Does microfinance reduce rural poverty?
Evidence based on household panel data from northern Ethiopia
Guush Berhane and Cornelis Gardebroek
Agricultural Economics and Rural Policy Group, Department of Social Sciences,
Wageningen University, The Netherlands
Hollandseweg 1
6706 KN, Wageningen, The Netherlands
guush.berhane@wur.nl
koos.gardebroek@wur.nl
Contributed Paper prepared for presentation at the International Association of
Agricultural Economists Conference, Beijing, China, August 16-22, 2009
Copyright 2009 by Guush Berhane and Cornelis Gardebroek. All rights reserved. Readers may
make verbatim copies of this document for non-commercial purposes by any means, provided
that this copyright notice appears on all such copies.
Abstract
This paper evaluates the long-term impact of microfinance credit from the intensity of
participation in borrowing. We use a four-round panel data set on 351 farm households that had
access to microfinance in northern Ethiopia. Over the years 1997-2006, with three-year intervals,
households are observed on key poverty indicators: improvements in annual consumption and
housing improvements. The relatively long duration in the panel enables to measure household
poverty changes between consecutive periods and see the long-run effects of exposure to
microfinance from the intensity of participation borrowing. The fixed-effects model is
innovatively modeled to account for potential selection biases due to both time-invariant and
time-varying unobserved individual household heterogeneities. Results show that microfinance
borrowing indeed causally increased consumption and housing improvements. A more flexible
specification that allows for the number of times the household has been in borrowing also shows
that repeated borrowing is effectively increasing consumption: the longer the borrowing
relationship the larger the effect partly due to lasting credit effects. Impact estimates that do not
account for such dynamic effects may therefore undermine the effect of MFI borrowing.
Key words: Microfinance, treatment effects, trend model, panel data
1 Introduction
The microfinance revolution got considerable momentum around the world in the last two and
half decades. The potentials of microfinance as an effective tool to break the vicious circle of
poverty has been widely voiced. As a result, several microfinance schemes have gone operational
around the world, providing financial access to millions of poor people both in rural and urban
areas. Important questions are however if and to what extent microfinance credit over its long
time existence has contributed in reducing poverty.
Despite efforts to measure this impact, evidence on the poverty reduction effects of long
term microfinance credit remains unclear mainly due to the difficulty of measuring counterfactual
outcomes and the lack of follow up data spanning over sufficiently long periods to measure the
impact. Without experimental designs, evaluations based on simple comparisons between
participants and non-participants are subject to biases from two sources (e.g., Pitt and Khandker,
1998; Ravallion, 2001). The first bias is due to program placement and occurs because
microfinance institutions (MFIs) do not randomize over villages to place programs. They often
choose on village characteristics that may not be observable to the researcher. The second bias is
due to the tendency of individual borrowers to self-select into programs. From the nature of
borrowing it is evident that potential applicants can choose themselves to apply for a loan. When
selection into the program is based on unobservable individual attributes (e.g. entrepreneurial
ability) that simultaneously affect the impact outcome, attributing observed differences to credit
gives biased impact estimates.
But even if pre-designed experimental or quasi-experimental designs that randomize over
potential sources of selection are implemented, estimates based on one-shot observations may fall
short of capturing the complete picture because longer periods may be required before the full
effects from credit are realized (Karlan and Goldberg, 2007). A recent review of the evaluation
literature emphasizes the issue of ‘timing and duration of exposure to programs’ is as important
but relatively less studied than the identification problems that often attract much of researchers’
attention (King and Behrman, 2009). Long period data is, however, costly and largely
unavailable. As a result, most studies so far (e.g., Coleman, 1999; Pitt and Khandker, 1998)
exploited program specific designs and employed innovative quasi-experimental survey methods
to generate control and treatment groups from cross sectional data. A few exceptions are
Khandker (2005), Copestake et al. (2005) and Tedeschi (2008) who used two-period data to
estimate impacts. Long-term panel data, under certain conditions, allows to measure impact from
intensity of participation over time by overcoming selection biases. An attractive feature of panel
data is the possibility to deal with unobserved time-invariant individual and village heterogeneity
using fixed-effects. However, when the selection processes is based on time-varying
unobservables, such as individual motivation which is likely to change over time and borrowing
status, standard panel data methods like fixed-effects and difference-in-difference are biased
(Armendárize de Aghion and Morduch, 2005: 210). Other less frequently used panel data
techniques such as random trend, and flexible random trend models offer alternative approaches
to mitigate this problem by allowing an arbitrary correlation between time-invariant
unobservables as well as individual trends in time-varying unobservables to program
participation (Wooldridge, 2002: 317).
This paper uses unique four-round household survey data covering 1997-2006 to estimate
the impact of participation in microfinance credit on annual household per capita consumption
and housing improvements. The data comes from sixteen villages in northern Ethiopia. We first
investigate the impact of credit using fixed-effects approaches that is standardly applied to
account for time-invariant individual as well as village unobservables. Further, we use variants of
the random trend model due to Heckman and Hotz (1989) that mitigates both time-invariant and
individual trends in time-varying unobservables. We find that program credit has significant
impact on household consumption and housing improvements of participants compared to non-
participants. However, compared to the random trend approach, results from the standard fixed-
effects approach that does not account for individual trends in time-varying unobservables
overestimates credit impact. We also model program credit more flexibly by including the effect
of loan-cycles and individual specific trends and find that credit impact on per capita
consumption increases with frequency of borrowing. The effect of borrowing on the probability
of housing improvement is realized after one-cycle but declines sooner after the third cycle
borrowing. From the flexible approach, we conclude that borrowing effects last longer than one-
period and cumulative effects are best captured the longer the time covered in the analysis.
Besides, while household borrowing effects are multidimensional and cannot be captured by a
single household outcome, we also conclude that effects on household outcomes are not
monotonic over time. Impact estimates that do not account for such dynamic effects may
therefore underestimate the effect of MFI borrowing.
The rest of the paper is organized as follows. Section 2 provides a brief review of the main
approaches followed in the literature on impact assessment. Section 3 describes the nature of the
data and section 4 presents the empirical method used. Section 5 provides the estimation results
and section 6 concludes.
2 A review of microcredit impact studies
This section presents a brief survey of the main methodological approaches of mitigating
selection bias in microfinance impact evaluations.
Measuring the impact of microcredit programs is a challenging task because establishing
‘causality’ between credit effects and changes in the outcome of interest is complicated by the
