Policy Research Working Paper 9852 Becoming Legible to the State The Role of Identification and Collection Capacity in Taxation Oyebola Okunogbe Development Economics Development Research Group November 2021 Policy Research Working Paper 9852 Abstract Whereas the tax compliance literature emphasizes verifying property photographs) in tax notices more than quadruples tax liability amounts, this paper highlights two other tax the payment rate, from 2 percent to almost 10 percent, capacity dimensions of particular relevance in lower-income when the notice also details noncompliance penalties. A countries: identifying taxable entities and enforcing col- second experiment finds that signaling greater collection lection of known liabilities. Leveraging a newly-digitized probability to delinquent taxpayers further increases com- property database, a randomized experiment in Liberia finds pliance. These results highlight the importance of targeted that including identifying information (owners’ names and investments in tax capacity. This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The author may be contacted at ookunogbe@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 Becoming Legible to the State: The Role of Identification and Collection Capacity in Taxation Oyebola Okunogbe∗ Originally published in the Policy Research Working Paper Series on November 2021. This version is updated on November 2023. To obtain the originally published version, please email prwp@worldbank.org. Keywords: Taxation, state capacity, property tax, technology, enforcement JEL codes: H24, H26, O10 ∗ World Bank Development Research Group, ookunogbe@worldbank.org. This paper was funded by the International Growth Centre and the World Bank Research Support Budget. I am grateful to many individuals and groups for helpful feedback and support: Pierre Bachas, Natalie Bau, Clare Balboni, Lori Beaman, Kathleen Beegle, Jeremy Bowles, Anne Brockmeyer, Pascaline Dupas, Jonas Hjort, Asim Khwaja, Justine Knebelmann, Anne Laski, Luis Martinez, David McKenzie, Emily Nix, Adeyemi Okunogbe, Owen Ozier, Dina Pomeranz, Ingrid Sjursen, Gabriel Tourek, Victor Pouliquen, Mazhar Waseem, Jonathan Weigel, Eric Werker, Laurel Wheeler, and seminar participants at the Central Bank of Colombia, the Centre for the Study of African Economies, the National Tax Association, the Online Public Finance Seminar, Texas A&M University, the University of Maryland Baltimore County, the University of Michigan, the World Bank Tax Conference, and the Zurich Conference on Public Finance and Development. I thank Elfrieda Tamba, Thomas Doe Naa, Darlingston Talery, Isaac Stevens, Sebastian Weah, Lasana Kromah, Robert Pyne, Scholastica Nimley, Clara Goba, Jlateh Pajibo, Titus Zoneh, Alphonso Jallah, and the management, officials, employees, and contractors of the Liberia Revenue Authority for their research partnership and kind hospitality. Aditi Kumar, Nicolas Orgeira, Alpha Simpson, Joe Lelee Sloan, and Kairat Umargaliyev provided excellent field coordination and research assistance. This study was approved by the Harvard Institutional Review Board (IRB14-1833) and Health Media Lab IRB (556WBG18). It was registered in the AEA RCT registry (AEARCTR-0004272 and AEARCTR-0011158). The findings, interpretations, and conclusions expressed in this paper, including any errors, are entirely mine and do not necessarily represent the views of the World Bank Group, its Executive Directors, or the countries they represent. 1 Introduction Lower income countries, on average, collect less taxes than higher income countries, limit- ing their ability to invest in infrastructure and public services. This tax gap occurs despite comparable statutory rates and accounting for different levels of economic activity, suggest- ing the presence of deficits in lower income states’ tax administration capacity (Besley and Persson, 2014). This paper distinguishes three key components of tax capacity that states must invest in to collect taxes effectively: identifying taxpayers (identification capacity), de- termining how much they owe (detection capacity), and taking in these liabilities to the state coffers (collection capacity). Whereas traditional tax evasion models (e.g. Allingham and Sandmo (1972)) and a resulting large literature on third-party reporting emphasize detection capacity, this paper focuses on the two other dimensions of tax capacity, identification and collection, which are particularly relevant in lower income countries. Identifying taxpayers, a presumption in most developed economies, is a first order concern for tax collection in many lower income countries characterized by a large informal sector and the lack of proper identification systems for individuals, businesses, property and other taxable entities/ tax bases. For example, 45 percent of people in Sub Saharan Africa do not have an official proof of identity (World Bank, 2022), and majority of lower income countries have sparse and inaccurate taxpayer databases (Nyanga, 2021).1 Many contexts are reminiscent of Scott’s (1999) description of the premodern state as “in many crucial respects, partially blind; it knew precious little about its subjects, their wealth, their landholdings and yields, their location, their very identity.” In order to tax, the premodern state sought to make society “legible” through the creation of permanent last names, the establishment of cadastral surveys and population registers, etc. Similarly, modern governments seeking to understand their potential tax base must increase coverage of foundational identification systems, business registries and property registries. Detection capacity is closely related to identification and builds on it. After the presence of a taxpayer or tax base is known to the state, it must then verify the income, assets, and 1 Of the 61 Tax Administration Diagnostic Assessment Tool (TADAT) assessments conducted in lower income countries from 2013 to 2020, only 20 percent of tax authorities had a score of good or very good for knowledge of taxpayer database and 5 percent did so for accuracy of information in taxpayer registry. 2 transaction amounts that are subject to taxation. Detection capacity by modern revenue agencies relies on third-party information and the centrality of this information for tax col- lection has been a major focus of the recent empirical literature (e.g., Kleven et al., 2011; Kleven, Kreiner and Saez, 2016; Jensen, 2019; Pomeranz, 2015; Naritomi, 2019). Identifi- cation and detection can thus be considered as the extensive margin and intensive margin, respectively, of measuring the tax base. After taxpayers are identified and liabilities determined, there remains the challenge of collecting the revenue through administrative (billing and payment) systems and legal pro- cesses. For those who do not voluntarily comply, collection capacity includes the use of measures such as delinquent notices, penalties and interest charges, wage garnishment, and legal prosecution to recover unpaid taxes. In countries with strong collection capacity, some of these sanctions are automatically triggered and routinely applied to taxpayers. In low- capacity contexts, however, the majority of unpaid liabilities go uncollected for several years,2 and there is limited recovery of taxes even after an individual is found to be non-compliant through audits or third-party information (Best, Shah and Waseem, 2021; Carrillo, Pomeranz and Singhal, 2017).3 The key contribution of this paper is to explicitly test for and highlight identification and collection capacity, separate from detection capacity, as these three are typically conflated as “enforcement” interventions in the literature. Specifically, the paper uses two randomized experiments to answer the following question: what is the impact on tax compliance of increasing taxpayers’ perception of the state’s identification capacity and, separately, its collection capacity? I study this question in collaboration with the Liberia Revenue Authority (LRA) in Monrovia, Liberia, a setting where, outside a small subset of properties with high revenue potential, the vast majority of property owners were not paying real estate taxes, and many had never received a tax notice. In this context, in a bid to expand the tax net, the first- order concern of the LRA was the extensive margin of compliance, as opposed to verifying 2 See evidence of low compliance rates on known property tax liabilities from Pakistan (Khan, Khwaja and Olken, 2015), Ghana (Dzansi et al., 2022), Haiti (Krause, 2020), Senegal (Knebelmann, 2019), and the Democratic Republic of Congo (Weigel, 2020; Bal´ an et al., 2022). 3 Best, Shah and Waseem (2021) provide striking evidence of low collection capacity in Pakistan where less than 2 percent of amounts detected during audits are recovered at the time, with the remainder amount subject to a lengthy judicial and appeals process—leading to large amounts persistently stuck in litigation. 3 the property value; as such, the detection capacity channel is muted. Using a modified version of the Allingham and Sandmo (1972) tax compliance framework, the paper shows that taxpayers’ decision to pay a given tax liability depends not only on their perceived probability of their property being discovered by the tax authority and the associated penalties, but also on the perceived probability that the government will enforce the law and collect the taxes and penalties they owe. In the first phase of the project, I worked closely with the LRA to conduct a door-to-door property enumeration and develop a new electronic property database. Prior to this exercise, only 5 percent of properties in the city were on the tax register, and many records were not digitized. This exercise thus represented a significant technology investment for the LRA that made the tax base more legible. Experiment one examines the impact of signaling this newly acquired identification capac- ity and providing information on penalties to property owners who had not previously paid real estate taxes. The control group received a generic tax notice addressed “Dear Property Owners.” The notice provided information on the procedure for tax payment and requested compliance. A treatment group received a tax notice that, in addition, included identification information from the new database: the name of the property owner and a photograph of the property. This notice also included information on legal penalties for noncompliance. Two additional treatment groups received notices with either the identification information alone or the penalty information alone. The combination of identification and penalty information significantly increased property owners’ compliance: property owners were more than four times as likely to pay the tax. The payment rate increased from 2.2 percent of property owners in the control group to 9.7 percent in the treatment group. This effect persisted, as the treatment group continued to have a higher payment rate over the next three years of available data. The impact of the combined identification and penalty treatment is statistically different from the identification- only and penalty-only treatments, neither of which is statistically significant. These results indicate that providing taxpayers with evidence that their non-compliant property is known to the tax authority leads to increased compliance only when this evidence is accompanied 4 by information on the associated penalties for noncompliance. Experiment two investigates collection capacity. As an additional layer to the identifica- tion and penalty information studied in experiment one, it examines the impact of signaling to taxpayers a greater likelihood that the tax authority will enforce tax laws on them and collect the taxes and penalties they owe. The sample for this experiment was property owners who had paid real estate tax at least once in the past but had since become delinquent. The control group received a notice requesting payment. Much like the combined identification and penalty notice in experiment one, the control group notice included the name of the property owner, a photograph of the property, and information on legal penalties for non- compliance. Members of the treatment group received a similar notice with an additional notification that their property would be included in the next batch of properties to receive enforcement action if they did not comply. This intervention increased the payment rate by 7 percentage points (a 33 percent increase over a payment rate of 21 percent in the control group). This result underscores the impact of increasing taxpayers’ perception that the state will take action toward them, over and above their awareness that the tax authority has identified them as non-compliant and knowledge of the associated penalties. In particular, to the extent that providing identification and penalty information to the control group may already have signaled a greater likelihood of enforcement, this observed impact serves as a lower-bound estimate of the impact of increasing taxpayers’ perception of the state’s collection capacity. These findings enhance our understanding of state capacity and taxation, turning the spotlight on identification and collection capacity in contrast to traditional tax compliance models that implicitly assume that taxpayers are known to the tax authority and that collec- tion occurs once noncompliance is detected. To the best of my knowledge, this study provides the first evidence from a randomized experiment on the tax compliance impact of identifica- tion capacity, specifically, of alerting people that they have been identified by the state. Most closely related are studies of large government programs that match different databases to identify unregistered properties (Casaburi and Troiano, 2015) or businesses (Lediga, Riedel and Strohmaier, 2020) but these do not have exogenous variation in who is treated with 5 the identification information, and instead use the variation arising from the intervention to study spillover effects on other outcomes. This paper shows the important relationship between identification and collection capacity: whereas making the tax base legible is an essential first step, it is insufficient to significantly increase revenues; information on penal- ties for noncompliance as well as signals of a greater likelihood that those penalties will be enforced are key complements to translate discovered tax liabilities into tax revenues.4 These results directly inform policy. Distinguishing between identification, detection, and collection capacity is crucial for improving tax collection because it allows the tax author- ity to determine which capacity is binding and target its investments and policy instruments accordingly. For example, a lack of commensurate increases in detection and collection capac- ity may be responsible for the disappointing revenue impacts of mass taxpayer registration campaigns (Gallien, Occhiali and van den Boogaard, 2023; Mascagni et al., 2022; Moore, 2023). Whereas many policy conversations on tax administration reform are centered around adopting new technology and improved databases to better understand the tax base,5 these results emphasize the unavoidable need for investment in the state’s ability to enforce tax laws and collect unpaid liabilities, which are often more politically difficult policies to implement. More broadly, this paper also relates to a large literature on the challenges of making the tax base more legible through firm formalization and business registration initiatives (Bruhn and McKenzie (2014) and Floridi, Demena and Wagner (2020) review this literature). In addition, this paper contributes to our understanding of how technology can improve tax administration in emerging economies, complementing recent papers on how technology interventions in property taxation could improve the effectiveness of tax collectors (Dzansi et al., 2022) and the accuracy of property valuation (Knebelmann, Pouliquen and Sarr, 2022).6 Finally, this paper contributes to a vibrant literature on interventions conducted in collaboration with tax authorities to understand different approaches to promoting tax 4 Besides the potential revenue impacts from expanding the tax base, strengthening identification capacity may also improve horizontal equity, for example, by ensuring that taxpayers in the formal sector are treated similarly to those of same income in the informal sector. 5 For example, in 78 percent of countries with an approved World Bank tax project from 2010 - 2022, the project included a tax digitization component (author’s analysis of World Bank projects). 6 Okunogbe and Santoro (Forthcoming) provide a review of this literature and Knebelmann (2022) describes applications for property tax in particular. 6 compliance and expands this literature in two ways.7 First, unlike many experiments that give a generic nudge or reminder to taxpayers, this paper (specifically, experiment one) leverages a significant real investment in the identification capacity of the government. Second, whereas most papers in the literature can only study the immediate impact of a notice, this paper is able to show that the impact of identification and penalty information on compliance persists over three years.8 The rest of the paper proceeds as follows. Section 2 provides background information on Liberia and the administration of property taxes, while section 3 outlines a simple conceptual framework for thinking about these three aspects of tax capacity on tax compliance. Sections 4 and 5 describe the design of the two experiments and present the corresponding results. Section 6 concludes and provides policy implications. 2 Context Liberia has low compliance rates for real estate taxes (as well as other tax categories). Pro- longed civil conflict in the 1990s and early 2000s led to the erosion of state capacity across several sectors, including tax administration. In 2018, the real estate tax register contained fewer than 7,000 properties in Greater Monrovia, the largest urban metropolis in Liberia, with over 1 million residents. Even among properties on the tax register, timely tax collec- tion remains a challenge. After taxpayers enter the tax base in one year, over 60 percent of them do not pay taxes in the subsequent year (according to the author’s analysis of admin- istrative tax data from 2015 to 2018). Many property owners who pay are motivated by the desire to obtain a tax clearance needed for administrative purposes in another sector, such as to bid for a contract with the government or to import goods at the port. Several factors related to weak identification and collection capacity contribute to the low real estate tax compliance. In the absence of a cadaster that outlines all the properties in the city, the LRA adds properties to the tax register on a case-by-case basis, resulting in very low coverage, with about 5 percent of taxable properties on the tax register. Further, 7 Mascagni (2018), Pomeranz and Vila-Belda (2019), and Slemrod (2019) provide reviews of this literature. 8 Important exceptions are Brockmeyer et al. (2019) in Costa Rica and De Neve et al. (2021) in Belgium, who find sustained impacts of messages on tax compliance over two to three years. 7 it is difficult to locate registered properties for tax administration because there is no proper address system and the tax database is not geocoded. The addresses provided at the time of registration are often vague (e.g., “18th Street” or “Catholic Hospital community”). As a result, officials from the LRA typically rely on information from community members to identify buildings from the tax database, especially in cases where there is no readily available photograph of the property.9 Because of the lack of addresses, tax bills are not routinely delivered to property owners; instead, they are required to visit the LRA to request their bills. Besides the challenges in identifying properties, the LRA faces legal, political, and lo- gistical constraints to implementing the penalties for noncompliance. The most common consequence for noncompliance is that penalties (up to 25 percent of the tax liability per year) and interest (an average of 15 percent per year) are added to the tax liability. Al- though the revenue code allows for the closure and seizure of delinquent properties, these actions require a court ruling, which implies significant time and administrative costs. As a result, in practice, enforcement consists of repeated written notices and in-person visits to property owners and, at most, the posting of a delinquent notice on the property for 72 hours. Furthermore, given the limited staff and resources of the LRA, these enforcement actions are focused on select properties that have the highest revenue potential, leaving the majority of the tax base essentially free from enforcement. An additional key feature of real estate tax in this context is the tax rate. It is 1.5 percent of the property value per annum for commercial properties and 0.25 percent for residential properties.10 Given the relative illiquidity of the real estate market and the lack of transaction data, it is challenging to determine property values. To facilitate extensive margin compliance, property owners are allowed to self-declare their property value, subject to the caveat that the government reserves the right to purchase the property from them at said value. Commercial properties with high revenue potential are valued by a valuation firm 9 Although property owners are required to submit a photograph at the time of property registration, these records are typically kept in hard copy and not organized in a way that allows for easy matching with electronic payment records for enforcement purposes. 10 Commercial properties are those that are used for income-generating activities, such as a rented apartment, a store, or an office building, whereas residential properties are used strictly for housing and do not generate any income. 8 from a list of approved providers, and valuation audits are typically restricted to this subset of properties. The average annual tax liability for a residential property is US$70 per annum, about two-thirds of the one-month GDP per capita in Liberia of US$96. Further, for newly registered properties, the first bill is typically issued not only for the current tax year but also for the previous four years. To ease liquidity constraints, payment plans are available. The low compliance rate in Liberia is comparable to estimates from cities in both fragile and otherwise stable low and middle income countries, such as 5 percent in Carrefour, Haiti (Krause, 2020), 5-20 percent in Bengaluru and Jaipur, India (Government of India, 2017), 8 percent in Kananga, Democratic Republic of Congo (Bergeron, Tourek and Weigel, 2021), 9 percent in Dakar, Senegal (Knebelmann, 2019), 10 percent in Kampala, Uganda (Manwaring and Regan, 2023), and 12 percent in Madina, Greater Accra, Ghana (Dzansi et al., 2022). This suggests that the lessons from Liberia can be valuable for understanding tax compliance in a wide range of settings. 3 Conceptual Framework In the canonical model of tax evasion, the taxpayer decides how much of her true income to reveal to the tax authority by trading off the utility gains from a lower tax bill (in the event that the true amount is not detected) against the risk of being detected and punished by having to pay a penalty in addition to the tax (Becker, 1968; Allingham and Sandmo, 1972; Yitzhaki, 1974). In the case of property tax, where the tax base is visible and verifiable (at some cost to the tax authority), we model the taxpayer as deciding whether to pay a fixed tax liability (rather than as deciding how much tax to pay) given the perceived costs and benefits.11 As such, we abstract away from the standard detection problem and will focus instead on identification and collection.12 13 11 This framework can also be used to understand a taxpayer’s decision to pay income tax after declaring her income (Hallsworth et al., 2015, 2017). 12 Many local government settings with a flat tax on property units also abstract from the detection problem, such as described in Bergeron, Tourek and Weigel (2021). More broadly, given the visible and verifiable nature of real estate, relative to other tax bases, the detection problem can be more easily solved with appropriate technology such as by using satellite imagery to calculate square footage of buildings for valuation purposes (Ali, Deininger and Wild, 2020). 13 This framework can be extended from having a fixed tax liability to the standard detection problem by including three different probabilities and the relevant payoffs: first, the probability that the government 9 An agent with income Y (Y ≥ 0) owns property valued at V (V ≥ 0) and faces a tax rate t (0 ≤ t ≤ 1) on the property. If the agent pays the tax, he is left with a posttax income of Y − tV . If the agent does not pay the tax, there is a probability of p (0 ≤ p ≤ 1) that he is discovered as non-compliant. Thus, with probability 1 − p, his noncompliance is undiscovered by the tax authority, and he keeps his full income Y . Although real estate taxes have the advantage of an observable and immobile tax base, without effective data management tools that identify and track property ownership and payments, the noncompliance of individual property owners may nevertheless remain undiscovered. Further, the penalties for noncom- pliance are only enforced with probability q (0 ≤ q ≤ 1). Thus, there is a probability of p(1 − q ) that the agent keeps his total income when noncompliance is detected but penalties are not enforced. Finally, there is a probability of pq that his noncompliance is discovered and the penalties are implemented, in which case, the agent faces penalty aV (0 ≤ a) in addition to the tax tV . Faced with the choice of whether to pay the tax (i = 1) or not to pay it (i = 0), the agent therefore chooses i to maximize   Y − tV  if i = 1 Ui = (1)  (1 − p)Y + p((1 − q )Y + q (Y − tV − aV ))  if i = 0 . The agent will pay the tax when U1 > U0 —that is, when pq (t + a) > t . (2) From this condition, the agent’s likelihood of paying the tax is increasing in p, q , and a. Importantly, this condition highlights that p and q are multiplicative. This implies that even if p is very high or takes on its maximum value of p = 1—that is, if the agent knows with certainty that he has been identified as non-compliant—if q is very low or zero—that is, if he does not expect any sanctions to follow—then it is possible that q (t + a) ≤ t and the identifies the taxpayer; second, conditional on being identified, the probability of discovering the true tax liability; and third, conditional on discovering the true tax liability, the probability that penalties would be enforced. 10 agent does not pay the tax. The standard Allingham-Sandmo model, by contrast, does not include q , suggesting that the model implicitly assumes that discovered evasion is always penalized—that is, that q = 1 or that q is not distinct from p. Under the scenario where q = 1, if p = 1 (perfect identification), there is no incentive to evade any amount of tax liability if there are penalties for evasion—that is, if a > 0. More broadly, this condition implies that in settings where perceptions of p, q , and a are low, as is the case in many low capacity settings, increasing only one parameter would have marginal impacts. Experiment one aims to increase property owners’ perception of p (the probability that the tax authority has identified them as non-compliant) and approximate a scenario where p = 1. In addition, it provides information on a (the associated penalties) and examines the impact of these two interventions, separately and jointly, on tax compliance. Experiment two aims to increase property owners’ perception of q (the probability that the tax authority will enforce tax laws and make them pay their uncovered liability), relative to a control group that only receives information on p and a as in the combined treatment in experiment one, and examines the impact on tax compliance. These experiments thus attempt to manipulate separately taxpayers’ perception of the tax authority’s identification capacity and its collection capacity. We do not have an experimental arm to measure the impact of q in the absence of p and a because it is most policy relevant to examine the impact of increasing perceptions that the state will enforce collection of owed liabilities when taxpayers know that their liability has been discovered. As will be discussed in section 5.3, informing taxpayers about penalties for noncompliance may also indirectly signal a greater likelihood that these penalties will be enforced. Similarly, providing identification information may signal to taxpayers that the tax authority is able to target them for enforcement, affecting their perceptions of collection capacity as well. Either of these scenarios suggests that the estimates of the impact of increasing q in experiment two may be a lower-bound estimate. 11 4 Experiment One 4.1 Experiment Design This study took place among property owners in Congo Town, one of the tax zones in Monrovia, the capital city of Liberia. This area was identified as having many properties that could be brought into the tax net. The first activity was creating a property database. From April to June 2014, I worked closely with the LRA to conduct a “block mapping” project to collect information on property locations and features. Enrollment agents (short-term government contractors) went door- to-door in teams of two to raise awareness of the real estate tax using brochures and flyers. They also completed a brief survey with residents, using mobile devices to obtain information about each property and its owner. The data collected in this enumeration process resulted in an electronic property database for taxation. This was the first time the government had electronic records and digital photographs of these properties.14 High-value properties, such as hotels, banks, and other commercial properties, were ex- cluded from the enumeration list because the LRA did not want to include them in the experiment, given their revenue potential. In addition, 5 percent of the enumerated proper- ties were found to be on the tax roll already and were thus removed from the experiment sample. In all, 571 properties belonging to 464 owners made up the sample. Properties in the sample were randomly assigned to four groups. The control group re- ceived a generic notice that informed property owners of their responsibility to pay the real estate tax and gave instructions about how to comply. The notice did not contain any infor- mation specific to the property owner but was simply addressed “Dear Property Owners.” The notice for the first treatment group (the identification and penalty notice ) contained the same information about how to comply; however, it was personalized by including the owner’s name and a photograph of the property (obtained during the door-to-door enumer- ation). It also included the legal penalties for noncompliance, citing relevant sections of the Liberia Revenue Code. The notice for the second treatment group (the identification no- 14 Following this initial phase, the government extended data collection to over 40,000 properties across Monrovia a few years later. 12 tice ) contained only the personalized information (the owner’s name and a photograph of the property) and the compliance instructions. Finally, the notice for the third treatment group (the penalty notice ) had no personalized information but included the information on penalties. Appendix Figure A1 provides images of the notices. Due to the 2014 Ebola outbreak, field activities were suspended following the property enumeration. In March 2017, using the information from the property enumeration list, the LRA prepared the different notices for property owners according to their randomized group. Delivery agents were provided with “address books” containing the photograph and location of each property to hand-deliver the notices. If the property owner was not present, the notice was handed to the resident to deliver to the property owner. Given the three-year lag between the enumeration process and the delivery of the notices, there was some attrition in the sample. The team successfully delivered notices to 90 percent of the property owners in the sample, and this rate was not significantly different across the treatment and control groups.15 4.2 Data and Empirical Specification The data for this study come from two main sources: the baseline enumeration data and administrative records on tax payments. The baseline enumeration data include the property block and number (using a numbering system put in place by a pilot project of the postal office), the community name, the name and phone number of the resident with whom the enumeration survey was completed and his or her relationship to the property owner, the name and phone number of the property owner(s) (if they were not the survey respondents), and property characteristics, such as the number of rooms and bathrooms, the presence of a porch and/or garage, building materials, and whether the property is residential or commercial. The enumerators also measured and recorded the exterior dimensions of the property using sketches and took photographs of the exterior. The LRA determined property values using a formula that incorporates the square footage of the buildings calculated from the sketches and the construction materials observed from the images. 15 The reasons for nondelivery were relocation of the property owner, destruction of the property, or inability to locate the property. 13 Tax administration records provide data on whether the property owner visited the tax office in response to the notice, whether the owner paid the tax, and how much tax was paid. Visiting the tax office was the first step of the compliance process. For most property owners, this was their first interaction with the tax system, so they first had to obtain a personal tax identification number (TIN). To track responses, each notice included a unique number, and property owners were asked to bring the notice along when they came to the tax office. Data on tax payments come from the integrated tax administration systems that record all tax payments electronically and are available for this sample from 2017 to 2020. Although tax revenue leakage may occur in several ways, this paper focuses on the exten- sive margin of compliance, where it is possible to determine with certainty the compliance status of a property owner— that is, whether or not the property owner made a tax payment. In addition, the paper also reports whether the property owner responded to the tax notice at all and the total amount paid. The unit of observation for outcomes is at the owner level because payments are recorded by the property owner’s TIN; however, randomization was done at the property level. To address this discrepancy, we use the property-level randomization to infer owner-level treat- ment assignments using the cumulative exposure that each property owner had to different elements of the treatments, as follows: (i) property owners who have only one property are assigned to the treatment group for that property; (ii) property owners who have multiple properties that were all assigned to the same treatment group (or to a single treatment group and the control group) are assigned to that treatment group; (iii) property owners who have multiple properties and who have any property assigned to the identification and penalty treatment group are assigned to the identification and penalty treatment group because they are considered to have been exposed to both the identification and penalty messages; and (iv) property owners who have multiple properties and for whom some properties were assigned to the identification treatment group and others to the penalty group are assigned to the identification and penalty treatment group because they were exposed to both the identifi- cation and penalty messages. This process affects the 10 percent of property owners in the 14 sample who received notices with different treatments.16 One implication of using property-level randomization to determine owner-level treat- ment is that owners of multiple properties are significantly more likely to be assigned to the identification and penalty treatment group for the owner-level randomization (but not for the property-level randomization). Of the identification and penalty treatment group, 30 percent own multiple properties, compared to 10 percent in each of the other groups. To account for this differential propensity to be assigned to the combined treatment, we control for owning multiple properties in all regressions given that, conditional on the number of properties owned, assignment to treatment is random (Hardy and McCasland, 2023).17 As discussed in section 4.3, we also show comparable results from conducting the main analyses on a restricted sample of property owners who are not affected because they own only one property or received the same treatment for all their properties. Table 1 provides summary statistics for the different variables and shows that the ran- domization achieved balance across different characteristics.18 For property owners who own more than one property, we used the characteristics of the highest-value property. The aver- age property has three rooms and two bathrooms, and its average value is US$27,000. About 70 percent of property owners own solely residential properties, and 24 percent of property owners are women (or include a woman if there is more than one owner). We use the following equation to examine the impact of each of the three treatments in promoting compliance: Di = β0 + β1 T1,i + β2 T2,i + β3 T3,i + αXi + i, (3) where Di is one of three outcome variables—an indicator variable for whether a property owner responded to the notice, an indicator variable for whether a property owner paid real estate tax, and the amount of tax paid. β1 , β2 , and β3 estimate the causal effect of receiving 16 Of the sample, 84 percent owned only one property, while 6 percent owned more than one property but received the same treatment for all their properties. 17 Since very few individuals own more than 2 properties (Figure A2), we prefer the specification with a dummy variable for owning more than one property. The specification with dummy variables for the number of properties owned yields identical results, as shown in section 4.3. 18 Table A1 provides the analogous balance table for the property-level randomization. 15 each of the three treatments, respectively, on the outcomes. Xi are control variables for whether the property owner owns more than one property, the value of the property, whether the property has a female owner or co-owner, whether any of the properties is used for commercial purposes, and the location (one of three blocks) of the property.19 We report randomization inference p-values obtained using the ritest command in Stata. To mirror the reassignment protocol described above, in each iteration, we rerandomize at the property level and use the property-level treatment to determine owner-level treatment, holding fixed the ownership structure in the data. Since the delivery rate was balanced across the treatment and control groups (table A2), the subsequent analysis will focus on those property owners who received the notices. Ta- ble A3 shows that property/property owner characteristics are also balanced in this group. 4.3 Results and Discussion Table 2 shows the impact of the different treatments (figure 1 displays the results graphically). The identification and penalty notice leads to an 8.1 percentage point increase in the likelihood of responding to the notice, almost tripling the 4.4 percent response rate in the control group. Similarly, this combined intervention more than quadruples tax payment rates: it results in a 7.6 percentage point increase over the control mean of 2.2 percent. For the impact on the amount of tax paid, the coefficient (a US$66 increase) is 11 times the control mean of US$6. However, the standard errors are large given the low overall compliance rate, and we are unable to detect a statistically significant difference. The other two treatments—the identification notice and the penalty notice—do not have a detectable statistically significant effect on any of the outcomes. For the identification notice, we can reject a 5.1 percentage point increase and a 4.3 percentage point decrease in the likelihood of making a payment. For the penalty notice, we can reject a 2.6 percentage point increase in the likelihood of making a payment and a 5.2 percentage point decrease. Importantly, the increase in payment from the identification and penalty notice is statistically different from the impact of the identification notice (p = 0.014) or the penalty notice (p = 19 For most property owners with multiple properties, the properties were colocated. When not colocated, the location of the highest-value property was used. 16 0.001). Notably, we find a positive long-term effect on whether property owners make a tax payment due to bringing people into the tax net using this intervention (figure 2). Table 2, panel B shows that the identification and penalty treatment leads to a 4.3 percentage point increase in the likelihood that a property owner paid property tax at least once over the three years following the intervention (2018–20)—this represents a fivefold increase over a base of 1.1 percent in the control group. As a robustness check, we repeat the analysis on a restricted sample of property owners for whom we do not infer owner-level treatment from property-level randomization because they own only one property or received the same treatment for all their properties. These results (table A4, panel A) are comparable: we find a 6 percentage point increase in the likelihood of making a payment (p = 0.036). In addition, table A4, panel B shows that including dummy variables for the number of properties owned to account for the differential propensity to be assigned to the combined treatment group yields identical results to the main results (table 4) where we include a dummy variable for owning multiple properties. In all, these results reveal that taxpayers’ receiving graphic confirmation that the tax authority has identified them as noncompliant leads to a sustained increase in tax payment when accompanied by information on the associated penalties for noncompliance. 5 Experiment Two 5.1 Experiment Design One key goal of this paper is to distinguish the role of collection capacity, separate from identification (and detection) capacity. Specifically, once people know that they have been identified as non-compliant and they know the associated penalties, how does increasing their perception that the tax authority will enforce the stated tax laws increase their likelihood of compliance? Experiment two addresses this question. The sample for this experiment is delinquent property owners across different communities in the Greater Monrovia area. While the sample in experiment one was also non-compliant 17 in the sense that they had never paid the tax, this sample differs in that they had paid tax at least once, but had overdue taxes for one year or more. Their prior compliance suggests their ability to overcome barriers to compliance and thus makes them a more practical focus for targeted enforcement action. Similar to experiment one, this sample also excluded property owners with the highest tax liabilities (above US$1,000) as the LRA prioritized them for enforcement given their revenue potential. In practice, taxpayers below this threshold often effectively fall out of the radar of the tax authority and do not receive significant enforcement attention. To locate these delinquent properties, a verification team was dispatched with the name of each owner and the property address on file. The task of the team was to go to the broad location indicated on the property registration record (e.g., “18th Street” or “Catholic Hospital community”) and ask around in the community to identify the property. Once the property was verified, the team was to capture the GPS coordinates of the property and a photograph using a form on a mobile device. Of the 1,382 properties the team attempted to locate, 728 were identified. Property owners were randomized into two groups. The control group received a notice requesting the payment of delinquent tax liability. The notice was comparable to the com- bined identification and penalty notice in experiment one: it was personalized to include the name of the property owner and a photograph of the property and also stated the penalties for noncompliance. The treatment group received a similar notice (the enforcement notice ) with one addition: it included a sentence informing property owners that their property had been included in the current batch of properties designated for intensive follow-up and, if they did not respond by the deadline, that the LRA would proceed to the next stage of enforcement actions. Figure A2 provides an image of the notice. The randomization was stratified by two variables: location (communities across the city were divided into three broad areas) and payment history (whether the individual had only one year or more than one year of unpaid taxes). These variables were considered to be potential dimensions of heterogeneity in the responses of taxpayers to the intervention. The properties were located in October and November 2018, and the treatment notices 18 were sent out in May 2019. Before randomization, some of the identified property owners had become compliant and were thus excluded. The final sample to which notices were sent included 529 individuals who own 580 properties. Of these, 3 percent did not receive notices due to not meeting anyone at the property during all delivery attempts. The delivery rates for both the treatment and control groups were equal at 97 percent of property owners (table A5), so we use this subsample for the subsequent analysis. 5.2 Data and Empirical Specification The data for this study come from two main sources: the property location verification data and administrative records on tax payments. The property location verification data provide the name of the property owner, as well as the address (directions), a photograph, and the GPS coordinates of the property. They also include descriptions of external property characteristics provided by the verification team on the basis of their observations, such as the number of floors, the presence of a fence, the quality of landscaping, and the kind of road the property is on. The administrative tax records provide the payment history of each property owner from 2014 to 2020. In addition, clerks at the tax office made note of each property owner that visited the tax office in response to the notice: even if no payment was recorded, taxpayers might visit the office to set up a payment plan or dispute the tax notice. Table 3 provides summary statistics and shows that the randomization achieved balance across different characteristics. The average property value is about US$38,000. About 8 percent of property owners own more than one property, and over 90 percent own solely residential properties. About 84 percent of properties have only one floor, and 56 percent do not have a fence. For property owners with multiple properties, we use the characteristics of the highest-value property. We use the following equation to examine the impact of the enforcement notice in pro- moting the payment of property tax arrears: Pi = β0 + β1 Ti + αXi + λSi + i, (4) 19 where Pi is one of three outcome variables—an indicator variable for whether a property owner responded to the notice by visiting the tax office, an indicator variable for whether a property owner paid their real estate tax, and the amount of tax paid. β1 estimates the causal effects of the enforcement notice on the outcomes. Si is a vector of six strata indicator variables (one for each pair of the three areas in the city and whether the individual has more than one year of arrears). Xi are control variables for whether the property owner owns more than one property, the value of the property, whether there are any female owners, and whether the property owner owns any commercial properties. We report randomization inference p-values. 5.3 Results and Discussion Before turning to the impact of the enforcement notice, it is interesting to observe that the payment rate in the control group is much higher than the payment rate for owners who received the identical identification and penalty treatment in experiment one (21 percent versus 10 percent). This likely reflects the lower barriers to tax payment for the experiment two sample as they have paid at least once in the past. It is important to note that this analysis does not require the impact of manipulating the identification and penalty information on the notices to be the same across the two samples. Rather, this study examines whether there is an additional role for q in a setting where p and a are already in operation. To the extent that increasing perceptions of p and a in the control group may also have implicitly signaled greater collection capacity to recipients, the following results may be interpreted as a lower-bound estimate of the impact of this intervention. Table 4 shows the main results from experiment two (figure 3 displays the results graph- ically). Receiving a notice with the enforcement message leads to a 7.6 percentage point increase (p = 0.054) in responding to the tax notice, a 32 percent increase relative to the 23.7 percent response rate in the control group. It also causes a 6.9 percentage point increase (p = 0.070) in the likelihood of paying the tax, a 33 percent increase relative to the 20.6 percent tax payment rate in the control group. The average payment also rises by US$47 (p = 0.037), a 70 percent increase relative to US$67 in the control group. 20 These results indicate that the signal of higher enforcement probability sent by the tax authority was effective in causing more property owners in the treatment group to pay their arrears, relative to property owners in the control group who were notified that the tax authority was aware of the delinquency on their arrears and informed about the penalties for noncompliance. This finding demonstrates that the implicit assumption in standard models that q = 1 does not hold and that taxpayers’ expectation that any penalties associated with noncompliance will be enforced on them is an important parameter to consider in tax compliance models (alongside the probability of identification/ detection and penalties, which are typically included). Taxpayers’ low q may, in part, explain why tax compliance remains quite low overall, even in the combined identification and penalty group in experiment one, of which only 13 percent responded to the notice. Low q is also consistent with evidence from other studies (e.g., Carrillo, Pomeranz and Singhal, 2017) in which a large majority do not respond evidence sent by the tax authority that their tax declarations were inconsistent with data from third- party information. While we find an additional impact from the enforcement message, the response rate in the treated group is still only 31 percent, meaning that there is a large proportion of taxpayers for whom significant compliance barriers remain. These barriers may include a low perceived enforcement probability (despite the message), liquidity constraints, difficulty navigating the tax system, or political-economic factors, such as low tax morale (Prichard et al., 2019). We conduct heterogeneity analyses to examine whether certain recipients were more re- sponsive to the enforcement treatment, focusing on the two variables by which the random- ization was stratified. First, we find no differential response from taxpayers by the length of the delinquency (one year versus multiple years). Thus, the results do not support the hypothesis that taxpayers with longer delinquencies are more difficult to bring back to com- pliance. Second, as shown in table 5, we find evidence that the treatment effect is driven by two of the three tax areas (area two and area three). The three areas of the city differ along multiple dimensions, such as property value, commercial activity, degree of enforcement, etc. While it is difficult to disentangle the different factors, we find some evidence consistent 21 with the hypothesis that property owners who had been more exposed to prior enforcement activity by the LRA were more responsive to the enforcement message. In the absence of direct information on the different enforcement activities carried out by area, the billing and payment records provide some suggestive evidence of the outcome of enforcement efforts across areas. We use the delinquency recovery rate from the prior fiscal year as a proxy for the extent of prior enforcement efforts. Table A6, column 4, shows that for properties similar to the experimental sample (tax liability below $1,000 and with tax arrears), the payment rate in area two is 9 percentage points higher than in area one (p = 0.0003), and the rate in area three is 5 percentage points higher than area one (p = 0.057). This pattern is consistent with a greater effect of the enforcement treatment in area two due to higher credibility of enforcement based on past experience. Although the LRA did not systematically record follow up enforcement activities carried out against delinquent property owners in the treatment group that did not respond, there were no significant changes to their collection capacity around this time and the available evidence suggests that their actions, if any, were not substantial. We find no impact of the intervention beyond the first year (table 4 and figure 4). This result contrasts with the sustained compliance impact in experiment one, where there was a notable investment in identification capacity and indicates that shifting long-term beliefs about enforcement probability requires real changes in collection capacity. 6 Conclusion This paper offers a framework to study tax capacity of emerging economies as three different dimensions: identification, detection, and collection capacity. Given the well-established body of evidence on the importance of verifying the amount of tax liabilities, for example, through third-party information, this paper focuses instead on identification and collection. It highlights the importance of first of all identifying taxable entities and tax bases to make the tax base more legible to the state. However, it also underscores that legibility is insufficient and demonstrates the key complementary role of the state’s capacity to enforce tax laws and 22 collect the unpaid tax liabilities it detects. Experiment one shows that alerting property owners that their noncompliance has been discovered leads to increased compliance, when accompanied by information on penalties for noncompliance. Experiment two reveals that even with the identification of the tax base and information on penalties already in operation, increasing the perception that penalties will be enforced leads to additional compliance. Distinguishing identification, detection and collection capacity is imperative for policy making. Diagnosing which capacity is the binding constraint in a given setting and investing accordingly is key to achieving higher tax receipts. There is significant policy interest in investments in identification and detection capacity, such as third-party reporting, the use of electronic databases, and technology-based tools for understanding and monitoring the tax base. However, donors and governments may overprioritize investment in costly information technology projects to the detriment of the less attractive and often politically costly task of increasing traditional enforcement to ensure the payment of known tax obligations (Arewa and Davenport, 2022). Further, investments in collection capacity can be challenging as they are often outside the tax system and require coordination with other government agencies— for instance, by improving the efficiency of the legal system for processing court cases. This framework of identification, detection and collection capacity is useful for analyzing a broad range of issues in taxation and other compliance contexts. For example, to realize the self-enforcing compliance properties of value-added tax (VAT), the tax authority needs to 1) register VAT-eligible firms, 2) use data analytics to uncover mismatched reports between buyers and sellers, and 3) adopt policies to ensure detected liabilities are paid. On another note, the rising popularity of the controversial policy of taxing mobile money transactions in some African countries is linked to the ease of meeting the different tax capacity requirements: the amount transacted is easily recovered and can be collected at source by the mobile phone company.20 Beyond taxation, separating the compliance problem into identification, detection and collection can also be useful for improving compliance with environmental regulations and labor laws, among other applications. 20 Uganda, Ghana, Malawi, Tanzania, Cameroon, and the Republic of Congo have recently introduced or implemented legislation to impose taxes on mobile money transactions. 23 References Ali, Daniel Ayalew, Klaus Deininger, and Michael Wild. 2020. “Using satellite imagery to create tax maps and enhance local revenue collection.” Applied Economics, 52(4): 415–429. Allingham, M. G., and Agnar Sandmo. 1972. “Income Tax Evasion: A Theoretical Perspective.” Journal of Public Economics, 1: 323–38. Arewa, Moyo, and Stephen Davenport. 2022. “The Tax and Technology Challenge.” In Innovations in Tax Compliance: Building Trust, Navigating Politics, and Tailoring Reform, ed. Roel Dom, Anna Custers, Stephen Davenport, and Wilson Prichard. Chapter 7, 171–204. Washington, DC: World Bank. an, Pablo, Augustin Bergeron, Gabriel Tourek, and Jonathan L. Weigel. 2022. Bal´ “Local Elites as State Capacity: How City Chiefs Use Local Information to Increase Tax Compliance in the Democratic Republic of the Congo.” American Economic Review, 112(3): 762–97. Becker, Gary S. 1968. “Crime and Punishment: An Economic Approach.” In The Economic Dimensions of Crime, ed. Nigel G. Fielding, Alan Clarke, and Robert Witt. 13–68. London: Springer. Bergeron, Augustin, Gabriel Tourek, and Jonathan L. Weigel. 2021. “The State Capacity Ceiling on Tax Rates: Evidence from Randomized Tax Abatements in the DRC.” Working Paper. Besley, Timothy, and Torsten Persson. 2014. “Why Do Developing Countries Tax So Little?” Journal of Economic Perspectives, 28(4): 99–120. Best, Michael, Jawad Shah, and Mazhar Waseem. 2021. “Detection without Deter- rence: Long-Run Effects of Tax Audit on Firm Behavior.” Working Paper. Brockmeyer, Anne, Spencer Smith, Marco Hernandez, and Stewart Kettle. 2019. “Casting a Wider Tax Net: Experimental Evidence from Costa Rica.” American Economic Journal: Economic Policy, 11(3): 55–87. Bruhn, Miriam, and David McKenzie. 2014. “Entry Regulation and the Formaliza- tion of Microenterprises in Developing Countries.” The World Bank Research Observer, 29(2): 186–201. Carrillo, Paul, Dina Pomeranz, and Monica Singhal. 2017. “Dodging the Taxman: Firm Misreporting and Limits to Tax Enforcement.” American Economic Journal: Applied Economics, 9(2): 144–64. Casaburi, Lorenzo, and Ugo Troiano. 2015. “Ghost-House Busters: The Electoral Re- sponse to a Large Anti–Tax Evasion Program.” The Quarterly Journal of Economics, 131(1): 273–314. De Neve, Jan-Emmanuel, Cl´ ement Imbert, Johannes Spinnewijn, Teodora Tsankova, and Maarten Luts. 2021. “How to Improve Tax Compliance? Evidence from Population-Wide Experiments in Belgium.” Journal of Political Economy, 129(5): 1425–63. 24 Dzansi, James, Anders Jensen, David Lagakos, and Henry Telli. 2022. “Technology and Local State Capacity: Evidence from Ghana.” NBER Working Paper 29923, National Bureau of Economic Research, Cambridge, MA. Floridi, Andrea, Binyam Afewerk Demena, and Natascha Wagner. 2020. “Shedding light on the shadows of informality: A meta-analysis of formalization interventions targeted at informal firms.” Labour Economics, 67: 101925. Gallien, Max, Giovanni Occhiali, and Vanessa van den Boogaard. 2023. “Catch Them If You Can: the Politics and Practice of a Taxpayer Registration Exercise.” ICTD Working Paper 160, Brighton: Institute of Development Studies. Government of India, . 2017. “Economic survey 2016–2017.” Ministry of Finance, Gov- ernment of India. Hallsworth, Michael, John A. List, Robert D. Metcalfe, and Ivo Vlaev. 2015. “The Behavioralist as Tax Collector: Using Natural Field Experiments to Enhance Tax Compli- ance.” NBER Working Paper 20007, National Bureau of Economic Research, Cambridge, MA. Hallsworth, Michael, John A. List, Robert D. Metcalfe, and Ivo Vlaev. 2017. “The Behavioralist as Tax Collector: Using Natural Field Experiments to Enhance Tax Compliance.” Journal of Public Economics, 148: 14–31. Hardy, Morgan, and Jamie McCasland. 2023. “Are Small Firms Labor Constrained? Experimental Evidence from Ghana.” American Economic Journal: Applied Economics, 15(2): 253–84. Jensen, Anders. 2019. “Employment Structure and the Rise of the Modern Tax System.” NBER Working Paper 25502, National Bureau of Economic Research, Cambridge, MA. Khan, Adnan Q., Asim I. Khwaja, and Benjamin A. Olken. 2015. “Tax Farming Redux: Experimental Evidence on Performance Pay for Tax Collectors.” The Quarterly Journal of Economics, 131(1): 219–71. Kleven, Henrik Jacobsen, Claus Thustrup Kreiner, and Emmanuel Saez. 2016. “Why Can Modern Governments Tax So Much? An Agency Model of Firms as Fiscal Intermediaries.” Economica, 83(330): 219–46. Kleven, Henrik Jacobsen, Martin B Knudsen, Claus Thustrup Kreiner, Søren Pedersen, and Emmanuel Saez. 2011. “Unwilling or Unable to Cheat? Evidence from a Tax Audit Experiment in Denmark.” Econometrica, 79(3): 651–92. Knebelmann, Justine. 2019. “Taxing Property Owners in Dakar.” International Growth Centre Policy Brief 50415, IGC, London School of Economics. Knebelmann, Justine. 2022. “Digitalisation of property taxation in developing countries.” ODI Report. London: ODI (www.odi.org/publications/digitalisation-of-property-taxation- in-developing-countries-recent-advances-and-remainingchallenges). Knebelmann, Justine, Victor Pouliquen, and Bassirou Sarr. 2022. “Widening the Tax Net When Information Is Scarce: The Role of Agents’ Discretion.” Working Paper. 25 Krause, Benjamin. 2020. “Balancing Purse and Peace: Tax Collection, Public Goods and Protests.” Working Paper. Lediga, Collen, Nadine Riedel, and Kristina Strohmaier. 2020. “What You Do (Not) Get When Expanding the Net - Evidence from Forced Taxpayer Registrations in South Africa.” Working Paper. Manwaring, Priya, and Tanner Regan. 2023. “Public Disclosure and Tax Compliance: Evidence from Kampala Uganda.” Working Paper. Mascagni, Giulia. 2018. “From the Lab to the Field: A Review of Tax Experiments.” Journal of Economic Surveys, 32(2): 273–301. Mascagni, Giulia, Fabrizio Santoro, Denis Mukama, John Karangwa, and Napthal Hakizimana. 2022. “Active Ghosts: Nil-filing in Rwanda.” World Development, 152: 105806. Moore, Mick. 2023. “Tax obsessions: Taxpayer registration and the “informal sector” in sub-Saharan Africa.” Development Policy Review, 41(1): e12649. e12649 DPR-Nov-21- 3402.R1. Naritomi, Joana. 2019. “Consumers as Tax Auditors.” American Economic Review, 109(9): 3031–72. Nyanga, Maimbo. 2021. “Performance Outcome Area 1 Integrity of the Registered Tax- payer Base- Essentials of a High Integrity Taxpayer Register.” TADAT Insights. Okunogbe, Oyebola, and Fabrizio Santoro. Forthcoming. “The Promise and Limitations of Information Technology for Tax Mobilization.” World Bank Research Observer. Pomeranz, Dina. 2015. “No Taxation without Information: Deterrence and Self- Enforcement in the Value Added Tax.” American Economic Review, 105(8): 2539–69. Pomeranz, Dina, and Jos´ e Vila-Belda. 2019. “Taking State-Capacity Research to the Field: Insights from Collaborations with Tax Authorities.” Annual Review of Economics, 11(1): 755–81. Prichard, Wilson, Anna Custers, Roel Dom, Stephen R. Davenport, and Michael Roscitt. 2019. “Innovations in Tax Compliance: Conceptual Framework.” Policy Research Working Paper 9032, World Bank, Washington, DC. Scott, James C. 1999. Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. New Haven: Yale University Press. Slemrod, Joel. 2019. “Tax Compliance and Enforcement.” Journal of Economic Literature, 57(4): 904–54. Weigel, Jonathan L. 2020. “The Participation Dividend of Taxation: How Citizens in Congo Engage More with the State When It Tries to Tax Them.” The Quarterly Journal of Economics, 135(4): 1849–1903. 26 World Bank, . 2022. “ID4D Global Dataset- Volume 1 2021: Global ID Coverage Esti- mates.” World Bank Group, Washington DC. Yitzhaki, Shlomo. 1974. “A Note on Income Tax Evasion: A Theoretical Approach.” Journal of Public Economics, 3(2): 2011–202. 27 Figure 1: Impact of Identification and Penalty Notices on Registration, Payment, and Amount Paid, Experiment One Note : This figure shows the treatment effect and 95 percent confidence interval from OLS regressions of the outcomes on the different treatment groups in 2017. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and location (one of three subareas). N = 421. 28 Figure 2: Event Study Graph, Experiment One Note : This figure shows the treatment effect coefficient and 95 percent confidence interval from OLS regressions of the outcomes on the indicator for the treatment group for each year. The dotted lines indicate the year of the intervention. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and location (one of three subareas). N = 421. 29 Figure 3: Impact of Enforcement Notice on Response, Payment, and Amount Paid, Experiment Two Note : This figure shows the treatment effect and 95 percent confidence interval from OLS regressions of the outcomes on the indicator for the treatment group in 2019. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and strata fixed effects (the interaction of being located in one of three areas and owing more than one year of arrears). N = 511. 30 Figure 4: Event Study Graph, Experiment Two Note : This figure shows the treatment effect coefficient and 95 percent confidence interval from OLS regressions of the outcomes on the indicator for the treatment group for each year. The dotted line indicates the year of the intervention. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and strata fixed effects (the interaction of being located in one of three areas and owing more than one year of arrears). N = 511. 31 Table 1: Randomization Balance Table (Property Owners), Experiment One (1) (2) (3) (4) (5) (6) Difference between control group and Mean and Identification SD of P-value and Identification Penalty N Control of f-test Penalty group Property value (USD) 27722.224 151.764 -313.375 -3986.625 0.606 464 [24673.623] (3300.924) (4236.209) (3476.295) {0.964} {0.939} {0.256} Area 1380.313 55.810 90.184 -129.496 0.487 464 [1058.523] (140.327) (178.568) (142.096) {0.704} {0.620} {0.367} Rooms 3.349 0.204 0.112 -0.181 0.415 463 [2.541] (0.307) (0.314) (0.290) {0.505} {0.701} {0.538} Bathrooms 1.925 0.029 -0.267 -0.252 0.375 462 [1.516] (0.221) (0.222) (0.210) {0.901} {0.236} {0.235} Has Porch 0.840 0.002 -0.038 -0.046 0.692 464 [0.369] (0.048) (0.053) (0.053) {0.963} {0.453} {0.390} Has Garage 0.113 -0.034 -0.028 -0.057 0.541 464 [0.318] (0.039) (0.041) (0.038) {0.401} {0.521} {0.149} Commercial property 0.358 -0.031 -0.026 -0.038 0.933 464 [0.482] (0.063) (0.066) (0.065) {0.628} {0.697} {0.556} Block two 0.509 -0.124* -0.048 -0.034 0.234 464 [0.502] (0.065) (0.069) (0.069) {0.058} {0.443} {0.589} Block three 0.311 -0.130** -0.029 -0.060 0.117 464 [0.465] (0.058) (0.063) (0.062) {0.025} {0.678} {0.320} Has female owner 0.236 0.010 0.083 0.024 0.549 464 [0.427] (0.057) (0.062) (0.060) {0.855} {0.175} {0.673} Note : Column 1 gives the standard deviation in brackets. Columns 2, 3, and 4 give the coefficient and robust standard errors (in paren- theses) from an OLS regression of the property/property owner characteristic on treatment indicator variables, controlling for owning multiple properties. Randomization inference p-values are in brackets. Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 32 Table 2: Impact of Identification and Penalty Notices on Registration, Payment, and Amount Paid, Experiment One (1) (2) (3) (4) (5) Responded Paid in Amount Paid in Amount paid to notice 2017 paid (USD) 2018-2020 2018-2020 (USD) Panel A: Short-term results Panel B: Long-term results Identification and Penalty 0.081** 0.075** 65.266 0.043** 235.762 (0.035) (0.029) (57.993) (0.021) (207.185) {0.032} {0.007} {0.389} {0.042} {0.432} Identification 0.037 0.004 10.867 0.006 6.859 (0.036) (0.024) (27.070) (0.019) (55.743) {0.361} {0.894} {0.807} {0.755} {0.951} Penalty 0.017 -0.014 -0.503 -0.012 12.053 (0.033) (0.020) (10.502) (0.012) (32.897) {0.681} {0.638} {0.988} {0.580} {0.897} Identification and Penalty=Identification (p-value ) 0.244 0.014 0.421 0.095 0.391 Identification and Penalty=Penalty (p-value ) 0.090 0.001 0.317 0.012 0.386 Identification=Penalty (p-value ) 0.581 0.540 0.782 0.395 0.946 Control mean of dependent var. 0.044 0.022 5.929 0.011 2.691 Number of observations 421 421 421 421 421 Note : This table gives results of OLS regressions of the outcome variables on the treatment indicator variables. Robust standard errors are in parentheses. Randomization inference p-values are in brackets. The omitted group is the control group. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and location (one of three subareas). Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 33 Table 3: Randomization Balance Table (Property Owners), Experiment Two (1) (2) (3) Mean and SD Difference in N of control group treatment group Property value (USD) 37943.081 3887.978 528 [48799.840] (4359.233) {0.381} One floor 0.836 -0.032 528 [0.371] (0.032) {0.347} No fence 0.557 0.045 528 [0.498] (0.041) {0.329} Has landscaping 0.511 0.006 528 [0.501] (0.043) {0.934} Very good quality 0.706 -0.030 528 [0.456] (0.040) {0.506} On footpath 0.187 0.018 528 [0.391] (0.033) {0.649} Owner has 2+ properties 0.080 0.014 528 [0.272] (0.024) {0.634} Owner has commercial property 0.069 0.021 528 [0.253] (0.024) {0.422} Female owner 0.351 -0.084** 528 [0.478] (0.040) {0.034} One year of arrears 0.508 0.000 528 [0.501] (0.000) {1.000} Zone 2 0.294 0.000** 528 [0.456] (0.000) {1.000} Zone 3 0.355 -0.000 528 [0.479] (0.000) {1.000} Note : Column 1 gives the standard deviation in brackets. Column 2 gives the coefficient and robust standard errors (in parentheses) from an OLS regression of the property/property owner characteristic on treatment indicator variables, con- trolling for strata fixed effects. Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 34 Table 4: Impact of Enforcement Notice on Response, Payment, and Amount Paid, Experiment Two (1) (2) (3) (4) (5) Responded Paid in Amount Paid in Amount paid to notice 2019 paid (USD) 2020 in 2020 (USD) Panel A: Short-term results Panel B: Long-term results Enforcement notice 0.076* 0.069* 46.754** -0.002 10.924 (0.040) (0.038) (21.781) (0.028) (12.289) {0.054} {0.070} {0.037} {0.947} {0.504} Control mean of dependent var 0.237 0.206 66.521 0.107 21.521 Number of observations 511 511 511 511 511 Note : This table gives OLS regressions of the outcome variables on the treatment indicator variable. Robust standard errors are in parentheses. Ran- domization inference p-values are in brackets. The omitted group is the control group. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and strata fixed effects (the interaction of being located in one of three areas and owing more than one year of arrears). Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 35 Table 5: Heterogeneous Impact of Enforcement Notice on Response, Payment, and Amount Paid, Experiment Two Responded Paid in Amount to notice 2019 paid (USD) (1) (2) (3) Impact of treatment in area one -0.031 -0.016 56.458 (0.070) (0.067) (49.750) {0.654} {0.830} {0.264} Impact of treatment in area two 0.140** 0.128** 45.444* (0.064) (0.060) (24.974) {0.035} {0.042} {0.074} Impact of treatment in area three 0.124* 0.101 38.436 (0.069) (0.067) (33.532) {0.073} {0.129} {0.248} Area one= area two (p-value ) 0.071 0.128 0.844 Area one= area three (p-value ) 0.106 0.212 0.758 Area two= area three (p-value ) 0.852 0.784 0.871 Control mean of dependent var. 0.237 0.206 66.521 Number of observations 511 511 511 Note : This table gives the heterogeneous effect of the treatment in the three areas of the city. Robust standard errors are in parentheses. Randomization inference p-values are in brackets. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and strata fixed effects (the interaction of being located in one of three areas and owing more than one year of arrears). Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 36 Appendix Figure A1: Distribution of Number of Properties Owned Note : This figure displays the distribution of property ownership among the sample for Experiment one. 37 Figure A2: Experiment One Notices Note : The identification and penalty notice included both the letter and the property identification sheet shown above, stapled together. The penalty notice included only the letter above. The identification notice included the letter and the property identification sheet, but the letter did not include the boxed section containing penalty information. The control group received only the letter without the boxed section. 38 Figure A3: Experiment Two Notices Note : These are the front and back pages of the enforcement notice. The control notice did not include the boxed section of the letter. 39 Table A1: Randomization Balance Table (Properties), Experiment One (1) (2) (3) (4) (5) (6) Difference between control group and Mean and Identification SD of P-value and Identification Penalty N Control of f-test Penalty group Property value (US$) 26444.264 -2060.367 -4238.109 -135.445 0.469 571 [24002.115] (2712.522) (2838.611) (3539.482) {0.447} {0.965} {0.142} Owner has multiple properties 0.329 0.000 -0.005 -0.038 0.881 571 [0.471] (0.055) (0.055) (0.055) {0.993} {0.505} {0.940} Area 1356.580 -57.137 -188.388 67.002 0.257 571 [1061.960] (119.362) (118.068) (153.369) {0.626} {0.658} {0.114} Rooms 3.370 0.013 -0.263 0.003 0.510 570 [2.441] (0.272) (0.241) (0.277) {0.957} {0.991} {0.277} Bathrooms 1.932 -0.267 -0.313 -0.215 0.424 568 [1.745] (0.184) (0.194) (0.225) {0.153} {0.353} {0.107} Has Porch 0.829 0.024 -0.054 -0.045 0.286 571 [0.378] (0.043) (0.047) (0.047) {0.574} {0.340} {0.240} Has Garage 0.103 -0.036 -0.053* -0.021 0.363 571 [0.305] (0.033) (0.031) (0.035) {0.276} {0.561} {0.083} Commercial property 0.397 -0.089 -0.073 -0.084 0.362 571 [0.491] (0.056) (0.057) (0.057) {0.116} {0.130} {0.198} Block 2 0.521 -0.138** -0.063 -0.028 0.098 571 [0.501] (0.058) (0.059) (0.060) {0.012} {0.627} {0.277} Block 3 0.281 -0.073 -0.041 -0.020 0.508 571 [0.451] (0.050) (0.052) (0.053) {0.150} {0.736} {0.454} Note : Column 1 gives the standard deviation in brackets. Columns 2, 3, and 4 give the coefficient and robust standard errors (in parentheses) from an OLS regression of the property/property owner characteristic on treatment indicator variables, controlling for owning multiple proper- ties. Randomization inference p-values are in brackets. Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 40 Table A2: Delivery Balance across Treatment Groups, Experiment One (1) Notice delivered Identification and Penalty 0.052 (0.042) {0.224} Identification 0.064 (0.043) {0.134} Penalty 0.046 (0.044) {0.297} Identification and Penalty=Identification (p-value ) 0.745 Identification and Penalty=Penalty (p-value ) 0.872 Identification=Penalty (p-value ) 0.666 Joint F-test (p-value ) 0.510 Control mean of dependent var 0.858 Number of observations 464 Note : This table gives an OLS regression of the property/property owner characteristic on treatment indicator variables, controlling for owning multiple properties. Robust standard er- rors are in parentheses. Randomization inference p-values are in brackets. Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 41 Table A3: Randomization Balance Table (Property Owners- Delivered Notices), Experiment One (1) (2) (3) (4) (5) (6) Difference between control group and Mean and Identification SD of P-value and Identification Penalty N Control of f-test Penalty group Property value (USD) 27722.224 -437.852 -980.220 -7168.440** 0.060 421 [24673.623] (3665.449) (4628.924) (3301.363) {0.915} {0.827} {0.027} Area 1380.313 30.642 61.952 -231.491 0.131 421 [1058.523] (155.473) (194.465) (143.868) {0.856} {0.759} {0.103} Rooms 3.349 0.110 0.033 -0.270 0.471 420 [2.541] (0.344) (0.350) (0.328) {0.745} {0.922} {0.414} Bathrooms 1.925 0.037 -0.298 -0.310 0.287 419 [1.516] (0.244) (0.244) (0.231) {0.880} {0.226} {0.179} Has Porch 0.840 -0.020 -0.062 -0.064 0.569 421 [0.369] (0.050) (0.055) (0.055) {0.699} {0.270} {0.255} Has Garage 0.113 -0.034 -0.028 -0.069* 0.396 421 [0.318] (0.043) (0.045) (0.041) {0.428} {0.545} {0.108} Commercial property 0.358 0.004 0.030 0.002 0.960 421 [0.482] (0.065) (0.069) (0.068) {0.953} {0.649} {0.970} Block two 0.509 -0.132* -0.047 -0.043 0.238 421 [0.502] (0.069) (0.073) (0.073) {0.055} {0.503} {0.543} Block three 0.311 -0.125** -0.012 -0.051 0.145 421 [0.465] (0.063) (0.068) (0.067) {0.045} {0.865} {0.434} Has female owner 0.236 0.008 0.072 0.034 0.674 421 [0.427] (0.061) (0.065) (0.064) {0.899} {0.279} {0.583} Note : Column 1 gives the standard deviation in brackets. Columns 2, 3, and 4 give the coefficient and robust standard errors (in paren- theses) from an OLS regression of the property/property owner characteristic on treatment indicator variables, controlling for owning multiple properties. Randomization inference p-values are in brackets. Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 42 Table A4: Robustness Checks, Experiment One (1) (2) (3) Responded Paid in Amount to notice 2017 Paid (USD) Panel A: Unaffected sample Identification and Penalty 0.050 0.060* 72.822 (0.036) (0.031) (70.802) {0.171} {0.026} {0.325} Identification 0.041 0.007 9.221 (0.037) (0.024) (29.178) {0.318} {0.809} {0.807} Penalty 0.021 -0.011 -1.891 (0.034) (0.020) (10.774) {0.600} {0.735} {0.962} Identification and Penalty=Identification (p-value ) 0.836 0.052 0.409 Identification and Penalty=Penalty (p-value ) 0.434 0.010 0.313 Identification=Penalty (p-value ) 0.622 0.538 0.784 Control mean of dependent var. 0.044 0.022 5.929 Number of observations 371 371 371 Panel B: Controlling for dummies for number of properties owned Identification and Penalty 0.077** 0.071** 65.647 (0.035) (0.029) (58.731) {0.049} {0.013} {0.389} Identification 0.038 0.004 10.862 (0.036) (0.024) (27.205) {0.348} {0.904} {0.805} Penalty 0.019 -0.016 -0.474 (0.034) (0.020) (10.592) {0.627} {0.579} {0.987} Identification and Penalty=Identification (p-value ) 0.301 0.022 0.423 Identification and Penalty=Penalty (p-value ) 0.133 0.002 0.322 Identification=Penalty (p-value ) 0.608 0.494 0.783 Control mean of dependent var. 0.044 0.022 5.929 Number of observations 421 421 421 Note : This table gives results of OLS regressions of the outcome variables on the treatment indicator variables, similar to Table 2. Panel A includes dummies for number of properties owned (replacing a dummy variable for owning more than one property). Panel B restricts the sample to property owners who own only one property or whose multiple properties were all assigned to the same treatment group. Robust standard errors are in parenthe- ses. Randomization inference p-values are in brackets. The omitted group is the control group. Control variables are owning multiple properties, owning a commercial property, property value (the highest value for those with multiple properties), having a female owner or co-owner, and location (one of three subareas). Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 43 Table A5: Delivery Balance across Treatment Groups, Ex- periment Two (1) Notice delivered Enforcement notice 0.004 (0.015) {0.806} Control mean of dependent var 0.966 Number of observations 528 Note : This table gives an OLS regression of the property/property owner characteristic on the treatment indicator variable. Robust stan- dard errors are in parentheses. Randomization inference p-values are in brackets. Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 44 Table A6: Prior Year’s Tax Arrears Payment Rate across Areas of Monrovia, Experiment Two (1) (2) (3) (4) Below Below US$1000 Owes tax All US$1000 liability and arrears liability owes tax arrears Area two 0.101*** 0.086*** 0.115*** 0.091*** (0.012) (0.013) (0.024) (0.025) Area three 0.046*** 0.043*** 0.066** 0.053* (0.013) (0.014) (0.027) (0.028) Number of observations 7496 6390 2203 2067 Mean of area one 0.207 0.228 0.290 0.320 Note : This table shows the success of prior enforcement actions in different areas of Monrovia. Robust standard errors are in parentheses. The omitted group is the control group. Control variables are owning multiple properties, owning a commercial property, property value, having a female owner, and location (one of three subareas). Significance level: * = 10 percent, ** = 5 percent, *** = 1 percent. 45