This group of works contains studies that have been designed as difference in difference or panel data. They are bundled together because in all cases except Morduch (1998)
Participants or borrowers A
Eligible but non participants B
Non eligible C
Eligible D
Non eligible E
ProgramVillages ControlVillages
Figure 2 Household types by eligibility and villages. Source: Armendariz de Aghion and Morduch, 2010.
they were implemented in two periods of time. In this case, DID and panel FE should produce identical estimates.
Both approaches suppose that there are time invariant unobservables that are correlated with the covariates and that can bias the impact effect. In the case of microfinance, it is usual to assume that these unmeasured characteristics are entrepreneurship, negotiation skills, perseverance and others that would make borrowers systematically different from non-borrowers. Assuming that these characteristics are constant over time, by differencing the variables they are swept out and it is possible to get rid of the bias source.
In the simplest case of two groups (treated-control) and two time periods, the usual approach is to run a baseline survey for the whole sample before the intervention and a second survey after the intervention. The difference in gains from the two groups would be the impact of the program. The time invariant differences between the groups are swept out.
The first study quoted in this group is Morduch (1998), already briefly outlined above. It takes advantage of the particular design of the Bangladeshi database and tackles an approach that is similar to the Difference in Difference (DID) technique but without a time component, which is essential in DID. In his own words “a clean estimate of the average impact of access may be more useful than a biased estimate of the impact of participation” (Armendariz de Aghion and Morduch, 2010 p. 285). This is, actually, the impact of the Intention To Treat (ITT) as in Banerjee et al. (2009). He does not find an increase of household consumption due to access to microfinance but a lower disparity in consumption along seasons. Thus, he concludes that microfinance has a positive effect on household consumption smoothing.
Khandker (2005) uses a second round of the same Bangladesh survey that took place in 1998-99. This second round covered not only the villages of the first round but also included villages from three additional thanas. He estimates fixed effects panel data with the households that could be traced to the second round. In this second survey there were no control villages as the program had extended to all villages. His main findings for 1998-99 are that each additional 100 taka borrowed by females increase annual total expenditures by 20.5 takas, attributing 16.3 to past borrowing (1991-92
survey) and only 4.2 to present borrowing. This might be a sign of decreasing marginal returns to borrowing. Male marginal returns are statistically insignificant.
Given the long period in between, however, the assumption of time invariant unobservables is questionable. In order to avoid measurement errors and reverse causality problems, Khandker (2005) also attempts a panel IV approach. The instrument is the eligibility rule. The Wu-Hausman test rejects the endogeneity of the credit variable, so the DID model is adopted.
Roodman and Morduch (2009) challenges P&K, Morduch (1998) and Khandker (2005). It casts doubts over all of them. They try to replicate these studies and conclude that reverse or omitted-variable causation are leading to wrong estimates. They also contend that the instrumentation strategy is failing and that there is a substantial change in the different subsamples in the credit consumption relationship as well as in borrower’s sex and this can explain the differences in impact by gender. Analysis of these three papers leads them to conclude that in social sciences where the endogeneity problem is normally present, RCTs can provide a simpler and neater approach as long as they can get rid of this bias.
Two additional sources can be included in this group, although they are not connected to the Bangladeshi studies. They both use DID. Copestake et al. (2005), mixes qualitative and quantitative approaches. The dataset is gathered in Peru and the microfinance institution is Promuc. The sample is rather smaller than those of other studies mentioned so far. As in Coleman (2006), they first study the impact for the whole sample and then they split it into different groups. In this case it is not rank-and- file members versus committee members but the sample is divided into households below and above the median income level. They find an overall significant impact of microfinance on income. Also, they conclude that the impact of participation for wealthier individuals is around 80% higher compared with poorer individuals. Nonetheless, they clearly state that some selection bias issues might not have been properly addressed. Finally, this splitting and analysing by subsamples might bring about some truncation issues that could have been avoided applying a quantile regression.
The last piece of research in this group is a DID found in Bruhn and Love (2009). The study takes advantage of the simultaneous opening of all the branches of Banco Azteca,
using the premises of a well established chain of domestic appliances, Grupo Electra. The opening took place in all those municipalities in which Gupo Electra had a retail shop, as the branches were opened within the shops themselves. They obtained the dataset from the Mexican National Employment Survey (ENE). It contains information collected before and after the appearance of the bank branches and therefore is the only DID approach with baseline information described so far. The other particularity is that the impact is not studied at individual, household or microenterprise level but at municipality level.
It controls for initial differences between municipalities with and without an Azteca branch. The main findings are 7.6% increase in the proportion of informal business, although the increase is only statistically significant for male-owned enterprises. On the contrary, the increase in waged employees is significant only in the case of females. Overall, the increase in total employment, counting new businesses and new waged employees is about 1.4%. The figure is statistically significant but the outcome is not encouraging. It, however, reinforces other studies that find a positive impact and underpins the argument that access to credit in the informal sector can have a positive effect on the welfare of the individuals.