Do Illicit Financial Flows Hurt Tax Revenues? Evidence from the Developing World

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.


I. Introduction
Domestic tax revenues are a vital resource for the developing world.By contributing to the financing of investments in education, health, and various types of infrastructure (such as electricity, roads, ports, highways), they can be a crucial source of economic development (see e.g.Calderón & Servén, 2004).Given their importance, a recent strand of literature investigates their fragility notably with respect to illicit financial flows (IFFs), which-by reducing the tax base-can be a potential major source of domestic revenue losses for developing countries (see e.g.Kar & Cartwright-Smith, 2008, 2010;Kar & LeBlanc, 2013;Ndikumana & Boyce, 2003, 2011, 2012).‡   The goal of the present paper is to investigate precisely the link between IFFs and domestic tax revenues mobilization.Shedding light on this issue is of primary importance, given the magnitude of IFFs: according to Kar & Spanjers (2015), developing countries illicitly lost around 800 billion USD per year over the period 2004-2013, namely above the yearly net FDI inflows (nearly 650 billion USD), remittances (around 350 billion USD), or official development assistance inflows (ODA, around 82 billion USD).However, analyzing the IFFs-tax revenues relationship raises at least two major challenges.
First-from an empirical perspective-IFFs are viewed as "funds that are illegally earned, transferred, or utilized" (Global Financial Integrity, 2015).§ However, since they stem from various sources such as corruption (embezzlement, bribery, and theft), criminal activities (drug and human smuggling, bootlegging, etc.), and international trade (export under-invoicing and import over-invoicing), ** defining IFFs is not straightforward-as indicated by the existing debate around this concept both in academia (see e.g.Reuter & Truman, 2004;Baker, 2005) and in international institutions (see the discussion in Forstater, 2018).Therefore, given the conclusions of Buchanan (2004) who-after examining the money laundering process and the 2011).As a result, econometric estimations of the effects of IFFs on tax revenues could be criticized for their lack of external validity related to the contingency of IFFs data on a given definition and/or methodology.Second-from a methodological perspective-the effect of IFFs on tax revenues may be subject to endogeneity issues arising (in addition to measurement errors related to the mere IFFs' definition) from reversed causality.Indeed, poor fiscal institutions can result-aside from low collecting rates and therefore low tax revenues-into poor fiscal control and enforcement that may ultimately favor IFFs.
Addressing these issues looks like an unfeasible task, given that it requires finding timevarying reliable instruments for a variable whose mere measurement may be unreliable.
Consequently, in this paper we adopt a different perspective that exploits the fact that combating IFFs requires compliance with international standards, as defined by the Financial Action Task Force (FATF) recommendations on domestic tax revenue mobilization.† † The FATF recommendations formulate a comprehensive and consistent regulatory framework that has been endorsed and recognized since 2003 by more than 180 countries as the international standard for anti-money laundering (AML) and countering the financing of terrorism (CFT).
Conversely, the lack of compliance with these recommendations is considered as a serious signal for not combating IFFs, and is made visible through the inclusion of such countries on the FATF "blacklist".Therefore, we develop an impact assessment analysis in which the presence of a country on the FATF list is considered as a "treatment" variable signaling the presence of major IFFs, i.e. a Non-Cooperative country, as opposed to Cooperative countries that are not on the FATF list.Of course, since being on the FATF list can hardly be seen as a random event, we draw upon appropriate techniques in order to make this event exogenous conditional upon a wide set of covariates that may influence the decision of becoming a listed country.
Our results are as follows.First, propensity score matching estimations reveal that countries on the FATF list experience significantly lower tax revenues with respect to comparable countries that are not on this list.This causal effect is economically meaningful, as the tax revenue loss-estimated always above 2 GDP percentage points-represents between one-third and almost half of the standard deviation of tax revenues depending on the considered estimator.Second, a wide robustness section confirms the soundness of our finding with respect to alternative matching estimators (e.g.including several "doubly robust" estimators), the use of the system-GMM method as an alternative technique to tackle potential endogeneity, and different samples.Third, we reveal important heterogeneities in the effect of being on the FATF list on tax revenues, on two grounds.On the one hand, being a Non-Cooperative country is particularly damaging for indirect taxes-goods & services taxes, value added taxes, and excise taxes-which are reduced by as much as three-fourths of the standard deviation with respect to comparable Cooperative countries; while direct taxes are mostly not significantly affected, with the notable exception of the fairly distortive social contributions that are significantly higher in Non-Cooperative countries (possibly as a-ultimately modest-solution to compensate for the large loss of indirect taxes).On the other hand, estimations based on the control function approach reveal that the effect of being on the FATF list on tax revenues may change in various environments.In particular, the tax revenues gap between Non-Cooperative and comparable Cooperative countries is reduced by the presence of good institutions-which may improve tax collection-or a large finance sector that provides credits-which, by supporting economic activity, may increase the tax base.
Altogether, our findings unveil a damaging and sizeable causal effect of IFFs on domestic tax revenues mobilization.A straightforward policy implication is that developing countries could improve their domestic revenue mobilization by combating IFFs.Our sensitivity analysis offers some directions in this way: improving the quality of institutions or developing a sound financial system may be a way to overcome the detrimental effects of IFFs and foster domestic tax revenue mobilization.
The rest of the paper is organized as follows.Section II discusses the related literature, Section III presents the data, Section IV describes the methodology, Section V reports the main results, Section VI confirms their robustness, Section VII compares the effect of IFFs on indirect and direct tax revenues, Section VIII provides a sensitivity analysis that unveils some channels in the effect of IFFs on tax revenues, and Section IX discusses the implications of our findings from the broad economic development perspective, and concludes the paper.

II. Related literature
According to the neoclassical theory, financial movements are the result of rational portfolio decisions by economic agents, related to macroeconomic conditions such as exchange rate duality, public sector indebtedness, or political stability (Dooley, 1988).In developing countries, a large empirical literature discusses the determinants of financial flows, with a particular focus on the macroeconomic environment (Dooley, 1988;Pastor, 1990;Boyce, 1992;Lensink et al., 1998;Ndikumana & Boyces, 2003;Brada et al., 2013) and the political circumstances (Khan & Haque, 1985;Cuddington, 1986;Alesina & Tabellini, 1989).However, recent voices (see e.g.Purcell & Rossi, 2019;IMF, 2020) suggest that an important determinant of financial flows may have been largely underexplored-namely, outflows resulting from an illicit appropriation, such as theft, plundering of public resources, corruption, trade mispricing, and so forth.
According to our knowledge, only few contributions investigate the illicit part of financial transfers.Regarding their (i) determinants, Kar (2011) argues that IFFs are influenced by macroeconomic factors (fiscal deficit, exchange rate, inflation, real GDP growth, real rate of return, and external debt), structural factors ("un"-inclusive growth, international trade, and reforms), and institutional factors (corruption, informal economy, business environment, and political instability).In particular, Kar & Freitas (2011) suggest that the decrease of IFFs from developing countries just after the Great Recession is triggered by the collapse of international trade, Pérez et al., (2012) highlight a central role for FDI in facilitating illicit money flows, and Holzenthal (2017) points out that financial crime increases in the aftermath of political and economic changes.
Consequently, some contributions try to assess the (ii) consequences of IFFs.In early works on the macroeconomic implications of money laundering, Tanzi (1996) and Quirk (1997) show that money laundering affects the international allocation of resources and the stability of the international financial system, as well as currency and money balances, and economic growth (see also Buchanan, 2004).More recently, Ndikumana & Boyce (2012) underscore that the absence of illicit funds in the African continent would increase the capital stock and GDP per capita by as much as 60% and 15% respectively.‡ ‡ Finally, some studies investigate (iii) policies that may mitigate the consequences of IFFs.Verdugo-Yepes (2011) finds that while the quality of the domestic regulatory framework boosts compliance with anti-money laundering (AML) and combating the financing of terrorism (CFT), an opposite effect is found for corruption (and no effects of financial depth or country openness, among others).In addition, as explained by Ajayi & Ndikumana (2015), although rooted in governance, the persistence of IFFs depends on both domestic and international actors, and therefore on the international political economy and the structure and functioning of global financial, legal, and political organizations; as such, by eroding governance, grand corruption is a key explanation of the IFFs-governance nexus.‡ ‡ Relatedly, the African Tax Administration Forum states that-since around one third of Africa's wealth is held abroad-tax authorities are deprived by resources that could be used to reduce inequalities.
With respect to this literature, our paper is mostly linked with the contributions in the strands of literature described in (ii) and (iii), as follows.First, our main focus is on the consequences of IFFs in terms of tax revenues; in particular, this will require mobilizing information from the studies presented in (i), since our impact assessment analysis contrasts Non-Cooperative countries (i.e. that do not cope with international standards on combating IFFs) with Cooperative countries that present comparable characteristics including in terms of IFFs determinants.§ § Second, a sensitivity analysis explores the possible heterogeneity of the effect of IFFs on tax revenues in various environments; as such, we are interested in policies that may help mitigate the potentially-detrimental effect of IFFs on tax revenues.

III. Data
Based on IFFs data availability our study covers 58 developing and emerging countries during the period 2004-2013.Data on tax revenues comes from the International Centre for Tax and Development's (ICTD) Government Revenue Dataset (GRD) and the IMF's tax revenue dataset, and data on the treatment variable comes from the Financial Action Task Force (FATF).
The remaining variables come from various sources including the World Bank Group (World Development Indicators and Worldwide Governance Indicators), the IMF World Economic Outlook (WEO), the Global Financial Integrity (GFI), ICRG, Kose et al. (2017)  Simple descriptive statistics that compare countries before and after their inclusion in the FATF list reveal the following.*** First, Figure 1a shows that the inclusion of countries in the FATF (i.e.Non-Cooperative countries) signals a change of the IFFs trend: on average, IFFs steeply increase-from 0.41% to 2.38%-compared with the counterfactual decrease for the countries that are not on the list (i.e. the Cooperative countries).Second, as shown by Figure 1b, including countries in the FATF list equally signals a change in the trend of tax revenues: on average, these resources are preserved in Cooperative countries (around 16%), while they dramatically fall (from 18.23% to 13.94%) in Non-Cooperative countries.Although these § § As such, our analysis equally contributes to the substantial literature on the determinants of tax revenues (see e.g. the classical contributions of Lotz and Morss, 1967;Tait et al., 1979;or Tanzi, 1992, and the more recent work of Gupta, 2007;or Bird et al., 2008).According to this literature, the main determinants of tax revenues include income, international trade, agriculture share, natural resources, foreign debt, and particularly the quality of institutions.*** The cut-off dates for non-cooperative countries are the years of their inclusion in the FATF list.The cut-off date for cooperative countries is the mid-year period between the first year when a country is included on the FATF list (2004 in our sample) and the sample end year (2013), as in Miskin & Schmidt-Hebbel (2007).
simple statistics should by no means be considered as reflecting a causal effect, † † † they do not go against a correlation between the inclusion on the FATF list and weaker tax revenues.

IV. Methodology
To evaluate the impact of being a country on the FATF list (i.e. the treatment) on tax revenues (i.e. the outcome), we compute the Average Treatment on the Treated (ATT) with 1 i Y the value of tax revenues for a country in the FATF list, and 0 i Y the value of tax revenues in the same country had it not been on the FATF list.Unfortunately, the latter variable is not observable, and one convenient solution would be to replace it by the mean value of tax revenues in the countries that are not on the FATF list.However, the treatment-being included in the FATF list-is probably not random but dictated by some observable factors (political institutions, macroeconomic conditions, and so forth) that also determine tax revenues.
Therefore, as previously pointed out, comparing the mean value of tax revenues between Non-Cooperative and Cooperative countries can generate a "selection on observables" problem that would lead to biased results.
To address this problem, we compute the ATT by replacing the last term of (1) by the average of tax revenues in countries that are not on the FATF list but present a set i X of characteristics comparable with those of countries on the FATF list, namely (2) † † † The simple comparison of mean values between Cooperative and Non-Cooperative countries may be biased by a "selection on observables" issue-the fact that the treatment (i.e.being a Cooperative country or Not) may not be random but depending on some variables.Under the conditional independence assumption, ‡ ‡ ‡ the ATT capturing the difference in tax revenues between Non-Cooperative and Cooperative countries is attributed to differences in the treatment-being included or not on the FATF list.Nevertheless, implementing this method would be difficult with a large number of X covariates; thus, following Rosenbaum & Rubin   (1983), we concentrate all the information in X into the propensity score (PS) defined as the probability of being a Non-Cooperative country conditional on the observables i X , namely: . Under the intuitive common support assumption-requiring the existence of comparable control units for each treated unit, ( ) . (3) Table 1 shows the estimates of propensity scores using a probit model.§ § § We explain the probability of being included on the FATF list (i.e.non-compliance with FATF recommendations) using a wide set of factors (see Tables A2-A3 in the Online Appendix for descriptive statistics, and definitions and sources of all variables).First, higher past IFFs may signal that a country ignores international standards for combating IFFs, which increases the probability of being included on the FATF list (i.e.being a Non-Cooperative country).Second, in addition to eroding the real value of tax revenues (e.g.Tanzi, 1977), a high inflation rate may signal macroeconomic imbalances that translate into a higher probability of Non-Cooperation.
Third, as more open economies can face important sanctions if they do not comply with international standards, a higher international trade reduces the probability of Non-Cooperation.
Fourth, highly-indebted countries may be exposed to "debt intolerance" (Reinhart et al., 2003), and are less predisposed to comply with international standards given their fiscal profligacy.In the same vein, since it can lead a country to a debt overhang problem, the variable debt service is expected to increase the probability of Non-Cooperation.Fifth, better overall economic performances captured by real GDP growth are not found to significantly affect the probability of Non-Cooperation.Sixth, since the capital is usually invested in sound economies (see e.g. Verdugo-Yepes, 2011), higher FDI inflows are associated with a reduction in the probability of Non-Cooperation.Seventh, a lower level of economic development, captured by a higher share of agriculture in the GDP, is found to be associated with a higher probability of Non-Cooperation.Eighth, the effect of natural resources is a priori ambiguous since countries with ‡ ‡ ‡ This identifying assumption requires that-conditional to the set of observable factors-the outcome (tax revenue) is independent of the treatment variable.§ § § Caliendo & Kopeinig (2008) state that the choice between logit and probit models is not critical as they usually yield similar results; we report that results are indeed comparable when using a logit model.
high natural resources may establish sound institutions and increase their compliance with FATF recommendations, or-on the opposite-be confronted with various forms of the "Dutch disease" (e.g.Sachs & Warner, 1995); in our estimations, the former explanation seems to drive a positive impact of natural resources on the probability of Non-Cooperation.Finally, good institutions are expected to signal well-governed countries that generally meet their commitments regarding the international cooperation against IFFs.As shown by columns [1]- [6], whenever significant, a better institutional quality-measured by various indexes-is associated with a lower probability of Non-Cooperation.Based on propensity scores computed from Table 1, we estimate in the following the effect of being on the FATF list on tax revenues.

V. Results
Capitalizing on existing studies (e.g.Lin & Ye, 2007;Tapsoba, 2012), we compare tax revenues in Non-Cooperative and Cooperative countries-matched based on their propensity scoresusing various matching methods: the Nearest-Neighbor (NN) matching, the Radius matchingwith a large (r=0.05), a medium (r=0.01), and a small (r=0.005)radius, the regression-adjusted Local Linear matching (Heckman et al., 1998), and the Kernel matching.
Estimations of the ATT are reported in Table 2. Before presenting our main results, we report at the bottom of the table several tests to evaluate the quality of our fitting.First, we check if our estimations are affected by an eventual hidden bias arising from unobservables (Rosenbaum, 2002).The value of the Rosenbaum sensitivity test-performed at the common 5% significance level-is around 1.5, namely above the value used by other studies to conclude to the lack of such a hidden bias (for example, Guerguil et al., 2017, conclude with values around 1.2).Second, following Rosenbaum & Rubin (1985), the high p-values of the standardized bias confirm the statistical equality between the means of the various observed variables among countries from the treated and the control group.Finally, the low values of the pseudo R2 reveal that our matching procedure leads to balanced scores (see Sianesi, 2004).
Altogether, these results support the conditional independence assumption and the common support assumption.
Given the quality of our estimations, we can now discuss our main results reported in the top part of Table 2.As illustrated by line [1], all eight ATTs are significant and negative irrespective of the matching method.Consequently, we reveal that being a Non-Cooperative country-included on the FATF list-is associated on average with significantly lower tax revenues, between-in absolute value-2.066(2-Nearest Neighbor Matching) and 2.778 (Local Linear Regression Matching) percentage points.Moreover, this effect is economically meaningful, as it represents between one-third and almost half of the standard deviation of our measure of tax revenues (which is slightly above 6, see Table A2 in the Online Appendix).

VI. Robustness
We explore the robustness of our main finding with respect to three important issues: (i) different methods to perform the matching, (ii) a different way to tackle potential endogeneity, and (iii) different samples.

Alternative matching methods
In our benchmark analysis, we draw upon propensity score matching to evaluate the effect of being in the FATF list on tax revenues.Alternatively, we estimate this effect using various matching methods, whose presentation below is based on Cattaneo (2010), Drukker (2016), and Heß (2017).First, the bias corrected matching estimator (see Abadie and Imbens, 2011) combines matching-in order to match each treated unit with the unit with the closest propensity score-and regression-in order to mitigate a potential bias that may survives due to imbalances in the covariates.Second, regression adjustment (RA) controls for the estimated propensity scores in a linear regression that is used to estimate treatment effects.Third, the inverse probability weighted estimator (IPWE) explores the idea that more rare observations in the treated (control) group relative to observations in the control (treated) group-with respect to a covariate-provide good counterfactual information and should therefore receive larger weights (i.e. by creating a pseudo-population) when the matching is performed.Fourth, the augmented inverse probability weighted estimator (AIPWE) augments the IPWE estimator with a regression analysis-making it a "doubly robust" estimator that still leads to consistent estimations results if either the treatment or the outcome model is misspecified.Fifth, the inverse-probability weighted regression adjustment (IPWRA) "doubly robust" estimator draws upon inverse-probability weights-based on the coefficients of a regression-to calculate the means of the treatment-specific predicted outcomes.
As illustrated by Table 3, irrespective of the alternative method considered all five estimated ATT are significant and negative.Moreover, the size of the effect is estimated between (in absolute value) 2.1 (regression adjustment) and 2.5 (AIPW) percentage points, namely comparable with our benchmark findings.Finally, accounting for alternative measures of institutions in the bottom part of Table 3 leads to comparable conclusions.Consequently, the significant loss of tax revenues in Non-Cooperative countries with respect to comparable Cooperative countries is robust to the use of alternative matching methods.

A different look at endogeneity
The use of propensity score matching and-as illustrated by the previous subsection-of alternative matching methods revealed a robust effect of our treatment-being included on the FATF list-on tax revenues.Such an event analysis is an appealing way to estimate a causal effect in the absence of robust external instrumental variables.However, an alternative way to control for possible endogeneity in the effect of IFFs on tax revenues is to use the System-GMM method coined by Blundell & Bond (1998) that draws upon internal instrumental variables-it instruments the variables using their lagged values as instruments.
Following the recent literature (see e.g.Combes et al., 2018), when applying System-GMM we restrict and collapse the instrument set to avoid the proliferation of instruments (Roodman, 2009) and we correct standard errors for the finite sample bias (Windmeijer, 2005).
As illustrated by the bottom part of Table 4, in addition to an appropriate number of instruments (namely, lower than the number of cross-sections), the Hansen J-test does not reject the null hypothesis that our instruments are valid (i.e.uncorrelated with the error term).Next, the AR(2) test showing the absence of second-order autocorrelation of the error term supports the quality of our fitting.Moreover, the fairly large lagged coefficient of tax revenues (see the top of Table 4) supports our use of a System-GMM model to cope with the important persistence of tax revenues.In addition, whenever significant control variables present the expected sign, namely a positive effect of trade and GDP growth.Finally, regarding the main result, column [1] of Table 4 shows that the effect of the variable illicit financial flows is negative and significant.Since this variable is computed as a dummy variable equal to 1 if a country is Non-Cooperative and to 0 otherwise, our estimations suggest that on average Non-Cooperative countries present significantly lower tax revenues compared with Cooperative countries.Given the dynamic setup, the loss of tax revenues is estimated at (in absolute value) around 0.77 percentage point in the short run, and at around 3.35 percentage points in the long run-comparable with the magnitude of our matching estimations.Consequently, System-GMM estimations support the causal effect of being on the FATF list on tax revenues outlined by our treatment analysis.

Different samples
Finally, in addition to different methods, we now look at the robustness of our main findings when using different samples.Noting that the various tests reported in Table 5 support the quality of our propensity score matching estimations, our results are as follows.First, to reduce the influence of the Great Recession we drop the year 2009.As illustrated by line [1] in Table 5, the eight estimated ATTs are all significant and negative, consistent with our benchmark findings.Second, comparable results arise when we drop the years before 2006 to abstract of the saving glut period (see line [2]).Third, if we disregard-given their sometimes particular macroeconomic and political conditions-former communist countries (line [3]) and former USSR countries (line [4]), the magnitude of the negative effect is higher (in absolute) valuea revenue loss of up to 5 percentage points.Fourth, dropping fuel exporters leaves ATTs negative and significant irrespective of the matching method (line [5]).Fifth, estimations on lines [6] and [7] show that the unfavorable effect of being on the FATF list on tax revenues is not driven by observations in countries with high external debt (i.e.above 90% in ratio of GDP) or high inflation episodes (i.e.inflation rates above 40%).Lastly, ATTs remain negative and significant when excluding countries from monetary unions (line [8]).Consequently, our main findings are unaffected by the use of different alternative samples.
To summarize, the robustness analysis supports our benchmark findings: compared with cooperative countries, countries that do not comply with international standards against international illicit flows present on average significantly lower tax revenues.Given the soundness of this result, we investigate in the following its sensitivity in various contexts.

VII. Effects on tax components
Our estimations support-by and large-a significant difference in terms of tax revenues between countries that comply and those that do not comply with the FATF standards.
However, tackling IFFs may affect the various direct and indirect taxes in different ways.

Indirect taxes
Given that our sample is composed of developing and emerging countries, tax revenues are mostly collected from indirect taxes (see e.g.Stiglitz et al., 2006).Estimations in Table 6 present the difference between Non-Cooperative and Cooperative countries regarding general goods & services taxes, value added taxes, and excise taxes, respectively.
Regarding general goods & services taxes, all estimated ATTs reported on line [1] are significant and negative, suggesting that being a Non-Cooperative country results into a tax loss of around 2.6 percentage points on average with respect to comparable Cooperative countries.
Moreover, such a significant difference between countries that are included or not on the FATF list is found regarding value added taxes: estimations reported on line [2] reveal a loss of VAT of around 1.6 points on average.Lastly, as shown by all eight ATTs displayed on line [3], being on the FATF list is equally associated with lower excise taxes-around 0.6 percentage point on average.
Consequently, complementing our benchmark analysis, these results emphasize that failing to respect the international standards on combating IFFs has negative consequences on the mobilization of all types of indirect taxes.Based on our estimations, the strongest negative impact is on the goods & services taxes (almost three-fourths of their standard deviation), followed by value added taxes (around 60% of their standard deviation), and excise taxes (around one-third of their standard deviation).

Various direct taxes
We perform estimations on several types of direct taxes, namely income taxes, individual taxes, property taxes, corporate taxes, payroll taxes, and social contributions.Estimations reported in Table 7 unveil two important results.
On the one hand, all 40 estimated ATTs reported on lines [1]-[5] are not statistically significant, suggesting that most direct taxes are not different between Non-Cooperative and comparable Cooperative countries.Consequently, the overall damaging effect of being on the FATF list is driven by the loss of indirect taxes.
On the other hand, the estimated ATTs reported on line [6] are all positive and statistically significant, suggesting that social contributions are higher in Non-Cooperative countries with respect to comparable Cooperative countries.Consequently, countries included in the FATF list try to compensate the loss of indirect taxes by strongly increasing (i.e. the estimated effect is around one standard deviation) the revenues from the-more distortivesocial contributions; even so, with a narrower tax base and weaker tax compliance, they only partially compensate for the tax revenue loss from IFFs as illustrated by our benchmark results.

VIII. Exploring heterogeneity in the treatment effect
A large literature highlights notable heterogeneities in developing countries' macroeconomic conditions and institutional frameworks-see e.g.Acemoglu et al. (2005); Lin & Ye (2009); Tapsoba (2012); Minea & Tapsoba (2014); Balima et al. (2017).Consequently, in this section we investigate a potential heterogeneity of the effect of complying with international standards for combating IFFs on tax revenues.
Following Lin & Ye (2009) and Tapsoba (2012) we explore various possible sources of heterogeneity based on the control function approach-see Wooldridge (2015) for an excellent presentation of the method.To this end, we begin by presenting in column [1] of Table 8 the results of OLS regressions with Impact the compliance dummy within the common support, when controlling for the estimated propensity score obtained from our baseline probit model.
The statistically-significant coefficient of the propensity score confirms the presence of selfselection in the model; conditional on controlling for this self-selection, the significant and negative estimated coefficient indicates that Non-cooperative countries collect less tax revenue than comparable Cooperative countries, consistent with our benchmark findings.In the following, we augment this baseline specification with various variables that could affect the effect of the compliance dummy (Impact) on tax revenues mobilization (see Tables A2-A3 the Online Appendix for descriptive statistics, and definitions and sources of these variables).
First, regarding fiscal variables (see e.g.Sargent & Wallace, 1981;Reinhart et al., 2003;Ostry & Abiad, 2005), we include an interaction term between the compliance dummy and the difference between the estimated propensity score and its sample average.The significant and positive effect of this interactive term in column [2] suggests that, as the external debt increases, the difference in tax revenues between Non-Cooperative and Cooperative decreases, probably because high-debt contexts call for additional taxes for fiscal stabilization.Such an effect is confirmed when alternatively considering in columns [3]-[4] the ratios of debt held by nonresidents in total debt and short term external debt in reserves.However, the lack of significance of the respective interactive terms suggests that the government size and debt default do not seem to trigger heterogeneous effects (see Table A4 in the Online Appendix for estimations with all the variables for which the interactive term is not significant).
Second, regarding monetary and financial variables (see e.g.Tanzi, 1977;Eichengreen et al., 2003;Brada et al., 2013;Balima et al., 2017), contrary to the lack of significance of the interaction terms using the variables saving glut and high inflation (see Table A4), the interactive terms using the variables domestic credit to the private sector and credit provided by banks are significant and positive (columns [5]-[6]).By stimulating the activity, higher credits may allow collecting additional taxes that reduce the tax revenues gap between Non-Cooperative and Cooperative countries.
Third, the difference in tax revenues between Non-Cooperative and Cooperative countries is equally mitigated in a context of better institutional quality, and notably in countries with higher regulatory (column [7]) and stability (column [8]) indexes (while-despite presenting the expected positive sign-the interactive terms using the variables effectiveness, law, voice, corruption, and political risk are not significant, see Table A4).Such a result may reproduce the fact that better institutions can improve tax collection, with a favorable effect on tax revenues.Fourth, regarding international factors (see e.g.Burnside & Dollar, 2000;Demir, 2016), contrary to the lack of significant effect of the interactive terms with the variables FDI inflows and terms of trade (see Table A4), the presence of the official development assistance (ODA) reduces the tax revenues gap between Non-Cooperative and Cooperative countries (column [9]).Indeed, since ODA often comes with measures related to improved allocation and greater transparency, such measures can equally translate into various government activities including in terms of tax collection, which may ultimately improve tax revenue mobilization.
Finally, Table A4 in the Online Appendix equally reports estimations of interactive terms with other variables (see e.g.Tanzi, 1992;Sachs & Warner, 1995;Gupta, 2007;Corsetti et al., 2012;Debrun & Kinda, 2016), namely the time length since a country was included on the FATF list, the difference between the estimated propensity score and its sample average, the phase of the business cycle, education, and mineral and natural resources.These variables are not found to affect the tax revenues gap between Non-Cooperative and Cooperative countries, despite being consistent with what one may expect-for example, the coefficient of the interactive term with education is positive and with an associated p-value of 0.25.
Altogether, the results presented in this section complement our previous findings, as they emphasize different environments that may affect the impact of being on the FATF list on tax revenues.

IX. Concluding remarks
Recent work draws attention to the fragility of domestic tax revenues-a vital resource for the developing world-to illicit financial flows (IFFs).To cope with two major challenges in the relationship between IFFs and tax revenues-related with the mere measurement of IFFs and reverse causality-we exploited the Financial Action Task Force (FATF) data in an impact assessment analysis.Propensity score matching estimations revealed a significant tax revenue loss in countries associated with important IFFs with respect to comparable countries that are not signaled as being related to important IFFs.This causal effect-that was estimated to be economically meaningful-is supported by a large robustness analysis, and in particular remains unchanged when using several "doubly robust" estimators.Lastly, we unveiled heterogeneities in the impact of IFFs on tax revenues, related to the type of tax-significantly lower indirect taxes but not direct taxes-and the considered environment.
Our findings should be understood in two ways.First, it is not because the number of countries on the FATF list decreased that the damaging effect of IFFs on tax revenue mobilization was wiped off: although probably of a lower magnitude, IFFs still continue to erode tax revenues in the developing world.Therefore, there is still room for policies aimed at fostering tax revenue mobilization-that, for example, improve the quality of institutions or develop a sound financial system, as shown by our estimations.Second, according to our analysis, the presence of large IFFs over several decades deprived many developing countries of important tax revenues, which may have resulted in a shortage of various types of investments that are crucial for economic development.From this perspective, combating IFFs may be a virtuous solution for limiting persistent delays in the economic development process.
Consequently, our findings may provide additional motivation for exploring the impact of IFFs on several grounds.First, with the development of reliable cross-country data on IFFs, it would be interesting to analyze possibly-complex non-linearities in the relationship between IFFs and tax revenues.Second, a cost-benefit analysis that would compare the cost of fighting IFFs with the extra tax resources they engender would be equally important.Finally, in addition to their effect on taxes, analysis of the effects of IFFs on other economic variables-such as growth or investment-would also be worthwhile.We leave these topics for future research.

Figure 1 .
Figure 1.Tax revenue and IFFs before vs after inclusion in FATF list of Non-Cooperative countries (a) Average IFFs (b) Average Non-resources Tax revenue Finally, as shown by lines [2]-[6], using different measures of institutional quality does not affect ATTs, who remain significant and negative in all cases, with an adverse effect climbing up to (in absolute value) 3.589 percentage points line [4]).