Policy Research Working Paper 10610 Improving Consumption-Based Tax Compliance Evidence from Point of Sale Usage in Subnational Governments in Indonesia Muhammad Khudadad Chattha Prabaning Tyas Naranggi Pramudya Soko Raka Rizky Fadilla Governance Global Practice November 2023 Policy Research Working Paper 10610 Abstract This paper studies the impact of point of sale technology treatments led to a substantial increase in restaurant tax pay- adoption on local tax compliance by firms. The paper ments, ranging from 55 to 180 percent, while others did not exploits administrative data on monthly restaurant and increase tax revenues. The paper discusses the underlying hotel tax payments in the Indonesian districts of West drivers of the results and argues that the effectiveness of Manggarai and Gorontalo and combines this with infor- point of sales technology is conditional on (i) the devices’ mation on the point of sales distribution timeline from ease of use, (ii) the recipients’ technological aptitude, and 2018 to 2022. The findings show that certain point of sales (iii) the presence of information accessible by third parties. This paper is a product of the Governance Global Practice. 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 mchattha@worldbank.org, ptyas@worldbank.org, nsoko@worldbank.org, and rfadilla@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 Improving Consumption-Based Tax Compliance: Evidence from Point of Sale Usage in Subnational Governments in Indonesia∗ † ‡ Muhammad Khudadad Chattha, Prabaning Tyas, § Naranggi Pramudya Soko, Raka Rizky Fadilla¶ ∗ We thank Francisco Vazquez and Ruth Nikijuluw for their valued comments. We are grateful for the cooperation and assistance of the local government representatives in West Manggarai and Gorontalo, Maria Yuliana Rotok and Nuryanto. We thank Alma Kanani, Habib Rab, Daniel Ortega, J¨ urgen Blum and Roy Kelly for their guidance. We thank Gisella Lokopessy and Marcha Violeta for their support. The work was financed by the governments of Canada, Switzerland, and the European Union, through the World Bank’s Public Financial Management Multi-Donor Trust Fund for Indonesia. All potential errors are our own † World Bank, E-mail: mchattha@worldbank.org ‡ World Bank, E-mail: ptyas@worldbank.org § World Bank, E-mail: nsoko@worldbank.org ¶ World Bank, E-mail: rfadilla@worldbank.org 1 Introduction Government enforcement is a key factor determining tax compliance. The better the enforcement quality, the more it can minimize evasion (Slemrod, 2007). Transaction digitization through electronic payment technology is one such way to improve tax com- pliance. This is because it generates paper trails by both third parties and the government as a reliable basis for audits. Combined with well-functioning monitoring and audit sys- tems, such paper trails can deter under-reporting behavior (Allingham and Sandmo 1972; Kleven et al. 2011, Pomeranz 2015, and Naritomi 2019). Consequently, international or- ganizations have advised promoting the use of an electronic payment system to increase tax compliance (OECD 2018, World Bank, 2020). There is a growing body of literature on third-party reporting and its impact on gov- ernment revenue in advanced economies, as discussed by Gordon and Li (2009), and Kleven et al. (2016). However, there is little evidence on whether enacting transaction digitization programs among businesses can boost compliance and therefore increase tax revenue in developing countries, particularly in subnational settings. This study exploits unique administrative data on restaurant tax receipts in the Indonesian districts of West Manggarai and Gorontalo.1 We aim to explore the variation in restaurant tax payments before and after the distribution of the Point of Sale (POS) machines among hotels and restaurants located in the districts. Initiated by the Indonesian government and gradually implemented at sub-national levels following the recommendation of The Commission of Corruption Eradication (Komisi Pemberantasan Korupsi ), the program was meant to optimize local governments’ own- source revenues (Komisi Pemberantasan Korupsi, 2020). Once installed, POS devices enable the direct dispatch of sales records from targeted tax subjects to the local tax authority. This mechanism enhances the accuracy of users’ taxable transaction reports and minimizes the possibility of taxpayers’ underreporting. Down the line, the produced sales records can be accessed by the tax authority for compliance monitoring. We begin by discussing a conceptual framework to consider how surveillance enhancement may affect firm behavior. The POS distribution program improves the accuracy and timeliness of taxable sales records and has a deterrent effect on businesses’ noncompliance. However, for POS adoption to positively affect tax payments, the program’s beneficiaries must install and use POS units. When left idle and/or tampered with the purpose of altering tax liability, the government’s control function running through the system becomes muted. However, studying recipients’ behavior towards adopting the assigned POS is beyond the scope of our study. This study investigates the general link between a program’s application and the obtained tax payments. To empirically examine the influence of POS distribution on tax payment, we utilized ad- ministrative data on firm-level monthly restaurant tax receipts, the list of POS recipients, and the time of POS distribution to perform a difference-in-difference (DiD) estimation in the Indonesian districts of West Manggarai and Gorontalo. 1 Throughout this paper, ’Gorontalo’ refers to The City of Gorontalo (Kota Gorontalo), not to be mistaken with The Regency of Gorontalo (Kabupaten Gorontalo) or The Gorontalo Province. 1 The two districts use several types of POS machines, Regular POS (RPOS), Transaction Monitoring Device (TMD), and Web Service (WS). Gorontalo, however, also employs an additional POS type for hotel taxpayers called PAC Hotel and a modification of RPOS with connection to third-party data, namely RPOS-Grab. RPOS refers to a complete set of sales recording systems from the cash register machine and interface which helps users record transactions and print receipts. TMD, on the other hand, refers to a tapping device which is planted to an existing sales recording system and steals the transaction data. While WS is an API based software that reads users’ online sales system and sends the transaction records to the local governments. The POSs were distributed based on preexisting condition of the receivers, which we describe in detail in the next section of the paper. Our research demonstrates that the influence of various POS device types on tax compli- ance varies. In West Manggarai, both TMD and WS have successfully improved restau- rant tax compliance. We estimate that TMD and WS led to 180 percent and 55 percent increases of restaurant tax payments compared to the control group, respectively. WS, however, proved to have no significant effect on improving hotel tax compliance in the district. In Gorontalo, restaurants that obtained RPOS-Grab on average paid 68 percent higher restaurant tax compared to those in the control group. On the other hand, restaurants that received RPOS units without Grab connection and TMD have indifferent tax pay- ments on average compared to their control. TMD and WS significantly increased tax payments among observed hotels and restau- rants in both districts. Our findings shed light on the influence of business size and existing technological mastery on the efficacy of POS adoption in boosting tax payments. Larger businesses with better proficiency in using electronic transaction technology tend to pay higher taxes on POS assignment. That leaves regular POS (RPOS) as the only POS type with consistent statistically insignificant impact on tax receipt in both districts. Our study contributes to the literature on tax compliance by providing unique insights into how various transaction-monitoring technologies could affect tax compliance in devel- oping economies’ subnational settings. In doing so, we connected several sets of studies. The first comprises a set of studies discussing tax enforcement (Andreoni et al., 1998; Slemrod and Yitzhaki, 2002; and Slemrod, 2019). Second are those evaluating policy- induced third-party reports on value-added tax (VAT) liabilities (Bellon et al., 2019; Lovics et al., 2019; Fan et al., 2018). The third is a group of studies on consumers’ use of electronic payment technologies (Brockmeyer and Somarriba, 2022; Agarwal et al., 2007; Bolt et al., 2010). Unlike existing literature, which mostly focuses on the acceptance aspect of POS devices, this study is the first to connect the POS distribution program and local tax payment data to examine the program’s overall efficacy. The remainder of this paper is organized as follows. Sections 2 and 3 present the back- ground and data used in our study together with some conceptual foundations. Sections 4 and 5 explain the empirical strategy employed and results, respectively. Section 6 extends the interpretation of the results, and Section 7 concludes. 2 2 Background and Data 2.1 Domestic Revenue Mobilization in Indonesia Indonesia’s tax revenue as a share of GDP has been on a declining trend since 2008, reaching its lowest in 2020 at 8.3 percent (Figure 1, red line). The World Bank has estimated that the country’s tax gap in 2020 was 6 percent and widening (Figure 1, gap between Frontier and Actual), showing the declining tax effort over time (Chattha et al., 2023). Figure 1: Declining Tax Ratio & Widening Tax Gap Source: World Bank staff estimation As a decentralized country, Indonesia delegates most of the spending responsibilities to the subnational governments (SNG), mostly district governments. Since the decentralization reform in 2001, the share of SNGs’ spending has rarely gone below 50 percent of the total government expenditure excluding subsidies, interest payments and transfers (Figure 2). Figure 2: Public Expenditures by Level of Governments Source: data.worldbank.org Although districts’ reliance on the central government’s intergovernmental transfers has declined over the last two decades, intergovernmental transfers remain the main revenue source of SNGs. The increasing district’s own-source income (OSR) has played a role in reducing its dependency on the central government. OSR is a revenue source which is generated directly by SNGs. In 2020, OSR accounted for 23.7 percent of total district revenues, up from 15.5 percent in 2001 (Figure 3). 3 Across district governments, local taxes are the biggest chunk of district governments’ OSR, making local taxes an important revenue source for SNGs. Based on Law Number 23/2009, district governments in Indonesia are allowed to collect taxes from 11 different tax bases. The list includes hotels, restaurants, entertainments, advertisements, street lighting, non-metal minerals and rocks, parking, ground water, swallows’ nests, residential properties, and property transfers. This paper investigates the impact of Indonesian districts’ POS distribution on their acquired hotel and restaurant tax payments. Figure 3: Public Revenues by Components Source: data.worldbank.org 2.2 Domestic Revenue Mobilization in West Manggarai and Gorontalo 2.2.1 West Manggarai West Manggarai is one of the eight districts that comprise the island of Flores, located in the province of East Nusa Tenggara in Indonesia. It covers an area of 9,450 km2 and had a population of 274,000 people in 2019. In contrast to Gorontalo, West Manggarai is a regency instead of a city government. Regency status is generally linked with less development and vaster area than a city. This regency’s economy is mainly driven by agriculture and fisheries. The district primarily relied on transfers to finance spending in 2020. OSR is a small proportion of overall revenues. However, local tax revenues primarily dominated its OSR by almost 60% of total OSR. West Manggarai’s top three taxes are Street Lighting Tax, Property Tax and Restaurant Tax. Restaurant tax contributes IDR 1.7 billion or 12% of the district’s local tax revenue in 2021. The LG, however, has highlighted the importance of restaurant tax as the CG has just established Labuan Bajo, a town located in the district, as one of Indonesia’s Super Tourism Destination Priority of the country. 2.2.2 Gorontalo Gorontalo City is the capital city of Gorontalo Province, which is located in the north of Sulawesi island. The province was part of North Sulawesi Province, before its expansion in 2000. Based on the 2020 census, 1.7 million people live in its administrative area of 12,025 square kilometers. The city’s economy is dominated by retail and services, and public administration takes a big part of the economy as the third biggest sector in the city. 4 Figure 4: Revenue Sources of West Manggarai Figure 5: West Manggarai’s Local Taxes Compositions Source: data.worldbank.org Digging deeper into the public revenue side, intergovernmental transfers from the central government played a substantial part by taking more than 50% of the LG’s revenue in 2020. Own-source revenue (OSR) only contributed 25.6% of the total revenue. Moreover, 67.8% of OSR came from Other Legitimate PAD, which was generated by Gorontalo Local Hospital. Local taxes were the second best OSR contributor by generating 22.4 Restaurant taxes plays an important role in Gorontalo’s OSR. In 2020, restaurant tax was the second biggest tax revenue eleven different taxes in the district. It contributed around IDR 10.5 billion or 20% of the total local tax revenue. Although not as significant as restaurant tax, hotel tax was the fifth biggest tax by contributing 8% of the total local tax revenue. 5 Figure 6: Revenue Sources of Gorontalo Figure 7: Gorontalo’s Local Taxes Compositions Source: data.worldbank.org 2.3 The POS Distribution Program In Indonesia, Point of Sales (POS) devices have been distributed to business taxpay- ers—namely hotels, restaurants, entertainment providers and parking space operators—since as early as 2019 following the country’s anti-corruption agency (Komisi Pemberantasan Korupsi) recommendation to optimize local governments’ OSR. The objective of this policy innovation is to increase tax revenues from the aforementioned tax bases. In this study, outcome variables of interest are hotel tax and restaurant tax. The procurement of POS has been typically handled by district-level local governments under the collaboration with regionally owned banks. POS units were given to targeted taxpayers at no cost and usually came with user support packages. Recipients were entitled to the provider’s technical support, which has their personnel ready in the district where their service is available to ensure prompt responses. Users who were particularly new to POS technology used this kind of support daily, especially at the beginning when they were still acquainting themselves with the tools. POS devices work by recording transactions on a real-time basis and directly relaying this information to the local tax authorities’ system. The relayed records are then used for monitoring purposes. Theoretically, this flow is meant to minimize the chance for 6 taxpayers (POS users) to alter their true sales reports, hence enhancing transparency. On the surveillance side, the system has been designed such that tax authorities are warned for any indications of (i) not reporting some of their transactions through POS, and/or (ii) turning off the devices altogether. For the first indication, both provider and tax authority receive a warning on a dashboard whenever there are no recorded transactions through POS for a certain period. In both cases, the local tax authority would normally respond by reaching out to users, checking if the inactivity was due to technical problem or fraud. If it were the former, the tax authority would contact the provider for follow-up. However, if fraud occurred, the tax authority would send an officer to perform a direct check at the restaurants or hotels. The officer would then have a discretion to escalate the case to an audit if necessary. In general, three POS types—Regular POS (RPOS), Transaction Monitoring Device (TMD), and Web Service (WS)—have been distributed in both West Manggarai and Gorontalo. In West Manggarai, restaurants were given either RPOS, TMD, or WS, while hotels were exposed only to WS and not to the other POS types. On the other hand, treatment restaurants in Gorontalo were facilitated with either RPOS, RPOS with optional connection to Grab’s system (RPOS-Grab), and TMD. Treatment hotels in the district were either given TMD or access to the PAC Hotel application. Detailed explanations on each type of POS machine are presented in the later part of this section. Both in West Manggarai and Gorontalo, program’s beneficiaries were assigned to most suitable POS type according to their needs and existing situation (see Table 1). RPOS refers to the entire sales recording system from cash register machine and inter- face which helps users record all transactions electronically. RPOS was usually given to restaurants with rudimentary transaction recording system such as manual bookkeeping or offline cash register. The data recorded on the RPOS is then sent to the local tax au- thorities online. In Gorontalo, RPOS consists of both hardware parts (monitors, tablets, keyboards, cash drawers, etc.) and the software (application or program). However, in Manggarai Barat, the vendor only provided the software. Thus, either the district government or taxpayers must provide the hardware needed for the RPOS to run on. RPOS-Grab is just RPOS with direct connection to Grab’s omni service application, enabling the RPOS device to automatically record sales made via Grab’s online food delivery arm, Grabfood. This direct link to Grabfood’s system is an add-on which was fully optional to the RPOS recipients. Those who chose to use it usually looked for the convenience of not having to manually backup their Grabfood transactions. TMD, also referred to as Tapping Box or Interceptor Box, is a device plugged into users’ existing electronic transaction system. It works without imposing changes in users’ system flow, thus requires less adjustment in terms of technology adoption compared to RPOS. All that the device does is pass on the transaction records, using a parsing technique, from the existing system to local tax authorities. This type of POS is designed for those already with an electronic, but offline, cashier system. WS—also known as Client Reader—is basically a software with API that reads users’ 7 online sales system to extract transaction records to be passed on to tax authorities. It is designed for larger users which typically have their own online sales system like chain hotels or restaurants. In contrast to RPOS and TMD that require both hardware and software installation, WS procurement is much simpler and cheaper as it merely requires installing WS software into users’ system. Lastly, PAC Hotel refers to an application specifically for recording hotel transactions in Gorontalo. The application is installed on the taxpayers’ device which then can be used for recording transactions. To record the transactions, taxpayers need to input the transaction data manually into the application. Local tax authorities usually give this application to small hotels who already have a computer but do not have their own transaction recording system. This treatment is equivalent to RPOS given to restaurants, only PAC does not come with the hardware. Similarly, PAC Hotel also sends transactions data to local tax authorities in real-time basis. Table 1: Targeted taxpayers before and after being assigned POS machines Taxpayer Before POS Machines Taxpayer After POS type Manual recording or bookkeeping RPOS (West Manggarai) RPOS or RPOS-Grab (restaurants-Gorontalo) PAC Hotel (hotels-Gorontalo) Electronic recording (cash register) Offline system POS TMD Own online sales system WS 2.4 Data To study the effect of POS on tax payment, we constructed a unique administrative data set on restaurant and hotel tax receipts in the regency of West Manggarai and district Gorontalo. Our data on restaurant tax receipt was paired with the list of restaurants in the two districts that received any of the POS installations during the program as well as the time they received it. Extracting from the information, we can thus identify treatment group, control group, as well as intervention time for our analysis purpose. Given the rudimentary stage of the POS distribution program’s scalability, the number of POS recipients is relatively small compared to those who have not received one–as the Data section shows. That results in small-sized treatment groups in our estimations, which may compromise the statistical power, and either mute or exaggerate the estimated effects. This could be a caveat, however with the near-complete administrative data on such a program, such a limitation is rather natural and leads to a minimum bias in the results. Therefore, the policy insights that this study offers remain valid. 8 Table 2: Summary Statistics Sample Obs. Mean* Std. Dev. Min* Max* West Manggarai: Restaurants All 18,282 1.889 15.252 0 629.083 RPOS 662 1.822 5.434 0 64.805 TMD 424 10.370 15.436 0 85.146 WS 236 57.744 111.881 0 629.003 Untreated 17,506 0.184 0.596 0 4.724 West Manggarai: Hotels All 11,280 4.968 45.170 0 1,659.119 WS 37 250.881 417.513 0 1,659.119 Untreated 11,243 4.159 35.904 0 1,332.909 Gorontalo: Restaurants All 14,730 4.259 14.415 0 301.368 RPOS 2,762 2.219 2.968 0 25.219 RPOS-Grab 230 2.103 2.030 0.006 8.921 TMD 580 30.961 30.464 0 180.854 Untreated 11,158 3.421 13.601 0 301.368 Gorontalo: Hotels All 14,337 2.257 17.103 0 386.770 PAC Hotel 109 6.580 2.624 0.490 13.823 TMD 218 96.938 86.998 0.570 386.770 Untreated 14,010 0.750 6.334 0 168.110 *in millions of Rupiah. ’0’ indicates a nonpayment. 9 2.4.1 West Manggarai Data on West Manggarai spans from January 2019 to December 2022. It was extracted from administrative data from more than 400 restaurants with a local tax registration number (NPWPD), which reduced to a balanced panel dataset comprising 399 restaurants (NPWPDs) for 48-month period post-cleaning. In the sample, a total of 30 restaurants were given POS units with each of them receiving only one type of POS (RPOS, TMD, or WS)—14 received RPOS, 9 received TMD, and 7 received WS. All of them obtained the units at separate times throughout the observed timeframe. The control group in each estimation is identical and comprises of 369 restaurants (NPWPDs) that did not receive any POS devices. The data set also includes 225 hotels with 6 of them being facilitated with WS, thus leaving the remaining 219 in the control group. Unlike restaurants, hotels in West Mang- garai were not subject to distribution of RPOS and TMD (Table 3). As the enactment of POS distribution program in West Manggarai came in waves, our data naturally carries varying treatment times, which we accommodate in the form of panel data analysis. 2.4.2 Gorontalo Compared to West Manggarai, Gorontalo’s data is more comprehensive. It spans longer from January 2018 to December 2022 and covers more businesses—up to 1,081 restaurants in total. Among them, 92 received RPOS, 14 received RPOS with added connection to Grab’s system (RPOS-Grab), and 7 received TMD. With 5 restaurants using 2 types of POS, the treatment group reduced to 107 restaurants and left the remaining 974 as controls. In contrast to West Manggarai, we found no information regarding the WS assignment to restaurants in Gorontalo. There are 400 hotels in the districts included in our Gorontalo observation. Two of them were given access to the PAC Hotel application and 3 were equipped with TMD devices. This boils down to 5 being in the treatment group and 396 in the control one. Intervention time in the district also varies (Table 3). 10 Table 3: Targeted taxpayers before and after being assigned POS machines Panel A: West Manggarai Restaurant Treatment Intervention Timeline N Treatment RPOS Oct. 2019, Nov. 2019, May 2021, March 2022 14 TMD October 2019 9 WS Dec. 2020, June 2022, and Nov. 2022 7 Hotel Treatment Intervention Timeline N Treatment WS June 2022 and Nov. 2022 6 Panel B: Gorontalo Restaurant Treatment Intervention Timeline N Treatment November 2020, February 2021, March 2021, April 2021, May 2021, June 2021, Aug. 2021, RPOS 92 Oct. 2021, Nov. 2021, Feb. 2022, March 2022, June 2022, and Aug. 2022 RPOS- November 2020 and February 2021 6 Grab January 2021, February 2021, June 2021, July TMD 2021, August 2021, September 2021, January 14 2022 and June 2022 Hotel Treatment Intervention Timeline N Treatment PAC Hotel March 2021 and May 2021 2 TMD March 2021, June 2021, July 2021, June 2022 3 11 3 Conceptual Framework Our analysis is guided by a conceptual framework on how the probability of detection can affect the evasion decision by taxpayers, as has been discussed in Allingham and Sandmo (1972). We proceed with the baseline case in which government monitoring acts as sole enforcement as cited in Naritomi (2019). Suppose a risk-neutral firm generates reported revenue Y ≥ 0 and makes proportional tax payment t ∈ [0, 1] out of it. The firm’s reported revenue, Y , could differ from its ¯ is the total evasion by the firm—in the case of actual pretax revenue Y . E (Y ) = Y − Y under-reporting, it becomes Y < Y . Let p(AE ) ∈ [0, 1] be the probability of audit in the event of evasion and d ∈ [0, 1] the government’s ability in detecting evasion in an audit. Firm faces detection probability p ≡ p(AE )d, where p′ (E ) = p′ (AE )d > 0. When firms get caught for evading, a fine θ ≥ 0 applies in proportion to the size of evasion t(Y − Y ¯ ). Therefore, firms maximize π by reporting revenue Y as follows: ¯ − θt(Y ¯ − tY )(1 − p) + [(1 − t)Y π = (Y ¯ − Y )]p, (1) Where firm chooses the interior optimal solution Y ∗ that satisfies first-order condition dπ/dY = 0 and [p(AE ) + p′ (AE ).E ]d(1 + θ) = 1. In practice, POS usage acts as an extra surveillance tool to monitor tax compliance, which enhances government’s evasion detection ability d. As d increases, the chance of getting caught, p (which is a function of d), rises. Given the convexity of Ep(AE ), Y ∗ will be rising in d-and therefore in p. In other words, when POS device is used properly and quality audit is in place, we can also expect the magnitude of evasion E also shrinks. Therefore, it reduces firms’ tax- evading behavior and drives the optimum reported revenue Y ∗ upward—manifested in overall higher tax receipts. 4 Empirical Strategy We evaluate the impact of giving POS installations to hotels and restaurants by running balanced panel difference-in-difference (DiD) specification for each of the POS types cov- ering a 48-month window throughout 2019 – 2022 (West Manggarai) and a 60-month one during 2018 – 2022 (Gorontalo). Hotels and restaurants which received any of the POS in or before month-of-year t belong to the treatment group, while the rest belong to the control group. By design, each recipient hotel and restaurant were given only one type of POS, therefore the observed impact of different types of POS on hotel/restaurant tax payment were not likely to overlap with each other. Following the nature of our data on West Manggarai, treatment time varies among restau- rants within the treatment group as POS distribution came in waves. An exception applies for the case of tapping box in the districts’ data where all restaurants in the treatment group received their tapping box installation in about the same month (Octo- 12 ber 2019). Such variation in treatment time also applies in Gorontalo. We performed DiD estimations separately for RPOS, TMD, and WS (West Manggarai) and RPOS, RPOS-Grab, and TMD (Gorontalo) to get a picture of POS impact on restaurant tax receipts. Likewise, we run separate estimation for WS (West Manggarai) and PAC Hotel and TMD (Gorontalo) on our hotel observations to get a sense of POS’ impact on hotel tax receipts. We opted to run the estimation routine separately for each type of POS devices instead of a single estimation on multiple treatments sample set due to a number of fundamental reasons. Firstly, pooling all the treatment types into a single estimation apparently requires a stricter assumption, that is, the effect of at least one type of treatment needs to be homogenous across observations (Fricke, 2017). In our case, the assumption requires, for example, the expected impact of RPOS distri- bution on supposedly TMD-targeted restaurants should be identical to RPOS’ impact on RPOS-targeted restaurants. Likewise, the expected effect of WS distribution among supposedly RPOS-targeted restaurants should be about similar to that of originally WS- targeted restaurants, and so on. That is not a plausible assumption given the non- random treatment assignment in most policy interventions, including in our case. More often than not, choice of treatments and targeted treated observations are usually fixed considering the varying efficacy of different treatments on different treated observation groups. Furthermore, we also detected a strong tendency of behavioral variation among POS recipients of different sizes, scale, and technological mastery which again signals the weak foothold for such an identifying assumption. Secondly, we do not have a clear a priori view on the relative intensity among different POS types discussed in the paper. In other words, the POS types are not clearly ordered in terms of effect strength on the outcome–they are distinct, and each works differently. This prevents us to confidently infer from a single estimation with multiple treatments as a lower bound for ATETs magnitude, as has been laid out in Fricke (2017). Throughout our estimation routine, control groups in each estimation excluded hotel and restaurants which received other treatments (for instance, when estimating RPOS treatment, the control group has excluded those which received TMD or Web Service). This was done in anticipation of potential underestimation that would arise if restaurants receiving different POS from the one being examined remain in the control group. We attempt to compare treated hotels and restaurants (re: those that received POS) to the control ones—pre-treatment to post-treatment period. For each POS type, we concretely estimate: ln Ri,t = α + µi + λt + βT reati × P osti,t + ui,t , (2) where ln Ri,t is the log of our outcome variable, recorded amount of hotel and restaurant tax receipt of hotel/restaurant i in a month t; α is a constant, µi are restaurant/hotel fixed effects, and λt are time dummies. The variable T reati = 1 if restaurant i received the POS 13 machine in question and T reati = 0 if it did not receive one at all. The term uit stands for residuals. Since we allow for varying treatment time within the treatment group, the term P osti t has both subscript i and t. Consequently, P osti t = 1 if restaurant i belongs to the treatment group (hence T reati = 1) and t is equal to or later than the treatment time for each i. Subscript i represents the panel identifier of each hotel/restaurant i, while subscript t indicates month i in the observed period. Our coefficient of interest, β , is the estimated DiD coefficient usually coined as the Aver- age Treatment Effect of the Treated (ATET), which measures percentage change of paid tax because of facilitating restaurants with POS installations. Since the data contains a lot of zeros signifying zero tax payment in each month, we performed hyperbolic-sine transformation to avoid data distortion problem which would arise if we did simple loga- rithmic transformation instead. Another form of logarithmic transformation, log (z + 1), is also presented for robustness check. Our baseline specification assumes businesses in the treatment group (1) to have been given POS machines on the first day of the month and (2) to use the units/devices imme- diately. Such an assumption implies a contemporaneous effect of giving POS machines to hotels/restaurants on their respective tax payments. Nevertheless, we are aware that they did not necessarily receive their POS machines in the beginning of the month (it could be in the middle or the end of the month). In that instance, POS usage might not impose noticeable impact—if any—in that same month. Delayed use of POS machines was also quite possible, especially if recipient businesses were new to computerized cashier system, therefore requiring sometimes to internalize POS machines into their operations. To take these possibilities into account, we also run a lagged DiD specification as follow: ln Ri,t = α + µi + λt−1 + βT reati P osti,t−1 + ui,t−1 (3) Where ln Ri,t stands for monthly tax receipt of hotel or restaurant i in month t. Unlike the baseline specification, it allows POS usage in any month t − 1 to take sometimes before it can affect tax receipt in the following month t. 5 Results 5.1 West Manggarai Our main DiD results for West Manggarai are shown in Table 4 and Table 5. Each row pertains to DiD estimates of a different type of POS, as denoted by coefficient β of the interaction term T reati × P ostit from equation 1. Each row pertains to a different treatment (POS) type. In the case of West Manggarai, we found that distribution of TMD and WS units to be impactful on restaurant tax receipts. It was estimated that restaurant tax receipts generally increased by about 180 percent (1.8 times higher than that of the control group) 14 after the provision of TMDs. The distribution of WS also imposed a positive impact on restaurants tax receipt, albeit a less pronounced one when compared to TMD. WS allotment brought about 55 percent higher restaurant tax payment compared to those in the control group. The estimates are statistically significant at 5 percent and 10 percent, respectively. Figure 8 Panel A depicts our main DiD results on restaurant observations. Overall, the observed wedge between treatment and control group for TMD far exceeded those ob- served in RPOS and WS. TMD’s stand out impact on tax compliance among restaurants is highly visible, as treatment group’s estimated tax payment diverged further away from that of control group upon TMD’s distribution in October 2019 (Figure A.14 Panel A). WS, on the other hand, seems to impose a less significant impact as visible divergence between treatment and control group. Lastly, the wedge between treatment and con- trol group in RPOS’ case is much less sizeable throughout the observation time, both in terms of outcome value (Figure A.14 Panel A) and growth (Figure A.14 Panel B), which manifests in a quite thin gap between its DiD estimates graphs (Figure 8 Panel A). WS’ moderate impact on tax compliance is also shown among our hotel observations (Tabel 5). WS-receiving hotels paid tax by around 51 percent higher compared to their control counterparts. This result is statistically significant at 10 percent. Such a sizeable impact of WS among hotel samples in West Manggarai is visible, both in terms of trend and change overtime (see Figure A.16). 15 Table 4: Baseline Difference-in-difference Estimation: West Manggarai (Restaurants) Control vs. POS recipient restaurants Dependent Variable: (1) (2) (3) Log (restaurant tax receipt) RPOS TMD WS ATET (given RPOS vs. not) 0.216 (0.117) ATET (given TMD vs. not) 1.814*** (0.222) ATET (given WS vs. not) 0.582** (0.295) Constant 0.00870 0.00883 0.00899 (0.0259) (0.0258) (0.0297) Observations 18,186 17,930 17,790 Adj. R2 0.0093 0.1028 0.0799 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 Table 5: Baseline Difference-in-difference Estimation: West Manggarai (Hotels) Control vs. POS recipient hotels Dependent Variable: (1) Log (hotel tax receipt) WS ATET (given WS vs. not) 0.517* (0.311) Constant 0.002 (0.050) Observations 10,800 Adj. R2 0.002 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 16 Figure 8: Effect of POS Distribution on Tax Compliance in West Manggarai Control vs POS Recipient Difference-in-Difference Estimation (a) Restaurant 17 (b) Hotel Notes : These graphs present the log (hyperbolic sine) form of predicted tax receipts of treatment and control group for each POS type using coefficients from the main DiD result. Red, vertical, reference lines refer to the actual treatment/intervention times. Panel A depicts the results for restaurant observations, while Panel B depicts the results for hotel observations. 5.1.1 Gorontalo Our estimation using data on Gorontalo suggests that RPOS-Grab distribution imposes a positive and significant impact on restaurant tax payment (Table 6). Restaurants that obtained RPOS and opted for integration with Grab application on average paid taxes by 68 percent higher compared to those in the control group, which showed visually in Figure 9 Panel A. Meanwhile, restaurants which obtained RPOS units but opted out from direct connection with Grab interface made only indistinguishable tax payments on average compared to their control counterparts. That applies also in the case of TMD, which generated statistically negligible effect on restaurants’ tax compliance. Tax receipts had already widely differed between treatment and control group long before the POS distribution program being enacted (see Figure A.18). Table 7 presents the main estimation results for Gorontalo’s hotel observations. According to the table, neither access to PAC Hotel nor TMD units imposed statistically significant effect on the hotels’ tax compliance (as measured by their tax receipts). Hotels that have been given access to PAC Hotels are relatively small, hence there was a chance that (1) in practice, they did not utilize PAC Hotel or (2) they used it, but they just ended up paying the equally low amount of tax to the authority due to their naturally low revenue from the start. Another possible explanation for this could be the low statistical power given the minuscule size of treatment group in both PAC Hotel and TMD estimation. These all are visualized into rather narrow payment gaps between the treatment and control groups in 9 Panel B. 18 Table 6: Baseline Difference-in-difference Estimation: Gorontalo (Restaurants) Control vs. POS recipient restaurants Dependent Variable: (1) (2) (3) Log (restaurant tax receipt) RPOS RPOS- TMD Grab ATET (given RPOS vs. not) 0.134 (0.081) ATET (given RPOS-Grab vs. not) 0.682** (0.335) ATET (given TMD vs. not) 0.155 (0.150) Constant 0.949*** 0.895*** 1.015*** (0.031) (0.033) (0.032) Observations 13,920 11,388 11,738 Adj. R2 0.000 0.008 0.000 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 Table 7: Baseline Difference-in-difference Estimation: Gorontalo (Hotels) Control vs. POS recipient hotels Dependent Variable: (1) (2) Log (hotel tax receipt) PAC TMD Hotel ATET (given access to PAC Hotel vs. not) 0.141 (0.129) ATET (given TMD vs. not) 0.213 (0.153) Constant 0.325*** 0.373*** (0.008) (0.008) Observations 14,119 14,228 Adj. R2 0.002 0.006 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 19 Figure 9: Effect of POS Distribution on Tax Compliance in Gorontalo Control vs POS Recipient Difference-in-Difference Estimation (a) Restaurant 20 (b) Hotel Notes : These graphs present the log (hyperbolic sine) form of predicted tax receipts of treatment and control group for each POS type using coefficients from the main DiD result. Red, vertical, reference lines refer to the actual treatment/intervention times. Panel A depicts the results for restaurant observations, while Panel B depicts the results for hotel observations. 5.2 Robustness Tests We conducted several robustness checks, with the main one being a placebo test. The test is typical in DiD analysis routine, with the objective to test for results’ validity. In the test, we picked both ‘fake’ intervention time and ‘fake’ observation time. A fake observation time is a time window prior to the implementation of the real treatment—that is, we excluded the observations from the real treatment time when conducting this test. Meanwhile, a fake intervention/treatment time is a time point when each treatment did not really take place. We picked fake intervention time for each POS type so that we have enough observations pre and post the chosen fake treatment time points (see Table 8 below). Table 8: Treatment and Observation time choice for placebo test West Manggarai Restaurants Treatment Earliest Intervention Fake Treatment Time Fake Observation Time RPOS October 2019 May 2019 Jan. 2019 – Sept. 2019 TMD October 2019 May 2019 Jan. 2019 – Sept. 2019 WS December 2020 May 2019 Jan. 2019 – Sept. 2019 Hotels Treatment Earliest Intervention Fake Treatment Time Fake Observation Time WS June 2022 Sept 2021 Jan. 2019 – May 2022 Gorontalo Restaurants Treatment Earliest Intervention Fake Treatment Time Fake Observation Time RPOS November 2020 June 2019 Jan 2018 – Dec 2019 RPOS-Grab November 2020 June 2019 Jan 2018 – Dec 2019 TMD January 2021 June 2019 Jan 2018 – Dec 2019 Hotels Treatment Earliest Intervention Fake Treatment Time Fake Observation Time TMD March 2021 June 2019 Jan 2018 – Dec 2020 PAC Hotel March 2021 June 2019 Jan 2018 – Dec 2020 The logic of placebo test is quite straightforward: if the test produces insignificant estimates, then we can be confident that our baseline results are attributable to the treatments. This is because the absence of significant estimates during fake observation time—before and after fake treatment time—indicates that no other factors imposing significant impact to the outcome variable (Angrist and Pischke, 2009). Otherwise, we cannot confidently attest the estimated impact to our treatments. The test can increase our confidence that the parallel trends assumption between treatment and control group whichis crucial for DiD estimation is likely to be true (Fredriksson et al., 2019). 21 5.2.1 West Manggarai Our placebo test result shows that our findings for West Manggarai are reliable. The test confirmed that there was no significant effect of POS circulation on restaurant tax receipt prior to the earliest intervention time of each POS types (or the ‘fake’ intervention time), which are reflected in insignificant ‘fake’ DiD estimates in Table 9. In other words, significant increases of restaurant tax payments detected in the baseline result in Table 4 can be attributed entirely to the usage of POS. Graphical representations are pictured in Figure 10. With the alternative logarithmic transformation of z = log (x + 1), our baseline results hold except with slightly lower point estimates across POS types and period (Figure A.1 Panel A). We also run a lagged specification as stated in equation (2) to see whether lagged effects of POS would differ from those in the baseline result. Table A.1 in the Appendix shows that the lagged estimates are consistent in the case of TMD. This suggests that TMD has both concurrent and asynchronous impact on restaurants’ tax compliance. As for RPOS and WS, the ATET turns statistically insignificant, which appears in a form of narrower gap between treatment and control, especially after May 2022 (Figure A.14 Panel B). As for hotel observations, the parallel trend assumption is not likely to hold true as the placebo test detects a statistically significant source of difference between treatment and control groups prior to the earliest WS distribution among hotels in West Manggarai in December 2020 (see 10 Panel B). A sizable gap of predicted tax receipts has been present even before the intervention took place (see A.16 panel B). Consequently, WS-recipient hotels’ higher tax payment could be attributed to other factor(s) other than POS usage. 22 Table 9: Placebo Test: West Manggarai (Restaurants) Control vs. POS recipient restaurants Dependent Variable: (1) (2) (3) Log (restaurant tax receipt) RPOS TMD WS ATET (given RPOS vs. not) -0.00657 (0.00424) ATET (given TMD vs. not) -0.00656 (0.00423) ATET (given WS vs. not) -0.00508 (0.00328) Constant 0.00891***0.00902***0.00909*** (0.00257) (0.00260) (0.00262) Observations 1,920 1,895 1,880 Adj. R2 0.004 0.055 0.072 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 Table 10: Placebo Test: West Manggarai (Hotels) Control vs. POS recipient hotels Dependent Variable: (1) Log (hotel tax receipt) WS ATET (given WS vs. not) 3.086*** (0.358) Constant 0.002 (0.048) Observations 10,350 Adj. R2 0.063 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 23 Figure 10: Placebo Test: West Manggarai (a) Restaurant (b) Hotel 24 Notes : Panel A visualizes the placebo test of our main DiD estimation for restaurant observations, in which observed timeframe being cut up to June 2019 to exclude the actual intervention and post-intervention period. Panel B displays the placebo test of our main DiD estimation for hotel observations, in which observed timeframe being cut up to May 2022. The grey vertical reference lines mark the fake treatment/intervention time. The stated DiD coefficients refer to ATETs in the respective placebo tests. Statistically significant DiD coefficients indicates that the parallel trends assumption does not hold for the main estimates pictured in Figure 4, vice versa. 5.2.2 Gorontalo Our placebo test results show that our DiD estimates on RPOS-Grab adoption among Gorontalo’s observed restaurants are not statistically robust, so we reject the parallel trend assumption (Table 11, column 2). Such result indicates the dissimilarity of trends between restaurants in treatment and control group prior to the POS distribution pro- gram, which is also a sign of the presence of other possible trigger(s) to restaurant tax payment changes in Gorontalo aside of the POS adoption. This is displayed graphically by a rather tiny gap between predicted tax receipts of treatment and control group after the fake intervention time in June 2019 and before the actual intervention time in November 2020 (see Figure 11 Panel A). The ATETs of specification with lagged dependent variable (Table A.3 column 2) turns insignificant, indicating a contemporaneous impact of POS distribution as well as other factors on both restaurant and hotel tax payments. Given the intervention times were in January 2020, the later-occurring peak of the global pandemic might have negated the possible positive impact of POS adoption among the restaurants. Also, once the pandemic receded, the economic bounce-back might have acted as a structural boost to tourism sector—and hence tax payment—but not the POS usage. There is a chance that the same reason might explain both RPOS and TMD’s insignificance in altering tax compliance. We also reject the hypothesis of parallel trend among our hotel observations in Gorontalo (see Table 12)., which means there was other possible factor(s) which affected hotel tax receipts during the timeframe other than the POS usage. 25 Table 11: Placebo Test: Gorontalo (Restaurants) Control vs. POS recipient restaurants Dependent Variable: (1) (2) (3) Log (restaurant tax receipt) RPOS RPOS- TMD Grab ATET (given RPOS vs. not) 0.149** (0.059) ATET (given RPOS-Grab vs. not) 0.343*** (0.105) ATET (given TMD vs. not) -0.055 (0.098) Constant 0.913*** 0.832*** 0.939*** (0.025) (0.026) (0.026) Observations 5,969 5,086 5,201 Adj. R2 0.001 0.005 0.001 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 Table 12: Placebo Test: Gorontalo (Hotels) Control vs. POS recipient hotels Dependent Variable: (1) (2) Log (hotel tax receipt) PAC TMD Hotel ATET (given access to PAC Hotel vs. not) - 0.134*** (0.017) ATET (given TMD vs. not) -0.501*** (0.025) Constant 0.321*** 0.371*** (0.008) (0.007) Observations 8,330 8,394 Adj. R2 0.000 0.019 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 26 Figure 11: Placebo Test: Gorontalo (a) Restaurant 27 (b) Hotel Notes : Panel A visualizes the placebo test of our main DiD estimation for restaurant observations, in which observed timeframe being cut up to December 2019 to exclude the actual intervention and post-intervention period. Panel B displays the placebo test of our main DiD estimation for hotel observations, in which observed timeframe being cut up to December 2020. The grey vertical reference lines mark the fake treatment/intervention time. The stated DiD coefficients refer to ATETs in the respective placebo tests. Statistically significant DiD coefficients indicates that the parallel trends assumption does not hold for the main estimates pictured in Figure 5, vice versa. 6 Result Interpretation Overall, we find that the distribution of POS in West Manggarai and Gorontalo has partially led to rises in tax compliance among hotels and restaurants. In West Manggarai, assigning TMD and WS worked better with tax-paying restaurants. As for hotels, WS did not prove to have significant effect. Meanwhile in Gorontalo, RPOS-Grab worked only subtly in boosting tax compliance among restaurants while no POS types seemed to increase hotel tax payment. The only POS type with consistent statistically insignificant impact on tax receipts in both districts is regular POS (RPOS). We now discuss the main explanations for the findings. First, it is crucial to always keep in mind that POS units are distributed in an accom- modative manner. Targeted tax-paying restaurants were handed the POS type most suitable to their existing operation and the stage of technological adoption. Recall that RPOS were typically given to businesses with either no previous experience in using electronic payment devices (entirely manual) or a very basic sales point tool without extensive and detailed sales and tax records. For such types of restaurants, adopting a novel technology might have been a hassle at first, thus requiring time for familiarization. Furthermore, the adoption of POS for these restaurants does not give any personal benefit to them. On the contrary, the targeted restaurants need to reallocate their resources to accommodate the POS (such as allocating manpower dedicated to the POS machine, providing electricity, space, and sacrificing time efficiency for the POS usage). Added by fees risks for cashless payments, cash and unrecorded transactions were much preferable (Visa, 2016). After all, their businesses were doing well before the POS adoption. In such instance, we may expect their reported tax receipts to be lower than actual—either because of unintentional errors or deliberate noncompliance. Either way, those in turn deducted the supposedly positive effect of RPOS distribution on tax receipts. It is important to note that, adverse attitude toward electronic payment systems—such as deliberate tampering with POS units to evade tax, in our case—is a strong predictor of low compliance, as pointed out in Night and Bananuka (2020). Low uptake of RPOS, if it is really the case, may be an indication of inherently low tax morale among the businesses which may necessitate a whole new strategy beyond distributing POS. Given these possibilities, perhaps RPOS could work when combined with better enforcement or regular spot checks. On the other hand, despite its weak robustness, RPOS-Grab’ had a relatively higher impact on restaurant tax payment in Gorontalo. This might signal the importance of third-party’s access to transaction information in improving compliance. While RPOS- Grab users did self-select into utilizing the direct connection to Grab’s application, they were probably aware of the possibility that their transaction records could be revealed to tax authority by the counterpart in case audit happens. Whether or not the third-party crosscheck really takes place, such an awareness eventually deters underreporting and even increases the tax paid. Second, TMD and WS were allocated for restaurants that already have a bookkeeping system in place. Adopting TMD and WS did not disrupt their usual business routine, 28 as both merely ‘latched’ ointo restaurants’ incumbent existing system and automatically relayed taxable sales records to local tax authorities. With such minimum required ad- justment, deviances were less likely to occur, therefore manifested in better compliance in general in the form of boosted tax receipts upon TMD and WS installation. Third, we may need to highlight the centrality of business size in influencing our results. Generally, less noncompliance observed in larger businesses (Slemrod, 2007). Likewise, we found that bigger hotels and restaurants tend to be more tax compliant. That is particularly visible among West Manggarai’s restaurant observation, where TMD and WS recipients were mostly larger businesses (Figure A.1). Consistent with the Allingham- Sandmo (1972) and Naritomi (2019) tax evasion framework, this compliance is because they—particularly those that are part of a restaurant/hotel chain—have higher perceived probability of detection stakes to underreport due to their visibility and number of third- party stakeholders able to attest to their report credibility. This high probability of being caught evading partly contributes to the upward impact of TMD and WS observed in our baseline estimation. On the other hand, smaller restaurants receiving RPOS normally have more leeway to underreport as they are generally part of the informal sector. Their scales of business enables them to lay low and alter their transaction reports with minimum risks. Due to the cash-based nature of the informal sector, it is simpler for these businesses to avoid having any third-party data associated with them, making it difficult for the tax authority to confirm their transactions. In Gorontalo, TMD’s muted effect on both restaurant and hotel tax compliance could be explained by the businesses’ relatively high compliance—as measured in their tax payment—prior to receiving TMD units. As the provincial capital, the city of Gorontalo is bigger and more densely populated compared to West Manggarai, and with such different characteristics we may expect better technology sophistication among bigger businesses. With more avid use of electronic payment technology even before the acquisition of TMD devices, there could be two possible outcomes: the businesses were just equally compliant following the use of TMD or they found a way around installed TMD to hold back their increased taxable transactions. Either way, the final tax payment did not change substantially. It is important to note, however, that our estimates might stem from the lack of statistical power given the small-sized treatment group. It is important to also note that the absence of significant effect of POS among hotel observations in both West Manggarai and Gorontalo might signal the industry-specific characteristic which shapes the way it responds to compliance-monitoring tools like POS. Hotel industries in both districts are generally made up of larger businesses with relatively high nominal average tax payments (see Table 2), which made up a large bulk of our hotel observations. Large hotels are usually part of international franchise or the local sizable management—both tend to be very visible and are subject to public audits. This made them were just as compliant even before the arrival of the POS distribution program. In addition to that, opening hotel businesses alike requires licenses and involves more officiation even for small to medium-sized ones. This is very contrast to F&B industry, which usually are quite loose in terms of license if the business size is small. 29 Given those factors, the extent to which the POS distribution program can be scaled up depends on several things. First is the quality of compliance enforcement—how often and how effective audits take place as well as the penalties for incompliance. Increasing the probability of detection or the penalty if detected will reduce tax evasion (Slemrod 2007, Slemrod 2019). Strong enforcement will signal to taxpayers for a high risk of getting caught, which in turn can deter potential evasion. Thus, district governments must have an effective audit and arrear collection system in place. Second is the economic aspect of enforcement effort: how cost-effective it is to continu- ously audit the reports. Enforcement entitles entails costs. Therefore, when seen from the government’s perspective, it would be central to consider and compare the marginal cost of tax enforcement (costs of conducting audit) to its marginal benefit (additional tax revenues, and probably some incurred fines). This calls for a risk-based audit system which aims for a positive net benefit. The point is particularly true for smaller districts with more limited resources. While procuring POS devices is indeed not cheap by any means, the corresponding comple- mentary monitoring activities once the devices are distributed and installed would also entitle costs. Therefore, in less-metropolitan districts such as West Manggarai, muted im- pacts of POS usage could also result from the absence of commensurate active monitoring due to lack of resources. In bigger districts with larger economic sizes, we could expect better efficacies in POS as they tend to have better resources for conducting quality and regular monitoring. Third, district government should have a clear legal framework that forces POS receivers to fully utilize and take care of the given POS machine. Most POS distribution programs at the district level rely financially on local government’s bank and SOE’s corporate social responsibility (CSR). While this approach is favorable for the district government’s budget, it may result in the district’s inability to keep recipients from damaging the device. As the devices are not government assets, district government cannot claim indemnity for damaged POS units. Fourth is the cost-benefit of POS adoption. POS adoption comes with costs, consisting initial capital and monthly maintenance fee. One of the POS vendors revealed that they charged IDR 150,000 per month per restaurant for regular POS maintenance, excluding the price of the hardware. While regular POS machines were given to small restaurants, a sound cost-benefit analysis is needed to measure the actual benefits of using POS for district governments. Last is the availability of technology enabling tampering with the POS devices, such as sales suppression software rampant in countries with mature use of POS. Given the recency of POS distribution program in Indonesia, we may believe that the use of sup- pression technology remains rare, albeit possible for larger taxpayers with advanced un- derstanding of information technology. However, if we look forward to the upscale of the program, then risks coming from this technology would need to be considered. It is because demands for it may grow as it is enticed by the spreading of nationwide POS usage. In that case, tax authorities may 30 consider having collaboration with IT providers to attempt building antidotes to such act with consideration on relative marginal cost and marginal benefit. 7 Conclusion This paper examines the impact of POS distribution on tax payment among hotels and restaurants in the Indonesian districts of West Manggarai and Gorontalo. Using Alling- ham and Sandmo’s tax compliance framework as cited in Naritomi (2019), we find that the POS distribution has various impacts on tax compliance. While some POS machines have proven to improve tax compliance in both districts, RPOS has consistently showed insignificant impact on tax compliance. We have identified several key takeaways from this varying impact of POS devices. First is the capacity of the receiver matter. POS receivers must sacrifice their resources to accommodate the machines. Without any personal benefit gained from the device, it is hard to imagine that taxpayers with low resources will voluntarily utilize the POS ma- chine given to them. Second, interventions that require minimum adaptation are likely to succeed. Minimum adaptation means less resources needed to operate the POS machines. Again, without personal benefit, taxpayers tend to be more open to devices that require minimum resource allocation. Finally, third-party data and informality play a vital role in increasing perceived proba- bility of detection. This explains why big hotels and restaurants, which are TMD and WS users, were more compliant than small hotels and restaurants, which are RPOS receivers. We suggest that national scalability should be considered based on several factors. First, district governments should not solely rely on POS machines to increase compli- ance. Without a strong and effective tax enforcement system, POS distribution will not be effective. Given their limited resources, district governments should manage their enforcement efficiently and effectively. This calls for the second, that is utilization of a risk-based enforcement system. Next, district governments ought to have a clear legal framework on POS distribution. This can help district governments to enforce POS usage and oblige receivers to maintain the devices properly. Fourth, district governments must conduct a thorough cost-benefit analysis to determine the actual benefits of implementing POS, before deciding to finance such intervention. This analysis needs to be done because POS implementation comes with costs. Finally, the governments need to be aware of the technology that makes it possible to tamper with POS systems. The risk from such technology will rise along with the utilization of POS machines. Future research may need to examine closely the POS uptake pattern in the subnational setting, as this will confirm whether the muted effect of the POS distribution program studied in this paper stems from businesses’ resistance in adopting POS units. The potential findings will provide a clearer picture of an effective incentive structure that the local governments might want to consider in encouraging POS usage. Also, with better data availability on the observed businesses, one can have more covariates to produce 31 more robust estimates by performing more heterogeneity checks with higher statistical power. 32 References Agarwal, S., Liu, C., & Souleles, N. S. (2007). The reaction of consumer spending and debt to tax rebates—evidence from consumer credit data. Journal of Political Economy, 115 (6), 986–1019. https://doi.org/10.1086/528721 Allingham, M. G., & Sandmo, A. (1972). Income tax evasion: A theoretical analysis. Journal of Public Economics, 1 (3), 323–338. https : / / doi . org / 10 . 1016 / 0047 - 2727(72)90010-2 Alm, J. R., & Torgler, B. (2004). Culture differences and tax morale in the united states and in europe. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.562861 Andreoni, J., Erard, B., & Feinstein, J. (1998). Tax compliance [Publisher: American Economic Association]. Journal of Economic Literature, 36 (2), 818–860. Retrieved June 16, 2023, from http://www.jstor.org/stable/2565123 Angrist, J. D., & Pischke, J.-S. (2008, December 15). Mostly harmless econometrics: An empiricist’s companion. Princeton University Press. https://doi.org/10.2307/j. ctvcm4j72 Bellon, M., Lima, F., Khalid, S., Dabla-Norris, E., Villena, P., Rojas, E., & Chang, J. (2019). Digitalization to improve tax compliance: Evidence from VAT e-invoicing in peru [OCLC: 1133286728]. International Monetary Fund. Bolt, W., Jonker, N., & Van Renselaar, C. (2010). Incentives at the counter: An empirical analysis of surcharging card payments and payment behaviour in the netherlands. Journal of Banking & Finance, 34 (8), 1738–1744. https://doi.org/10.1016/j. jbankfin.2009.09.008 Brockmeyer, A., & Somarriba, M. S. (2022). Electronic payment technology and tax compliance: Evidence from uruguay’s financial inclusion reform. Callaway, B., & Sant’Anna, P. H. (2021). Difference-in-differences with multiple time periods [Themed Issue: Treatment Effect 1]. Journal of Econometrics, 225 (2), 200–230. https://doi.org/https://doi.org/10.1016/j.jeconom.2020.12.001 Chattha, M. K., Blum, J. R., & Kelly, R. (2023). Personal income tax piggybacking. World Bank, Washington, DC. https://doi.org/10.1596/39554 Fan, H., Liu, Y., Qian, N., & Wen, J. (n.d.). The dynamic effects of computerized VAT invoices on chinese manufacturing firms. Fredriksson, A., & Oliveira, G. M. D. (2019). Impact evaluation using difference-in- differences. RAUSP Management Journal, 54 (4), 519–532. https://doi.org/10. 1108/RAUSP-05-2019-0112 Fricke, H. (2017). Identification based on difference-in-differences approaches with multi- ple treatments. https://doi.org/https://doi.org/10.1111/obes.12178 Gordon, R., & Li, W. (2009). Tax structures in developing countries: Many puzzles and a possible explanation. Journal of Public Economics, 93 (7), 855–866. https://doi. org/10.1016/j.jpubeco.2009.04.001 Kleven, H. J., Knudsen, M. B., Kreiner, C. T., Pedersen, S., & Saez, E. (2011). Un- willing or unable to cheat? evidence from a tax audit experiment in denmark. Econometrica, 79 (3), 651–692. https://doi.org/10.3982/ECTA9113 Kleven, H. J., Kreiner, C. T., & Saez, E. (2016). Why can modern governments tax so much? an agency model of firms as fiscal intermediaries. Economica, 83 (330), 219–246. https://doi.org/10.1111/ecca.12182 Korupsi, K. P. (2020). Rencana strategis komisi pemberantasan korupsi 2020-2024. https: //cms.kpk.go.id/storage/429/Renstra-FINAL-for-User.pdf 33 Lovics, G., Szˆ oth, C. G., & V´ ıoke, K., T´ an, B. (n.d.). The effect of the introduction of online cash registers on reported turnover in hungary. Naritomi, J. (2019). Consumers as tax auditors. American Economic Review, 109 (9), 3031–3072. https://doi.org/10.1257/aer.20160658 Night, S., & Bananuka, J. (2019). The mediating role of adoption of an electronic tax system in the relationship between attitude towards electronic tax system and tax compliance. Journal of Economics, Finance and Administrative Science, 25 (49), 73–88. https://doi.org/10.1108/JEFAS-07-2018-0066 OECD. (2013). Electronic sales suppression: A threat to tax revenue. OECD. Pomeranz, D. (2013, July). No taxation without information: Deterrence and self-enforcement in the value added tax (w19199). National Bureau of Economic Research. Cam- bridge, MA. https://doi.org/10.3386/w19199 Slemrod, J. (2004). The economics of corporate tax selfishness. National Tax Journal, 57 (4), 877–899. https://doi.org/10.17310/ntj.2004.4.06 Slemrod, J. (2007). Cheating ourselves: The economics of tax evasion. Journal of Eco- nomic Perspectives, 21 (1), 25–48. https://doi.org/10.1257/jep.21.1.25 Slemrod, J. (2018, July). Tax compliance and enforcement (w24799). National Bureau of Economic Research. Cambridge, MA. https://doi.org/10.3386/w24799 Slemrod, J., & Yitzhaki, S. (2000, January). Tax avoidance, evasion, and administration (w7473). National Bureau of Economic Research. Cambridge, MA. https://doi. org/10.3386/w7473 Torgler, B., & Schneider, F. (2005). Attitudes towards paying taxes in austria: An em- pirical analysis. Empirica, 32 (2), 231–250. https://doi.org/10.1007/s10663-004- 8328-y World Bank Group. (2020). Electronic payments acceptance incentives literature review and country examples. World Bank Group. 34 Appendices Table A.1: Difference-in-difference estimation: robustness check with lagged outcome variable - West Manggarai (Restaurants) Control vs. POS recipient restaurants Dependent Variable: (1) (2) (3) Log (restaurant tax receipt at t+1) RPOS TMD WS ATET (given RPOS vs. not) 0.016 (0.079) ATET (given TMD vs. not) 1.223*** (0.194) ATET (given Web Service vs. not) -0.193 (0.255) Constant 0.006 0.006 0.006 (0.021) (0.021) (0.025) Observations 18,167 17,929 17,789 Adj. R2 0.002 0.044 0.045 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 Table A.2: Difference-in-difference estimation: Robustness check with lagged outcome variable - West Manggarai (Hotels) Control vs. POS recipient hotels Dependent Variable: (1) Log (hotel tax receipt at t + 1) WS ATET (given Web Services vs. not) -0.345 (0.305) Constant 0.000 (0.051) Observations 10,799 Adj. R2 0.000 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 A.1 Table A.3: Difference-in-difference Estimation: Robustness check with lagged outcome variable - Gorontalo (Restaurants) Control vs. POS recipient restaurants Dependent Variable: (1) (2) (3) Log (restaurant tax receipt) RPOS RPOS- TMD Grab ATET (given RPOS vs. not) 0.099 (0.079) ATET (given RPOS-Grab vs. not) 0.658* (0.339) ATET (given TMD vs. not) 0.079 (0.141) Constant 0.968*** 0.915*** 1.034*** (0.031) (0.033) (0.033) Observations 13,919 11,389 11,740 Adj. R2 0.000 0.008 0.000 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 Table A.4: Difference-in-difference Estimation: robustness check with lagged outcome variable - Gorontalo (Hotels) Control vs. POS recipient hotels Dependent Variable: (1) (2) Log (hotel tax receipt) PAC TMD Hotel ATET (given access to PAC Hotel vs. not) -0.029 (0.143) ATET (given TMD vs not) 0.063 (0.145) Constant 0.324*** 0.379*** (0.007) (0.007) Observations 14,119 14,227 Adj. R2 0.000 0.001 Robust standard errors in parentheses *** p < 0.01, ** p < 0.05, * p < 0.1 A.2 Figure A.1: Distribution of Restaurant Tax Receipt by POS Type: West Manggarai Figure A.2: Distribution of Hotel Tax Receipt by POS Type: West Manggarai A.3 Figure A.3: Distribution of Restaurant Tax Receipt by POS Type: Gorontalo Figure A.4: Distribution of Hotel Tax Receipt by POS Type: Gorontalo A.4 Figure A.5: Distribution of Restaurant Tax Receipt: West Manggarai, RPOS Figure A.6: Distribution of Restaurant Tax Receipt: West Manggarai, TMD Figure A.7: Distribution of Hotel Tax Receipt: West Manggarai, WS Figure A.8: Distribution of Restaurant Tax Receipt: West Manggarai, WS A.5 Figure A.9: Distribution of Restaurant Tax Receipt: Gorontalo, RPOS Figure A.10: Distribution of Restaurant Tax Receipt: Gorontalo, RPOS-Grab Figure A.11: Distribution of Restaurant Tax Receipt: Gorontalo, TMD A.6 Figure A.12: Distribution of Hotel Tax Receipt: Gorontalo, PAC Hotel Figure A.13: Distribution of Hotel Tax Receipt: Gorontalo, TMD A.7 Figure A.14: Restaurant Tax Receipts in West Manggarai Control vs POS Recipient (a) Time Trend A.8 (b) Intertemporal Change Notes: These graphs compare tax receipts between restaurants which did not receive any POS units (control) and those which did (treated) around the POS distribution timeline. The red reference lines indicate times of POS handover: October 2019, November 2019, May 2021, and June 2022 for Regular POS (RPOS); October 2019 for Transaction Monitoring Device (TMD); and December 2020, June and December 2022 for Web Service (WS). Panel B shows the time trends of the outcome variable, tax receipts. Panel II shows the percentage change of monthly tax receipts. Figure A.15: Robustness Checks: Restaurant Observations in West Manggarai (a) Baseline specification 2 (using log(z+1) formula) A.9 (b) Lagged outcome variable Notes: Panel A depicts predicted tax receipts of treatment and control group for each POS type using coefficients generated when the independent variable is transformed using the z = log (x + 1) formula, while Panel B does that of specification using lagged outcome variable from equation (2). Figure A.16: Hotel Tax Receipts in West Manggarai Control vs POS Recipient (a) Time Trend A.10 (b) Intertemporal Change Notes: These graphs compare tax receipts between hotels which did not receive any POS units (control) and those which did (treated) around the POS distribution timeline. The red reference lines indicate times of POS handover in June and November 2022. Panel A shows the time trends of our outcome variable, tax receipts. Panel B shows the percentage change of monthly tax receipts. Figure A.17: Robustness Checks: Hotel Observations in West Manggarai Control vs POS Recipient Difference-in-Difference Estimation (a) Baseline specification 2 (using log (z + 1) formula) A.11 (b) Lagged outcome variable Notes: Panel A depicts predicted tax receipts of treatment and control group for each POS type using coefficients generated when the independent variable is transformed using the z = log (x + 1) formula, while Panel B does that of specification using lagged outcome variable from equation (2). Figure A.18: Restaurant Tax Receipts in Gorontalo Control vs POS Recipient (a) Time Trend A.12 (b) Intertemporal Change Notes: These graphs compare tax receipts between restaurants which did not receive any POS units (control) and those which did (treated) around the POS distribution timeline. The red reference lines indicate times of POS handover: November 2020, February 2021, March 2021, April 2021, May 2021, June 2021, August 2021, October 2021, November 2021, February 2022, March 2022, June 2022, August 2022 for RPOS; November 2020 and February 2021 for RPOS-GRAB; January 2021, February 2021, June 2021, July 2021, August 2021, September 2021, January 2022 and June 2022 for Transaction Monitoring Device (TMD). Panel A shows the time trends of our outcome variable, tax receipts. Panel B shows the percentage change of monthly tax receipts. Figure A.19: Robustness Checks: Restaurant Observations in Gorontalo Control vs POS Recipient Difference-in-Difference Estimation (a) A. Baseline specification 2 (using log(z+1) formula) A.13 (b) Lagged outcome variable Notes: Panel A depicts predicted tax receipts of treatment and control group for each POS type using coefficients generated when the independent variable is transformed using the z = log(x+1) formula, while Panel B does that of specification using lagged outcome variable from equation (2). Figure A.20: Hotel Tax Receipts in Gorontalo Control vs POS Recipient (a) Time Trend A.14 (b) Intertemporal Change Notes: These graphs compare tax receipts between hotes which did not receive any POS units (control) and those which did (treated) around the POS distribution timeline. The red reference lines indicate times of POS handover: March 2021 and May 2021 for PAC; March 2021, June 2021, July 2021, June 2022 for Transaction Monitoring Device (TMD). Panel A shows the time trends of our outcome variable, tax receipts. Panel B shows the percentage change of monthly tax receipts. Figure A.21: Robustness Checks: Hotel Observations in Gorontalo Control vs POS Recipient Difference-in-Difference Estimation (a) A. Baseline specification 2 (using log (z + 1) formula) A.15 (b) Lagged outcome variable Notes: Panel A depicts predicted tax receipts of treatment and control group for each POS type using coefficients generated when the independent variable is transformed using the z = log(x+1) formula, while Panel B does that of specification using lagged outcome variable from equation (2).