Jerom

Intro#

I want to study how mobile money usage by microfinance institutions (MFIs) affects the extent to which they service disadvantaged groups (rural, poor and female borrowers) and the total number of active borrowers. The first three measures are related to a concept called depth of outreach, whereas the fourth is a measure of breadth of outreach. In research focusing on the performance of microfinance institutions, depth and breadth of outreach typically combine into what is called social MFI performance

The motivation for your research is very strong: social performance has been neglected, whereas it is arguably more important than finance performance of MFI. .

As such, especially the poorest and most remote borrowers were hurt, at a time when scholars were already starting to doubt the benefits of microfinance for those most disadvantaged communities.

Can you give some examples/expand upon which scholars (and why) they think the benefits of microfinance for the most disadvantaged is questionable? In general, I think a short paragraph recapitulating the costs/benefits of microfinance would be a good addition to the introduction, which you could add before you start talking about your own contribution on pp. 3.

Central methodological issue in your paper: people who use mobile phones (and hence, are eligible to use mobile money) are already different (e.g. less poor) than those that do not use phones. Hence, there is a selection into the treatment (i.e. mobile money usage). If you have a plausible control group (the selection bias is constant over time), this can still give a good estimate of the effect of mobile money usage on outcomes.

Literature Review#

Definition of microfinance: I would think, limit yourself to what the literature employs, and introduce your own definition/choices in the methods section.

In general, your literature review is full of empirical illustrations (as it should be!), but it leaves me wondering about how researchers manage to identify the impact of e.g. “how M-PESA has increased per capita consumption levels (a critical indicator in economic development studies) and lifted an estimated 194,000 households, or 2% of the total of Kenyan households, out of poverty.”

Hypotheses#

H1: Mobile money is positively related to the relative amount of rural borrowers

H2: Mobile money usage is negatively related to average loan size divided by GNI per capita

The second hypothesis is enunciated in terms of variables, not in terms of concepts. (Why not just: mobile money usage is negatively related to average loan size?)

Main argument: it becomes much easier for banks to screen/monitor because of mobile money usage

H3: Mobile money usage is positively related to the relative amount of female borrowers

H4: Mobile money usage is positively related to the number of active MFI borrowers

I think they are well argued by means of a (sometimes loose) application of theory. It makes sense to keep the explanation at this level of generality, because your research questions are very broad, and you are not testing a specific theory of MFI.

Methodology#

Your data is at the MFI level. So in terms of the methodological problem as I formulated it, MFI’s that choose to offer mobile money services are different from those that do not.

If you estimate model (1), you will most likely not tackle the methodological challenge adequately: there might exist difference between country, and differences over-time.

  • Question: do you have data for the period before the first MFI’s started to introduce mobile money services?

  • My advice: employ the two-way fixed effects estimator (time-dummies and country-dummies) to investigate the influence of $S_{ijt}$ on $SP_{ijt}$. Then, your inferences are within-country, and corrected for shocks at a certain time that affect all countries.

  • There is also the De Chaisemartin-D’Haultfoeille estimator, to take into account possible heterogeneous treatment effects. (Link to paper)

  • You can check for parallel trends in the dependent variable before the introduction of MFI (and graphically show it)

    • No Aschenfelter dip, plausible counterfactual

We test both random and fixed effects to control for this, in consonance with Bibi et al. (2017) using the Hausman test. We include time fixed effects in our models. Furthermore, we have employed a generalized least squares model instead of ordinary least squares, since the first has smaller standard errors and a better linear unbiased estimator than the latter, when correcting for autocorrelation and heteroskedasticity according to panel data estimation literature (Bibi et al., 2017; Lee, 2005).

I don’t really get this argument (yet).