Our study was conducted in Kerala's district of Wayanad. The population is largely dependent on agriculture with the main crops being coffee and paddy. The study site was a single Panchayat (territorial decentralised unit) with a land area of 31.75 sq km and a population of 16,110 individuals. Forty-three percent of households were classified as below the poverty line at the time of the study in 2003. The classification of households into poor and non poor in India is not based on a direct assessment of income. Households are classified below the poverty line according to a series of indicators, including education, presence of disability, social group affiliation, dwelling type, land and livestock owned, formal training for skill development of household members, and consumption expenditure.

Women's health is unevenly distributed in the Panchayat; the prevalence of poor health was higher among women of low socioeconomic position and with low caste affiliations 11. There is a high density of public and private health care facilities and geographical access to care is evenly distributed among the population. The poor, however, often face difficulties in obtaining services because of financial barriers. Rapidly rising health care costs have been observed in Kerala 17. And although the services are supposedly free of charge, deficiencies in the quality of the public system force the poor to turn to the private sector to purchase medication and other services.

In 1995, SHGs began operating in the area. This initiative emerged from within the community, initially led by a small group of women with local NGO support. Following this, there was an incremental expansion of SHGs. Later, in 2002, kudumbrasree was introduced. Both networks follow similar procedures: small groups of women engage in savings and loan activities; their weekly contributions are deposited in a commercial bank. After an initial savings period (typically 6 months), members are eligible to take loans. Each SHG sets their interest rates and procedures for loan allocation. Any woman 18 years and older can participate, whether she is from a household that is below or above the poverty line. Due to the similarities of the NGO supported groups and kudumbrasree, we do not discriminate between the networks in this study (hereafter both networks are collectively referred to as SHGs). Basic information on participation and loans are presented in Table 1.

Cross-sectional data were used from a household survey implemented in 2003 as part of a our (Centre for Development Studies and Université de Montréal) action research project. The project obtained ethical approval by the Université de Montréal Ethics Committee on April 25, 2003. Trained local female surveyors canvassed all 3,352 households identified in the Panchayat. The household questionnaire has several modules, including questions pertaining to demographics, socioeconomic characteristics, health, and SHG participation. One woman from each household, the head or spouse of the head, was also invited to participate in a women's well-being module, which collected information on markers of mental health and women's decision-making agency. To maintain privacy, women were encouraged to respond to this module separately from other household members.

The study sample included non-elderly females (18 to 59 years) who responded to the women's well-being module for themselves. Paniya women (N = 647), a particularly deprived and socially marginalized group, were not included in the study. Large socioeconomic and cultural differences between groups can increase difficulties in comparing health status when using self reported health 18. We detected that Paniyas were rating their health in a distinct way that underestimated their health; therefore, we did not include Paniya women in our analysis 11. Finally, because our hypotheses relate specifically to poor women, we only included women from households below the poverty line.

SHG participation was measured as a three level variable in order to consider duration of participation – the longer a woman has participated, the more likely she may benefit 16. The categories were: early joiners (members for more than 2 years), late joiners (members for 2 years and less), and non-members. In our models for exclusion to health care, we further distinguished non-members between women with and without another household member participating in a SHG because exclusion was measured at the household level; therefore, women who have another household member participating in a SHG could indirectly benefit. We gathered additional background information on SHG participation, including the number and purpose of loans of each woman and the reasons for non-participation at the household level.

A variety of indicators can be used to measure socioeconomic position. We used three indicators that we believe are pertinent for Indian agrarian societies: education, employment status, and size of land holdings 11. We measured the size of household landholdings in cents (100 cents equals one acre). We also included the specific caste of the household, which was surveyed and categorized using the conventional three-way classification system adopted in Kerala, which ranks Hindu castes and other religions. The first category, at the bottom of the caste hierarchy, includes both Scheduled Castes (SC) and Scheduled Tribes (ST), or "SC/ST". Next is a residual category of lower castes and Muslims, known as other backward castes or "OBC". The highest ranking group are the upper or "forward castes", including Christians. Finally, we controlled for age and the position of women in the household (i.e. head or spouse of the head of household).

Data analysis followed a three step process (Figure 1). To illuminate potential selection bias in our sample we began by examining the socioeconomic and demographic characteristics of SHG participation (S1). Basic cross-tabs with chi-squares were used to develop a portrait of the characteristics of participation, followed by multinomial logistic regression analyses to test whether SHG participation (early joiner, late joiner, non-participant) varied according to these characteristics. We also assessed reasons reported for non-participation and the number of and purposes for loans among SHG members.

We then tested the study's hypotheses: SHG participants face less exclusion to health care, lower health risks, and are less likely to engage in male dominated decision-making (A1); SHG participants have greater health achievements, mediated by the determinants of health examined in A1 (A2). Due to the nature of the variables (categorical or scores with little distribution), a series of binomial logistic regressions were performed for each of the dependent variables. Associations between participation and health determinants (A1) adopted a three step approach to modelling. In step one, only SHG participation was included (Model 0), allowing us to assess the independent effects of SHG participation. Steps two and three adopted a sequential approach: socioeconomic characteristics and caste were entered in the model (Model 1), then, SHG participation was added (Model 2). We examined the sequential models in two ways. First, we employed goodness-of-fit tests to observe whether SHG participation significantly improved Model 1 by testing whether the deviance was statistically significant between Model 1 and Model 2. Second, we examined the odds ratios to assess the associations between being an early or late joiner and each of the women's characteristics and the dependent variable.

Modelling of health achievements (A2) followed a similar approach using four steps. In step one, only SHG participation was entered (Model 0). A sequential approach was then followed: socioeconomic characteristics and caste were entered in the model (Model 1), SHG participation was added (Model 2), and finally the determinants of health (exclusion, decision-making, health risks) were added as explanatory variables (Model 3). This approach allowed us to test the effect of our variable of interest, SHG participation, on our dependent variables, while assessing the influence of other determinants of health. Deviances comparing Model 1 and Model 2 indicate whether or not SHG participation significantly contributes to the model and deviances comparing Model 2 and Model 3 indicate whether, globally, the block of determinants of health significantly improve the model. Odds ratios were examined as in A1. All multivariate analyses controlled for age and women's position in the household. Data analysis was performed using SPSS version 14.0 20.

There are 336 women who did not participate in a SHG, of which 100 women are from households already containing a SHG participant. The main reason reported for non-participation was financial barriers (over 50%), confirming the need to control for socioeconomic characteristics in our analyses. Three percent of households reported ill-health as a reason for non-participation. Sensitivity analysis showed that including women from these households did not affect our results; therefore, these women were retained in the analysis.

Over half of the women are members of a SHG (150 early joiners and 442 late joiners). Almost 75% of members had received at least one loan (91% of early joiners and 67% of late joiners). In addition to productive activities, 42% of women who had ever received a loan reported health consumption purposes.

Women are less likely to be a SHG member if they are in the youngest or oldest age category and had less than a high school education (Table 3). There are no significant associations between SHG participation and employment status, household landholdings, or caste.

The results of the models of SHG participation are presented in Table 4. Having no education is associated with a lower odds of being an early or a late joiner, compared to non-participants (reference group). The odds ratio is lower for early joiners than it is for late joiners, suggesting a gradient between education and SHG duration. Having a primary education is associated with a lower odds of being an early joiner compared to non-participants. Women belonging to the middle age category (31 to 44 years) have a higher odds of being an early or late joiner than younger women (under 30 years). These results indicate that education and age are influencing women's self selection into a SHG and confirms the need to control for these characteristics in our analyses.

The results of the models of exclusion to health care, exposure to health risks, and male decision-making are shown in Tables 5, 6, 7.

Thirty-five percent of women come from households reporting at least one episode of exclusion to health care. The odds ratios in Model 0 suggest that SHG participation is associated with lower rates of exclusion: compared to the reference group, which in this case is non-participants living in a household without a SHG member, the odds of facing exclusion is significantly lower among early joiners (OR = 0.56, CI = 0.36–0.86), late joiners (OR = 0.57, CI = 0.41–0.79), and non-participants who live in a household with a SHG member (OR = 0.58, CI = 0.35–0.94). Belonging to a household with small landholdings (i.e. less than 50 cents of land) and having OBC affiliations is associated with exclusion (Model 1). Notably, there were no statistically significant differences between forward caste and SC/ST women – who rank the lowest on the caste hierarchy. This finding may be attributed to the poorest tribal group (the Paniyas) not being included in the sample. Model 2 shows that after adjusting for women's characteristics, SHG participation is a significant factor for exclusion to care (deviance = (24.2 (3)). The odds ratios are similar to the estimates of Model 0: the odds of facing exclusion is significantly lower among early joiners (OR = 0.58, CI = 0.41–0.80), late joiners (OR = 0.60, CI = 0.39–0.94), and non-participants who live in a household with a SHG member (OR = 0.53, CI = 0.32–0.90) (M2).

Perceived exposure to health risks was reported by 22% of the women. Exposure to health risks is not significantly associated with any women's characteristic (Model 1) or SHG participation (Model 0 and Model 2).

Globally, there appears to be a high level of decision-making agency, only 12% of women reported male decision-making. As an independent predictor, SHG participation was not significantly associated with decision-making (Model 0). There is a lower odds of reporting male decision-making if a woman is engaged in paid employment (Model 1). Models One and Two are significantly different (deviance = 10.4(2)). After adjusting for women's characteristics, we found a lower odds of reporting male decision-making if women are late joiners (OR = 0.62, CI = 0.39–0.97), but contrary to our expectations, we found no significant associations between decision-making and being an early joiner (Model Two).

This section presents the results of the binomial logistic regressions for A2 for each of the four health achievements (self perceived health, limitations in ADLs, emotional stress, life satisfaction).

Results for models of self assessed health are presented in Tables 8 and 9. Thirty-five percent of women reported bad health. We found no associations between self perceived health and SHG participation (Model 0). There are significantly greater odds of reporting bad health if a woman is not engaged in paid employment and comes from a household with small landholdings (Model 1). Even after adjusting for women's characteristics, SHG participation is not associated with perceived health (Model 2). As a block, the determinants of health significantly contribute to the model (deviance = 366.0(3)). There are robust associations between bad health and all three health determinants (Model Three). A woman has a significantly greater odds of reporting bad health if she faces exclusion to health care, is exposed to health risks, and reported male decision-making. Limitations in ADLs were reported by 41% of women. The logistic regression results for limitations in ADLs are comparable to those for bad health.

Table 10 and 11 shows the results for the markers of mental health. Eighty-eight percent of women reported emotional stress. SHG participation as a sole independent variable was not found to be associated with emotional stress (Model 0). There is a significantly greater odds of reporting emotional stress if a woman is engaged in paid employment and if she comes from a household with small landholdings (Model One). After controlling for women's characteristics, SHG participation significantly improves the model for emotional stress (deviance = 18.6(2)). Inspection of the odds ratios highlight a striking result: although we find no associations between emotional stress and being a late joiner, the odds of reporting emotional stress is significantly lower for early joiners compared to non-participants (OR = 0.52, CI = 0.30–0.93) (Model Two). The determinants of health significantly improve the model (deviance = 55.8 (3)).