Our estimation strategy follows Angrist and Pischke , who advocate the use of two alternative approaches for dealing with omitted variable bias in panel data. The first employs a lagged dependent variable to account for the possibility that the subsidy was targeted based on pre-treatment trends in party support. The second uses a difference-in-difference model that eliminates possible confounding by time-invariant individual characteristics and controls for a number of time variant factors that might be correlated with the treatment and party preferences. In conjunction, these approaches are useful for bracketing the estimated effect when the potential sources of omitted variable biases are unknown . The two approaches produce estimates of a statistically-significant treatment effect ranging from 6.2% to 7.5%. The first specification, which we refer to as the lagged-dependent variable model, accounts for a large number of potential confounding variables,vertical gardening in greenhouse drawing both on existing studies of electoral preferences in Africa and on the analysis presented in the previous section. We include all variables from the analysis of targeting in the previous section, as well as several other covariates.
We briefly explain the rationale for their inclusion. Unless otherwise specified, we measure these variables using data from the 2008 survey, prior to the distribution of the 2009/10 subsidy. First and foremost, it is important to include the measures of party support from the first survey round. We speculate that those who did not feel close to any party in 2008 might be more responsive to the subsidy, given weaker pre-existing party ties. Likewise, we expect that opposition supporters might be particularly resistant, given their pre-existing attachments. It is also important to account for ethnicity, given that members of some communities might be more likely to become DPP supporters for reasons other than the subsidy program. In particular, members of the president’s own ethnic group might be more likely to follow Mutharika to the DPP than members of other communities. The Chewa, who have long been associated with the Malawi Congress Party , might be particularly disinclined to become DPP supporters. The Yao might also be particularly disinclined to become DPP supporters given the acrimonious split in 2005 between president Mutharika and former president Bakili Muluzi, a Yao. Finally, since the demise of the AFORD party following the 2004 elections , the Tumbuka have been less tied to a particular political party and might therefore be more likely to move toward the DPP. In addition, we include a measure of whether individuals come from minority groups within their villages.
We also include a number of variables found to be associated with subsidy reception based on our analysis of targeting in Figure 1. We account for the possibility that individuals with local political ties might be more likely to become DPP partisans by controlling for membership on district development committees, village development committees, and the chief’s council. We also account for individual economic shocks – loss of crops, loss of income source, and the death or illness of a family member – that might reduce support for the incumbent party, based on research from the United States that shows that voters punish incumbents when their personal well-being is affected by natural disasters and other unforeseen events . We include measures of these shocks in both 2008 and 2009 . We also include variables that track participation in the many other government anti-poverty programs in Malawi to account for possible correlation with subsidy reception. Specifically, we account for participation in the following programs : free food/maize distribution, food-for-work, inputs-for work, scholarships for secondary education; scholarships for tertiary education, and direct cash transfers. We include a measure of subsidy reception in the previous year to account for the possibility that the 2009/10 subsidy may have targeted individuals whose views of Mutharika were in transition due to the prior year’s subsidy.Finally, we account for demographic factors – age, education, farm size, income, and households headed by women – that could affect the strength of pre-existing partisan ties and therefore the likelihood of changing partisan allegiances. In the previous section, our analysis showed minimal evidence of targeting with regard to variables measured in 2008.
However, because multicollinearity between variables could reduce the significance of variables in the targeting model, we test for differences between control and treatment groups on each variable individually and add those that were not included in the analysis of targeting. Following Ho et al. , we test for differences of means and differences in distributions . The balance statistics reveal statistically significant differences on several covariates, indicating the need to control for these factors. We estimate a logit model of DPP support in 2010 that controls for all variables described above and includes village fixed effects to account for possible targeting across villages in our survey area. We cluster standard errors by household. The results, presented in column 1 of Table 3, show an estimated treatment effect that is significant at the p<.05 level. Full logit results are shown in Table A2 in the on-line appendix. As an alternative way to address covariate imbalance, we employ matching before estimating the effect of the subsidy on party preferences using the LDV approach. For this we use the Coarsened Exact Matching approach developed by Iacus, King, and Porro . Matching works by creating matched pairs between those who received the 2009/10 subsidy and those who did not that are similar along observed covariates. Respondents that are not matched are excluded from the analysis, thereby improving overall balance on relevant factors between the treated and untreated groups. The advantage of matching is that one can account for possible confounds through pre-processing rather than controlling for confounds in a parametric model. The parametric approach relies on assumptions about the functional form between confounds and the outcome variables, which if incorrect can bias the estimate. Matching, by contrast, makes no such assumptions . We match on variables that we consider to be most relevant based on theoretical importance and the balance statistics shown in Table A1 in the on-line appendix: region, prior partisanship, membership in the village development committee,greenhouse vertical farming whether respondents experienced an illness or death in the family within the last two years, female-headed household, and age. We limit the matching to this set of variables because including additional variables greatly reduces sample size and because we are able to improve imbalance by matching on this set of variables . We use matching with replacement, which has the advantage of producing better matches and dropping fewer observations than one-to-one matching .We account for remaining imbalance by including all covariates in the estimation of the treatment effect, as recommended by Ho et al. . Column 2 in Table 3 presents the results from a logit model employing the matched data. The model produces a similar estimate of the treatment effect , comparable to the estimate in the pre-matching results, and again the estimated effect is statistically significant. The second estimation strategy uses a difference-in-difference approach designed to account for all time-invariant individual-level factors that could be correlated with both subsidy reception and party preferences. For this, we estimate a pooled OLS model that includes a dummy variable for treatment condition, a dummy for the time period, and the interaction of the two.
Specified in this way, the model is equivalent to a two-way fixed effects model that includes both individual fixed effects and a period dummy. With this specification, the only potential omitted variables of concern are time-varying factors that might be correlated with both treatment status and partisan preferences. As in previous models, we include measures of economic shocks that occurred between the two survey rounds and which might be correlated with both subsidy reception and party preferences. These include measures of whether the respondent’s household experienced a loss of crops or livestock, the loss of an income source, or the death or serious illness of an adult family member. We also include measures of household participation in a host of other government-sponsored anti-poverty programs: food distribution, food-for-work, inputs-for-work, scholarships for secondary and tertiary education, and cash transfers. All time-invariant factors from previous models – such as gender, education, ethnicity, and village – are excluded by design as these factors are accounted for by the specification. The results, shown in column 3 of Table 3, indicate an estimated treatment effect of 6.2% that is again significant at conventional levels .In this section we discuss the limitations of our estimation strategy and relate our findings to relevant literatures. With regard to the methods used to identify the effects of the subsidy on political orientations, the main limitation is that because the program was not randomly distributed, we cannot entirely rule out omitted variable bias with regard to unobserved factors. To address this concern, we use the sensitivity test developed by Rosenbaum to estimate the extent to which our results are potentially driven by one or more omitted factor. This test estimates how large an effect one or more omitted variables would have to be to overturn the estimated treatment effect. For this analysis, we re-estimate the treatment effect using the LDV approach after conducting one-to-one matching on the same set of covariates used above . The results again indicate a statistically significant effect of the subsidy on preferences, with the estimated size of the effect being somewhat larger. The Rosenbaum bounds test shows that one omitted variable would have to increase the likelihood of respondents receiving the subsidy by 16%, after having already accounted for the rich set of covariates we use as controls, in order to overturn the finding. There is no agreed standard for evaluating the results of Rosenbaum bounds. We can, however, compare the results to other known factors that affect assignment. Looking at the marginal effects of factors found to be statistically significant in our analysis of targeting , we find that those who had experienced an illness or death in the household in the last two years were 7.8% less likely to benefit, and households headed by females were 10.4% less likely to receive it. Thus, to overturn the positive finding on the effect of the subsidy on preferences, one or more omitted variables would have to exert a larger effect on assignment than these covariates for which we have measures, after having accounted for these factors and all others included in the estimation. While not impossible, it seems unlikely that our findings are due to omitted variable bias, given that none of the measured covariates exerts an effect of this magnitude. A second limitation relates to the short duration of the period under study. Ideally we would like to know the full extent to which the program affected partisan attachments and electoral behavior. Our data, of course, only allows us to examine the effects of the program across a two-year time span. It is reasonable to believe that the effects we identify likely hold more broadly across the program’s implementation. However, it is also possible that the effects will diminish over time as the program becomes a more routine aspect of Malawian life. It is also possible that the program will have less of an effect during periods when the party system is more stable. It is particularly important to note with regard to the enduring effects of the subsidy that the death in office of president Bingu wa Mutharika in 2012, and the subsequent struggle over succession created a disruption to the political continuity of claiming credit for the AISP , even though the program itself continued. In the 2014 election, two of the main parties promised to continue some version of the AISP going forward. Mutharika’s brother and the DPP presidential candidate Peter Mutharika boasted during the 2014 election campaign that his party had a “good track record” in managing the program and promised not just to continue the AISP, but to abolish the coupon program and expand the subsidy so as to “make the subsidized fertilizer available to every maize subsistence farmer who needs it” . Then-president Joyce Banda and her ruling party proposed during the election campaign that in the next administration the fertilizer program would be scaled back, offering fertilizer loans instead of subsidies . Though covered in major party manifestoes, the AISP was not a central issue in the 2014 election campaign. Instead, voters and politicians alike were pre-occupied with a major corruption scandal implicating the Banda administration .