Does Taxing Mobile Money Harm the Poor: Evidence From the 10 Per Cent Excise Fee Introduction in Kenya

Mobile money has been a key innovation in Africa that has increased the �nancial resilience of vulnerable households by providing an easy way to send and receive remittances. While recent literature has focused on the link between mobile money and its function as a social protection mechanism for the vulnerable population, less is known about the extent to which the costs of using the service affect the transaction behaviour of these people. By exploiting a natural experiment in the form of an excise fee that was imposed on mobile-money transactions in Kenya, this paper estimates the differential effect of a price hike in mobile-money transaction fees on transaction behaviour of households with a daily income below 1.25 USD and households above this threshold. The study �nds that households with an income below 1.25 USD reduced their monthly mobile-money transactions volume by 25 per cent compared to households above this threshold. Additionally, the paper �nds suggestive evidence that the tax also led to a relative reduction in mobile-money remittances received by households below an income of 1.25 USD. The loss in received remittances was not substituted by an increase in remittances sent by cash or in-kind.


Introduction
Mobile money has been a key innovation in Africa in the recent decade; it has been attributed to increased nancial resilience of vulnerable households by providing an easy way to send and receive remittances (Jack and Suri, 2014;Riley, 2018). In Kenya, the rst mobile-money service was launched in 2007 by Safaricom, a subsidiary of Vodafone (Jack and Suri, 2014). By 2013, 74 per cent of the Kenyan population above the age of 15 had a mobile-money account (Van Hove and Dubus, 2019).
Mobile money has reduced transaction costs, increased the geographical reach of nancial transactions and guaranteed a speedy and safe arrival of money to recipients (Aron, 2018). Before the introduction of mobile money, sending money was a dif cult and costly endeavour because fewer than 23 per cent of people had a bank account in Kenya in 2009 (Jack and Suri, 2011). For example, cash was sent through a trusted person or in-kind in the form of goods (Zollmann, 2014). In contrast, mobile money has enabled account holders to send each other digital values of money directly by way of text messages without the need for an internet connection. Apart from sending and receiving money, mobile money can be used to deposit and withdraw money at a mobile-money agent or to directly pay for goods (Van Hove and Dubus, 2019). Mobile-money operators fund their expenses by charging one-off fees for sending and withdrawing money. The charge varies between 0.2 per cent and 20 per cent of the transaction value according to the value and type of transaction (Safaricom, n.d.).
The main drawback of the survey is that only about 1 per cent of the transactions (8334 transactions by 278 households) accounted for mobile-money transactions. Similarly, the survey picked up merely 1369 remittances made by mobile money and received by 183 households, 4332 remittances made by cash and received by 246 households, and 4681 remittances made in-kind and received by 229 households, which can be used for the analysis. It would have been best to analyse only households that were represented in all above categories so that relative comparisons between the categories was fully possible. However, this would have reduced the sample size even further. Given that the difference-in-differences estimator is an unbiased estimator and selection bias into remitting and transacting by mobile money is not the focus of analysis, the paper uses all above observations.

Summary statistics
The above-mentioned observations are aggregated to the volume of monthly mobile-money transactions per household and the volume of monthly remittances received per household, broken down by whether they were received by mobile money, cash or in-kind. Furthermore, observations are also aggregated by the total monthly value of mobile-money transactions per household and the total values for remittances received broken down by mobile money, cash and in-kind transactions. For the main results, the paper focuses on the volume, rather than the total value, of transactions, given that the excise fee was levied on a unit -rather than value -basis. However, results are also discussed on the effect that the excise fee had on total monthly transaction values.  (Mbiti and Weil, 2011). In terms of total transaction values, households made mobile-money transactions worth approximately 75 USD per month. On average they received remittances via mobile money and cash worth 30-35 USD per month and in-kind remittances worth approximately 25 USD per month. It is noteworthy that all aggregate transaction measures are strongly skewed to the left, which is visible by the fact that the maximum value for each of the above variable is very large compared to the mean.
The key explanatory variable is the indicator variable 'below 1.25USD', which equals one if the income of the household calculated as total median income over a calendar month per adult equivalent (OECD method) per day was below 1.25 USD, and zero otherwise. Out of 298 households, 204 (68 per cent) in the survey made less than 1.25 USD per adult per day.  income grew slower than that of households with a higher income, leading to a 10.9 per cent gap between the transaction volumes of the two groups by August 2013. This seems to suggest that the tax adversely affected the mobile-money transaction volume of households with a daily income below 1.25 USD relative to households with a higher income. All visual inspections remain intuition rather than causal evidence.
Figures for monthly remittance volumes broken down by transaction method (i.e. mobile money, cash and in-kind) are provided in the Appendix.

Methodology
This paper applies a difference-in-differences estimation strategy to estimate the causal impact that the 10 per cent excise fee had on the monthly mobile-money transaction volumes of households with an income below 1.25 USD per person per day. [1] The difference-in-differences estimator in this paper captures the difference between the actual monthly mobile-money transaction volume made by households with an income below 1.25 USD and the counterfactual transaction volumes these households would have made had the excise fee affected them as it affected households with an income above 1.25 USD per person per day.
This gives an estimate of how much more households with an income below 1.25 USD were affected by the tax than households above this threshold. This estimator is also referred to as the average treatment effect (ATE) (Angrist and Pischke, 2009: 228).
The main difference-in-differences regression model (referred to as main speci cation hereafter) which this paper estimates by OLS is: Reinvention: an International Journal of Undergraduate Research 15:2 (2022) where mobile money transaction (vol) it is the monthly mobile-money transaction volume of household i in the month of year t (e.g. December 2012); below 1.25USD i is an indicator variable that equals one for households with an average income below 1.25 USD per day and is zero otherwise; Post tax t is an indicator variable equalling one if the data point is for an observation of monthly mobile-money transaction volume made after the excise fee introduction on 1 February 2013 and zero otherwise. The interaction term (below 1.25USD i X Post tax t ) it equals one if the mobile-money transaction volume belongs to a household with an income below 1.25 USD after 1 February 2013. The coef cient estimate on the interaction term shows the differential growth in mobile-money transaction volume for households with an income below 1.25USD after the tax introduction compared to transaction volumes of all other households after the tax introduction.
This coef cient size gives an estimate of the ATE. α i denotes the household-xed effect, γ t denotes the yearmonth xed effect and ϵ it is the error term. In line with Angrist and Pischke (2009) and Bertrand et al. (2004), standard errors are clustered at the household level to account for clustering and serial correlation in the error term.
For the ATE to be a causal effect, it is important that the monthly mobile-money transaction volumes of households with an income below and above 1.25 USD respond as similarly as possible to external factors, except for the policy of interest. In other words, the mobile-money transaction volumes of the households with an income below 1.25 USD and the transaction volumes of households with an income above that threshold should move parallelly to each other up until the excise fee is introduced. This condition is referred to as parallel trends assumption . Otherwise, the post-tax mobile-money transaction volumes of the households above the 1.25 USD threshold do not provide a good counterfactual to estimate the ATE because it will be dif cult to believe that transaction volumes of households below 1.25 USD income would have grown at the same rate as those of households above the threshold had the tax not been implemented (Angrist and Pischke, 2009: 230). however, parallel trends are not visible. Therefore, the paper treats the OLS regression results using remittances as a dependent variable with caution.
To further investigate whether it is reasonable to assume parallel trends, in line with Clair and Cook (2015), a pre-treatment speci cation test and an event study is performed for mobile-money transaction volumes and values in Table 2   In Table 2, the main speci cation (Column 5) is compared against the regression results of the pre-treatment speci cation test. Column (1) runs the main speci cation, but Post tax is an indicator variable that equals one if the transaction was made after 1 December 2012, and is zero otherwise. Column (2) runs the same regression as Column (1), however Post tax is an indicator variable that equals one if the transaction was made after 1 November 2012, and zero otherwise. Furthermore, the sample size is restricted to observations made before the tax introduction between September 2012 and January 2013. By picking a tax introduction date that is prior to the actual date in a sample before the tax was introduced, Columns (1) and (2) function as a pre-treatment speci cation test that can pick up any diverging trends between the two household groups before the excise fee was introduced. The coef cient of interest is the coef cient on the interaction term of Post tax and below 1.25 USD income, which should be as close to zero as possible if poor and wealthy households were trending similarly pre-treatment. Indeed, for both Columns (1) and (2), the interaction term is far closer to zero than in the main speci cation and is statistically insigni cant. In that respect, Table 2 provides further evidence that the volume and values of mobile-money transactions for both treatment groups were trending similarly before the tax introduction.
Further evidence for parallel trends comes from the event study in Figures 5 and 6 (Appendix). The event study plots the coef cient size and its 95 per cent con dence interval for the interaction between the indicator variable 'below 1.25USD' and every year-month observation. In Figure 5, the coef cients are plotted for mobile-money transaction volumes, and in Figure 6, they are plotted for total monthly values.
The vertical line is present at February 2013, the month in which the excise fee was introduced. Both gures show that the coef cient sizes of the interaction terms were close to zero before the tax introduction and dropped to negative levels after the tax was introduced. This further provides evidence in favour of the parallel trends assumption. Furthermore, the drop in the size of the coef cient post-tax introduction indicates that the ATE estimate in the main speci cation is likely to be negative, con rming the economic intuition presented in the introduction section. Table 3 shows the main results of this paper. For all results, standard errors are clustered to the household level. In Column (1), mobile-money transaction volumes is regressed on 'below 1.25USD'. The coef cient on 'below 1.25USD' shows that households below 1.25 USD income make, on average, 2.6 mobile-money transactions per month fewer than households that have a higher income. This result is statistically Reinvention: an International Journal of Undergraduate Research 15:2 (2022) signi cant at the 1 per cent level. In absolute terms, households with an income below 1.25USD made on average 2.8 transactions (see 'Constant'-'below 1.25USD') per month, while households with higher income made on average 5.4 transactions (see 'Constant'), which is a large difference. One explanation for the usage difference is that poor households were less likely to own a mobile phone, which is essential for the usage of mobile money (Jack and Suri, 2011;Van Hove and Dubus, 2019). Mbiti and Weil (2011) estimate that mobile phone owners made three times as many mobile-money transactions as non-owners. Non-owners were far more likely to come from a low-income background (Aker and Mbiti, 2010). In Column (2), 'Post tax' is added to the regression. The coef cient on 'Post tax' shows that mobile-money In Column (3), the interaction term between the 'below 1.25USD' and 'Post tax' is added. The coef cient on the interaction term suggests that post-excise-fee introduction, households with an income below 1.25USD

Main results
The addition of the urban control variable addresses this potential omitted variable problem (Wooldridge, 2016). The size of the coef cient and the statistical signi cance for the interaction term change only slightly, suggesting that the coef cient is not driven by these household-speci c factors.
In Column (5), household-xed effects replace the control variables in Column (4). The reason for the usage of xed effects is that they are able to account for any further time-invariant household-speci c characteristic that may affect the transaction volume of mobile money (Wooldridge, 2009: 413). [2] The size of the coef cient and the statistical signi cance for the interaction term changes only slightly, suggesting that the coef cient is not driven by household-speci c factors. This provides further evidence that the drop in transaction volumes for households with an income below 1.25 USD is due to the tax introduction.
In Column (6), a month-year xed effect is added to the speci cation to account for possible common shocks over time at a national level. The addition does not affect the coef cient size and standard errors of the interaction term, suggesting that the estimate of the interaction term is also not driven by common shocks over time. Given that this speci cation is the most demanding in the sense that it accounts for all household-speci c characteristics and common shocks over time, it is used as the main speci cation to which results from further robustness checks will be compared.
As mentioned in the data section, the transaction volume of mobile money is strongly left-skewed. Because the large values on the right tail can bias the OLS estimator (Wooldridge, 2009: 296), in Column (7) Another concern to the main speci cation is that the composition of household classi ed as 'below 1.25USD' could have changed over the course of the study period. For example, households that were just below the cut-off of 1.25USD on average, and were therefore classi ed as 'below 1.25USD', could have had a higher income post-tax introduction for an idiosyncratic reason and bias the results through their behaviour aligning more with the households with an income above 1.25 USD. To account for that, in Column (8), the main speci cation is run with households that are in the bottom quartile in terms of income calculated as total median income over a calendar month per adult equivalent (OECD method) compared with households in the top quartile. The alternative classi cation has the advantage of excluding the two quartiles in the middle of the income distribution. Given that it is less likely that any household in the top quartile would drop to the bottom quartile within the study period and vice versa, this speci cation is less prone to bias resulting from composition change. The size and signi cance of the interaction term stay relatively similar to the main speci cation, providing further robustness to the result.
Finally, in Column (9), the main speci cation is run with the monthly total mobile-money transaction values per household rather than monthly volumes of mobile-money transactions. While the coef cient on the interaction term is not statistically signi cant, the negative size of the coef cient is in line with the results of the main speci cation.
All in all, the robustness checks provide further evidence that the size and signi cance of the ATE estimate in the main speci cation are consistent. Table 4 turns to address the effect of the excise fee on the remittances. Remittances made by mobile money, cash and in-kind remittances made up 99.3 per cent of total remittances received by households in the survey, thereby capturing nearly the entire ow of remittances. By comparing the effect of the excise fee on remittances sent by mobile money with remittances sent by cash and in-kind, it is possible to grasp both the overall ow of remittances and the substitution between the transaction methods that the excise fee may have induced. It should be noted that results are only brie y discussed as they do not hold much explanatory power due to limited evidence in support for parallel trends. Table 4 reports estimates from running the main speci cation for remittances received by mobile money (Columns (1) and (2)), cash (Columns (3) and (4)) and in-kind (Columns (5) and (6)) separately. In Column (1), (3) and (5), results in terms of volumes -that is, number of remittances received over a month -are reported, while Columns (2), (4) and (6) show results in terms of total value of remittances received over the speci c months.

Financial resilience
In Columns (1) and (2), the estimate of the interaction term is negative and statistically signi cant at a 10 per cent level. This suggests that the excise fee not only affected mobile-money transactions more generally, but also remittances sent via mobile money more speci cally.
In contrast, there is little evidence suggesting that the excise fee impacted the volumes and total values of remittances received as cash and in-kind. There is no indicative evidence that households substituted remittances received by mobile money with cash or in-kind transactions as an alternative transaction method. However, this conclusion should be seen with caution as the sample size is smaller than for the main results and there is less evidence that the parallel trends assumption holds for remittances.

Conclusion
This paper has investigated whether the 10 per cent excise fee on mobile-money transactions levied in Kenya on 1 February 2013 disproportionately affected mobile-money transaction volumes of households below an income threshold of 1.25 USD per person a day. The study is the rst of its kind to utilise a difference-indifferences estimation strategy to investigate the effect of a price hike in transaction fees on transaction behaviour of vulnerable populations in Kenya. The paper provides evidence that mobile-money transaction volumes of households with an income below 1.25 USD dropped by 25 per cent more than for households above the threshold. The study further indicates that these households would have made 25 per cent more transactions over the study period, had the excise fee not been introduced. Furthermore, the paper offers suggestive evidence that the excise fee also reduced remittances received via mobile money for households below an income of 1.25 USD compared to households with an income above this threshold. The reduction in remittances made by mobile money did not lead to a substitution with other transaction methods, indicating a potential slowdown in the growth of remittances ows to vulnerable households with an income below 1.25 USD compared to households with a higher income.
The ndings of the paper highlight the importance of researching the extent to which the price of mobilemoney transaction fees affects the transaction behaviour of vulnerable populations. This is because, for mobile money to function as a social protection mechanism, it must be accessible to the most vulnerable people. From a policy perspective, this study also provokes Kenya and other countries introducing mobilemoney excise fees to rethink the bene ts of such taxes with respect to the effect they might have on the social protection mechanism that the mobile-money service provides.
However, the above ndings should also be interpreted with caution. One concern of the study is whether the ATE can be called a causal effect. As mentioned throughout the paper, evidence for parallel trends in some series was limited. Furthermore, it is noteworthy that the results attained in Kenya in 2013 might not apply to other countries and may also not apply to Kenya today as the economy transforms at a rapid pace.
Future studies could replicate the estimation using a control group of households outside of Kenya. With a larger sample size, it may be possible to analyse the ow of remittances with more statistical power and analyse the characteristics of those who send remittances and how they might change with the onset of the excise fee. Furthermore, the advantage of picking a control group outside of Kenya may be that the total effect of the tax on mobile-money usage could be estimated. Households from the CGAP Financial Diaries 2014-2015 (Anderson and Ahmed, 2016) -which tracked transactions of 86 households in Tanzania, a neighbouring country of Kenya, and 93 households in Mozambique, another East African country -may be a promising dataset to pick the control group from. Acknowledgement I would like to thank Dr Thomas Martin for his constant support and clear advice. I am also grateful to Dr Gianna Boero for valuable guidance during discussion on the research project.    Table 3: Regression results -Impact of the excise fee on volume and total value of monthly mobile-money transactions Table 4: Regression results -Impact of excise fee on remittances received

Notes
[1] The same methodology is applied to the analysis of remittances broken down by transaction method (mobile money, cash and in-kind).
[2] The Hausman test con rms the choice of a xed effects over a random effects model.