Scale and Sustainability: The Impact of a Women’s Self-Help Group Program on Household Economic Well-Being in India

Abstract Microfinance groups are a prominent source of small-scale rural credit in many developing countries. In India, evidence of the impact of the now ubiquitous women-only savings and credit self-help groups (SHGs) on household consumption and asset accumulation is inconclusive and based on small-scale interventions. Further, little is known about the sustainability of impacts at scale. We use panel data on close to 2500 households from five states in India to estimate the impact of SHG membership on household expenditure and asset ownership. Over four years, we find small but significant impacts of SHG membership on household expenditure and livestock ownership. Membership duration has a modest effect, suggesting that initial impacts may taper off as the program scales up, though small sample sizes limit our ability to draw inferences. Accompanying evidence on pathways is compelling; related work shows that SHG participation improves information, empowerment, and access to entitlements. While the direct impacts of SHG membership may not suffice to fill gaps in access to credit faced by the rural poor, impacts along these additional pathways could intensify the benefits of these groups.


Introduction
Since 2010, a large proportion of microfinance in rural India has been provided through women's savings and credit-based self-help groups (SHGs) which actively target women from poor and marginalized communities that are excluded from formal credit systems (Baland, reported positive impacts, the available evidence suggests that group-based microcredit programs, including those based on SHGs, fall short of being the one-stop solution to rural poverty once hoped for. Our paper contributes to this body of evidence by providing causal estimates of the impact of participation in SHGs on measures of household well-being captured by per capita household consumption expenditure, household wealth and asset accumulation, and ownership of small, medium, and large livestock. These results, coming shortly after the nationwide expansion of SHG membership under the NRLM, provide reasonable estimates of impact during a period of rapid scaling up. We use data from a sample of households at three points in time: a baseline in 2015 Q4, a midline in 2017 Q4, and an endline in 2019 Q4. With pre-existing SHGs and in the absence of any clear targeting criteria, we are unable to experimentally test the impact of the introduction of SHGs. Instead, we use matching techniques to construct a comparison group and track changes in outcomes from baseline to endline using difference-in-difference (DID) models. Treatment was primarily defined as the woman respondent being a member of an SHG (see section 3), and treated and non-treated households matched to each other using baseline characteristics.
SHGs in our sample are supported by a range of organizations. In areas where PRADAN has an active presence, they work with the state livelihoods missions to support SHG formation and functioning. In areas where PRADAN does not operate, SHGs are supported by state actors, sometimes with other smaller NGOs playing a role as well. For brevity, we call these 'SHGs in PRADAN areas' and 'SHGs in non-PRADAN areas', respectively. Consistent with literature pointing to the role of 'missions' as incentives to improve employee motivation and performance (Banuri & Keefer, 2016;Besley & Ghatak, 2005;Cassar & Armouti-Hansen, 2020), it is plausible that the employees of a development goal-oriented NGO-here, PRADAN-are more mission-oriented than the employees of the state-led programs, or even of smaller, less well-established NGOs. This, combined with the longer duration of PRADAN's operation and their greater saturation in some areas, could lead to better outcomes among households with women who are members of SHGs in PRADAN areas. This hypothesis remains to be tested.
We use three comparisons in our analysis, keeping in mind possible spill overs from SHG women to non-SHG women within villages. We start by examining the impact of SHG membership relative to non-membership, first in the full sample, and then in PRADAN areas only. We then test the relative benefits of the two SHG models by comparing outcomes for women who are members of SHGs in PRADAN areas to women who are members of SHGs in non-PRADAN areas. Next, under the twin assumptions that SHGs are beneficial and that SHGs in PRADAN areas have greater spill overs to nonmembers than SHGs in non-PRADAN areas (though greater community engagement, higher saturation, and so on), we compare outcomes for women who are members of SHGs in PRADAN areas to women who are non-members residing in areas where PRADAN does not operate. Finally, we examine impacts related to the duration of membership, a proxy for the degree of exposure to the intervention.
We find positive impacts of membership in any SHG on per-capita monthly food expenditures and expenditures on animal-sourced foods, and on total livestock units. The impacts on livestock ownership are consistent across all comparisons; that is, regardless of the organization supporting the SHG, being an SHG member leads to a positive impact on the woman's total and medium-sized livestock ownership. Being a member of an SHG in a PRADAN area also confers an added benefit over and above membership in an SHG in a non-PRADAN area in expenditure on animal-sourced foods, asset ownership, and livestock ownership. Finally, compared to non-members in non-PRADAN areas, households with women SHG members in PRADAN areas experience positive impacts on per-capita total monthly expenditures, per-capita expenditures on food and animal-sourced food items, and on asset ownership.

Context
Founded in 1983, PRADAN currently works in seven states in India to form and strengthen women's SHGs and higher-level collective organizations, provide administrative and capacitybuilding support for groups to undertake internal savings and credit activities, open deposit accounts, access formal credit, and, more recently, to avail of government-sanctioned funds for groups through the NRLM and other development programs of the central and state governments. PRADAN works predominantly in the most remote areas and among the poorest caste and tribal groups; this orientation, along with their clear emphasis on working with women, comes from their belief that even the most disadvantaged in society can be drivers of change.
Before the launch of the NRLM in 2011, PRADAN was often the sole organization supporting SHG formation and engagement in the areas where it was working. However, after 2011, State Rural Livelihood Missions (SRLMs), state bodies tasked with implementing the NRLM guidelines, were instituted, and SHG formation was scaled up rapidly throughout India. Where PRADAN already had a presence ('PRADAN areas'), it worked alongside the SRLMs, thereby contributing to a greater intensity of organizational support and engagement. In areas where PRADAN was not operational, SHGs were supported by the SRLMs. Other NGOs were also present and worked with the SRLMs in some select areas, but these were both smaller in scale and relatively recent engagements.
Regardless of the supporting organization, the core function and structure of SHGs remain broadly the same. Members meet weekly for 1-2 h to deposit a small individual amount (INR 10-20 2 in most cases) and discuss topics of common interest. In the initial months of operation, groups are provided support and training on basic group functioning, including responsibility, timeliness, how to hold respectful conversations and conduct group meetings. Subsequently, groups shift their focus to regularizing the savings and credit process, developing the base of land and water assets like irrigation, horticulture, land husbandry, and stabilizing livelihoods, that is, moving from subsistence to income generation, especially in agriculture and allied activities. Depending on the group's progress and its members' interests, more mature groups can advance to discussions and activities around other themes, such as nutrition, gender, and rights and entitlements.

Data
Our data come from a larger impact evaluation of nutrition interventions PRADAN introduced in 2015. The evaluation assessed the impact of the additional nutrition intensification efforts on improving maternal and child nutrition outcomes over and above the standard PRADAN livelihoods model, and, secondarily, compared both standard and nutrition intensification interventions against the alternative, the absence of any PRADAN intervention.
PRADAN teams that expressed an interest in and commitment to rolling out the nutrition interventions were included in the study, in other words, the selection of study districts and blocks within those districts where the nutrition intensification efforts were to be introduced was pre-determined by PRADAN. The study was conducted in eight districts in the five eastern and central states of West Bengal, Chhattisgarh, Odisha, Jharkhand, and Madhya Pradesh ( Figure 1). Two districts each were selected from Madhya Pradesh, Jharkhand, and Odisha, and one each from Chhattisgarh and West Bengal. Within each district, three blocks were selected: one where PRADAN rolled out its nutrition interventions (pre-determined), one where PRADAN conducted its standard SHG operations, and one control block where PRADAN Scale and sustainability 493 did not have a presence. 3 The impact evaluation design matched the standard PRADAN and the control blocks to the pre-determined block where the nutrition interventions were conducted using a set of demographic, economic, infrastructure, standard of living, and agriculture indices. The block level matching was conducted in consultation with PRADAN, and further details on the matching process are in Appendix A. Baseline data from the larger impact evaluation shows that the three arms of the study design were very well balanced (Table A2).
In each control block, we conducted a full listing of all villages. In the two PRADAN blocks in each district, we conducted a listing of all the villages PRADAN was working in. We then selected villages at random from these lists-five from each PRADAN block, and seven from control blocks. 4 We conducted a full listing of all ever-married women between the ages of 15 and 49 years in each selected village and sampled 20 women per village at random for inclusion in our study. SHG membership was not a criterion for inclusion in the sample as PRADAN encourages message diffusion from SHG members to non-members in the same community. 5 Our evaluation sample consisted of 1600 women in the two PRADAN arms (800 each), and 1120 women in the control arm, for a total of 2720 women. Due to slight oversampling, 2744 women were included in the study at baseline. We conducted interviews with the respondent woman and her husband or another adult male member of the household at three points in time: a baseline in 2015 Q4, a midline in 2017 Q4, and an endline in 2019 Q4. We also conducted community (village) level interviews with knowledgeable persons, such as the village head, service providers, and schoolteachers. This dataset forms the basis for the present analysis, although we are interested in the impacts of SHG participation, and not the nutrition intensification efforts covered in other ongoing work.

Ethical review
Our study underwent an ethical review from our institutional as well as a local review body. All study participants provided written or oral consent to being interviewed. The impact evaluation study was registered in 3ie's Registry for International Development Impact Evaluations before this analysis (RIDIE-STUDYID-5d567e7e8b967; https://ridie.3ieimpact.org/).

Methods
Our main research questions are 2-fold: first, what is the impact of SHG membership of the respondent woman in the household on short-and medium-term measures of household economic well-being? Second, do outcomes differ for members of SHGs that are in PRADAN areas compared to non-PRADAN areas?
To answer these research questions, we test four sets of comparisons ( Figure 2):

Scale and sustainability 495
Comparison D: SHG in PRADAN areas vs. non-SHG in non-PRADAN areas: SHG membership in an area where PRADAN is present, compared to those who are not members of any SHG and reside in an area where PRADAN is not present.
SHG membership impacts could also vary in intensity depending on the length of the membership. We use self-reported duration of SHG membership to examine heterogeneities along this dimension, as follows: If a woman was a member at baseline but not midline, we use the membership duration reported at baseline, thereby assuming she stopped being a member immediately after baseline. If a woman was a member at midline but not at baseline, we used the duration reported at midline but top-coded the data at maximum 2 years, assuming she became a member after baseline and that the reported inconsistencies in membership were data entry or recall error. Lastly, if a woman was a member at both baseline and midline, we used the lower reported duration of membership. In each case, therefore, we are using the lowest estimate of membership duration.
The length of membership in an SHG is endogenous and both the timing of the decision to join an SHG and the decision to remain an active member are likely to be affected by individual unobserved characteristics. Thus, the duration analysis is intended to provide a descriptive analysis of whether impacts intensify or attenuate with time. These estimates should not be interpreted as being causal.

Matched difference-in-difference
We use difference-in-difference models with nearest neighbor matching (DID-NNM) to provide causal estimates of the impacts of the four types of treatments defined above. Within a district and depending on the comparison being made, treated individuals are matched to non-treated individuals using a set of pre-specified baseline characteristics. We then estimate the impact using difference-in-difference models on this matched sample, comparing changes in outcomes from baseline to endline (2019) between the appropriately defined treated and non-treated individuals. Note that since our treatment definition includes SHG membership at both baseline and midline there is some variation in the duration of membership among the treated; we return to this later.
We use nearest-neighbor matching (NNM) (Abadie & Imbens, 2006) to match each treated individual to one or more non-treated individuals based on pre-intervention individual, household, and community characteristics. NNM selects non-treated individuals that minimize the average difference in characteristics from the treated individual using a multidimensional metric to determine the weights for constructing the average. The treatment effect is measured as the average difference in the outcome for each treated individual from the average outcome among its matched non-treated individuals. Unlike other matching methods, NNM permits the calculation of analytical standard errors. Since we match on more than two continuous variables, we use the bias correction method proposed by Abadie and Imbens (2006).
As a robustness check to the DID-NNM estimates, we also present the estimates from a difference-in-difference estimation using inverse probability weighting (DID-IPW). This method statistically balances a large set of pre-intervention characteristics across treated and nontreated groups and involves weighting each observation (individual) in the non-treated group based on its predicted probability of being in the treated group (or the estimated propensity score). Each observation in the treated group gets a weight equal to one. The impact is then estimated by running an ordinary least squares model on the weighted data. Intuitively, nontreated individuals with similar observable characteristics to treated individuals are assigned higher weights and those that are less similar receive lower weights. The main disadvantage of this method of weighting, as opposed to matching methods, is the higher variance of the estimator (Freedman & Berk, 2008). We address this issue by including the covariates used to estimate the propensity score directly in the weighted regressions estimating treatment effects. This improves the precision of the estimate without affecting the magnitude of the point estimate of the treatment effect (Hirano, Imbens, & Ridder, 2003).
The DID-NNM and DID-IPW methods help account for individual-level selection into SHGs. The non-random placement of SHGs in communities is another source of potential bias, but we believe this is likely to be small in comparison. Following the introduction of the NRLM in 2011, SHGs have rapidly proliferated across the country. In 2015, when the baseline round was conducted, more than 50 million households had already been mobilized into SHGs, with millions more added each subsequent year. With this massive scale-up, the placement bias that might have been a concern earlier in the program is less of a concern. There might still be placement bias in the location of PRADAN's SHGs, since their engagement often predated NRLM. Again, we believe this is not a significant concern for two reasons. First, Table A2 in the Appendix A shows that the block-level matching was successful at achieving balance across both PRADAN and non-PRADAN arms in key individual and household characteristics. Second, PRADAN deliberately entered the most marginalized and vulnerable communities. If the block-level matching exercise did not account for this placement bias, it would be expected to bias our impact estimates downward. Our estimates could then be interpreted as lowerbound estimates of actual impact.

Covariate selection
To match individuals from the treated and non-treated groups, we use the baseline levels of individual, household, and village-level covariates that are likely to affect both the selection into treatment and the outcomes of interest. At the respondent woman level, we include age and its square, years of schooling and its square, binary indicator for being married, number of years married, and whether the mother and father-in-law are members of the household. At the household level, we include indicators for household head being from a Scheduled Tribe (ST) or Scheduled Caste (SC) and from the Other Backward Classes (OBC) categories, 7 the number of female and male members in the household, and the ratio of the number of household members <15 or more than 55 years old to those between the ages of 16 and 55. At the village level, we include the average years of education among women (calculated using the sample of 20 women in each village), indicators for having at least one government primary school, at least one private primary school, at least one Anganwadi centre, having electricity in all areas, and being <20 km from the nearest town. All matching specifications also include district-level geographic dummies.

Outcomes
We consider three groups of outcomes, all expressed as changes from baseline to endline: 1. Household per capita monthly expenditures-total, non-food, food, and animal-sourced foods-each deflated using the Consumer Price Index in 2015 for the general rural population. 2. Household asset ownership-total number of types of assets, number of types of productive assets, and the household wealth score, calculated as the first component of the principal component analysis (PCA) of household asset ownership. The full list of assets has 26 items, including mattress, pressure cooker, chair, cot/bed, bed net (for flies/mosquitos),

Attrition
Because attrition from baseline to endline could bias our results, we present some checks for sample attrition. We interviewed a total of 2744 women at baseline in 2015-2016; of these, 2484 were also interviewed at midline (2017-2018) and endline (2019-2020). Of these 2484, information on outcomes and covariates was available for 2358 women, for an overall attrition rate of 14 per cent over four years. We investigate whether there was differential attrition by treatment status for each of our four comparisons. Table SA.1 presents the results of the probit models estimating the likelihood of re-interviewing the household at endline by relevant individual woman, household, village, and location characteristics (the base model), and by these characteristics along with their interaction with the relevant treatment indicators (fully interacted model). The fully interacted models show that the coefficients on the interaction terms are largely insignificant across the four types of comparisons. It is also clear from the kernel densities of the inverse attrition probabilities for each of the four comparisons used that these probabilities are very tightly distributed close to 1, with an average of 0.92 or above in each case ( Figure SA.1 in the Supplementary Appendix). This gives us confidence that differential attrition is unlikely to be a large concern and that incorporating attrition weights would have a small impact on our estimates.

Results
Table 1 provides the full set of descriptive statistics for each of the four comparisons. Since the results are broadly similar to the other comparisons, we discuss the statistics comparing SHG members and non-members using the full sample. In the full sample, about 61 per cent (1448 of 2358) of women were SHG members either at baseline or at midline. Women who were SHG members at either baseline (2015) or midline (2017) were, on average, about 1.4 years older at baseline than those who were non-members (33.6 vs. 32.2 years). The women in our sample had only slightly over two years of education. About three-fourths of them were employed, based on self-reported occupation status, and almost all were married. Close to 80 per cent of the women were SC or ST, this proportion was significantly higher in the non-SHG sample. The dependency ratio (the number of household members younger than 15 years or older than 55 compared to those aged 16-55 years) was significantly higher in SHG member households.
Most villages had at least one government primary school and at least one Anganwadi centre and about 80 per cent had electricity in all areas. There were no significant differences between villages where the households of SHG members and non-members were located. Table 2 presents the results from the DID-NNM estimations for outcomes listed in Section 3.4, and for the four sets of comparisons between treated and untreated individuals. All tables in this paper follow the same format. Each comparison is in a separate column identified by the same letter (A, B, C, D) used in the definitions of the four treatments above; the numbers presented in the table are treatment effects estimated using a nearest neighbor matching model with outcomes expressed as changes from their baseline levels. These estimates are to be interpreted as the coefficients on a time-treatment interaction term. We discuss the results by type of outcome with each point estimate expressed as a percentage of the mean of that outcome in the comparison group at endline.    Scale and sustainability 501

Household expenditure
The DID-NNM estimation methods in Panel I of Table 2 show that membership in any SHG leads to an increase in per capita monthly food expenditure of INR 51.5 (which translates to 11.7% of the mean per capita monthly food expenditure in the comparison group at endline, p < 0.05) and an INR 16.9 increase in expenditure on animal-sourced food (71.5%, p < 0.05) relative to not being an SHG member (Column A). Restricting the comparison to PRADAN areas, we find the impact of SHG membership on animal-sourced food increases to INR 25.8 (107.5% increase, p < 0.01), while other differences between SHG members and non-members in PRADAN areas on expenditure measures are negligible (Column B). Members of SHGs in PRADAN areas have higher expenditure on animal-sourced foods than SHG members in non-PRADAN areas, Comparison C above (Column C). These differences are likely a result of the larger set of nutrition-related interventions that were implemented in half the PRADAN areas that emphasized the importance of animal-sourced foods, indicating a redistribution from nonanimal-source to animal-source foods rather than an increase in overall food expenses. The last comparison, between women who were members of SHGs in PRADAN areas and those who were neither in PRADAN villages nor members of any SHG at either baseline or midline (Column D) shows the largest impacts of INR 176.2 on total monthly expenditure (20.4%, p < 0.05), INR 103.4 on per capita monthly food expenditure (25%, p < 0.01) and INR 38.1 on per capita monthly expenditure on animal-sourced foods (164%, p < 0.01).

Household asset ownership
Panel II in Table 2 shows the impact of the treatments on household asset ownership and the household's wealth PCA using DID-NNM estimation. We do not find any significant differences between SHG members and non-members either in the full sample or in the sample restricted to PRADAN areas only (Columns A and B). Compared to SHG members in non-PRADAN areas, SHG members in PRADAN areas see a positive impact on asset ownership of 0.3 (Column C, 5% increase, p < 0.05), and a slight positive impact on the wealth PCA. Finally, SHG members in PRADAN areas also see a positive impact on asset ownership of 0.3 compared to non-members in non-PRADAN areas (Column D). The positive impacts on asset ownership suggest that increases in income resulting from being an SHG member are reinvested, potentially contributing to future income streams; that this is seen only in PRADAN SHGs indicates that these groups provide additional inputs or support absent in non-PRADAN SHGs.

Livestock ownership
Households with women SHG members see a significant change in ownership of medium livestock of 0.6 (41% increase, p < 0.01) and in the TLU of 0.15 (10%, p < 0.05), compared to women who are not SHG members (Table 2, Panel III, Column A). The same patterns are observed in the comparison of SHG members and non-members in PRADAN areas only (Column B). The change in ownership of large livestock and TLUs in households in PRADAN areas with SHG members was significant when compared to households with SHG members in non-PRADAN areas (Column C). The final column once again shows the largest differences, with SHG members in PRADAN areas seeing a significant impact on ownership of both large and medium livestock and on overall TLUs when compared to non-members in non-PRADAN areas. Livestock is among the types of assets that women are more likely to own and can more easily accumulate, compared to land. Within this asset category, women are more likely to accumulate small and medium types of livestock (poultry, sheep, and goats) compared to cattle, which are usually considered as 'men's assets' (Quisumbing et al., 2015). The positive impact on large livestock ownership suggests shifting patterns of gendered asset ownership.
Before we discuss the results on the duration of SHG membership, we acknowledge the concern that our study might be under-powered, especially since our sample was not selected with these four comparisons in mind. To check this, we calculate the minimum detectable difference for each outcome across the four comparisons and under two alternate values of power, 0.8 and 0.9 (Supplementary Appendix Table SA.2). In general, we are sufficiently powered to detect the differences we note above for food expenditures and livestock ownership but are under-powered for the asset ownership measures. Our results for asset ownership outcomes should, therefore, be taken as indicative.

Descriptive analyses by duration of membership
Next, we investigate the heterogeneity of our main results to restrictions in the sample based on the duration of SHG membership. Membership duration could affect our outcomes of interest in either direction. Immediately after an SHG is formed, the group establishes norms of group functioning and of regular deposits into the group account; only after the group demonstrates its ability to meet and save regularly are other functions initiated, including lines of credit from formal institutions. While impacts of SHG membership may intensify with the length of membership, initial impacts may also be larger as initial existing constraints on credit are relaxed, and these tapers off over time as SHG saturation increases (Wydick, 2016).
SHG members in our sample differ in the duration of their group membership. The average length of membership among members is 2.8 years, ranging from one month to more than 20 years. There is variation by PRADAN presence-SHG members in PRADAN areas have an average duration of membership of 2.9 years compared to 2.6 years in non-PRADAN areas-   and by the district. Because the duration of SHG membership is likely to be endogenous, these results only suggest the possible direction of association with membership length. Average changes in outcomes between baseline and endline for the full sample, non-PRADAN areas, and PRADAN areas are depicted in Figures 3-5. We divide the sample into four categories: (1) non-members, with zero years of SHG membership; (2) those with some but <2 years of membership, followed by (3) those who have between two and four years of membership, and (4) those with more than four years of membership. These categories were chosen to ensure that each category had a reasonable number of women; they also align roughly with the time intervals between our three rounds of data collection. Nevertheless, the small sample sizes within each category limit our statistical power.
Expenditures tend to increase with SHG membership duration (Figure 3) across all types of comparisons, except for expenditure on animal sourced foods in PRADAN areas. However, expenditure on animal-sourced foods declines between baseline and endline in all categories and could be driven by other factors not captured here. A higher slope in non-PRADAN areas vs. PRADAN areas suggests that returns to the duration of membership accrue more in the former, though with our small sample sizes in each category we cannot test this formally.
Asset ownership shows similar trends (Figure 4), especially for the number of types of all assets owned, which rises monotonically with the duration of membership. Women in non-PRADAN areas who have been members for more than four years own fewer productive assets than those in the same SHGs but with a lower duration, but again, sample sizes are small and confidence intervals large. Wealth PCA also increases with duration, except for 4þ years of membership, where the sharp fall in the point estimate for those in PRADAN areas is likely also driving the fall in the first treatment, that of membership in any SHG. Once again, our study is not sufficiently powered to determine if the declines are statistically significant.
Finally, Figure 5 shows the trends in livestock ownership across membership duration categories. Ownership seems to shift from large livestock to medium and small livestock as the duration of membership increases, especially in PRADAN areas. Women SHG members may be acquiring the means to purchase livestock (or greater control over household income from all sources), since smaller livestock and poultry are easier for women to accumulate. It could also reflect the possibility that secular increases in cropping intensity and mechanization make large livestock, particularly bullocks, less easy to maintain and less crucial for livelihoods. In contrast, small livestock remains a buffer to fluctuations in farm income and require less time and labor for maintenance.
We then investigate the role of duration within the same difference-in-difference with the matching framework. While matching on baseline characteristics could reasonably account for endogenous variation in SHG membership, the decision to remain an active member of an SHG could be influenced by unobservable household characteristics that matching cannot correct for. These estimates should therefore be treated as suggestive rather than causal. Table 3 presents the DID-NNM estimates for the same sets of outcomes and comparisons, but with treatment now defined as being an SHG member for at least 2 years. We compare these results with those in Table 2. Stronger results for those who have been SHG members for at least 2 years when compared to non-members would suggest that the effects of SHG membership compound over time.
Unlike the suggestive trends in the figures, duration does not seem to matter much among those with more than 2 years membership. Using DID-NNM, point estimates on most expenditure indicators (Table 3, Panel I) are either very similar to or even slightly smaller than those for the full sample of SHG members we saw in Table 2. The point estimates on asset ownership (Panel II), where significant, are slightly larger than the corresponding estimates in Table 3. We see stronger associations with the ownership of medium livestock (Panel III) in all four Table 3. DID-NNM impact estimates for baseline to endline change in expenditure outcomes, treatment defined as those who have been SHG members for at least 2 years Baseline (2015) or Midline (2017)  Scale and sustainability 507 comparison groups, this also seems to drive larger impacts on total livestock units. Taken together, our results suggest that the duration of SHG membership may not matter for expenditure and assets beyond a two-year threshold but might have implications for livestock ownership. One hypothesis is that impacts in the early years of participation are concentrated on shorter-term household level outcomes, like expenditures and consumer durables that meet a more immediate need. Over time, women's increased autonomy and control of their finances allow them to instead accumulate more assets they can control, such as small livestock.

Alternate estimation models
As a final check on the robustness of our main results, we re-estimate impacts using differencein-difference with inverse-probability weighting (DID-IPW). Tables SA.3 and SA.4 in the Supplementary Appendix present the DID-IPW results for the same outcomes and the same sets of comparisons as Tables 2 and 3. With the full sample, the DID-IPW estimates on the expenditure outcomes are slightly larger in magnitude (Panel A, Table SA.3), except for expenditure on animal-sourced foods. The DID-IPW results for both asset and livestock ownership are comparable in magnitude, but with a larger number of statistically significant differences between treatment and control than with DID-NNM. The same patterns also hold for Table SA.4 where the sample is restricted to those who have been SHG members for more than 2 years; compared to the corresponding estimates in Table 3, DID-IPW estimates are broadly similar in magnitude and statistical significance.

Discussion
We show conclusively that belonging to an SHG of any type has positive impacts. Compared to households where the woman was not a member of an SHG, households with SHG members experience larger changes over time in household per capita monthly expenditure on food, especially animal-sourced foods, and livestock ownership, driven mostly by household ownership of medium-sized livestock. However, there were no statistically significant differences between members and non-members in other expenditure and asset-related outcomes, although the absence of significant results for assets may be driven by a lack of statistical power. We also note some differences in impacts depending on the organization supporting the SHGs. Membership in a PRADAN SHG positively affects per capita monthly expenditure on animal-sourced foods, the total number of types of assets owned and the number of large livestock owned, compared to membership in a non-PRADAN SHG. The impact on animal-sourced foods could be driven by the additional health and nutrition-related interventions conducted in half of the PRADAN areas, the subject of the larger impact evaluation from where our data is drawn, or by women's increased ownership of livestock, which increases access to these foods. As expected, the contrast is greatest when comparing PRADAN SHG members with non-members in non-PRADAN areas. Impacts of SHG membership also appear to increase with the length of membership, up to a certain threshold, although our sample of members is too small to test this rigorously.
What about SHG membership potentially leads to improved welfare outcomes, such as increase in food expenditures, expenditure on animal source foods, and livestock ownership? Several pathways exist from SHG membership to well-being outcomes (Kumar et al., 2018), through enhanced incomes from savings and access to credit, improved agriculture and livelihoods, changes to health and nutrition knowledge and behavior, and an enhanced understanding of rights and entitlements. In addition, there are cross-cutting pathways to building social capital, improving collective action, and enhancing women's empowerment. Examining evidence along each of these pathways is beyond the scope of this paper, instead, we draw on our other work using data from the same evaluation. Kumar et al. (2019) show that women who Scale and sustainability 509 are SHG members are more likely to be aware and make use of entitlement programs, cash or in-kind transfers, and workfare programs; SHG members also have wider social networks and greater mobility compared to non-members. Raghunathan, Kannan, and Quisumbing (2019) show that participation in SHGs improves women's access to information, their participation in making certain agriculture-related decisions, and household-level measures of production diversity. Kumar et al. (2021) find that SHG members exhibit greater control over income, greater decision-making over credit, and greater and more active involvement in groups within the community relative to non-members.
Since these studies use baseline, midline, or endline data from the same evaluation that the analysis in this paper draws on, they provide compelling support to possible mechanisms: over time, participation in SHGs confers the benefits of improved access to information, including on own rights and entitlements, greater participation in decision-making on agriculture and livelihoods, wider social networks, greater mobility, and higher overall levels of empowerment. The savings and credit 'core' SHG activities have a direct link to the expenditure, asset, and livelihood outcomes we discuss in this paper. The additional pathways described above could serve to enhance these impacts and improve the use of additional resources.
Other work on SHGs-in India, South Asia and globally-supports the evidence generated by this evaluation on the indirect pathways through which SHGs can enhance individual and household well-being. Notably, Brody et al. (2017) discuss the evidence of the positive impact of women's groups on various aspects of empowerment, and Diaz-Martin, Gopalan, Guarnieri, and Jayachandran (2020) find that financial groups (including SHGs) increase women's participation in economic activities through improved access to credit or assets. Pandey et al. (2019) also find positive impacts of SHG membership on women's labour force participation in India. In the context of interventions in other countries, women's groups have been shown to provide mutual insurance, thereby permitting consumption smoothing in times of shocks (Malde & Vera-Hern andez, 2022 in Malawi) and to successfully manage common property resources, attributed at least in part to the trust and transparency they inspire (The Water Trust, 2020 in Uganda). This illustrates the impacts of SHGs beyond credit provision.
Our analysis has several limitations. Since SHGs had already been functional in these areas before baseline data collection, we use quasi-experimental methods to attribute causality. We do not have data before our baseline to test parallel trends and rely instead on evidence of baseline balance across treatment and control arms. Importantly, our sample was not originally designed to answer these specific questions about the impact of SHG membership. This analysis was possible because we had a relatively balanced sample of SHG members and non-members, but we are under-powered to detect impacts on some outcomes. Nor can our duration analysis be interpreted as causal due to small sample sizes. Finally, we did not collect information on some variables that would have improved our understanding of mechanisms, such as intrahousehold labour supply or individual-level consumption expenditures.
Despite these caveats, our findings add to growing global evidence of the effectiveness of women's groups in achieving positive outcomes beyond savings and credit. Given the ubiquity of SHGs, their constant evolution and repurposing as a means of delivering a range of interventions, and the amount of money national and state governments and international donor organizations have committed over the last decade, evaluations of the impact of SHGs on household and individual welfare are needed to inform whether and how to support these platforms. Tracking the short-, medium-and long-run impacts of these groups, especially as they scale up or approach saturation, is particularly valuable to understanding how impacts intensify or change over time, both as the microfinance model becomes increasingly embedded into policies and programs, but also from accompanying shifts in decision-making, mobility, networks, and asset accumulation patterns as women gain experience within the program. The potential for SHGs to improve other dimensions of household welfare is enormous-how to best harness that to achieve outcomes at scale is a promising area for future research. 3. We do not analyze the impact of the additional health and nutrition-focused training here; see Scott et al. (2022) for details. 4. Control blocks were oversampled to allow for better matching across arms at the time of analysis. 5. Spillovers across blocks-which are much larger administrative units-are far less likely, especially since the blocks chosen for inclusion in the study arms are not always contiguous. 6. A caveat for this comparison is that PRADAN's approach to programming is one where their SHG members become agents of change and disseminate messages throughout their communities. Therefore, if this leads to spillovers to non-SHG members then this comparison will underestimate the impact of SHG membership. 7. The Scheduled Caste and Tribe groups, SCs and STs, are officially designated groups delineated in the Constitution of India. These groups are recognized as being historically marginalized and have been accorded political, occupational and educational reservations. These groups typically lag behind the Other Backward Class and the General caste categories in a range of socioeconomic and demographic criteria. 8. We only collect information on whether a household owns at least one of each type of asset and not the number of each type owned, so this is a lower bound on household asset ownership. 9. The common standard used for one TLU is one cattle with a body weight of 250 kg. In Asia, the TLU weight applied for the quantity of cows is 0.5 while the one applied for the quantity of goats is 0.1. This means 5 sheep of 25kg will consume as much as 1 cow of 125kg in Asia (Njuki et al., 2011). agriculture domains was taken from the Indicus Analytics (2013) report. This report draws on nationally representative secondary data sources, such as the Census of India, the Agricultural Census, and the National Sample Survey Organization's Consumption Expenditure Survey to rank districts as poor (bottom tercile), average (middle tercile), or high (upper tercile) on each domain, using a comprehensive set of indicators listed in Table A1 combined using principal component analysis. Given the small number of blocks used for the matching exercise, we used a simplified matching process. We scored poor, medium, and high levels on each domain as 1, 2, and 3, respectively, and constructed an equally weighted sum of scores across domains. Using these scores, we first selected from among the blocks where PRADAN was conducting its standard SHG interventions that block with the closest overall score to the (pre-determined) block where the nutrition intensification interventions were conducted. We followed the same procedure to select the control block with the closest score to the two PRADAN blocks. Where there was more than one such control block (or where data on one or more domains was missing), we consulted PRADAN and chose the best match based on their advice.  Notes: Authors' calculations. The sample includes SHG members and non-members and all those sampled at baseline and is indicative of the success of the block-level matching exercise. Superscripts indicate statistically significant differences (at 10% or less) across arms in pairwise comparisons, where a: Control vs.

Standard, b:
Control vs. Nutrition Intensification, and c: Nutrition Intensification vs. Standard. Where sample sizes differ due to conditional responses, they are indicated after the variable name in the same order as the columns. These treatment arm definitions are not identical to those used in the paper, as these relate to the larger impact evaluation.