Katerina Tertytchnaya et al. Our main dependent variable asks respondents how RESULTS much they generally trust the president.Relying on this item,we construct a variable,presidential trust, Models 1 through 3 in Table 2 explore the relationship which ranges from 1 to 4,with higher values denot- between changes in remittances and changes in incum- ing greater trust in the president.In line with other bent support.The dependent variable in Models 1 to 3 work.we use trust in the incumbent,here the president. is the difference in levels of trust in the President across as a proxy of incumbent approval (e.g.Williams 1985, survey waves,while our measure of changes in remit- Ahmed 2017).Based on this measure we construct a tances differs across models.Model 1 shows the results variable Change in Trust in the President()by sub- for changes in the amount of remitted income,Model tracting the trust in the president at wave t-1 from trust 2 the results for changes in the frequency with which op//s in the president at wave t.This variable ranges from a respondents receive remittances,and Model 3 the re- minimum value of-3 and a maximum value of 3.Over- sults for changes in our index.All results are based all,we expect to find a positive effect for a change in the on a panel data model that accounts for repeated ob- amount and the frequency of remittances received or in servations of individuals and includes both household the Remittance Index,and a negative effect for reduc. and survey wave fixed effects.The results show that an tions in remittances on change in trust in the president. increase in remittances coincides with an increase in The analyses also include a set of control variables trust in the president,while a decrease in remittances including gender,age,education,marital status,ethnic decreases trust in the president.For example,when a ity,intention to migrate,life satisfaction and attitude respondent experiences a change in remitted income to risk.Controlling for ethnicity in the context of Kyr- from the minimum to maximum amount.trust in the gyzstan is important as ethnicity captures one of the president increases by 2.8 points on a seven-point scale most salient political divides in the country.It serves These results are robust against different ways to mea- somewhat as a proxy for partisanship in a context with sure changes in remitted income.different estimation a traditionally weak party system,and where ethnic and methods and different model specifications(see Tables clan divisions are very important(Fumagalli 2016).We C.4-C.7 in the SI). also control for other sources of income or wealth,such Still we might be concerned that changes in the Kyr- as being in paid employment,household income and a gyz economy could be driving both the changes in re- wealth index.Summary statistics of all variables used mittances and incumbent approval.To deal with this in the analysis of the Lik data and question wordings concern,we employ an instrumental variables (IV)re- are provided in Tables A.2-A.3 of the SI. gression.As an instrument,we use annual change in Our dataset pools observations for respondents unemployment in Russia,the major destination coun- nested in households across the four different waves. try for Kyrgyz migrants,interacted with (or weighted To deal with both the temporal and nested data.we em by)the share of women in each household.The share ploy two different estimation strategies.First,we per- of female household members is a key household leve form a panel data generalized least squares(GLS)es- characteristic correlated with the receipt and amount timation with household and panel wave fixed effects of remittances,but not with incumbent approval.The as well as random effects varying across individuals instrument we employ incorporates the idea,common Second,we use a hierarchical linear model (HLM)to in previous studies,that growth in the migrant host 。101g deal with the fact that individuals are nested in house- country is likely to be a key driver of remittances(e.g. holds and waves (Snijders and Bosker 2012).We esti- Barajas et al.2009;Singer,2012).The results,which mate a model consisting of two levels,one is the respon- are reported in Table D.1 of the SI,show that remit- dent level and the other is a household*wave level.The tance effects on approval remain robust,even when we HLM results are reported in Tables C.4,C.8 and C.11 in instrument remittance amounts using annual changes the SI.In both estimations,we regress changes in trust in unemployment in Russia weighted by the share of in the president on changes in remittances.To check women in the household. the robustness of our results,we also employ nearest Before turning to the analysis of the mechanism driv- neighbor matching(NN matching)based on our di- ing the connection between changes in remittances and chotomous measure of a decline in remittances.These approval,we investigate how a decline in remittances results are reported in Tables C.6,C.7,C.10,C.13 and stacks up against other shocks that individuals face.We C.14 of the SI,and show that the effect of remittances do so by comparing the effect of experiencing a reduc- remain robust even when we match respondents on a tion in remittances on incumbent approval against the series of covariates that could predict their susceptibil- effects of other adverse income shocks that individuals ity to experiencing a decline in remittances in the first may experience,such as the effect of having suffered a place. loss in agricultural income or being affected by land- slides.Respondents in the LiK surveys were asked if they had suffered agricultural loss,for example through diseases in crop or livestock,or were affected by a land- L slide.The variables Reduction in Remittances,Agricul- tural Loss and Affected by Landslides take on a value of 5 In the LiTS II survey of 2010,measures of trust are highly cor- 1 when respondents experienced this kind of shock and related with evaluations of government performance.The bivariate correlation between government approval and trust in the national a value of 0 if not.While 28.2 percent of respondents government is 0.5,statistically significant at the 0.01 level. who received remittances across the four waves stated 766Katerina Tertytchnaya et al. Our main dependent variable asks respondents how much they generally trust the president. Relying on this item, we construct a variable, presidential trust, which ranges from 1 to 4, with higher values denot￾ing greater trust in the president. In line with other work, we use trust in the incumbent, here the president, as a proxy of incumbent approval (e.g. Williams 1985, Ahmed 2017).5 Based on this measure we construct a variable Change in Trust in the President t-(t-1) by sub￾tracting the trust in the president at wave t-1 from trust in the president at wave t. This variable ranges from a minimum value of -3 and a maximum value of 3. Over￾all, we expect to find a positive effect for a change in the amount and the frequency of remittances received or in the Remittance Index, and a negative effect for reduc￾tions in remittances on change in trust in the president. The analyses also include a set of control variables, including gender, age, education, marital status, ethnic￾ity, intention to migrate, life satisfaction and attitude to risk. Controlling for ethnicity in the context of Kyr￾gyzstan is important as ethnicity captures one of the most salient political divides in the country. It serves somewhat as a proxy for partisanship in a context with a traditionally weak party system, and where ethnic and clan divisions are very important (Fumagalli 2016).We also control for other sources of income or wealth, such as being in paid employment, household income and a wealth index. Summary statistics of all variables used in the analysis of the LiK data and question wordings are provided in Tables A.2-A.3 of the SI. Our dataset pools observations for respondents nested in households across the four different waves. To deal with both the temporal and nested data, we em￾ploy two different estimation strategies. First, we per￾form a panel data generalized least squares (GLS) es￾timation with household and panel wave fixed effects as well as random effects varying across individuals. Second, we use a hierarchical linear model (HLM) to deal with the fact that individuals are nested in house￾holds and waves (Snijders and Bosker 2012). We esti￾mate a model consisting of two levels, one is the respon￾dent level and the other is a household∗wave level. The HLM results are reported in Tables C.4, C.8 and C.11 in the SI. In both estimations, we regress changes in trust in the president on changes in remittances. To check the robustness of our results, we also employ nearest neighbor matching (NN matching) based on our di￾chotomous measure of a decline in remittances. These results are reported in Tables C.6, C.7, C.10, C.13 and C.14 of the SI, and show that the effect of remittances remain robust even when we match respondents on a series of covariates that could predict their susceptibil￾ity to experiencing a decline in remittances in the first place. 5 In the LiTS II survey of 2010, measures of trust are highly cor￾related with evaluations of government performance. The bivariate correlation between government approval and trust in the national government is 0.5, statistically significant at the 0.01 level. RESULTS Models 1 through 3 in Table 2 explore the relationship between changes in remittances and changes in incum￾bent support. The dependent variable in Models 1 to 3 is the difference in levels of trust in the President across survey waves, while our measure of changes in remit￾tances differs across models. Model 1 shows the results for changes in the amount of remitted income, Model 2 the results for changes in the frequency with which respondents receive remittances, and Model 3 the re￾sults for changes in our index. All results are based on a panel data model that accounts for repeated ob￾servations of individuals and includes both household and survey wave fixed effects. The results show that an increase in remittances coincides with an increase in trust in the president, while a decrease in remittances decreases trust in the president. For example, when a respondent experiences a change in remitted income from the minimum to maximum amount, trust in the president increases by 2.8 points on a seven-point scale. These results are robust against different ways to mea￾sure changes in remitted income, different estimation methods and different model specifications (see Tables C.4-C.7 in the SI). Still we might be concerned that changes in the Kyr￾gyz economy could be driving both the changes in re￾mittances and incumbent approval. To deal with this concern, we employ an instrumental variables (IV) re￾gression. As an instrument, we use annual change in unemployment in Russia, the major destination coun￾try for Kyrgyz migrants, interacted with (or weighted by) the share of women in each household. The share of female household members is a key household level characteristic correlated with the receipt and amount of remittances, but not with incumbent approval. The instrument we employ incorporates the idea, common in previous studies, that growth in the migrant host country is likely to be a key driver of remittances (e.g. Barajas et al. 2009; Singer, 2012). The results, which are reported in Table D.1 of the SI, show that remit￾tance effects on approval remain robust, even when we instrument remittance amounts using annual changes in unemployment in Russia weighted by the share of women in the household. Before turning to the analysis of the mechanism driv￾ing the connection between changes in remittances and approval, we investigate how a decline in remittances stacks up against other shocks that individuals face.We do so by comparing the effect of experiencing a reduc￾tion in remittances on incumbent approval against the effects of other adverse income shocks that individuals may experience, such as the effect of having suffered a loss in agricultural income or being affected by land￾slides. Respondents in the LiK surveys were asked if they had suffered agricultural loss, for example through diseases in crop or livestock, or were affected by a land￾slide. The variables Reduction in Remittances, Agricul￾tural Loss and Affected by Landslidestake on a value of 1 when respondents experienced this kind of shock and a value of 0 if not. While 28.2 percent of respondents who received remittances across the four waves stated 766 Downloaded from https://www.cambridge.org/core. Shanghai JiaoTong University, on 26 Oct 2018 at 03:53:04, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0003055418000485