Financial inclusion helps rural households address climate risk

Financial inclusion plays an important role in helping households manage risks, but its role in mitigating climate risks is unexplored. Access to formal financial institutions in regions with high climate risks increases households’ access to liquidity that they need to buffer against climate shocks. Using longitudinal data from 1082 rural households located in the semi-arid tropics in India, we find that households facing high climate risks hold a higher proportion of assets in liquid form. Access to formal financial services, however, reduces the need to keep liquid assets to be able to respond to high climate variability. Our results suggest that expanded financial inclusion in regions with high climate variability can reallocate resources held in unproductive liquid assets to invest in climate adaptation.


ICRISAT-VDSA Household Dataset
The International Crops Research Institute for the Semi-Arid Tropics (ICRISAT) conducts village level studies to track rural poverty in households in village economies in South Asia under the Village Dynamics studies in South Asia (VDSA) initiative (24). The ICRISAT-VDSA dataset provides granular household-level data about resource endowments, cultivation input and output, employment, livestock, cropping pattern, financial transactions, and development program participation. It also has information about the demographics of each household such as caste, landholding, and household size. The presence of a strong panel component, temporal granularity of data, village-level household identification, high quality of refinement, and the ability to track households across 5 years makes it a suitable dataset for our study. We excluded households that were not present in the data for all five years as well as households that were subsequently included as replacements. We have a balanced panel of annual data for 1082 households spanning across 30 villages in semi-arid tropics (states of Andhra Pradesh, Maharashtra, Madhya Pradesh, Gujarat, Bihar, Jharkhand, and Orissa in India).
The VDSA survey includes the following two questions for all households: (1) Did you experience any severe drought/flood/pest/ diseases that affected your livelihoods during the last year? (2) Conditional on experiencing a problem, list the coping mechanisms adopted in order of importance. The distribution of households facing climate shocks over the 5-year period (Fig. S1) shows that these rural households face substantial climate risk. 59% households experienced climate shocks -drought, flood, cyclone, heavy rainfall, or frost -in at least one year. 13% of households experienced climate shocks in more than 2 of the 5 years. Out of the 633 households who have faced climate shock at least once, for a majority (57%) the primary coping mechanism is their own savings (Fig. S2).
Given dependence on own savings in the face of climate risk, our main variable of interest is the proportion of liquid assets held by the households. We define the primary variables used in this study in Table S1.
A follow up question on coping mechanism in the survey asks households to rank the individuals/ institutions they approach for assistance based on the reliability of these sources in the event of a drought or a flood. The geography of our study being a semi-arid region, the households are much more likely to face drought: out of 633 households that reported facing climate shocks, 514 households have experienced drought at least once in our sample. Thus, we focus on reliable sources of assistance in the event of drought. 'Kin and relatives' and 'Friends' are the top 2 reliable sources of assistance for the households in our sample. (Fig. S3). Bank, the 5th most common source of assistance, features in the top 3 reliable sources for 552 households. When households can rely on formal financial institutions, they rely on the moneylender less (Fig. S4). We create a variable BANK_TOP3 equal to 1 if a household in a year reports bank in its top three reliable sources of assistance. This is one of our indicators of financial inclusion. The second variable, HH_BANKED is 1 if the households borrow from or save with a bank or a formal financial institution. Table S2 reports the descriptive statistics of all our main variables.

Climate Variability
For creating climate risk variables, we use rainfall and temperature variables from India Meteorological Department (IMD) Data from 1951Data from -2014 by the Ministry of Earth Science, Government of India (33-34).
Since our sample consists of rural households primarily dependent agriculture, we focus on the climate variability during monsoon, the primary growing season in India. For each village using annual data from 1951 to 2014 we calculate the coefficient of variation of total rainfall and mean daily temperature for the 120 days following the onset of the monsoon.

Village Infrastructure and Economic Activity
In addition to the household and climate variables, we use infrastructure and economic activity indicators at the village level as instruments or controls in our estimation. Based on 2001 Village Census (35), we create indicators whether the village has a paved road or a power connection within 5 km. To control for economic activity, we use the Defence Meteorological Satellites Program -Operational Linescan System (DMSP-OLS) night time lights (NTLs) dataset (36) to create village-level night lights variables as measured in 2010.

Comparison: Our Sample and Broader Sample of Rural Households
We compare the distribution of our key variables to a broader sample NSSO Debt and Investments Survey conducted in 2013 (37). We consider all the districts in the NSSO sample in 12 Indian states in the semi-arid tropical geography (Andhra Pradesh, Bihar, Chhattisgarh, Goa, Gujarat, Jharkhand, Karnataka, Madhya Pradesh, Maharashtra, Odisha, Rajasthan, and Telangana). We restrict our sample to households who have 30% or less land irrigated to make our sample comparable to the VDSA sample. Fig S5 shows the distribution of Liquidity Ratio (%) and the climate risk variables for the VDSA and the NSSO samples. The distributions of the variables across the two samples are generally similar, indicating that the VDSA sample is representative of the broader sample of rural Indian households.

Estimation Approach
We want to investigate how proportion of saved assets responds to climate risk, financial inclusion, and the interaction of the two. However, there are two potential problems in drawing causal inference based on an OLS regression. First, the household's participation in the formal modes of financial savings and credit can depend on many of its observed and unobserved characteristics. While we control for relevant observed ones (including fixed effects for household landholding and caste), we cannot control for the unobserved attributes. Second, there is the possibility of reverse causality -households with a certain level of desired liquid assets may choose to access financial institutions. Hence, high proportion of liquid assets may drive financial inclusion instead of (or in addition to) financial inclusion influencing the level of liquidity. To address these problems, we use an instrumental variable regression. Our objective in conducting an instrumental variable regression is capture "supply" of formal financial institutions by a variable that is uncorrelated with "demand" at the household level. Following the approach suggested by (38), applied (39-41), we use village indicators as our instrument. The argument is that whether a household is financially included by a bank depends on the relevant characteristics of the household as well as the bank. The inclusion also depends on distribution of households and proximate bank branches in each village. (38) argue that this distribution is exogenous and recommend use of a set of geography-based instruments. In order to satisfy the exclusion restriction, appropriate instruments for our regression specifications should affect the selection into financial inclusion but should not directly be related to the outcome (household liquidity) after adding the relevant control variables.
One concern is using village indicators as instruments is that some village level characteristics in particular economic prosperity and climate risk might affect financial inclusion and also be directly related to level of household liquidity. We address this concern by including village night lights and climate risk variables as controls in our second stage regression. Additionally, we include district fixed effects to absorb unobserved characteristics such as district-level economic activity that may not be fully captured by night lights.
We follow Procedure 18.1 for a binary endogenous variable, described in Chapter 18.4 of (42). Specifically, we run a probit regression of Bank, the financial inclusion indicator, as a function of village indicators and all the other explanatory variables (climate risk and controls) and calculate the predicted value BANK_HAT. Then we run an instrumental variable regression with INVNORM_LIQASSETS as the dependent variable and BANK_HAT, BANK_HAT*RAINFALL RISK, BANK_HAT*TEMPERATURE RISK as instruments for Bank, Bank*Rainfall Risk, Bank*Temperature Risk. INVNORM_LIQASSETS is the inverse normal transformation of the liquid assets fraction.
We use this transformation so that any predicted value when converted back to liquid assets fraction will always lie between 0 and 1. Specifically, our estimation equations look as follows: Probit: ( ℎ = 1) = Φ( 0 + 1 + 2 + 3 + BANK_HAThvy = Fitted value of BANKhvy from the probit estimation.   Table S3, it is negative and mostly significant in the alternative specifications discussed below. Thus, we find that financial inclusion mitigates the need to hold liquid assets as a coping mechanism for climate risk. Table S4 presents the first stage results of the models in Table S3. We see that pseudo R-squared for the probit model ranges from 0.236 to 0.269, providing support, in addition to the F-statistic reported in Table S3, for the strength of the instrument. Further, we see that in the first stage regressions the coefficient for BANK_HAT, BANK_HAT*RAINFALL RISK, and BANK_HAT*TEMPERATURE RISK in regression equations of BANK, BANK*RAINFALL RISK, and BANK*TEMPERATURE RISK, respectively, is positive, highly significant and close to 1, indicating the validity of the instruments.

Alternative Specifications
To investigate robustness of results we estimate various alternative specifications. We use one climate risk variable at a time (Table S5). We use OLS instead of instrumental variable (Table S^) and within OLS we include village fixed effects (Columns 3-4 of Table S6). All these results broadly support the conclusions that i) households hold more liquid assets when facing greater climate risk ii) financial inclusion reduces the need to hold liquid assets in response to climate risk.
We note that the coefficient of the BANK variable is positive and significant in the OLS specification (Table S6). This is line with the pattern in Figure 1D. However, the coefficient is negative and significant when we use the instrumental variable approach (Tables S3 and S5). Thus, the positive relationship observed in Figure 1D and Table   S6 is likely to be due to households with greater need or ability to save choosing to be banked. Once we control for this household-level selection through instrumental variables, we see that financial inclusion results in lower liquidity ratio, in line with our expectation.

24.
A        Table S1 for variable definitions. Rainfall and Temperature Risk are standardized to have 0 mean and unit standard deviation for the sample. Robust standard errors are in the parentheses. *, **, and *** indicate significance at 1%, 5%, and 10% respectively. The last 5 rows provide additional details of the specification, including whether Bank (financial inclusion) is measured using BANK_TOP3 or HH_BANKED.  Table S1 for variable definitions. Rainfall and Temperature Risk are standardized to have 0 mean and unit standard deviation for the sample Robust standard errors are in the parentheses. *, **, and *** indicate significance at 1%, 5%, and 10% respectively. The last 4 rows provide additional details of the specification, including whether Bank (financial inclusion) is measured using BANK_TOP3 or HH_BANKED.