3. Instrumentos
3.3. La Guitarra y el requinto
In 2009 we collected data from 430 households in 28 communities, across four agro-ecological zones in Kenya. These data are summarized in Table D1 (Appendix D). The survey covers various important aspects of household livelihoods. First, we collected a comprehensive measure of household income, Yi for household i. Household income sums crop income (value of total output less input costs), livestock income (value of livestock and livestock product sales less production costs), salaried income, remittances, business income, and income from casual labour and dividends. To control for differences in family size across households, we computed per capita income levels. Average per capita income in 2009 was KSh 32,564 (USD 434). We use the log of per capita income as our dependent variable.
Next, we collected information on four dimensions of institution quality; vector Ij for village j. To attenuate endogeneity concerns in our income model and to reflect that institutional quality is a characteristic of the community (and not of individual villagers) we consistently measure institutions at the village level — sometimes by aggregating household responses and creating a community average. Vector Ij includes measures of (i) the quality of
local governance, (ii) social capital, (iii) political participation, and (iv) trust in local government institutions.
Information on governance quality was collected through focus group discussions (FGDs). We asked respondents to rate, on a five-point scale, the performance of the sub-chief41 regarding (a) working in the best interest of the community; (b) working in the best interest of the people in general; and (c) the level of confidence in the sub-chief.
Social capital is a fuzzy term, and typically captures trust, norms and networks that reduce transaction cost and enable collective action (Bowles and Gintis, 2002; Knack and Keefer, 1997). We quantified two dimensions of social capital; community-level trust and cooperation. To measure trust, we followed the World Value Survey approach, and asked respondents to rate their level of trust in household members, extended family, and community members. Trust was measured on a ten-point scale, and scores for the sub-indices were added, normalized between zero and one, and aggregated at the community level. On average, the trust level was rather high with a score of 0.7. To measure the level of cooperation, respondents were asked whether most people in their village would help them in case of need. On average, 83% of the households provided a positive response.
Political participation may matter for development as well. Arguably political participation fosters accountability of village leaders, and facilitates the flow of information from the people to the leaders. We measure participation as the average proportion of adults that voted in the 2005 referendum and 2007 general elections. Because these two measures are highly correlated (ρ=0.91), we use participation in the 2005 referendum in our models.
Our final institutional dimension trust in local government institutions — is based on the stated level of confidence in the justice system, police, and local political authorities.
These were measured at the household level on a ten-point scale. The scores for the sub-indices were added and normalized between zero and one, and aggregated at the community level. The average score across communities equals 0.42, indicating modest trust.
To what extent do these different institutional variables indeed capture different dimensions of the institutional framework? To probe into this issue we computed the correlation between these variables, and found that they were not highly correlated. The correlation coefficient ranges from ρ=-0.004 to ρ=-0.495, suggesting that each variable picks up something that is ―distinct.‖ The greatest correlation exists between social capital (trust level) and political participation (ρ=-0.495 and significantly correlated at the 1% level). This
41 A chief is a government representative at the lowest administrative unit — the location. A sub-location is made up of a cluster of villages or communities.
implies multicollinearity may emerge when both proxies are included in one regression model.
Next, we turn to the set of variables inspired by the alternative hypothesis that geographical factors explain income (differences). Our vector with geography variables (Gi) contains data from primary and secondary sources. We collected GPS coordinates for each household in 2009, and combined these coordinates with existing data for a number of geography variables. Following the literature, we include altitude, latitude, rainfall, temperature, indicators of soil quality (different landscape attributes and soil types), normalized differential vegetation index (NDVI), malaria risk, distance to markets, distance to major towns, distance to good road, and population density (from the 1999 census). The rainfall data were collected by the National Weather Service Climate Prediction Centre as part of a Famine Early Warning System.42 We calculated average main season rainfall for each household as well as the variance of rainfall. However, when controlling for rainfall levels, the variance never enters significantly (so this variable is dropped in what follows).
For latitude, we use the absolute value measure of latitude (distance from the equator). Soil data are based on the Exploratory Soil Map and Agro-climatic Map of Kenya.43 We used the new malaria risk data based on the work by Noor et al. (2009), which is exogenous to per capita incomes as it is based on climatic conditions.
We also collected information on a number of household and community controls.
Household controls, vector Xi, are the age, gender and education of the household head, and per capita land holdings (in acres). The vector of community controls (Cj) includes population density, ethnicity, availability of clean water, distance to health facilities, availability of transport, having a primary school within the community, number of years of mobile phone coverage, religious composition, perceived security changes (subjective assessment of whether security had improved, stayed the same or worsened over the last 12 years), and the presence of illegal ―ngangs‖ (mainly Mungiki ngangs; travelling gangs of robbers).44
The religion variables warrant extra discussion. Below we explore whether the religious composition of a village affects the quality of local institutions. There are various mechanisms via which religion may affect local institutions. For example, as religion may
42 This data has been compiled by the Tegemeo Institute of Agricultural Policy and Development.
43 For details on the actual soil types and the associated information see the documentation ―Exploratory Soil Map and Agro-climatic Zone Map of Kenya, 1980‖.
44 Mungiki is a political-religious group and a banned criminal ngang in Kenya. The group operates mainly in Central, Nairobi and parts of the Rift Valley, and involved in criminal activities including illegal taxation of businesses, murders among others.
prescribe certain behaviors and impose norms of conduct, it helps villagers to form consistent expectations about the responses of their peers — facilitating cooperation and trust. Religious values or traditions can also influence the views of individuals about community members, and indeed people in general. Moreover, participation in religious groups (direct interaction) builds social capital. Religion can, therefore, via several channels enhance intra-group trust, and facilitate collective action (see also Welch et al., 2004).45 Religion can also potentially influence civic and political participation via certain social norms. Civic participation is facilitated by social capital, as reflected in social networks characterized by norms of reciprocity and trust. In addition, churches are avenues for providing political information. In Kenya for instance, some churches are more vocal in political matters than others. Religious institutions, therefore, hold the potential to reconnect people to politics, and provide political information and participatory opportunities to their followers (Greenberg, 2000). They can make demands upon the state, and influence political outcomes so that its potential effect on income may be indirect. We will empirically explore these issues below.
We calculate the proportion of the adult population in our sample belonging to different religious groups: Catholics, Protestants, Pentecost, and other religious affiliation (Muslim and others; the omitted category in our regression models). These data were collected at the household-level and aggregated at the community level. On average, there were more protestants than other religious groups. The average proportion of Protestants was 0.53, followed by Catholics (0.26), and Pentecostals (0.17). A small minority were Muslims or members of other religious groups.
In addition to controlling for the religious composition of villages, we also tried to control for the ethnic composition. Ethnolinguistic diversity (fractionalization or polarization) may directly hinder economic development and indirectly shape the underlying institutions and policies that influence economic development. Evidence suggests ethnically fragmented societies tend to suffer from reduced social cohesion and a smaller supply of public goods.46 In rural Kenya, however, villages tend to be ethnically homogenous, even if the ethnic identity of villagers varies from one region to the next. Hence we do not include an ethnic fractionalization variable. We did set out to control for the ethnic identity of the villages. Our study sites spread across four major ethnic communities: Kikuyu (Central region), Luhya and
45 In Italy, Putnam (1993) attributes the prevailing lack of trust toward others in the South to the strong Catholic tradition, which emphasizes the vertical bond with the Church and tends to undermine the horizontal bond with fellow citizens.
46 The incumbent leader may implement policies aimed at expropriating resources from ethnic losers, restricting the rights of other groups, and discouraging the growth of industries or sectors that might threaten the position of the ruling group (e.g., Alesina et al., 1999; Easterly and Levine, 1997)
Luo (Western region), and Kamba (Eastern region). The Kikuyu, Luyha, and Luo communities comprised some 28% each of our sample, while the rest (16%) were Kamba.
When including ethnicity dummies to represent these groups, we found they were correlated with some geographical variables. Hence we do not include them explicitly in most analysis (but including them does not affect our main results; details available on request).
Finally, we turn to our identification strategy, which is inspired by standard macro level income regressions. Indeed, we follow Pande and Udry (2005) who recommend to exploit synergies between research on specific institutions (based on micro data) while addressing the ―big macro questions‖ of growth. We first explore correlations between geography, institutions and income using simple OLS models, specified as:
ij community-level institutional variables, and Gij is a vector of geography variables. Finally, Xij and Cj are vectors of household and community controls, respectively. We enter Ij and Gij
separately and in combination. Virtually all geography variables are exogenous to household income. This is perhaps not true for the institutional variables. While income is measured at the household level and institutions are measured at the community level — attenuating concerns about reverse causality — we cannot rule out endogeneity concerns due to measurement error or omitted variables. For that reason we augment our OLS analysis with a series of 2SLS models. For the IV model, our first stage regression is specified as:
ij influence income other than through its impact on institutions (i.e., our religion variables are correlated with local institutional quality, but not with the error term of the income model).
We explore the empirical basis for these assumptions below. The second stage of the IV model is akin to model (1), but replaces our vector of institutional variables Ij by their predicted values, Ij* (as predicted in accordance with the model in (2)).
As mentioned above, when testing for multicollinearity we found high correlations between some community control variables, notably population density, ethnicity, water source, distance to health facilities, availability of transport, and geography variables (e.g.
malaria endemicity). We only included community controls that are not highly correlated
with the geography variables. Since some of our soil variables were highly correlated, we estimated models with soil type and landscape attributes separately. Based on the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) we decided to retain models with landscape attributes. However, our results are robust to including all these additional controls.
Finally, and as a robustness analysis, we have also done a factor analysis to reduce the dimensionality of our institutional vector. We selected three factors for inclusion in a series of estimates.