount poverty in households without children (percent) Child scale Adult scale parameter e parameter 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 b 0.01 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.02 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.03 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.04 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.05 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.06 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.07 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.08 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.09 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 0.1 2.8 3.2 3.5 3.9 4.2 4.7 5.4 6.5 8.6 11.3 Panel C. Absolute difference in poverty headcount between households with and without children (percent) Child scale Adult scale parameter e parameter 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 b 0.01 -2.5 -2.3 -2.3 -2.2 -1.4 0.3 2.1 4.4 7.3 11.8 0.02 -2.5 -2.4 -2.3 -2.4 -1.7 0.1 1.2 3.4 6.5 10.1 0.03 -2.6 -2.5 -2.5 -2.5 -1.9 -0.7 0.7 3.0 5.3 8.4 0.04 -2.6 -2.5 -2.5 -2.5 -2.1 -1.2 0.2 2.2 4.4 7.2 0.05 -2.6 -2.6 -2.5 -2.5 -2.2 -1.6 -0.1 1.6 3.3 5.9 0.06 -2.6 -2.7 -2.6 -2.6 -2.4 -1.7 -0.4 1.0 2.3 5.2 0.07 -2.6 -2.8 -2.6 -2.8 -2.6 -2.0 -0.8 0.3 1.5 4.1 0.08 -2.6 -2.8 -2.6 -2.8 -2.6 -2.2 -1.1 -0.3 1.1 2.7 0.09 -2.6 -2.8 -2.6 -2.8 -2.6 -2.4 -1.3 -0.4 0.8 1.7 0.1 -2.6 -2.8 -2.6 -2.9 -2.7 -2.5 -1.7 -0.9 0.0 1.1 Note: Panel C provides the absolute differences in the poverty rate for the same combination of child and adult scale parameters shown in Panels A and B. Absolute poverty line is defined as a minimum regional subsistence level per person for each year. Both the poverty thresholds and household income are converted to constant 2011 rubles using regional CPI indices provided by the Rosstat. 41 Table A.7. The effect of expectations, RLMS 1994-2017 Main variables Pooled Ordered Logit BUC 0.623*** 0.406*** Log of household income (0.011) (0.079) -0.404*** -0.166*** Log of household size (0.022) (0.034) 0.050*** 0.020** Children*Log of household size (0.008) (0.010) Expectations in T-1 period (base – “Nothing will change”) -0.509*** -0.139*** You will live worse (0.016) (0.017) 0.448*** 0.152*** You will live better (0.013) (0.014) Number of observations 237,395 712,448 Log pseudolikelihood -400,970 -263,550 Note: Robust standard errors are in parentheses, controlling for two-way clustering (i.e., at the individual for the POL model and at the household-wave level for the BUC models). All regressions include the same controls as in Table 1. 42 Table A.8. The effect of reference group, RLMS 1994-2017 Main variables Pooled Ordered Logit BUC 0.484*** 0.282 Log of household income (0.021) (0.213) -0.241*** 0.007 Log of household size (0.034) (0.117) 0.028** 0.003 Children*Log of household size (0.011) (0.022) Relative income (base – 1st quartile) 0.161*** 0.071 2nd quartile (0.021) (0.084) 0.310*** 0.203 3rd quartile (0.025) (0.144) 0.477*** 0.323 4th quartile (0.033) (0.234) Number of observations 140,026 366,511 Log pseudolikelihood -237,167 -135,630 Note: Robust standard errors are in parentheses, controlling for two-way clustering (i.e., at the individual for the POL model and at the household-wave level for the BUC models). All regressions include the same controls as in Table 1. 43 Table A.9. The effect of welfare definition, RLMS 1994-2017 Satisfaction with economic Dependent variable Satisfaction with life conditions Pooled Ordered Pooled Ordered Main variables BUC BUC Logit Logit 0.606*** 0.431*** 0.863*** 0.719*** Log of household income (0.010) (0.077) (0.013) (0.112) - Log of household size -0.462*** -0.284*** -0.790*** 0.569*** (0.022) (0.035) (0.025) (0.046) Children*Log of household 0.071*** 0.018* 0.069*** 0.015 size (0.007) (0.011) (0.008) (0.012) Number of observations 240,383 525,832 211,032 446,937 Log pseudolikelihood -329,968 -196,405 -286,373 -169,660 Note: Robust standard errors are in parentheses, controlling for two-way clustering (i.e., at the individual for the POL model and at the household-wave level for the BUC models). All regressions include the same controls as in Table 1. 44 Table A.10. The effect of measurement error, RLMS 1994-2017 Pooled Ordered Logit BUC 0.558*** 0.372*** Log of household income (0.010) (0.070) -0.318*** -0.127*** Log of household size (0.022) (0.032) 0.050*** 0.021** Children*Log of household size (0.008) (0.010) Number of observations 242,768 733,754 Log pseudolikelihood -413,485 -271,759 Note: Robust standard errors are in parentheses, controlling for two-way clustering (i.e., at the individual for the POL model and at the household-wave level for the BUC models). All regressions include the same controls as in Table 1. 45 Table A.11. The effect of sample restriction, RLMS 1994-2017 Main variables Pooled Ordered Logit BUC 0.640*** 0.396*** Log of household income (0.011) (0.079) -0.369*** -0.121*** Log of household size (0.022) (0.033) 0.013* -0.005 Children*Log of household size (0.007) (0.009) Number of observations 245,777 746,710 Log pseudolikelihood -418,712 -276,200 Note: Robust standard errors are in parentheses, controlling for two-way clustering (i.e., at the individual for the POL model and at the household-wave level for the BUC models). All regressions include the same controls as in Table 1. 46 Table A.12. The effect of pensioners, RLMS 1994-2017 Main variables Pooled Ordered Logit BUC 0.657*** 0.409*** Log of household income (0.011) (0.079) -0.368*** -0.144*** Log of household size (0.029) (0.043) 0.042*** 0.019* Children*Log of household size (0.008) (0.011) -0.022 -0.044 Pensioners*Log of household size (0.023) (0.051) -0.040 0.079 Number of pensioners (0.027) (0.060) Number of observations 231,972 692,336 Log pseudolikelihood -394,096 -256,312 Note: Robust standard errors are in parentheses, controlling for two-way clustering (i.e., at the individual for the POL model and at the household-wave level for the BUC models). All regressions include the same controls as in Table 1. 47 Table A.13. Total effect of pensioners by household composition, RLMS 1994-2017 Total number of hh members Number of pensioners Pooled Ordered Logit BUC -0.056*** 0.049* 2 1 (0.01) (0.03) -0.065*** 0.033* 3 1 (0.01) (0.02) -0.129*** 0.067* 3 2 (0.02) (0.03) -0.071*** 0.022 4 1 (0.01) (0.02) -0.142*** 0.043 4 2 (0.03) (0.04) -0.076*** 0.013 5 1 (0.02) (0.03) -0.152*** 0.026 5 2 (0.03) (0.06) Note: Standard errors in parentheses are calculated using delta-method 48 Table A.14. Results for non-parametric scales, RLMS 1994-2017 Parameter Coefficient 1.422*** a1k1 (0.233) 1.243*** a2k0 (0.105) 1.374*** a2k1 (0.146) 1.259*** a2k2 (0.161) 1.454*** a3k0 (0.143) 1.704*** a3k1 (0.187) 1.731*** a4k0 (0.198) 1.836*** a4k1 (0.241) 1.968*** a5k0 (0.312) 1.836*** a5k1 (0.326) Number of observations 74,627 Log pseudolikelihood -237,360 Note: Robust standard errors are in parentheses, controlling for two-way clustering. All regressions include the same controls as in Table 1. 49 Figure A.1. Estimation sample distribution of subjective welfare variable, RLMS 1994-2017 .25 .2.15 Density .1 .05 0 1 2 3 4 5 6 7 8 9 50 Figure A.2. Scale Factors and Headcount Poverty Rate, RLMS 1994-2017 80 Per capita income 70 Equivalent income 60 Headcount poverty (%) 50 40 30 20 10 0 1994 1995 1996 1998 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 Note: Population weights are applied. Standard errors are adjusted for complex survey design. Poverty line is defined as a minimum regional subsistence level per person for each year provided by the Rosstat. Household incomes are adjusted with equivalence scale weights from Table A.7 using BUC model (where baseline elasticity equals 0.407 and every child has a weight 0.048). Both the poverty thresholds and household incomes are converted to constant December prices of 2011 using regional CPI indices provided by the Rosstat. Real values of household incomes are also adjusted for regional differences in the cost-of- living. 51 Figure A.3. Sensitivity of chronic poverty to scale factors, using other definitions of chronic poverty, RLMS 1994-2017 Note: Panels A and B use Foster’s (2009) measure of chronic poverty. Panels C and D use Duclo et al.’s (2010) measure of chronic poverty. Estimates are provided with α=2. Absolute poverty line is defined as a minimum regional subsistence level per person for each year. Relative poverty line is set on 60% of household size-weighted median equivalized income for each year. Both the poverty thresholds and household income are converted to constant 2011 rubles using regional CPI indices provided by the Rosstat. 52 Figure A.4. Scale Factors and Income Mobility, RLMS 1994-2017 53