How tax morale is related to remittance dependency

A consensus in the literature examining how migrant remittance affects the political behaviours of those left behind is that dependence on migrant remittance significantly affects the recipient's political behaviour. One of the causal pathways is predicated on the presumption that dependence on migrant remittance weakens the efficacy of vote buying as the additional income it raises the recipient's income and lowers the marginal utility of targeted electoral transfers (Ahmed, Reference Ahmed2017).Footnote ⁴ Ultimately, this undermines the remittance recipient's predilection for political engagement. Goodman and Hiskey (Reference Goodman and Hiskey2008) provide evidence in support of this view. Using a municipal and individual-level data from Mexico, they found that high migration municipalities exhibit lower voter turnout rates, while individuals in high migration areas report lower levels of political efficacy as well as participate less in politics. Similar evidence has also been documented by López García (Reference López García2017, Reference López Garcia2018).

While the foregoing is suggestive that migrant remittance results in political apathy among remittance dependents, the additional income migrant remittance brings also makes the recipient wealthier. Hence, migrant remittance dependents may have relatively fewer economic grievances and hold a more positive view of current economic conditions which they attribute to the incumbent's favourable performance.Footnote ⁵ This view has received empirical support from Germano (Reference Germano2013) who found that individuals who depend more on remittance were less likely to lobby officials for economic assistance and consequently less likely to punish the incumbent party in elections for poor economic performance. A study by Ahmed (Reference Ahmed2017) based on a sample of Latin American countries also showed empirically that at increasingly higher levels of dissatisfaction with the incumbent, a remittance dependent is more likely to vote for the incumbent.

One of the recurring arguments in the above literature is that migrant remittance dependents rely less on government goods and patronage because the remittance empowers them to enter the private market and substitute state provisions. This underscores two causal remittance effects vis-à-vis the income effect and substitution effect that are often discussed in the literature (see Acevedo, Reference Acevedo2020; Duquette-Rury, Reference Duquette-Rury2014; López García and Maydom, Reference López García and Maydom2021, Reference López García and Maydom2023). On the one hand, the income effect is associated with an increase in the consumption level of remittance dependents due to the general income increase remittance brings. In this case, the recipient begins to demand those goods and services [s]he wasn't demanding before and is able to easily afford the private goods alternative. The substitution effect, on the other hand, is associated with the remittance dependent's changes in consumption pattern or preferences (such as a change from preferring lower- to higher-quality goods and services) that are induced by the additional income embodied in remittance. In this case, the additional income remittances enable remittance dependents to substitute public goods and services with private alternatives they believe are of better and higher quality. These factors put together underscore a widely held view in the literature that individuals and households in developing countries use migrant remittance to access both basic and public utilities (e.g. food, shelter, security, sanitation, water, electricity, education and roads) in the private market (Adida and Girod, Reference Adida and Girod2011; Doyle, Reference Doyle2015).Footnote ⁶ In the case of public utilities, remittance dependents become self-providers of public goods and services, and therefore more autonomous from (or less dependent on) the state (López García and Maydom, Reference López García and Maydom2021, Reference López García and Maydom2023: 1351).

In line with the foregoing, we conjecture that migrant remittance can also affect the tax morale of individuals that depend more on remittance through a substitution effect. Tax morale is the intrinsic motivation to pay tax, and it has been argued to have a stronger and sizeable role in explaining tax compliance behaviours (Luttmer and Singhal, Reference Luttmer and Singhal2014). One of the long-standing views in the public finance literature is that the quality and quantity of public goods provisioning are important determinants of tax morale. This view is formally enshrined in the psychological tax contract thesis by Feld and Frey (Reference Feld and Frey2007) or Wicksell's voluntary exchange theory, which posits that taxes are voluntary payments by individuals in exchange for public goods and services (Christian, Reference Christian1978; Musgrave, Reference Musgrave1939).

The psychological tax contract or the voluntary exchange theory considers the act of tax paying as a quasi-voluntary one and portrays the existence of the state as a social contract that defines the relationship between the government and the governed. This contract involves duties and rights (Feld and Frey, Reference Feld and Frey2007) such that ‘while the government taketh away, it also giveth’ (Alm et al., Reference Alm, Jackson and Mckee1992), and breaking this contract creates a vicious circle (Hug and Spörri, Reference Hug and Spörri2011). Against this backdrop, migrant remittance allows recipients greater access to the private market alternatives to public goods and services and would in turn affect their predisposition to pay tax.

The negative effect of migrant remittance on tax morale via the substitution effect is further exacerbated by the fact that it can incentivize policymakers to decrease public spending or divert resources away from citizens. The small increases in remittance could shift the allocation of government expenditures on public goods and services to patronage. This is because remittances constitute a form of private subsidy for the provisioning of public goods and services. Hence, the government has more incentive to reduce and divert resources for private gains because it thinks that remittances will do the ‘job’, while access to remittance income makes remittance receivers reluctant in holding the government accountable because it makes political patronage less costly for them to bear (Ebeke, Reference Ebeke2012).Footnote ⁷ This view is also in line with the literature showing that remittance reduces government spending as well as deteriorates governance (Abdih et al., Reference Abdih, Chami, Dagher and Montiel2012; Ahmed, Reference Ahmed2013; Doyle, Reference Doyle2015). Consistent with the ‘psychological tax contract thesis’ or the ‘voluntary exchange theory’ espoused above, this will negatively reduce the tax morale of citizens. Moreover, an existent literature indicts such reduction in government spending and bad governance as causes of lower tax morale (see Boly et al., Reference Boly, Konte and Shimeles2021; Kogler et al., Reference Kogler, Muehlbacher and Kirchler2015; Lago-Penas and Lego-Penas, Reference Lago-Penas and Lego-Penas2010).

The foregoing discussions, therefore, lead to the following hypothesis:

The negative effect of migrant remittance on tax morale hinges on the implicit presumption that public goods and services are underprovided either in quantity or quality such that migrant remittance offers an opportunity of a substitute. The question then that emerges is what happens when public goods and services are adequately provided, and the quality of governance is good? In this case, the additional income embodied in migrant remittance may still induce a substitution from public goods and services to their private alternative. However, this will be for other reasons, such as the ‘positional’ value of the private goods and services as consuming them signals status, wealth or success. Furthermore, the additional income can be channelled to strictly private goods and/or private investments. Other things equal, therefore, we expect the tax morale of the remittance dependents with positive valuations about the quality of public goods and services to be unaffected as such income effect is only expected to increase consumption of private goods and drive private investments. Conversely, we expect migrant remittance to negatively affect the tax morale of remittance dependents with negative valuations about the quality of public goods and services because they use remittance to self-provide for themselves these public goods and services. The foregoing argument is consistent with the public finance literature, which highlights citizens' discontent with public goods and service delivery as one of the prominent factors that increase tax resistance and widespread tax evasion (Ali et al., Reference Ali, Fjeldstad and Sjursen2014; Fjeldstad and Semboja, Reference Fjeldstad and Semboja2001; Russo, Reference Russo2013). This leads to the following hypothesis:

Finally, it is worth noting that while the preceding hypotheses focus largely on the substitution/income effect, migrant remittance can also affect the tax morale of remittance dependents through the ‘norms-transfer’ channel as espoused in the introduction. As we do not have the requisite data to test this causal pathway, we do not provide hypothesis for them.

Data description

Our empirical analysis primarily depends on the Afrobarometer data, which contain a collection of nationally representative surveys across 34 African countries. The surveys measure public opinion on economic, political and social aspects relevant to development, including citizens' attitudes and behaviour towards paying taxes. For the baseline analysis, we use round 7, conducted between 2016 and 2018, because it is the only round that provides information on whether the respondents depend on remittance. Rounds 4 and 6 will be used to assess the robustness of the results as they only inform whether the respondents received remittance but does not necessarily inform how dependent the respondents are on receiving remittance.

To assess whether the respondents depend on remittance, we refer to the question in round 7 surveys that asks the respondents how dependent they are on receiving remittance. The possible responses range from not at all to a lot. We thus create a dummy variable, ‘remittance dependence’, that equals one if the respondent depends on remittance and 0 otherwise.Footnote ⁸ In robustness checks, we also consider an ordinal variable that equals 0 for those who don't depend on remittance, 1 for those who depend on a little bit and 2 for respondents who depend on a lot or somewhat.

Table A2Footnote ⁹ in the online appendix shows the percentage of people who depend on migrant remittances by country. The countries are listed in ascending order of the percentage of respondents who depend on remittance. On average, around 21% of the people interviewed reported being dependent on remittance in the full sample. However, we observe some heterogeneity across the countries. For example, Gambia has the highest proportion of people who depend on remittance, with a percentage of 47%. It is followed by Lesotho and Cape Verde with percentages of 37 and 31%, respectively. The country with the lowest proportion is Kenya, where only 5% of the respondents depend on remittance, followed by Tanzania, with a proportion of around 9%, and Mauritius, with 12%. We also report in Table A2 the amount of remittances as a share of GDP in the countries. We observe that countries that have remittance inflows representing more than 10% of their GDP have at least 20% of their respondents depending on remittances. These countries include Liberia, Senegal, Cape Verde, Lesotho and Gambia.

To measure support for taxation, we rely on the question of the surveys that asks whether the respondents agree that people must pay taxes or revenue owed to governments. Table A3 shows the distribution of the respondents into these three categories: ‘agree/strongly agree’, ‘neutral’ and disagree or ‘strongly disagree’. As shown in Table A3, around 75% of the respondents agree or strongly agree that people must pay taxes, while there is 20% of people who disagree or strongly disagree with this statement. The remaining 4% of the respondents neither agree nor disagree that people must pay taxes. The table also shows some heterogeneity across countries, indicated by a high standard deviation of around 9%. Malawi records the lowest percentage (38%) of people who support taxation, and Sierra Leone is the country with the highest level of support for taxation, where 94% of the respondents agree that citizens must pay taxes. For the empirical analysis, we create a dummy variable support_tax that equals 1 for respondents who agree or strongly agree that citizens must pay taxes and 0 for all the other responses. Missing values are allocated for the respondents who did not or refuse to respond to the question.

The third set of critical variables for our analysis are the variables that assess the quality of public services as perceived by the respondents. These variables will be used as determinants of the classification of the respondents into the different classes/subtypes that will be detected in the data if any. We assess whether the probability that an individual belongs to a class with a specific relationship between remittance dependency and support for taxation depends on how well or badly the same individual assesses the quality of public services. To this end, we rely on the questions of the survey that ask the respondents how well or badly they think their government handles the following matters: health, educational needs, the provision of water, sanitation services and electricity supply. In Table A4, we show for each of the public services the proportion of people that have a positive appraisal about how their governments handle the matter. Overall, about 54% of the respondents have a positive appraisal of how their government handles the provision of health services in the sample. This number is 57% for the provision of education needs. However, we have lower percentages for the provision of water and sanitation and for the provision of electricity for which we have 46 and 48%, respectively. In the empirical analysis, we will first create a dummy variable that equals 1 for respondents who have a positive appraisal of any of the public services and 0 otherwise. Second, we create four different dummy variables for each public service to test if some of them matter more than others. Finally, we also create an index based on how many of the public services a respondent holds a positive appraisal about.

We also control for a range of individual socioeconomic characteristics, including gender, education, location, age, access to information, poverty index, paying a bribe to receive official documents, household services or to avoid problems with the police. The inclusion of these variables is primarily motivated by the extant literature on the drivers of tax morale (see Ali et al., Reference Ali, Fjeldstad and Sjursen2014; Alm and Torgler, Reference Alm and Torgler2006; Boly et al., Reference Boly, Konte and Shimeles2021; Fjeldstad and Semboja, Reference Fjeldstad and Semboja2001; Kogler et al., Reference Kogler, Muehlbacher and Kirchler2015; Kountouris and Remoundou, Reference Kountouris and Remoundou2013; Lago-Penas and Lego-Penas, Reference Lago-Penas and Lego-Penas2010; Martinez-Vazquez and Torgler, Reference Martinez-Vazquez and Torgler2009; Rodriguez-Justicia and Theilen, Reference Rodriguez-Justicia and Theilen2018). Because some of these additional controls may be correlated with the two variables of interest – remittance dependency and quality of public services – their exclusion may induce omitted variable bias. For instance, urban dwellers tend to have higher tax morale, but at the same time have better access to quality public goods and services. An individual's employment status as well as poverty level are also strong determinants of the person's tax morale, but also correlated with whether the person will depend on remittance as a sole source of income (Acevedo, Reference Acevedo2020). Therefore, the omission of the urban variable could confound the effect of quality of public services we document, while the omission of poverty level and employment status could confound the effect of remittance dependency. We also control for country-level variables such as GDP per capita, control of corruption, remittance inflows as a share of GDP and the weighted average of income taxes in the host countries.Footnote ¹⁰ The inclusion of the country-level variables enables us to isolate the effect of the variables of interest from these country-specific confounding factors.

Empirical specification

Let's define T _ij as the response of an individual i living in country j for whom T _ij equals 1 if she/he supports taxation and 0 otherwise. R _ij is the remittance dependency response of individual i from country j, which equals 1 if the individual depends on receiving remittance and 0 if no. n _j is the total number of individuals interviewed in country j such that $N = \sum\nolimits_{j = 1}^n {n_j}$, where N is the total number of observations in the data and n is the total number of countries. In the analysis, we also control for varying individual-level variables, X _ij, and country-level variables, Z _j. For simplicity, we define ω, the vector of all the individual- and country-level variables.

A standard OLS or multilevel model would assume that all the observations fall into a single class/subtype and that the effects of the control variables, including the effect of remittance dependency on support for taxation, are similar for all the observations in the data. This would mask any potential unobserved heterogeneity in the data, increasing the risk of biased estimates on the relationship between remittance dependency and support for taxation. To deal with unobserved heterogeneity, we employ the finite multilevel mixture of regressions model (Henry and Muthén, Reference Henry and Muthén2010; Muthén and Asparouhov, Reference Muthén and Asparouhov2009). The approach enables us to detect endogenously hidden classes or subtypes of individuals that exist in the data such that the conditional density of the dependent variables given all the explanatory variables is class-specific. The latent classes are defined by qualitative differences in the relationship between tax compliance and all the control variables, including remittance dependency. Unlike traditional supervised or unsupervised clustering methods that identify groups of similar observations based on the unconditional density of one or more variables, the finite multilevel mixture approach identifies groups of similar observations based on the conditional density of the dependent variable given a set of explanatory variables. In other words, this method endogenously identifies classes or groups of individuals that have similar patterns on how the dependent variable is affected by the explanatory variables introduced in the analysis, without a priori assumption on the sign, size or significance of the estimates across the classes.

To define our multilevel finite mixture model, let us denote class _ij the latent class variable at the individual level. We also consider the possibility that latent groups at the country level may exist and could affect the classification of the individuals into the identified individual-level classes. Thus, we define gclass _j as the latent class at the country level. We assume that the total number of latent classes at the individual level varies between 1 and K while the country-level latent classes vary between 1 and L.

The general form of the density of the dependent variable, tax compliance (T), given all the explanatory variables and the parameters to estimate can be defined by f(T|R, ω;θ). For a multilevel finite mixture model, this density can be explicitly written as follows:

(1)$$\eqalign{f(T|R,\;\omega ,\;\theta ) = &\prod\limits_{{\kern 1pt} j = 1}^n {\left[ {\sum\limits_{l = 1}^L {\pi _l} } \right.} (gclass_j = l)\left[ {\prod\limits_{i = 1}^{nj} {\sum\limits_{k = 1}^K {\pi _{ijlk}} } } \right.(class_{ijl} = k|gclass_j) \cr \left. {\left. { \times P(T_{ij} = 1|class_{ijl} = k,\;R_{ij},\;\omega _{ij},\;\theta )} \right]} \right]} $$

where π _l(gclass = l) is the probability that a given country j belongs to the country-level latent class l, and n is the total number of countries in the sample. π _ijlk(class _ij = k|gclass _j) is the probability that an individual i from country j belongs to the individual-level latent class k given that its country belongs to gclass l with a probability π _l(gclass = l). For each individual, the sum of the estimated probabilities to belong to the different classes is equal to 1. The component P(T _ij = 1|class _ij = k;R _ij, ω _ij;θ) in equation (1) is the probability that an individual i from country j supports taxation given that he/she belongs to class k.

To examine whether the quality of public services determines class membership, we endogenize the probability to belong to a given class k, i.e. the parameter π _ijlk(.) by defining it as a function of variables commonly named concomitants which help to explain inter class heterogeneity. In this paper, the concomitant variables are the variables that capture the respondents' appraisal of how their governments handle public services such as water and sanitation, health, education and electricity. Let us denote ψ _ij the vector of concomitant variables. Hence, by incorporating the concomitant variables in equation (1) we obtain the following equation (2):

(2)$$\eqalign{\,f( {T{\rm \vert }R, \;\omega , \;\psi ; \;\theta , \;\emptyset } ) = & \mathop \prod \limits_{\,j = 1}^J \left[{\mathop \sum \limits_{l = 1}^L \pi_l( {gclass_j = l} ) \left[{\mathop \prod \limits_{i = 1}^{nj} \mathop \sum \limits_{k = 1}^K \pi_{ijlk}( {class_{ijl} = k, \;\psi_{ij}; \;\emptyset } ) } \right.} \right. \cr & { { \times P( {T_{ij} = 1\vert class_{ijl} = k; \;R_{ij}, \;\omega_{ij}; \;\theta } ) } \Bigg ] } \Bigg ]} $$

Vector θ includes the parameters on the explanatory variables, including dependence on remittance, while the vector ∅ consists of the parameters on the concomitant variables. To estimate the parameters θ and ∅, we maximize the log of f(T|R, ω;θ) using the expectation-maximization (EM) algorithm (Dempster et al., Reference Dempster, Laird and Rubin1977). An important question is how to set the maximum numbers of individual- and country-level latent classes to estimate, i.e. class and gclass. While there is no rule of thumb, it is important to note that the number of parameters to estimate increases with the number of classes. Also, higher number of parameters to estimate more data is needed to reach convergence in the estimations. Our approach is to estimate as many models with different combinations of individual- and country-level classes as possible until we encounter serving convergence issues or observe classes with no or too few observations. In our estimations, we start observing some convergence issues or empty classes when K = 7 and L = 5. For the rest of the paper, we hence assume that there are between 1 and 7 latent classes/subtypes of individuals and that countries may be grouped between 1 and 5 groups.Footnote ¹¹

To select the model that best fits the data, we employ statistic criteria based on the log-likelihood of the estimations: the Bayesian information criterion (BIC), the consistent Akaike information criterion (CAIC) and the Schwartz Bayesian criterion (SBC). These statistics help us select the model with the combination of gclass and class that best fits the data. The lower are these statistics greater is the goodness of fit.

Estimation results

Identifying latent classes/subtypes of individuals in the data

We estimate different models, changing the number of latent classes/subtypes while also accounting for possible unobserved heterogeneity at the country level that could affect individual responses. Recall that in the previous section, we defined gclass as the number of latent groups at the country level and class as the number of latent classes at the individual level. We run each model with many random starting points and several iterations to ensure that we reach convergence. For each combination of country- and individual-level latent classes, we report the values of the BIC, CAIC and SBC. Our best model should be the one with the lowest values on all or at least two of the statistics criteria.

Table 1 reports the goodness of fit of the different models estimated, showing the statistic criteria (BIC, CAIC and SBC). Among all the models estimated, the model where gclass = 1 and class = 1 is the one with the highest values on all three statistic criteria, indicating that this model has the lowest goodness of fit. In fact, this is the model for which we assume that there are no hidden latent classes and that the effects of remittance dependency on support for taxation are the same for all the observations in the data. Table A5Footnote ¹² in the appendix shows this model's estimation results and highlights that remittance dependency has a negative and statistically significant effect on support for taxation. This finding is consistent with the evidence in López García and Maydom (Reference López García and Maydom2021, Reference López García and Maydom2023) that shows a negative relationship between receiving remittance and support for taxation. But because this model has the lowest goodness of fit, we confidently suspect that there is unobserved heterogeneity in the data that needs to be addressed to best estimate the effect of dependence on remittance on support for taxation.

Table 1.

Goodness of fit of the multilevel mixture models

Note: This table reports the goodness of fit for the different multilevel mixture models estimated, using different values for the number of clusters. gclass refers to the number of groups at the country level, while class refers to the number of classes at the individual level. Selected model in bold.

In the rest of Table 1, we show the goodness of fit for the models where we relax the hypothesis that there is a single class/type of individuals in the data while accounting for possible unobserved heterogeneity by using a varying combination of gclass and class. It is worth noting that the models where we only account for the heterogeneity at the individual level and leave aside potential heterogeneity at the country level (i.e. when class > 1 and gclass = 1) have a weaker fit than the models where gclass is higher than 1. This indicates that the grouping of the countries also improves the classification at the individual level. We highlight in bold our best model – i.e. the model that records the lowest values on the statistics criteria. As we can observe, the best model is the one that has two distinct latent classes/subtypes at the individual level and four latent groups at the country level. Another important remark is that regardless of the number of groups fixed at the country level, the model with two latent classes at the individual level is superior to any other models with more than two classes. This result confirms Konte's (Reference Konte2016) findings, who also detected two classes in the relationship between receiving remittance and support for democracy in Africa.

Remittance dependency and support for taxation across the two classes

Table 2 reports the results of the selected model, showing the estimated coefficients across the two latent classes identified. In terms of the distribution of the individuals across the two classes, we find that most of the respondents have a higher probability of being in class 1 than in class 2. In fact, 62% of the respondents are in class 1, while 38% of the respondents are in class 2. Regarding the estimated coefficients of the variable of interest, the results show that the effect of remittance dependency on support for taxation depends on the class/subtype of individuals that we consider. In class 1, the estimated coefficient on remittance dependence is positive but not statistically significant, while in class 2, it is negative and statistically significant at the 5% significance level. These findings indicate that in class 1, remittance dependents in that class are as likely as non-dependents to either support or not support taxation. In class 2, however, remittance dependents are more likely to have lower support for taxation than non-remittance dependents. This result corroborates the idea espoused in section 2 that remittance allows individuals to substitute public goods and services with private ones, leading to a reduction in the tax morale of remittance dependents than non-dependents because the latter may not have the opportunity to access private goods and services.

Table 2.

Remittance dependence, support for taxation and public services quality (glass = 4, class = 2)

Notes: This table shows the results of the selected model in bold in Table 1. The dependent variable is the probability to agree that a citizen must pay taxes. Standard errors in brackets. ***p < 0.01, **p < 0.05, *p < 0.1. Year and country fixed effects are included in the model.

Keeping all the additional control variables constant, the estimated odds ratio of a remittance dependent over a non-remittance dependent in class 2 is lower than one and equal to 0.9.Footnote ¹³ This indicates that the likelihood of being tax compliant is reduced by 10% for a remittance dependent that belongs to class 2. In class 1, even though the estimated coefficient on remittance dependent is not statistically significant, the odds ratio is higher than 1 with a value of 1.01, meaning that a remittance dependent that belongs to this class is 1% more likely to be tax compliant than a non-remittance dependent.

Looking at the other control variables, the effect of gender is only statistically significant in class 2 where being a female decreases support for taxation. Education only matters in class 2 where an increase in the level of education is associated with higher support for taxation. Being located in an urban area increases tax compliance in class 1 while it does not have any effect in class 2. Access to information is statistically significant in both classes but with opposite effects. Particularly, whilst it reduces support for taxation in class 1, it enhances it in class 2. Asset-based poverty harms support for taxation in both classes. Regarding the country-level variables, we also find differing effects across the classes. For instance, in class 1, individuals who live in countries with a higher level of income per capita and higher remittance inflows relative to GDP tend to have higher support for taxation. In class 2, however, we obtain opposing results for these variables. Also, while an increase in the weighted average of income taxes in the host countries affect leads to an increase in the support for taxation in class 1, it has no significant effect in class 2.

To date, extant studies are yet to consider the heterogeneous effects of these conventional drivers of tax morale across latent classes/subtypes of individuals. However, the results on the control variables show that the effects of most of these variables vary across latent classes/subtypes of individuals. For instance, consider the effect of gender, which is only statistically significant in class 2 where being a female decreases support for taxation in this class. This result is inconsistent with extant literature suggesting that women are more tax compliant than men, perhaps due to the large differences in honesty between men and women and because women are more prosocial than men (Alm and Torgler, Reference Alm and Torgler2006; Torgler and Valev, Reference Torgler and Valev2010). Yet, it is important to highlight that these studies, unlike in our analysis, do not consider its heterogeneous effect across latent classes/subtypes of individuals. The differences in the results are therefore explained by this source of heterogeneity. Along this line, the result on gender in class 2 is very informative if we consider that females are more negatively affected by poor provision of the public goods and services that we identified as significant determinants of class membership. In fact, the result on gender in class 2 shows that females who are less satisfied with the quality of public goods and services are less likely to support tax than their male counterpart. Conversely, the result on gender in class 1 shows that there is no statistically significant difference in the support of taxation among men and female who feel equally satisfied with the quality of public goods and services.

Remittance dependency and support for taxation: the role of the quality public services

The second main objective of this paper is to explore whether the quality of public services as perceived by the citizens explains class membership. To do so, we endogenize the classification of the individuals across the two classes by allowing the variable quality of public service delivery to be a concomitant variable. The results are shown at the bottom of Table 2. As a retrospection, the concomitant variable quality of public services is a dummy that equals 1 if the respondent thinks that the government handles well at least one of the following: education, health, water and sanitation, and electricity. To estimate the effect of the quality of public services, we use class 1 as the control group. The results show a negative and statistically significant coefficient on the concomitant variable. This means that people who have a positive appraisal of how their governments handle at least one of the public services (health, education, water and electricity) are less likely to be in the second class, where remittance dependency reduces support for taxation. This result is consistent with the psychological tax contract thesis or Wicksell's voluntary exchange theory, positing that taxes are voluntary payments by individuals in exchange for public goods or services. Our result suggests that, other things equal, if individuals have a positive valuation of government provision of public services, their dependence on remittance would have little or no effect on their support for taxation. Keeping everything else constant, the value of the odd ratio of belonging to class 2 over class 1 indicates that having a positive valuation of the quality of public services decreases the probability of belonging to class 2 by 33%.

In Table 3, we further show the descriptive statistics of the remittance dependency, support for tax morale and quality of public services variables across the two identified classes, as well as the descriptive statistics for the individual- and country-level variables. As we can see, there is no large difference between the percentages of respondents who depend on remittance across the two classes (20% in class 1 against 23% in class 2). However, respondents in class 1 have higher support for taxation compared to respondents in class 1. In fact, in class 1, around 95% of the respondents support taxation, while in class 2 this percentage is down to 45%, yielding a difference of almost 50 percentage points. The distribution of how people evaluate the quality of public services also varies across the two classes. In class 1, 83% of the respondents positively assess how their governments handle the quality of public services against 69 in class 2. Regarding the other variables included in the analysis, we do not observe any large differences across the two classes.

Table 3.

Descriptive statistics by class

Next, Table 4 shows in each country the percentage of respondents who have a higher probability of being classified in the second class where remittance dependency decreases support for taxation. The countries are ranked in descending order of the percentage of people that belong to class 2. The table shows that eight out of the 34 countries in our sample have more than 50% of their respondents belonging to class 2. There are seven countries that have all their citizens having a higher probability to be in class 2. These countries are Benin, Cote d'Ivoire, Niger, South Africa, Sudan, Togo and Morocco.

Table 4.

Percentage of respondents by country sorted in class 2

Table 5 shows the distribution of how people evaluate the quality of public services across the two classes. Among people who have a negative appraisal of the quality of public services, 48% are in class 1, and the remaining 52% are in class 2. Sixty-six per cent of those who have a positive appraisal of the quality of public services are in class 1 and 34% in class 2. In sum, this table shows that the two classes identified do not coincide with an ad hoc grouping of the data based on individuals' valuation of the quality of public services. This means that besides the quality of public services, there may be other important factors that also explain class membership. Hence, examining the heterogeneous effect of migrant remittance by grouping the data based on the quality of public services would fail to correct for unobserved heterogeneity because different individuals with similar (different) valuations of the quality of the public services fall into different (similar) classes.

Table 5.

Distribution of the valuation of the quality of public services across the two classes

This table is drawn using our baseline model classification.

Robustness checks

In this section, we subject our analysis to a battery of sensitivity checks to ensure the robustness of our results. Table A6 in the appendix reports the estimation results for the second and third best models. Overall, the results align with the best baseline model where the effect of remittance dependency on support for taxation is class-specific, with a significant negative coefficient in the second class and a non-statistically significant coefficient in the first class. Consistent with the previous results, we also find that a positive assessment of the quality of public services decreases the probability to belong to the second class.

Next, we probe whether some of the public services are more important than others in explaining the classification of the respondents across the two classes. The results of this exercise are reported in Table A7 where we estimate six different models. In model 1, we jointly include all the measures of public services as concomitant variables.Footnote ¹⁴ We find that the effect of remittance dependency on support for taxation is not statistically significant in class one but is negative and statistically significant in the second class. As per the concomitant variables, the estimated coefficients of all the measures of public services are negative and statistically significant. Hence, a positive appraisal of how the government handles education, health, water and sanitation, and electricity significantly decreases the likelihood of being a subtype 2 citizen. In models 2–5, we include the public service variables separately one by one in the estimations. The results are consistent with the previous finding regardless of the public service we consider. Finally, model 6 shows the results where we use an index measure of public services as a concomitant variable. Again, the findings are consistent with those obtained using a binary variable to measure satisfaction with public goods and service provision.

In our baseline model, the remittance dependency variable is dichotomous where 0 is coded for people who do not depend on remittance and 1 for respondents who depend on remittance regardless of the extent to which they depend on remittance. For robustness checks, we recode the variable into a categorical variable that ranges between 0 and 2 where 0 is coded for respondents who do not depend on remittance, 1 for those who depend on remittance but just a little and 2 is allocated for all the other respondents that depend somewhat or a lot on remittance. Hence, a higher value means a higher dependence on remittance. Table 6 reports the results of this exercise. As shown in the table, whereas an increase in remittance dependency does not affect support for taxation in the first class, it does in the second class. In line with the previous findings, the results support that a positive assessment of how the government handles the basic public services significantly decreases the probability to be in class 2 where remittance dependency reduces support for taxation.

Table 6.

Remittance dependence (ordinal variable), support for taxation and public services quality

Notes: The dependent variable is the probability to agree that a citizen must pay taxes. The variable must pay taxes is ordinal ranging from 0 to 2 where a higher value means higher support for taxation. Each model contains the same controls as in the baseline model. Standard errors in brackets. ***p < 0.01, **p < 0.05, *p < 0.1.

Although we focus on the quality of public services which fits into the psychological tax contract or voluntary exchange thesis, we follow up the prior argument about other factors explaining class membership by controlling for additional concomitant variables, including gender, education, employment status, bribe payment, access to information, poverty index and institutional trust (i.e. trust in the president, parliament and local councillors).Footnote ¹⁵ This exercise also enables us to ascertain that the effect of quality public services is not confounded by other factors that could serve as potential concomitant factors. The results are reported in Table A8. Although we observe a drop in the size of the coefficient on quality public services from its previous value of 0.40 in Table 2 to 0.23 in Table A8, the coefficient remains significantly negative at the 1% significance level. Hence, the result confirms our baseline findings where the quality of public services is an important determinant of the classification of the observations across the two classes. In this case, we are confident that the effect of perceived quality of public services as a determinant of class membership where remittance dependency reduces support for taxation is not confounded by other factors such as perceived level of corruption, institutional trust and demographic and socio-economic factors that may serve as potential concomitant factors.

As a final robustness check, we use the surveys in round 4 [2008–2009] and round 6 [2014–2016] that include a question on how often a respondent received migrant remittance in the past 12 months. Unlike round 7, these two rounds do not allow an assessment of whether the respondents depend on remittance. We argue that the number of times a person receives remittances does not necessarily reflect how the person depends on remittance. In fact, some people may receive remittances frequently but with a small amount, while others may receive them less frequently but with a substantial amount that may help households cover major expenses such as education at the beginning of the school year. The findings are consistent with the baseline results (see Table A9). Particularly, we find that in the first class, receiving remittance does not have a significant effect on support for taxation, while in the second class, it reduces the probability of having a positive attitude towards paying taxes. Furthermore, we find negative and significant coefficients on all the public service variables, meaning that a positive assessment of the quality of public services reduces the probability of belonging to the remittance tax-reducing class. It is also worth noting that in Table 5 we have an equal distribution of the respondents across the two classes. We have roughly 50% of the respondents that belong to each of the two identified latent classes.

Finally, we acknowledge that we are not ruling out potential endogeneity that may still exist in our analysis due to omitted variables that could explain, on the one hand, why some people have relatives who migrate and thus receive remittances and, on the other hand, differences in support for taxation. Therefore, we do not interpret our results as causal. However, the mixture of regressions model is implemented to correct or reduce significantly unobserved heterogeneity that could bias the results. Individuals who share similar patterns on how the omitted variables may affect their support for taxation are more likely to fall into the same classes. If significant omitted variables bias still existed within the two classes, the goodness of fit would show that the best model is a model with more than two classes.