Does Microfinance Work as a Recovery Tool After ... - Econ-Jobs.com

17 downloads 4515 Views 515KB Size Report
We evaluate the effectiveness of microfinance as a recovery tool after tsunami by testing the .... (2008), using data from rural Bangladesh, finds that BRAC's.
World Development Vol. 39, No. 6, pp. 898–912, 2011 Ó 2011 Elsevier Ltd. All rights reserved 0305-750X/$ - see front matter www.elsevier.com/locate/worlddev

doi:10.1016/j.worlddev.2009.10.020

Does Microfinance Work as a Recovery Tool After Disasters? Evidence from the 2004 Tsunami LEONARDO BECCHETTI Universita´ di Tor Vergata, Roma, Italy

and STEFANO CASTRIOTA * University of Perugia, Italy Summary. — We evaluate the effectiveness of microfinance as a recovery tool after tsunami by testing the impact of an equity injection from foreign donors which recapitalizes a Sri Lankan MFI and allows it to refinance borrowers seriously damaged by the calamity. We find that loans obtained from the MFI after the catastrophic event have a positive and significant effect on the change in real income and in weekly worked hours, and that the impact on performance variables is significantly stronger for damaged than non-damaged borrowers. Results hold after controlling for selection effects and for heterogeneity in both the timing of the intervention and the characteristics of treatment and control samples. Ó 2011 Elsevier Ltd. All rights reserved. Key words — Asia, Sri Lanka, tsunami, natural catastrophe, crisis recovery, microfinance

1. INTRODUCTION

generally not the case for developing countries (Fafchamps & Gubert, 2007) and after natural catastrophes (Skees, Varangis, Larson, & Siegel, 2002). This is why an important recovery tool is represented by loans provided by traditional banks and MFIs. With respect to donations and charity, credit has the advantage that it does not affect income in the mere short term and that, if the loan is paid back, it perpetuates the financial flow and satisfies new investment opportunities. It is important to notice that natural hazards tend to be accompanied by liquidity squeezes since, in spite of a boom in credit demand due to the need to restore destroyed and damaged buildings and economic activities, banks are often forced to reduce the supply of loans because of the sudden worsening in the quality of their assets. For this reason recapitalizing MFIs after calamities may be crucial. The recent historical evidence documents that microcredit programs contributed to reduce the vulnerability of the poor by assisting them to re-build assets and by providing emergency assistance after natural disasters. There are many examples of MFIs active in post-conflict and post-disaster countries whose loans have been claimed to be a useful recovery tool. The widespread diffusion of MFIs in Uganda and Bosnia, among other countries, in the post-war periods is a clear example. Another case refers to Thailand in the post-tsunami period where, as reported in the USAID’s web site, 2 “a USAID-financed microfinance fund in a tiny, tsunami-ravaged village has proven so

One of the main obstacles to economic development for the poor is the lack of access to traditional credit markets due to the scarce availability of collateral resources and the high screening, monitoring and enforcement costs incurred by financial intermediaries when lending to them (Hermes & Lensink, 2007). Microfinance tries to circumvent these problems with a mix of solutions. Assessing the impact of microfinance programs is not easy. First, empirical studies may incur in self-selection bias since those who borrow may have better unobservable traits than the control sample, mainly as a consequence of the same bank screening process. Second, undocumented village-level differences could influence the demand for/use of credit, thereby leaving space for placement bias (Hulme & Mosley, 1996). Third, comparing old and new clients might be subject to attrition bias, with survived old clients being of “better type” than new ones, as underlined by Karlan (2001). Fourth, data collection is difficult and costly, especially when repeated across time. Nevertheless, a number of studies have found positive effects of microcredit programs on clients’ income, women empowerment, contraceptive use and nutrition (for a survey, see Goldberg (2005) and Armendariz de Aghion and Morduch (2005)). The focus of our research is on the relatively less explored topic of the relevance of microcredit as a recovery tool after a natural catastrophe such as the 2004 Asian Tsunami. 1 After a catastrophe occurs, the first financial source used to recover is self-insurance (savings and accumulated assets). Unfortunately, a relevant share of the poorest do not have enough savings after a natural disaster when the latter destroys their few available resources. For this reason, especially in low-income rural areas, it is common practice to form risk-sharing networks. The problem with them is that they work at best when members’ incomes are uncorrelated or negatively correlated,

* We thank Robert Cull, Rafael Di Tella, Rita Ferrer-i-Carbonell, Robert Lensink, Enrico Longoni, Craig McIntosh, and Bruce Wydick for their useful comments and suggestions. We wish to thank David Berno, Salvatore Morelli, and Francesca Palermo for valuable research assistance and Laura Foschi, Francesca Lo Re, Marco Santori, and all Etimos team for their support. The usual disclaimer applies. Final revision accepted: October 30, 2009. 898

EVIDENCE FROM THE 2004 TSUNAMI

successful that it has been accepted as an associate member of the Credit Union League of Thailand.” 3 Apart from this anecdotal information, rigorous and sound empirical evidence on the usefulness of microcredit as recovery tool after natural catastrophes is still scarce. Khandker (2007) studies the coping strategies adopted by rural households during the 1998 flood in Bangladesh and assesses their impact on welfare. The author concludes that “the presence of microcredit programs increased the amount of borrowing coping. Household borrowing also increased household welfare by raising both consumption and asset holding” (p. 179). Hoque (2008), using data from rural Bangladesh, finds that BRAC’s micro-credit program may increase participating households’ abilities to cope with economic hardships following floods and other natural disasters. Nevertheless, the author concludes that “further research to much more systematic information needs to be conducted about micro-credit program before conclusive results can be reached.” The purpose of our research is to contribute to fill in this gap. Our specific focus is the evaluation of the effects of a donors’ intervention aimed at recapitalizing a Sri Lankan MFI 4 which reported post-tsunami certified losses for 24.4% of its credit portfolio due to the serious damages to the business of a large part of its borrowers. This implies that, in case of adoption of a standard capital adequacy rule of, say, 10%, the MFI’s losses would amount to 250% of its capital. The recapitalization program was aimed to provide the MFI with the capital necessary to grant new loans to the damaged borrowers enabling them to start back their activity and proved to be much cheaper for the donors 5 than more classical aid schemes. Our goal is to evaluate whether such intervention acted as an effective liquidity injection for the damaged borrowers enabling them to restore their economic activity and to significantly improve their economic conditions with respect to the immediate posttsunami levels. The advantage of our framework is that the tsunami event in fact creates two “randomly selected” groups: a treatment group (borrowers directly hit by the tsunami shock) and a control group (borrowers from the same MFI not affected by it). Our study can, therefore, be assimilated to a (quasi) natural experiment in which the exogenous shock makes a difference between the two above mentioned groups which are ex ante not significantly different in terms of borrower’s quality or seniority characteristics, overcoming the standard selection bias problem in microfinance impact analyses. We exploit this unique opportunity by focusing on the effects of post-tsunami MFI refinancing. More specifically, we evaluate its impact (measured by the size of the loans obtained after the tsunami scaled on the borrower’s post-tsunami pre-refinancing monthly income) on two performance variables (percent change in income and in worked hours after refinancing) by carefully taking into account problems related to heterogeneity of loan timing and endogenous size and timing of the loan (see Section 3). The paper is divided into five sections including introduction and conclusions. Section 2 provides details on our survey. Section 3 describes the dataset, explains the methodology adopted and provides summary statistics. Section 4 presents the estimation approach and comments descriptive and econometric evidence. Section 5 concludes. 2. AGRO MICRO FINANCE AND THE SURVEY Agromart Foundation is a Sri Lankan NGO founded in 1989 to carry out grassroots work with a large number of com-

899

munities in Sri Lanka. The Head Office is located in Colombo with nine other provincial offices in Uva, the Southern, North Western, and Eastern provinces. The core of its mission is strengthening the competencies of its members through participatory trainings. In order to achieve this goal Agromart Foundation created self-help groups in rural areas through the provision of technical assistance and education. In 1994 the Foundation broadened its activity by working as a microcredit institution for its clients, but in 2000 it decided to fund Agro Micro Finance (AMF) and to delegate this task to it. Even if the respective fields of action remain separated, the links between the two organizations are strong. Agro Micro Finance lends only to members of community based organizations which received for at least six months self-employment, entrepreneur development and literacy trainings from the Agromart Foundation. Seventy-two percent of AMF borrowers are women. In March 2005 the MFI’s loan portfolio was of 295.000€. After the tsunami, Agro Micro Finance and Agromart Foundation certified direct and indirect losses on 620 clients in the district of Galle, Matara, and Hambantota. The estimated corresponding financial needs to cover such losses amounted to almost 24.4% of the MFI loan portfolio at the tsunami date. This evidence documents the importance of foreign intervention to avoid the MFI financial distress and the consequent restriction to credit access for the MFI borrowers. The liquidity provided by foreign institutions allowed AMF to avoid credit restrictions and the risk of default. Support to AMF refinancing needs came from USAID, UNDP, and an Italian MFI (Etimos). To evaluate the impact of post-tsunami MFI refinancing we randomly selected from the bank records a sample of 305 borrowers: 200 with at least one type of damage (which we define as treatment group) and 105 with no damages (which we define as control group). We created a treatment group larger in size since part of our analysis specifically focuses on subsamples of the treatment group which differ by damage typologies in addressing some of the above mentioned research questions. A questionnaire was administered to both groups in April 2007. The interviews were conducted face to face by one of the authors of the paper with the help of two more researchers and three translators with economic degree (the questionnaire was translated in Sinhalese). Some borrowers were interviewed at their homes, some during the monthly society meetings and the remaining during some extra meeting arranged by AMF for this purpose. Since the tsunami was unexpected we could not organize a panel survey with observations repeated in time, before and after tsunami, and, therefore, had to rely on a retrospective panel data approach (see McIntosh, Villaran, & Wydick, in press) specifically tailored for our case. In April 2007 respondents were asked to declare the current and remember the past levels of memorable variables by making reference to four different periods. We selected periods easy to remember due to the occurrence of memorable events. The four time windows we consider are: (P1) the six month interval before the first microfinance loan ever obtained; (P2) the period going from the first microfinance loan to the tsunami date (December 26, 2004); (P3) the period between the tsunami date and the first microfinance loan after tsunami; (P4) the period from the first microfinance loan after tsunami to the survey date (April 2007, see Figure 1). Our approach is not free from critical points which we conveniently address. A first methodological issue in this analysis is the heterogeneity in time windows of the four different periods since only two time points (the tsunami date, December 26, 2004, and

900

WORLD DEVELOPMENT

Figure 1. Time schedule in our tsunami study (P1 = six month interval before first AMF financing; P2 = period ranging from first AMF financing to the tsunami date; P3 = period ranging from the tsunami date to AMF refinancing; P4 = period ranging from the AMF refinancing to the survey date; dotted lines indicate non-overlapping window borders, continuous lines coincident window borders).

the month in which the survey was administered, April 2007) are common to every interviewed borrower. Consequently, only the first time interval (six months before the first MFI financing) is fixed in length, even though not coincident for all respondents. It is, however, important to remark that our focus on the effect of post-tsunami refinancing (periods three and four) limits the heterogeneity of time windows to the refinancing date 6 since the other two extremes (beginning of period three and end of period four) are common to all respondents. Nevertheless, there remains a potential problem of heterogeneity of time windows which will be controlled for in two ways. 7 First, we add the length of the windows in months as controls in our regressions. Second, and more important, we perform robustness checks of post-tsunami refinancing effects by reducing the sample to individuals within a one year difference in monthly length (restriction “only P4 < 12 months” in Tables 7a, 7b, 8, 9, 10). This implies that we finally handle a level of heterogeneity not different from that of analyses based on yearly data and, by knowing the monthly timing of the refinancing event and controlling for it with the window length, we are actually doing more than what is usually done in such studies. A second methodological issue is the quality of information on lagged variables in two respects: the quality of respondents’ memories (events which are more distant in time, such as the first loan ever obtained from AMF, are more difficult to remember) and the absence of interview biases. From this point of view it is important to underline that part of the information we use does not come from survey data since the timing and amounts of all loans released by AMF have been obtained from the MFI official records. Furthermore, we take into account the limits, remarked by McIntosh et al. (in press), of memories on indicators such as income by looking at an alternative performance indicator such as the more easily memorable number of weekly worked hours. Our assumption is that it is easier to remember with precision the length of the working day (approximated by the discrete number of worked hours) than one’s own past income. We preliminarily verify that there is no significant change in productivity (real income per hour worked) across the two relevant sample periods and

within subsample groups. The change in worked hours, therefore, appears to be a good (and more memorable) proxy of real income. 3. SUMMARY STATISTICS With survey data collected on the field we obtain information on the respondents’ socio-demographic characteristics, on hours worked and on a series of economic indicators. Table 1a and b describe the socio-demographic characteristics and economic variables used in our study. Table 2a reports summary statistics for selected socio-demographic characteristics of AMF sample borrowers. We can see in this table that slightly less than half of the sample has house and economic activity within one kilometer from the coast. Most clients work at home or very close to it, to save money and time (only a minority of families hold a motorbike, almost nobody has a car). Eighty-five percent of the sample is composed of women. Most of the clients are married and aged around 40 or 50, with complete primary or incomplete secondary education. Twenty-three percent of them (the sum of the men and the widowed women) claim to be the head of the household. Over the four time windows, most borrowers (94%) are self-employed while only 2% are unemployed. A large number of them are involved in trade (46%) and manufacturing (39%), with a significant share (21%) working in the agricultural sector. 8 The average family size is 4.6, with 2.3 children currently living at home. Table 2b reports summary statistics for the economic variables. Real and equivalent 9 income exhibits a high variability, while equivalent income in PPP is 5.26$ per day, well above the symbolic 1 or 2$ threshold. On average, over the four analyzed periods, 13% of clients had problems to provide daily meals to their family, with a peak of 26% in the third period after the tsunami shock. Most of the families either are unable to save money or save very little. However, a lot of money is often invested to start or improve a business activity, thus the actual savings given by investments plus net savings should be higher.

Table 1. (a) Socio-demographic characteristics and (b) economic variables (a)

(b) Dummy variable (DV) = 1 if the house is on the coast DV = 1 if the business activity is on the coast DV = 1 if the province is Galle DV = 1 if the province is Matara DV = 1 if the province is Hambantota DV = 1 if the gender is female Age of the respondent in years DV = 1 if the marital status is single DV = 1 if the marital status is married DV = 1 if the marital status is widow DV = 1 if the marital status is divorced DV = 1 if the marital status is separated DV = 1 if the marital status is cohabitant DV = 1 if head of the household DV = 1 if the education level is incomplete primary DV = 1 if the education level is complete primary DV = 1 if the education level is higher than primary DV = 1 if the employment status is employed full time DV = 1 if the employment status is employed part time DV = 1 if the employment status is self-employed DV = 1 if the employment status is unemployed DV = 1 if the employment status is student DV = 1 if the employment status retired DV = 1 if the sector of activity is agriculture DV = 1 if the sector of activity is fishery DV = 1 if the sector of activity is manufacturing DV = 1 if the sector of activity is trade DV = 1 if the sector of activity is something else

Number of children Real income Equivalent real income Equivalent real income PPP Standard of living Problems with meals Savings Van Tractor Motorbike Bicycle Hours worked Damages to family Damages to the house Damages to office buildings Damages to working tools Damages to raw materials Damages to the market Number of damages Savings withdrawn Remittances Subsidies Donations and grants Other charity/fundings Etimos Loan to income ratio

People in the house

Number of people living in the house

Length

Seniority

Number of children currently living in the house Real monthly income in April 2007 Sri Lankan Rps. Equivalent real monthly income in April 2007 Sri Lankan Rps. (see footnote 10) Real equivalent monthly income in April 2007 PPP USD (see footnote 10) Standard of living in terms of consumption goods DV = 1 if the respondent declares to have problems in providing daily meals Amount of savings from 0 (not at all) to 4 (very much) DV = 1 if the respondent owns a van DV = 1 if the respondent owns a tractor DV = 1 if the respondent owns a motorbike DV = 1 if the respondent owns a bicycle Number of hours worked per week DV = 1 if the respondent reported damages to the family DV = 1 if the respondent reported damages to the house DV = 1 if the respondent reported damages to the office buildings DV = 1 if the respondent reported damages to the working tools DV = 1 if the respondent reported damages to the raw materials DV = 1 if the respondent reported damages to the market Number of types of damage from 0 to 6 DV = 1 if the tsunami forced the respondent to use the savings after the tsunami DV = 1 if the respondent received remittances from foreign countries DV = 1 if the respondent received governmental subsidies DV = 1 if the respondent received donations and grants DV = 1 if the respondent received other forms of charity DV = 1 if the respondent was refinanced by AMF through ETIMOS funds Amount of the first loan obtained after the tsunami divided by post-tsunami prerefinancing monthly income Number of MFI loans ever obtained from the MFI since the beginning of the relationship Length of the fourth time window (after the tsunami) in months

EVIDENCE FROM THE 2004 TSUNAMI

House on coast Business on coast Galle Matara Hambantota Female Age Single Married Widowed Divorced Separate Cohabitant Head of family No education Primary education Secondary education Full time Part time Self-employed Unemployed Student Retired Agriculture Fishery Manufacturing Treade Other sector

901

902

WORLD DEVELOPMENT Table 2. (a) Socio-demographic characteristics and (b) economic variables

a

(a)

Obs.

Mean

Std. Dev.

Min

Max

House on coast Business on coast Galle Matara Hambantota Female Age Single Married Widowed Divorced Separate Cohabitant Head of family No education Primary education Secondary education Full time Part time Self-employed Unemployed Student Retired Agriculture Fishery Manufacturing Trade Other sector People in the house

1,220 1,220 1,220 1,220 1,220 1,220 1,160 1,220 1,220 1,220 1,220 1,220 1,220 1,220 1,220 1,220 1,220 1,219 1,219 1,219 1,219 1,219 1,219 1,219 1,219 1,219 1,219 1,218 1,216

0.44 0.46 0.31 0.52 0.17 0.85 48.48 0.08 0.82 0.09 0.00 0.01 0.00 0.23 0.35 0.48 0.16 0.02 0.02 0.94 0.02 0.00 0.00 0.21 0.02 0.39 0.46 0.09 4.61

0.50 0.50 0.46 0.50 0.37 0.35 10.15 0.27 0.39 0.28 0.06 0.08 0.00 0.42 0.48 0.50 0.37 0.13 0.14 0.23 0.15 0.06 0.03 0.41 0.15 0.49 0.50 0.28 1.57

0 0 0 0 0 0 23 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1

1 1 1 1 1 1 73 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 12

(b)

Obs.

Mean

Std. Dev.

Min

Max

Number of children Real income Equivalent real income Equivalent real income PPP Standard of living Problems with meals Savings Van Tractor Motorbike Bicycle Hours worked Damages to family Damages to the house Damages to office buildings Damages to working tools Damages to raw materials Damages to the market Number of damages Savings withdrawn Remittances Subsidies Donations and grants Other charity/fundings Etimos Loan to income ratio Senioritya Lengtha

1,216 1,122 1,118 1,108 1,219 1,219 1,200 1,219 1,219 1,219 1,214 1,220 305 305 305 305 305 305 305 300 305 304 305 305 305 261 289 285

2.38 19,277 8,351 5.26 2.27 0.13 0.86 0.05 0.03 0.21 0.51 49.94 0.04 0.19 0.25 0.27 0.32 0.49 1.56 0.32 0.06 0.32 0.27 0.03 0.34 8.94 3.44 10.51

1.48 13,540 6,067 4.45 0.96 0.34 1.01 0.21 0.17 0.40 0.50 27.30 0.19 0.39 0.43 0.45 0.47 0.50 1.67 0.47 0.23 0.47 0.44 0.18 0.48 11.17 1.25 5.93

0 1080 327 0.21 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 .12 2 1

7 120,000 47,817 39.02 4 1 4 1 1 1 1 100 1 1 1 1 1 1 6 1 1 1 1 1 1 89.51 8 26

Data from bank records, apart from income which comes from the survey.

Many clients reported damages from Tsunami to raw materials (32%), tools (27%), buildings (25%), and house (19%). Only 4% had family members injured or dead in the catastrophe. Half of the people in the sample declared that their business was indirectly damaged by the tsunami because of the worsened macroeconomic situation. One-third of the sample was forced to use their savings immediately after the tsunami to buy food or repair the damages. After the disaster the government, international organizations, and NGOs tried to help the populations by providing food, raw materials, medicines, money, etc. In our random sample, only looking at the third window, 6% of people could rely on remittances from relatives abroad, 32% on governmental subsidies (in most cases a fourmonth check of Rs. 5,000 per month to buy food), 27% on donations from international organizations, foreign governments, and NGOs and 3% on other forms of charity or funding. The last four variables in Table 2b are from bank records, apart from the denominator of the ratio loans (t)/income (t  1) which comes from the survey, and refer to the fourth time window only, which is the period of interest for our econometric analysis. MFI borrowers received after the tsunami loans equivalent to the income gained in almost nine months, with an average number of 3.44 loans obtained since the beginning of the relationship with Agro Microfinance. 4. EMPIRICAL ANALYSIS Our empirical analysis focuses on the effects of MFI refinancing on damaged borrowers’ recovery with preliminary

descriptive evidence on the damage and recovery in the first place and, after it, a series of econometric tests on the impact of refinancing which try to solve the usual methodological problems arising in this kind of impact analyses. (a) Descriptive analysis of the fall and recovery Even though the focus of the paper is on the role of microfinance as a recovery tool after calamities and the econometric analysis will be exclusively run on the period going from the tsunami pre-refinancing to the post refinancing period, we extend the descriptive analysis to the situation before the tsunami. We believe that this is a necessary premise to provide a description of the background in which the tsunami shock occurs and to give some insights on the counterfactual (what would have happened to the damaged if they were not) in two ways showing how (i) in the “normal” (shock free) periods AMF microfinance borrowers tend to improve their economic conditions; (ii) damaged borrowers interrupt the trend when moving from the second to the third (post-tsunami pre refinancing) period and, after post-tsunami loan refinancing, have to catch up versus non-damaged who did not suffer the shock. We start by looking at average changes in the selected variables period by period over the four windows (see Table 3a–c). Among the several possible variables, we consider monthly real household income (net of loan repayments), daily equivalent income in PPP, and the number of hours worked. An indirect measure of poverty is created with answers to the question: “did you have problems in providing daily meals to your family?.” With this respect an important consistency check in our data is the strict correspondence between

EVIDENCE FROM THE 2004 TSUNAMI

903

Table 3. (a) Changes in mean of selected indicators, full sample. (b) Changes in mean of selected indicators, non-damaged respondents only. (c) Changes in mean of selected indicators, damaged respondents only Variable (a) Change in real income Change in Eq. Real Income PPP Change in hours worked Change in problems with meals (b) Change in real income Change in Eq. real income PPP Change in hours worked Change in problems with meals (c) Change in real income Change in Eq. real income PPP Change in hours worked Change in problems with meals

P2–P1

P3–P2

P4–P3

P4–P2

4273.118 (7.18) 1.3792 (7.49) 7.006557 (6.50) 0.0263158 (1.89)

5556.833 (7.04) 1.675444 (7.07) 9.203279 (7.13) 0.1836066 (7.49)

4441.441 (7.22) 1.409149 (6.62) 11.10164 (8.65) 0.1704918 (7.26)

1066.862 (1.51) 0.2225023 (0.93) 1.898361 (2.36) 0.0131148 (0.78)

3972.47 (5.33) 1.257967 (5.46) 8.342857 (4.85) 0.0380952 (1.65)

1255.463 (1.55) 0.4536367 (1.80) 1.933333 (1.48) 0.0380952 (1.42)

2908.87 (3.00) 0.8078155 (3.43) 0.7904762 (0.58) 0.047619 (1.68)

1795.856 (1.70) 0.4296515 (1.38) 1.142857 (0.70) 0.0095238 (0.38)

4431.439 (5.40) 1.442423 (5.69) 6.305 (4.59) 0.0201005 (1.16)

8037.377 (7.24) 2.381209 (7.16) 13.02 (7.28) 0.26 (7.78)

5333.118 (6.79) 1.759619 (5.77) 16.515 (9.71) 0.235 (7.42)

2529.337 (2.79) 0.5557458 (1.72) 3.495 (4.06) 0.025 (1.15)

Note: Change in monthly real income is expressed in 2007 Sri Lankan Rupees, while the change in equivalent real income in PPP in US Dollars. The change in weekly worked hours is the difference between hours of current and previous periods. The variable capturing the change in food problems is equal to one if in the previous period providing daily meals to family members was not a problem while in the current period it was. T-statistics are in parentheses. Variable legend is in Table 1a and b.

equivalent household income and declared existence of problems in providing daily meals to the family. The average equivalent income in PPP for those declaring such problems is 2.74$ against a value of 5.58$ for those not declaring them (the difference in means being highly significant). The first approach we use here is simply a test on the significance of the change in the indicator from a period to the next one (t-statistics are reported in brackets). In Table 3a we observe, for the overall sample, a (slight) amelioration in several economic indicators in the second period (P2), a fall after the tsunami (P3) and a recovery in the last time window (P4). This pattern is obviously stronger if we look at the subsample of damaged individuals (Table 3c), while the tsunami effect almost disappears in the non-damaged subsample (Table 3b) since economic indicators of the latter stop growing but do not display any reduction in the third period. 10 The strong impact of the tsunami shock on the full sample is evident from the descriptive findings presented in Table 3a. Real average income falls by Rs. 5,556 (a 25% reduction with respect to sample average), daily equivalent income in PPP by 1.67$ (a 27% reduction) and the probability of having problems in providing daily meals rises by 18% (from 8% in P2 to a worrying 26% in P3). Worked hours fall by nine units. The fourth column of Table 3a provides a test on the statistical significance of the changes from the second to the fourth period for the full sample. This can be considered as a test to verify whether the economic indicators in the fourth period (P4) recovered to pre-tsunami levels (P2). The real income in P4 is not statistically different from that in P2 and the standard

of living in terms of consumption goods improves. Non-damaged people significantly improve their economic situation, while their hours worked remain unchanged. Damaged individuals increase the number of hours worked but do not fully recover the pre-tsunami purchasing power. It is also important from a descriptive point of view to examine not just average values but also the dynamics of the entire distribution of selected well being indicators across the four time intervals. Figure 2 clearly documents the downward shift of the cumulative distribution of real household income and weekly worked hours in the third period for the whole sample and the subsample of the most damaged people. The comparison of cumulative distributions in the four periods shows that the shock and the catching up effect do not act only on the mean of the subsample distribution of the selected economic indicators, but almost on any point of the distribution, with special reference to its low tail. In fact, it is clear that the poorest are both the most damaged and those registering the most significant recovery across the tsunami and refinancing periods, respectively. In the third time interval all points of the cumulative distribution function are not above those of the second time interval so that first order stochastic dominance of the former on the latter is evident just from this picture. By focusing on the subsample of people with at least three damages the fall in economic indicators after the tsunami and the more pronounced drop in the left part of the distribution is even more evident. Table 4 tests period by period the difference in the mean of each variable between damaged and non-damaged

904

WORLD DEVELOPMENT

Figure 2. Cumulative distributions of real income and hours worked. Note: Ventiles of monthly real income and weekly worked hours are calculated with respect to the pre-microfinance (P1), the pre-tsunami (P2), the tsunami (P3) and the refinancing (P4) time windows.

Table 4. Difference in mean of selected indicators between non-damaged and damaged respondents Variable Real income Eq. real income PPP Hours worked Problems with meals

P1

P2

P3

P4

1037.945 (0.65) 0.5749012 (1.08) 0.9738095 (0.29) 0.0007657 (0.02)

566.9911 (0.30) 0.4018226 (0.65) 1.064048 (0.36) 0.0183333 (0.56)

7139.124 (5.03) 2.192877 (5.05) 12.15071 (3.50) 0.2402381 (4.68)

4905.54 (2.93) 1.326885 (2.29) 3.57381 (1.15) 0.0528571 (1.52)

Note: Monthly real income is expressed in 2007 Sri Lankan Rupees, while the change in equivalent real income in PPP in US Dollars. Weekly hours worked are average number of worked hours per week. The variable capturing the food problems is equal to one if the respondent had problems in providing daily meals. T-statistics are in parentheses. Variable legend is in Table 1a and b.

respondents. Our two main questions here are whether the two groups were significantly different before and after the tsunami and whether there has been a complete recovery in the fourth period. The homogeneity between damaged and non-damaged borrowers, all clients of the same MFI, is confirmed by the fact that all the indicators were not significantly different among the two groups at 5% level before the tsunami (P1 and P2).

In the third period all the means become strongly different, while in the fourth period there is a partial convergence of damaged people to the levels of non-damaged ones. In fact, the number of hours worked and the probability of having problems in providing daily meals of the two subsamples are not different at 5% level, while the difference still exists for the remaining variables. However, it must be underlined that

EVIDENCE FROM THE 2004 TSUNAMI

905

Table 5. Percent change in real household income and weekly worked hours for individuals whose loan to income ratio was above and below the median Variable

a

Full sample

Damaged borrowers only

Non-damaged borrowers only

Below the mediana

Above the mediana

Below the mediana

Above the mediana

Below the mediana

Above the mediana

(1) Percent change in real income Mean Min. Conf. Int. Max. Conf. Int. Obs.

21.8 13.4 30.1 131

86.40 56.30 117.20 130

30.97 16.72 45.22 83

101.96 63.33 140.59 82

11.44 0.66 22.22 48

29.58 7.90 51.26 48

(2) Absolute change in worked hours Mean Min. Conf. Int. Max. Conf. Int. Obs.

4.9 2.5 7.2 130

15.6 11.7 19.5 130

9.3 5.0 13.1 81

21.6 15.4 29.3 81

2.96 2.79 8.71 49

7.26 8.27 22.78 49

Average (first post tsunami) loan to (post-tsunami pre-refinancing monthly) income ratio below the median in period P4.

the gap has enormously reduced: the size and the statistical significance of the difference in the variables are much lower in P4 than in P3 documenting a partial, but still incomplete, catch up. (b) The post-tsunami refinancing effect of microfinance In order to evaluate the effectiveness of microfinance as a recovery tool after tsunami we test its impact measured by the loan to income ratio, that is, the amount of the first loan obtained after the tsunami divided by the post-tsunami prerefinancing monthly income on performance. As performance indicators we consider two variables: (i) the change in real income from the post-tsunami pre refinancing to the post refinancing period (from P3 to P4); and (ii) the change in the number of hours worked in the same interval. Since we are aware of the criticism on the use of past income in retrospective panel data (McIntosh et al. (in press)) we introduce the second indicator in addition to income. Our assumption is that it is much easier to remember the length of the working day (approximated by the discrete number of worked hours) than one’s own past income. 11 Before using the change in hours worked we test whether there are significant reductions in productivity (real income per hour worked) across the sample period. We indeed find that differences are not significant also between and within treatment and control group in post tsunami pre and post refinancing periods. We are, therefore, encouraged to consider changes in hours worked after post tsunami refinancing as good proxies of changes in income. (i) Descriptive evidence on the effect of MFI refinancing We start in Table 5 by evaluating from a descriptive point of view the correlation between the loan to income ratio and our performance variables. The percent change in income from P3 to P4 for those with a loan to income ratio below the sample median is 21.8% against 86.4 for those above it (and goes up to 120% for borrowers with a loan size to income ratio above the 75th percentile). In the same direction, the (absolute) change in weekly worked hours for those receiving after tsunami a loan to income ratio below the sample median is 4.9, against 15.6 for the complementary sample, the difference being significant at 1% (and up to 18.8 if we consider borrowers with a loan size to income ratio above the 75th percentile). Since we are arguing that the MFI loan is a useful recovery instrument for tsunami victims, as a robustness check we limit

the sample to individuals who reported damages. This helps us also to control for heterogeneity problems between treatment and control samples which may be related to the size and the use of the loan. Our descriptive results on the two performance variables of interest remain strong and significant. By calculating the new median loan to income ratio for the subsample of treatment (damaged) group we find that the change in real income is around 110% for the individuals whose loan to size ratio is above the median against around 31% for those below the median. Similarly, the change in (weekly) worked hours is 21.6 for people above and 9.3 for people below the median. The effect of MFI refinancing described here with simple differences in means requires to be controlled for composition effects, heterogeneity between treatment and control groups and in time windows, and various selection effects. This is what we will try to do in the following sections. (ii) Econometric evidence: the simultaneous pattern of effects When evaluating econometrically the effect of the loan to income ratio on performance indicators we need to correct for concurring factors affecting it. Intuitively, we want to avoid our findings to be driven by spurious effects related to a third unobserved factor, correlated with both the loan to income ratio and the performance indicator, which may generate a spurious positive correlation between the two. A strong suspect may be borrower’s seniority which affects both timing and size of the loan and may proxy unobservable borrower’s quality. A second concern is about selection effects on the timing and relative size of the loan. With regard to the first, we find that timing is affected by the number of borrower’s previous loans with the MFI and the sum of damages. As for the second, we observe that the loan to income ratio is affected by three factors: the number of damages, the post-tsunami pre refinancing level of income (negatively), and the number of previous microfinance loans. 12 This implies that, on the one side, the lender looks at reputation and previous record of the borrower while, on the other side, he is not insensitive to the tsunami emergency and tries to counteract two of its effects (low income and number of damages). 13 To sum up, the most damaged and those with longer seniority receive first and more with respect to their income. However, as we will see in the next paragraph, while the loan to income ratio affects performance, timing is not significant. Moreover, both timing and loan to income ratio are not related to proxies of past performance or positively related to productivity of the pre-tsunami period (the relative size is

906

WORLD DEVELOPMENT Table 6. SUR estimates

Variable

(i) Coef.

Eqn. (1)

(ii) z

Coef.

Dep. Var.: log diff. of real income

Galle Matara Agriculture Fishery Manufacture Age Female Primary education Secondary education Number of children Number of damages Remittances Subsidies Donations and grants Other charity Ln (Loan to income ratio) Length Seniority Constant

0.16 0.06 0.12 0.05 0.03 0.00 0.02 0.09 0.02 0.00 0.03 0.13 0.10 0.29 0.04 0.25 0.01 0.15 0.09

Eqn. (2)

Dep. Var.: log of loan to income ratio

Female Number of damages Real income (t  1) Seniority Constant

0.44 0.08 0.84 0.77 8.65

Eqn. (3)

z

Dep. Var.: log. diff. of worked hours

1.77 0.72 1.59 0.28 0.44 0.31 0.26 1.43 0.26 0.16 1.27 0.64 0.87 1.09 0.64 7.97 0.94 2.78 0.45

0.22 0.13 0.11 0.22 0.06 0.00 0.23 0.00 0.59 0.02 0.15 0.19 0.06 0.57 0.23 0.17 0.01 0.13 0.72

1.00 0.62 0.62 0.49 0.40 0.46 1.14 0.02 2.72 0.43 2.90 0.39 0.19 0.89 1.47 2.25 0.56 0.95 1.41

Dep. Var.: log of loan to income ratio

3.34 2.39 12.21 10.93 12.47

0.45 0.07 0.79 0.78 8.18

Dep. Var.: time window length

3.35 2.02 10.90 10.67 11.18

Dep. Var.: time window length

Number of damages Number of loans Constant

0.51 3.72 10.80

2.40 7.64 11.97

0.49 3.69 10.83

2.12 7.31 11.48

Equation

Obs.

R2

Obs.

R2

Eqn. (1) Eqn. (2) Eqn. (3) Residuals correlation

247 247 247 Correl.

0.22 0.55 0.20

230 230 230 Correl.

0.13 0.53 0.19

Correlation (e1, e2) Correlation (e1, e3) Correlation (e1, e3)

0.53** 0.28** 0.84**

0.19** 0.06** 0.84**

Note: SUR regressions make use of heteroskedasticity-robust standard errors. ** Significant at 99%. Variable legend: see Table 1a and b.

indeed weakly negatively correlated with such productivity variable). 14 In order to take into account this complex pattern of links among variables we estimate a simultaneous (SUR) 15 three equation system which takes into account the impact of determinants of loan to income ratio, timing of the refinancing and our performance variable and the possible correlation of residuals among the three single equations. We, therefore, estimate the following system: lnðY i;t Þ  lnðY i;t1 Þ ¼ a0 þ a1 lnðLi;t =I i;t1 Þ þ a2 Lengthi;t1 X þ a3j X ð1Þji;t þ e1i;t ð1:1Þ j

lnðL=IÞi;t ¼ b0 þ

X k

b1k X ð2Þk;it þ e2i;t

ð1:2Þ

Lengthi;t ¼ c0 þ

X

c1l X ð3Þl;it þ e3i;t

ð1:3Þ

l

where Y is the real household income (Table 6, columns 1 and 2) or, alternatively, weekly worked hours (Table 6, columns 3 and 4) and the dependent variable in 1.1 is the log difference of Y, L is the size of the first AMF loan obtained after the tsunami, I is the real household income, Length is the length of the post-tsunami pre-refinancing time window (the distance in month from the tsunami date to the first MFI refinancing) and X(1), X(2), and X(3) are the other regressors of the three equations, respectively, as described in Table 6 (and detailed in the legend of Table 1a and b). A first concern is whether a strong correlation among regressors may distort the magnitude of our coefficients. We, however, find that there are no serious multicollinearity problems in our estimate. The highest VIF 16 factor among

EVIDENCE FROM THE 2004 TSUNAMI

907

Table 7a. Robustness checks on the SUR model (dependent variable: log difference of real income) Check Nr.

1 2 3 4 5 6 7 8 9 10 11

One year refinancing window

Damaged borrowers only

Lagged real income included

(From general) to specific

Obs.

Pseudo R2 of Eq. (1.1)

No No No Yes Yes Yes Yes No No Yes Yes

No Yes Yes No No Yes Yes No Yes No Yes

Yes Yes No Yes No Yes No Yes Yes Yes Yes

No No No No No No No Yes Yes Yes Yes

247 156 156 180 180 115 115 258 164 189 121

0.33 0.43 0.26 0.36 0.27 0.49 0.33 0.28 0.33 0.32 0.39

Loan to income ratio Coef.

T-stat

0.05 0.05 0.30 0.10 0.29 0.05 0.35 0.07 0.10 0.09 0.13

1.96 1.94 7.47 2.17 7.8 1.95 7.62 5.7 2.74 2.74 3.21

Note: Robustness checks are built on regression in Table Table 6, column 1, with additional restrictions. The possible restrictions are (i) only individuals whose P4 window is shorter than 12 months, (ii) only people damaged from the Tsunami, (iii) ln of monthly real income of the previous period included as control and (iv) only significant regressors included in the regression. Loan to income ratio: amount of the first loan obtained after the tsunami divided by post-tsunami pre-refinancing monthly income. The pseudo R2 refers to the first of the three equation of the SUR model. Estimates make use of heteroskedasticity-robust standard errors.

Table 7b. Robustness checks on the SUR model (dependent variable: log difference of hours worked) Check Nr.

1 2 3 4 5 6 7 8 9 10 11

Only P4 < 12 months

No No No Yes Yes Yes Yes No No Yes Yes

Damaged borrowers only

No Yes Yes No No Yes Yes No Yes No Yes

Lagged worked hours included

(From general) to specific

Yes Yes No Yes No Yes No Yes Yes Yes Yes

No No No No No No No Yes Yes Yes Yes

Obs.

230 140 140 167 167 103 103 239 146 174 107

Pseudo R2 of Eq. (1.1)

0.36 0.40 0.18 0.39 0.16 0.45 0.24 0.30 0.33 0.30 0.36

Loan to income ratio Coef.

T-stat

0.09 0.20 0.30 0.07 0.23 0.21 0.41 0.12 0.17 0.18 0.26

1.94 1.90 2.40 1.75 2.25 1.97 2.60 2.07 1.92 2.22 2.21

Note: Robustness checks are built on regression in Table 6, column 2, with additional restrictions. The possible restrictions are (i) only individuals whose P4 window is shorter than 12 months, (ii) only people damaged from the Tsunami, (iii) number of worked hours of the previous period included as control and (iv) only significant regressors included in the regression. Loan to income ratio: amount of the first loan obtained after the tsunami divided by posttsunami pre-refinancing monthly income. The pseudo R2 refers to the first of the three equation of the SUR model. Estimates make use of heteroskedasticity-robust standard errors.

variables used as regressors in the three equations is below three (in the first equation). 17 (iii) Significance and magnitude of our results Estimate results from the SUR in Table 6 show that residuals of the three system equations are significantly correlated thereby justifying the simultaneous estimate. The loan to income ratio affects positively and significantly the change in real income (Table 6, columns 1 and 2) confirming evidence from subgroup difference in means shown in Table 5. The magnitude of its effect is such that, by doubling the loan size to income ratio from its average level we obtain a change in income of 25% in the SUR estimate. Furthermore, the variables affecting timing and size of the post tsunami refinancing are all significant in the second and third equation of the simultaneous model, confirming the previously explored pattern of relationships. The results in column 1 are paralleled by an analogous significant effect of the loan to income ratio on the percent change in worked hours (Table 6, columns 3 and 4) with

17% elasticity (to correctly frame this result consider that the mean change from P3 to P4 is 11 hours and the average number of worked hours in P3 is 44.3). (iv) Robustness checks: lagged income, time window heterogeneity, and treatment sample only Since we are confident that all other potential factors supporting convergence of damaged to non-damaged borrowers are conveniently captured by our regressors (which include remittances, grants, etc.) we do not include the lagged values of real income and weekly worked hours in our base specification presented in Table 6. As a first robustness check we, however, correct for the possibility of a convergence process not fully captured by our regressors by including such variable in our specification. As a consequence the impact of the loan to income ratio is still significant but lower in magnitude (Tables 7a and 7b). A second robustness check is on heterogeneity in the timing of MFI refinancing. When focusing on the effects of post-tsunami refinancing the problem of time window heterogeneity

908

WORLD DEVELOPMENT Table 8. The different refinancing effects on damaged versus non-damaged borrowers

Variable

Galle Matara Agriculture Fishery Manufacture Age Female Primary Secondary Number of children DAM Remittances Subsidies Donations and grants Other charity Loan to income ratio Ln [1 + Loan to income ratio  DAM] Length Seniority Constant Only P4 < 12 months Obs. (Pseudo) R2

Log difference of real income

Log difference of worked hours

(i)

(ii)

(iii)

(iv)

0.1412 (1.21) 0.0223 (0.93) 0.1142 (1.03) 0.0754 (0.71) 0.0215 (0.42) 0.0013 (0.62) 0.0051 (0.42) 0.0632 (1.21) 0.0012 (0.04) 0.0051 (0.11) 0.084 (0.89) 0.0241 (0.88) 0.1517 (1.66) 0.3163 (1.91) 0.0412 (0.35) 0.0891 (2.14) 0.1564 (2.38) 0.0083 (1.71) 0.052 (1.14) 0.936 (0.21) No 247 0.2813

0.1315 (1.80) 0.1545 (0.76) 0.1261 (0.91) 0.0064 (0.08) 0.0215 (0.71) 0.001 (0.76) 0.0352 (0.91) 0.0941 (1.21) 0.0062 (0.07) 0.0652 (0.76) 0.0072 (0.62) 0.0732 (0.73) 0.0752 (0.63) 0.2615 (1.62) 0.0744 (0.88) 0.1351 (2.51) 0.1281 (1.89) 0.0072 (0.25) 0.1062 (1.36) 0.1062 (0.91) Yes 180 0.3105

0.1532 (0.94) 0.1053 (0.92) 0.0764 (0.98) 0.2164 (0.56) 0.1048 (0.96) 0.0021 (0.25) 0.1981 (0.71) 0.0062 (0.08) 0.3625 (1.81) 0.0752 (0.91) 0.2514 (1.75) 0.1051 (0.81) 0.0863 (0.81) 0.5262 (1.62) 0.2512 (1.52) 0.0632 (0.61) 0.1504 (1.94) 0.0065 (0.51) 0.0523 (0.35) 0.4623 (1.41) No 230 0.1931

0.2518 (0.62) 0.2052 (0.93) 0.1962 (1.53) 0.5295 (1.81) 0.2514 (0.35) 0.0083 (1.01) 0.0732 (0.06) 0.0628 (0.52) 0.9526 (1.71) 0.1503 (1.03) 0.0963 (1.61) 0.3215 (0.91) 0.0521 (0.07) 0.5377 (1.02) 0.2142 (1.21) 0.0632 (0.61) 0.1841 (2.05) 0.0521 (0.58) 0.0352 (0.61) 1.2518 (1.61) Yes 167 0.1942

Note: OLS regressions with robust standard errors in parenthesis. Ln [1 + Loan to income ratio  DAM] is the log of a slope dummy variable, being the product of the loan to income ratio (amount of the first loan obtained after the tsunami divided by post-tsunami pre-refinancing monthly income) and a dummy variable (DAM) equal to 1 if the respondent was damaged by the tsunami. Variable legend: see Table 1a and b.

reduces to the timing of the post-tsunami loans since the other two time points (tsunami and interview dates) are given and equal for all respondents. To reduce it further we restrict the sample to borrowers receiving the first post-tsunami loan within a one year interval (between 10 and 22 months from the tsunami) dropping around 27% of the observations. Consider that, by doing so, we handle a level of heterogeneity not different from that of analyses based on yearly data when events are irregularly spaced during the same year. Furthermore, by knowing the monthly timing of the refinancing event and controlling for it with the window length, we are doing more than what is done in studies on yearly data. Tables 7a and 7b show that the significance of the loan to income ratio persists when we consider this restricted sample. The elas-

ticity on the change in income is around 10% when we include lagged income and 29% if we do not include it (Table 7a, checks 4 and 5). The corresponding numbers are 7% and 23%, respectively, for the change in worked hours (Table 7b, checks 4 and 5). Consider further that heterogeneity in the loan to income ratio and in the loan purpose between treatment and control samples may distort our previous findings. To avoid this we restrict the sample to those individuals reporting damages and find that elasticities are up to 30% for both performance indicators in the base specification without the lagged dependent variable (Tables 7a and 7b, checks 3 and 7), while are much smaller but still significant when we include it (Tables 7a and 7b, checks 2 and 6). When considering jointly the time

EVIDENCE FROM THE 2004 TSUNAMI

909

Table 9. Robustness checks on regressions with slope dummy variables for damaged people Dependent variable

Ln [1 + Loan to income ratio  DAM] t-statistic Ln of real income (t  1) Worked hours (t  1) Only P4 < 12 months Obs. R2

Log difference of real income

Log difference of worked hours

(i)

(ii)

(iii)

(iv)

0.1214 2.16 Yes – No 247 0.3315

0.1215 1.91 Yes – Yes 180 0.3516

0.1351 2.41 – Yes No 230 0.3721

0.1914 1.92 – Yes Yes 167 0.3806

Note: Results are robustness checks of regressions in Table 8, with lagged values of the log of monthly real income/weekly worked hours and restrictions on the time window length. Ln [1 + Loan to income ratio  DAM] is the log of a slope dummy variable, being the product of the loan to income ratio (amount of the first loan obtained after the tsunami divided by post-tsunami pre-refinancing monthly income) and a dummy variable (DAM) equal to 1 if the respondent was damaged by the tsunami. T-statistics are calculated according to heteroskedasticity-robust standard errors.

Table 10. Treatment effect models Robustness check

Ln [1 + Loan to income ratio  DAM] t-statistic Ln of real income (t  1) Only P4 < 12 months v2 (1)a Prob > v2 Obs. Wald v2

Log difference of real income

Log difference of worked hours

(i)

(ii)

(iii)

(iv)

(v)

(vi)

(vii)

(viii)

0.1275 2.90 Yes No 2.09 0.15 247 131.2

0.1553 3.16 No No 2.13 0.14 247 91.3

0.1235 2.90 Yes Yes 2.66 0.10 180 117.1

0.1311 3.23 No Yes 2.17 0.16 180 87.6

0.1732 2.52 Yes No 3.12 0.08 230 82.1

0.1362 2.13 No No 1.27 0.21 230 35.1

0.2159 2.51 Yes Yes 2.61 0.14 167 63.3

0.2642 2.05 No Yes 0.92 0.32 167 37.1

Note: Results are from treatment regression estimates where the main equations remain a those shown in (2) and ((20 )) in the text while in the selection equation (see Eqn. (3)), participation to the treatment group (the Damage dummy equal to 1 if the respondent was damaged from the tsunami) is regressed on Agriculture and House on coast, the two above mentioned variables found to be significantly different between the treatment and the control samples. Ln [1 + Loan to income ratio  DAM] is the log of a slope dummy variable, being the product of the loan to income ratio (amount of the first loan obtained after the tsunami divided by post-tsunami pre-refinancing monthly income) and a dummy variable (DAM) equal to 1 if the respondent was damaged by the tsunami. a LR test of independent equations (correlation of residuals of the two equations = 0).

window heterogeneity and treatment-sample-only restrictions we still have strong and significant results with elasticities above 20% in the specification without lagged dependent variable (Tables 7a and 7b, check 7). Finally, given the limited number of observations and the large number of potential controls included in our base estimates, as in a general to specific approach we re-estimate our model including only variables whose significance is above the 95 percent threshold to make the effect of our MFI refinancing variable cleaner. We do that for the overall sample and for all the above commented robustness checks. In all of these cases the impact of the loan to income ratio remains strong and significant (Tables 7a and 7b, checks 8–11). (v) The differential effect of the loan to income ratio on damaged versus non-damaged borrowers To evaluate the impact of microfinance we are finally interested to test whether, ceteris paribus, loans received by damaged borrowers were more productive than those received by non-damaged ones. If the null of no significant difference is rejected we can show that the loan is relatively more important when hit by natural calamities. To perform such test we estimate with OLS the first equation of the SUR model and introduce as additional regressor the size to income variable interacted with the damage dummy (DAM  (L/I)) where DAM takes value of one if the borrower suffered from tsunami damages or zero otherwise. 18 The null of no different impact

of the MFI loan on damaged versus non-damaged borrowers is rejected if the added variable is significant. More specifically we estimate the following specification: lnðI i;t Þ  lnðI i;t1 Þ ¼ a0 þ a1 lnðLi;t =I i;t1 Þ þ a2  ln ½1 þ DAM i  ðLi;t =I i;t1 Þ X þ a3 Lengthi;t þ a4j X ð1Þji;t þ e1i;t

ð2Þ

j

for real income (I) or: lnðH i;t Þ  lnðH i;t1 Þ ¼ a0 þ a1 lnðLi;t =Y i;t1 Þ þ a2  ln ½1 þ DAM i  ðLi;t =I i;t1 Þ X þ a3 Lengthi;t þ a4j X ð1Þji;t þ e1i;t

ð20 Þ

j

for worked hours (H) and test the following null hypothesis H 0 : a2 ¼ 0. If we reject the null on the positive size we have that the loan to income ratio affected significantly more the performance of damaged versus non-damaged borrowers. Such test is almost free from selection problems since we demonstrated that the tsunami acted as a random shock which divided a homogeneous set of borrowers from the same MFI into a treatment and a control sample. We know from Table 6 that the size of the loan is affected by some determinants, such as gender, number of damages, seniority, and lagged real

910

WORLD DEVELOPMENT

income. We also take the point that there may be some omitted variables which still affect both loan size and performance (i.e., better business opportunities). For this reason we devise in Table 8 a test which should be immune from the selection bias generated by the endogeneity of the loan to income ratio since we measure with the slope dummy (damage dummy  loan to income ratio) the presence of a significantly different effect of the loan for damaged versus non-damaged borrowers conditional to the same loan size. Findings from our estimate show that the base effect with the full sample estimates is around 9% to which is added an additional 16% for damaged borrowers (Table 8, column 1). The significance of our results is confirmed when restricting the sample to borrowers receiving the first post-tsunami loan in a one year interval in order to correct for time window heterogeneity (Table 8, column 2). Results on our second performance variable (the number of hours worked) are consistent with previous findings showing that the elasticity of the impact of the loan to income ratio for damaged people is now between 15% (Table 8, column 3) and 18% (one year window of loan refinancing, column 4). Finally, Table 9 repeats the exercise by adding lagged income or lagged weekly worked hours to regressions in Table 8, with findings which remain consistent with those shown in Table 8. (vi) Robustness checks with treatment regression Even though the tsunami acts as a random shock on an ex ante homogenous group of borrowers from the same MFI (see Table 4), two characteristics (house on the coast and agricultural activity) may still in principle discriminate between them. In order to check whether such differences create a selection problem we re-estimate the model with a treatment regression approach in which main equation and selection equation are jointly estimated. The main equations remain as those shown in (2) and (20 ) while in the selection equation, participation to the treatment group (the Damage dummy) is regressed on the two above mentioned variables which we have found to be significantly different between the treatment and the control samples: Damagei ¼ b0 þ b1 Agriculturei þ b2 House:on:coasti þ vi

ð3Þ

with Agriculture being a dummy for borrowers working in agriculture and House on coast for those living not farther than one km from the coast. 19 Note that, in order to meet the requirement of using selection variables not affecting our performance indicator, in both estimates we use regressors which revealed themselves not to be correlated with the dependent variable in single equation estimates. Two characteristics of our new results are important (Table 10). First, findings are not substantially different in magnitude with respect to those in the SUR. Second the null of no correlation of residuals between the main and the selection equations is not rejected. As a consequence, the only two factors which make treatment and control samples different (location of the house on the coast and agricultural activity) do not generate selection effects on performance. This reinforces the validity of our test on the differential effect of the

loan to income ratio on damaged versus non-damaged borrowers documented in Table 8. 5. CONCLUSIONS Our paper examines the role of MFI loans as an effective recovery tool after tsunami for a sample of 305 randomly selected clients of a Sri Lankan MFI. In order to reconstruct time series we devise an ad hoc retrospective panel data approach by asking interviewed borrowers to declare their current and remember their past economic levels. We carefully control for the reliability of our data by combining survey data and bank records, comparing results on income with those on a more memorable performance indicator like the number of weekly worked hours, looking at validating evidence across different indicators and controlling for heterogeneity of time windows and characteristics between treatment and control samples. Our main findings are that (i) the post-tsunami loan to income ratio has a significant effect on the borrowers’ recovery measured in terms of change in income or in worked hours and (ii) the effect of the loan to income ratio is significantly stronger for damaged versus non-damaged borrowers. To evaluate the robustness of our results we control for selection bias on loan to income size, heterogeneity in loan timing and characteristics between treatment and control borrowers. Based on our findings we identify four elements supporting the usefulness of the donors’ intervention on the MFI and of the microfinance loans after tsunami in our specific case: (i) the deterioration of the MFI portfolio after tsunami was such that it is hard to imagine that it could have continued to operate without recapitalization (this starting point is based on the official certified losses of the MFI); (ii) the MFI contrasted the negative effects of tsunami by granting to damaged borrowers and to borrowers with lower income higher loans in proportion to their level of income. This is far from being obvious (even though for individuals with lower income the loan to income ratio could tend to be higher) since damages and falls in income negatively affect the collateralization capacity and, therefore, the creditworthiness of the borrowers. Finally, our econometric test on the relative “productivity” of the loan shows that (iii) the loan significantly affected worked hours and real income for damaged borrowers after recovery and (iv) the loan is relatively more important when hit by calamities since the effect on damaged borrowers is higher than that on non-damaged ones. Our results provide food for reflection on new ways for using donor’s resources after calamities. In parallel to the direct provision of food, investment goods, or infrastructure, fundamental to address emergency needs and rebuilding, recapitalising MFIs under stress after calamities may provide an effective liquidity injection by acting as a sort of expansive monetary policy measure for the poor. Such measure can restart and stimulate economic activity with significant effects in terms of both worked hours and income creation.

NOTES 1. Developing countries are particularly vulnerable to the effects of climate variability, climate change (Smith, Klein, & Huq, 2003) and other natural catastrophes like earthquakes and tsunamis. In absolute terms, natural catastrophes disproportionately affect low and middle income countries. Available data (see World Bank, 2001) show that, between 1990 and 1998, 94% of natural disasters and 97% of deaths from natural catastrophes occurred in less developed countries (LDCs).

2. http://www.usaid.gov/locations/asia_near_east/tsunami/. 3. The number of families involved and the amount of loans provided are very similar to those of the MFI we considered in our study in Sri Lanka. Our empirical results on Sri Lankan borrowers confirm the qualitative evidence on the usefulness of microcredit as a recovery tool already fund in Thailand by USAID in a very similar setting.

EVIDENCE FROM THE 2004 TSUNAMI

911

4. Sri Lanka was severely affected by the tsunami of the December 26, 2004, even though it was not the most damaged country. The disaster was triggered by an earthquake measuring 9.0 on the Richter scale whose epicenter was located in the Indian Ocean near Banda Aceh in Sumatra. The giant wave struck a coastal area stretching over 1,000 kilometers, or two-thirds of the country’s coastline (from Jaffna in the North to the west coast, North of Colombo). Human losses were huge (over 35,000 dead and 443,000 displaced people, see Athukorala & Resosudarmo, 2005). Economic losses were also extremely severe since the wave damaged 24,000 boats (about 70% of the fishing fleet), 11,000 businesses and 88,500 houses, of which more than 50,000 were completely destroyed.

12. Estimates are omitted for reasons of space and available upon request. The above described pattern of relationships will be, however, visible in the SUR estimates which follow.

5. The donors still have property rights on the original fund which is now part of the equity of the MFI.

15. The seemingly unrelated regression (SUR), model (Zellner, 1962), is a well known technique for analyzing a simultaneous pattern of relationships among variables in a system of multiple equations with correlated error terms. We adopt this choice following Davidson and Mackinnon (1993) who argue that the univariate approach should be extended in order to take into account a more complex pattern of dependence among variables and consider SUR a possible solution when all dependent variables are modeled as correlated and simultaneous. The SUR is used to provide clearer details of the pattern of relationships among the different variables and to show how the loan to income ratio and window length are in turn affected by other factors. After doing this we address in a different way the two endogeneity problems with estimates in Table 8 where we control for heterogeneity in loan to income ratio and window lengths.

6. The length of the third interval is six month for the first quarter of the sample, 10 months for half of it and 15 months up to 75% of the sample. We do not observe significant differences between the two groups in terms of duration of the third time window. 7. The estimation of the significance of a common event in a sample of non-synchronous events is the typical focus of event studies in finance (for a standard treatment see Campbell, Lo, & McKinlay, 1997). In those studies non-synchronicity concerns the event date and abnormal returns are calculated on the basis of a normal return model estimated in the period preceding the event window. 8. Some people had more than one economic activity, thus the total sum exceeds one. 9. Under the current OECD rule, earnings are divided by a scale factor A, where A = 1 + 0.5(Nadults  1) + 0.3Nchildren. However, in our sample a large part of consumption is food consumption. It is, therefore, advisable to reduce the extent of economies of scale by increasing weights in the equivalence scale. We, therefore, decide to follow the standard suggestion in development studies of giving unit weights to each member (for a discussion of the methodological problems in creating equivalence scales see Deaton & Paxson, 1998). 10. We perform a robustness check with non parametric analysis of the difference in the change of the considered economic indicators between treatment and control samples which confirms significance of parametric findings. Results are omitted for reasons of space and available upon request. 11. Consider as well that, in this question, the interviewer made a clear distinction between paid and unpaid work and reduced respondents mnemonic effort by asking worked hours per working days and in weekends.

13. We cannot provide evidence whether the intention to support more the most damaged borrowers is affected by constraints posed by donors who provide equity capital to the MFI but we strongly suspect it. 14. Estimates are omitted for reasons of space and available upon request.

16. The VIF (variance inflation factor) formula is 1/1  R(x) where R(x) is the R2 when the independent variable is regressed on all other independent variables (Marquardt, 1970). If R(x) is low (tends to zero) the VIF test is low (equal to one). A VIF value below 10 (or, more restrictively, five) is considered acceptable by rules of thumb usually adopted in the literature. 17. VIF factor results and the correlation matrix are omitted for reasons of space and available from the authors upon request. 18. We control whether such introduction generates multicollinearity but we find that the maximum VIF is below 3.5 and the correlation between the interacted and non interacted loan to income ratio is around 50%. 19. The treatment-effects model evaluates the impact of an endogenously chosen binary treatment on another endogenous continuous variable, conditional on two sets of independent variables using either a two-step consistent estimator or full maximum likelihood (we opt for the second solution). In the two equations system (v) and (e) are the error terms of the selection and main equations and are bivariate normal random variables   r q with zero mean and covariance matrix . The likelihood function q 1 for the joint estimation is provided by Maddala (1983) and Greene (2003).

REFERENCES Armendariz de Aghion, B., & Morduch, J. (2005). The economics of microfinance. Cambridge, MA: The MIT Press. Athukorala, P., & Resosudarmo, B. P. (2005). The Indian Ocean tsunami: Economic impact, disaster management and lessons. Asian Economic Papers, 4(1), 1–39. Campbell, J. Y., Lo, A., & McKinlay, C. (1997). The econometrics of financial markets. Princeton, USA: Princeton University Press. Davidson, R., & MacKinnon, J. G. (1993). Estimation and inference in econometrics. New York, NY: Oxford University Press. Deaton, A., & Paxson, C. (1998). Economies of scale, household size and the demand for food. Journal of Political Economy, 106, 897– 930. Fafchamps, M., & Gubert, F. (2007). The formation of risk sharing networks. The Journal of Development Economics, 83(2), 326–350.

Goldberg, N. (2005). Measuring the impact of microfinance: Taking stock of what we know. Grameen Foundation, www.grameenfoundation.org. Greene, W. (2003). Econometric analysis: 5th edition. Prentice Hall. Hermes, N., & Lensink, R. (2007). The empirics of microfinance: What do we know?. The Economic Journal, 117, 1–10. Hoque, S. (2008). Does micro-credit program in Bangladesh increase household’s ability to deal with economic hardships? MPRA Paper No. 6678. Hulme, D., & Mosley, P. (1996). Finance against poverty: Vol. I, II. London, UK: Routledge. Karlan, D. S. (2001). Microfinance impact assessments: The perils of using new members as a control group. Journal of Microfinance, 3(2), 76–85. Khandker, S. R. (2007). Coping with flood: Role of institutions in Bangladesh. Agricultural Economics, 36, 169–180.

912

WORLD DEVELOPMENT

Maddala, G. S. (1983). Limited-dependent and qualitative variables in econometrics. Econometric society monographs in quantitative economics. Cambridge, UK: Cambridge University Press. Marquardt, D. W. (1970). Generalised inverses, ridge regression, biased linear estimation, and nonlinear estimation. Technometrics, 12, 591–612. McIntosh, C., Villaran, G., & Wydick, B. (in press). Microfinance and home improvement: Using retrospective panel data to measure program effects on fundamental events. World Development, 39 (6), 922–937.

Skees, J., Varangis, P., Larson, D., & Siegel, P. (2002). Can financial markets be tapped o help poor people cope with weather risks? World Bank Policy Research Working Paper No. 2812. Smith, J. B., Klein, R. J., & Huq, S. (2003). Climate change, adaptive capacity and development. London, UK: Imperial College Press. World Bank (2001). World development report 2000/2001: Attacking poverty. Oxford, UK: Oxford University Press. Zellner, A. (1962). An efficient method of estimating seemingly unrelated regression equations and tests for aggregation bias. Journal of the American Statistical Association, 57, 48–368.

Available online at www.sciencedirect.com