1. Microfinance Impact What are we trying to measure? How can we “accurately” evaluate the impact of microfinance? Attempts to measure impact thus far?

2. Microfinance may impact households in various ways via income for example: Income effect: → ↑demand for children → ↑health → ↑children’s education → ↑leisure Substitution effect →↓demand for children (i.e., women’s opportunity cost increases) →↓children’s education →↓leisure

3. Other channels Women’s bargaining power Social capital And, more direct interventions via services added to financial services: Education and training (i.e., Freedom for Hunger, Pro Mujer, BRAC) Health (i.e., “Health Banks”)

4. Let’s Narrow- down our search: impact on income Attributes Measurable: age, education, experience.. Non-measurable: entrepreneurial & organizational skills, valuable networks…. Challenge: disentangling the role of microfinance from measurable and non-measurable attributes Challenge even greater when the decision to participate in a microfinance program depend on those same attributes

5. T2 –T1 compared with C2 – C1: “difference-in-difference approach”

6. Problems: Make sure that control groups are comparable to treatment groups → need to consider who joins the microfinance program that cannot be compared to those who do not Why? Unmeasured attributes (i.e., entrepreneurial ability of those who join) → “selection bias” Potential solution: consider a similar village without microfinance Problem: again, unmeasured attributes of villagers that have not yet self – selected themselves →selection bias Potential solutions: a) measure the effects of microfinance “access” & b) try to identify potential borrowers in the control village

7. Well-known attempts: Bret Coleman (1999) & (2002) on Thailand Tries to address the selection bias by identifying potential borrowers in villages where microfinance is not yet present In particular: He gathers data in 1995 from 14 villages, 8 of which already have a microfinance program in operation, and the other 6 do not but would-be borrowers have already been identified Estimates:

8. Findings: After controlling for selection and program placement: - Impact not significantly different from zero - Some impact for the “wealthier” participants However: Thailand is relatively wealthy, village members have multiple sources of credit ….

9. Karlan (2001) on Peru Comparing “old borrowers” with “new borrowers” using cross – sectional Data →selection bias due to the timing of entry (early entrants may differ from late entrants, i.e., more entrepreneurial, more motivated…..) And Karlan’s experience in FINCA Peru, points out two additional biases, both due to dropouts: 1)Dropouts may be the “failures” → impact is overstated, or vice versa 2) Non random attrition: If dropouts are “failures” → pool of borrowers are richer on average → impact overstated, and vice versa Potential solutions: Hunt down the drop outs which is expensive, or estimate “predictors” which has a problem in that the “reweighing scheme” does not take into account the size of the impact

10. USAID AIMS on India, Peru, and Zimbabwe Use data collected at several points in time allowing for “before versus after” comparisons Control for nonrandom participation and nonrandom placement However, approach subject to biases due to unobservable attributes that change over time Nevertheless: Data collected from a random sample of participant households in several programs that were resurveyed two years later As for the control groups: random sample from nonparticipants ( India and Peru) Or A “random walk” (Zimbabwe)

11. Researchers followed dropouts to avoid attrition biases However: Researchers decided against analyzing difference- in – differences → Biases due to omitted variables that do not change over time In particular, researchers should have estimated: Yijt = Xijt α + Vj β+ Mij γ + Cijt δ + ηijt, (8.2) → Problem: Potential bias due to omission of unobservable variables that do not change over time

12. Addressing the problem: Yijt+1 = Xijt+1 α + Vj β+ Mij γ + Cijt+1 δ + ηijt+1 (8.3) And, subtract (8.2) from (8.3) to obtain: Δ Yij = Δ Xij α + Δ Cij δ + Δ ηij, (8.4) → a consistent estimate of the impact of δ Additional problem: reverse causality

13. BANGLADESH Population: 143.8 million Urban 23.9 million HDI Rank: 138 Adult illiteracy 58.9% Population < $1 36.0 million Largest Microfinance Programs ’98: Grameen, BRAC, RD-12 Serving the landless rural poor

14. Pitt and Khandker (1998) Attempt to measure the impact of microfinance participation, by gender on: - boys’ and girls’ schooling - household expenditures (consumption) - accumulation on non – land assets - women’s and men’s labor supply

15. Cross – Section Data: 1,798 households in 87 villages were surveyed in 1992 905 households were under a microfinance program← treatment 893 households were not ← control Results: Relative to credit provided to men, credit provided to women: (a) ↑Schooling (both boys and girls) (b) ↑Household expenditures (consumption) (c) ↑Non-land assets held by women (d) ↓Labor supply of men and women

16. Basic insight

17. How to address the biases? Find an IV: a variable that explains levels of credit received but has no direct relationship with the outcomes of interest In this case: Schooling, Household Expenditures, Non Land Assets, Labor supply “An eligibility rule: only “functionally landless” households (with < ½ of land) can have access to microfinance” The fact that there ineligible households (260) within villages with programs → there is another “control” group which helps to alleviate the bias Problem

18. An improved estimation strategy Compare: (a)Treatment with ineligible households living in the same village (b)Ineligible with “would be” eligible → households with access to microfinance are doing better than their ineligible neighbors relative to the difference in outcomes between functionally landless households in control villages versus their ineligible neighbors

19. Yij = Xij α + Vj β+ Eij γ + (Tij Eij) δ’ + ηij,(8.5) Disappointing results w/r to impact on household consumption But: Microfinance helps to diversify income streams so that consumption is less variable across seasons Also: Landholdings may not be “exogenous” On the other hand Successful borrowers were buying land → may explain why no impact on household consumption ☺ Moreover, debate over ineligible households that participated (25%). But Pitt- Khandker (1999) acknowledged the problem, made robustness checks and show that their results change very little ☺

20. Note that: Yij = Xij α + Vj β+ Eij γ + Cij δ” + ηij,(8.6) Where: δ” captures credit “access” Now, by expanding the set of instruments to Xij Tij Eij → there are as many instruments as there are X (education….) → δ” takes advantage of variation of how much credit households receive

21. Now, when comparing groups of men with groups of women Pitt-Khandker (1998) most cited result: For every 100 taka lent to a woman consumption ↑ 18 taka For every 100 taka lent to a man consumption ↑11 taka Now, another round of data was collected in 1998 – 1999 And Khandker (2003) look at some trends

22. 20 per cent poverty decline both participants and nonparticipants Pessimists: decline would have happened even without microfinance Optimists: impact of microfinance has had positive spillovers to nonparticipants

23. Khandker’s (2003) econometric estimates show that: Microfinance contributed to roughly ½ of the 20 percentage points decline in poverty For every 100 taka lent to a woman consumption ↑8 taka Ideally, another round of data collection should help Problem: microfinance in Bangladesh has spread far and wide → No more control groups!!!