The previous chapter provided an overview of the range of different CSR-initiatives which Bayer Crop Science has introduced in the two model villages since 2011 to address challenges at the Base of the Pyramid. The empirical analyses in this chapter will examine whether participation of the villagers in these different activities has contributed to an improvement of reported well-being of the villagers or not. In this paper, we focus on an aggregate evaluation of the MVP activities. Therefore, in the two model villages we will distinguish between “MVP participants” who report that they have participated in at least one of the MVP activities and “Non-participants” who report that they have not participated in any activity so far.
The empirical analyses are based upon indirect assessments (1) by comparing reported well-being changes in the model villages and the control villages and (2) by contrasting reported well-being development of participants and non-participants in the two model villages. The measurement of subjective well-being change from 2011 to 2014 is based upon a simple question asked to all villagers interviewed in the two model and the two control villages: “And you personally: would you say your personal well-being has improved or worsened in the last three years?” Villagers have answered this question on a 5-digit-scale ranging from 1 “very much improved” to 5 “very much worsened”. Thereby, the question has been asked in a general context which has not been related to any concrete CSR activities.
Descriptive evidence
The following results are based on those 1812 individuals from 922 panel households who have been interviewed in the four villages in 2011 and in 2014. In total, almost 70% of the villagers interviewed in all four villages answer that their personal well-being has a bit or even very much improved during the recent three years. 29% report that there has been no change, only slightly more than 1% are reporting a lower personal well-being. This positive development of perceived well-being in the four villages might be explained by a variety of different causes, among them e.g. a general positive economic development in India and in rural Karnataka in these years. However, if there are significant differences between the well-being development in the two model villages and the corresponding development in the two control villages, this can be interpreted as a hint for positive effects of the MVP activities.
The results of our empirical survey demonstrate that indeed the share of villagers who report well-being improvement in the two model villages is 78% and thus substantially higher than the corresponding share of about 62% in the two control villages. A simple test of equality of the shares proves that the difference of 16 percentage points is highly significant. This significant difference in shares can also be observed for each of the model villages in contrast to each of the control villages. (82% in Kadivala and 77% in Mangalagudda compared to 65% in Chimalaggi and 59% in Chikanal).
Within both model villages the MVP participants report a better well-being development than non-participants: while 84% of the participants in Kadivala answer that their well-being has a bit or even very much improved since 2011, it is only 66% of the non-participants. In Mangalagudda, the corresponding shares are 83% and 69%. Thus, in both model villages, the differences between the shares for participants and non-participants are highly significant.
To summarize, for both evaluation strategies our simple descriptive comparisons show that reported well-being development has been better in model villages compared to control villages and for participants in model villages compared to non-participants in the same villages. This can be interpreted as a first empirical hint that the MVP activities might have contributed to this difference. For a more profound analysis of a possible causal impact of the MVP activities on subjective well-being different types of multivariate discrete choice models will be estimated to be able to control for differences in socio-economic and socio-demographic characteristics.
Multivariate analyses
To analyze potential impacts of the MVP activities on the change of subjective well-being from 2011 to 2014 multivariate discrete choice models are estimated that take into account that the dependent variable is a categorical variable (see e.g. Greene
2018, chapters 17 and 18). To examine the robustness of our empirical findings we apply three different regression methods.
We will first estimate simple logit regression models in which the dependent variable is a binary variable. We therefore aggregate the answers of the relevant question on well-being change to only two categories 0 “no change of well-being or even worse” and 1 “well-being has a bit or very much improved”. Positive coefficients of the independent variables thus imply a positive impact of the variable considered on reported well-being changes.
Second, we alternatively estimate ordered logit models which are discrete choice models for three or more ranked outcomes. In this regression, the dependent variable “reported change of personal well-being from 2011 to 2014” is therefore condensed to the following three outcomes 1 “no change or even worse”, 2 “a bit improved” and 3 “very much improved”.
While the ordered logit model takes into account the ranking of the outcomes and estimates one coefficient for one explaining variable measuring the average impact of an explaining variable on the dependent variable with three or more outcomes, multinomial logit models are more flexible in this respect as they treat the dependent variable only as nominal. This implies that we will have the same three outcomes as for the ordered logit model, but that we allow the estimated coefficients to vary between different outcomes compared to a base category. “No change or even worse” is taken as base category in our following estimations. The estimation results of multinomial logit models will thus allow us to examine the impact of an explaining variable whether – compared to the base category – differs for the outcomes “well-being has improved a bit” and “well-being has very much improved”.
To assess a potential impact of the MVP activities on subjective well-being, the following explaining variable is created. Persons living in the control villages are treated as a reference group as they could not profit from any MVP activities. These are about 52% of the villagers interviewed. Additionally, we distinguish between the following groups of villagers:
-
Non-participants in the model village Kadivala (1.8% of all villagers)
-
Non-participants in the model village Mangalagudda (14.5% of all villagers)
-
Participants in the model village Kadivala (14.1% of all villagers)
-
Participants in the model village Mangalagudda (17.4% of all villagers)
To control for a possible impact of important socio-demographic and socio-economic variables that are known from existing studies on subjective well-being to possibly influence personal well-being (see e.g. Blanchflower and Oswald
2011) a set of control variables presented in Table
1 (with its descriptive statistics) will be incorporated into our models.
Table 1
Socio-demographic and socio-economic control variables as determinants of personal well-being, descriptive statistics from the estimation sample
Model vs. control village and participation in MVP activities | 1 = person from control village | 0.522 | 0.500 | 1812 |
2 = non-participant, Kadivala | 0.018 | 0.132 | 1812 |
3 = non-participant, Mangalagudda | 0.145 | 0.352 | 1812 |
4 = participant in at least one MVP activity, Kadivala | 0.141 | 0.348 | 1812 |
5 = participant in at least one MVP activity, Mangalagudda | 0.174 | 0.384 | 1812 |
Socio-demographic variables
|
Gender | 0 = male | | | |
1 = female | 0.536 | 0.497 | 1812 |
Age | 0 = 15- < 25 years | 0.144 | 0.351 | 1812 |
1 = 25 - < 40 years | 0.343 | 0.475 | 1812 |
2 = 40- < 55 years | 0.287 | 0.452 | 1812 |
3 = 55+ years | 0.226 | 0.419 | 1812 |
Marital status | 0 = never married, | 0.094 | 0.292 | 1812 |
1 = currently married, | 0.810 | 0.393 | 1812 |
2 = widowed, divorced, separated | 0.097 | 0.295 | 1812 |
Caste | 0 = Upper caste | 0,077 | 0.266 | 1812 |
1 = Other backward caste | 0.650 | 0.477 | 1812 |
2 = Scheduled tribes | 0.077 | 0.266 | 1812 |
3 = Scheduled castes | 0.149 | 0.356 | 1812 |
4 = Other or Muslim / Christian | 0.048 | 0.213 | 1812 |
Socio-economic variables
|
Low personal spendings | 0 = able to spend 420 INR per month individually | | | |
1 = not able to spend 420 INR per month individually | 0.450 | 0.498 | 1812 |
Land and livestock ownership of the household | 0 = owns land and livestock | 0.580 | 0.494 | 1812 |
1 = owns land or livestock | 0.260 | 0.439 | 1812 |
2 = owns neither land nor livestock | 0.160 | 0.367 | 1812 |
Debt problems sometimes or often | 0 = No | | | |
1 = Yes | 0.603 | 0.489 | 1812 |
No work | 0 = does work | | | |
1 = does not work | 0.095 | 0.294 | 1812 |
Illiteracy | 0 = can read and write | | | |
1 = cannot read and write | 0.632 | 0.482 | 1812 |
Illness | 0 = does have no illness often | | | |
1 = has at least one illness often | 0.211 | 0.408 | 1812 |
Other things equal, if there is a positive impact of the MVP initiatives on subjectively reported well-being, we should expect the following estimation results:
Hypothesis 1: Participants of MVP activities in the two model villages are expected to report better well-being changes than people in the two control villages.
Hypothesis 2: Participants of MVP activities in each of the two model villages are expected to report better well-being changes than non-participants in the two model villages.
Table
2 presents the results of the estimation of all different types of discrete choice regression models. Thereby, time-varying variables all refer to the base year 2011. As independence of observations is not given for people from the same household, our estimation of standard errors accounts for possible intra-household correlation of standard errors for all regression methods. For the multinomial logit model, a Wald test for combining the three alternatives concludes that the dependent categories are distinguishable and should not be combined.
Table 2
Results from multivariate analyses. Estimated coefficients
Variables on MVP participation |
Person from Kadivala, no participation | 0.284 | 0.295 | 0.255 | 0.631 |
(Reference group: person from control village) | (0.50) | (0.46) | (0.55) | (0.44) |
Person from Mangalagudda, no participation | 0.345 | 0.274 | 0.350 | 0.362 |
| (0.06)* | (0.11) | (0.06)* | (0.35) |
Person from Kadivala, participation in at least 1 activity | 1.276 | 1.431 | 1.112 | 2.468 |
| (0.00)*** | (0.00)*** | (0.00)*** | (0.00)*** |
Person from Mangalagudda, participation in at least 1 activity | 1.070 | 0.912 | 1.034 | 1.470 |
| (0.00)*** | (0.00)*** | (0.00)*** | (0.00)*** |
Socio-demographic variables |
Women | 0.192 | 0.164 | 0.188 | 0.249 |
| (0.11) | (0.13) | (0.12) | (0.25) |
25- < 40 years (reference: 15- < 25 years) | −0.405 | −0.274 | −0.411 | −0.355 |
| (0.07)* | (0.13) | (0.07)* | (0.33) |
40- < 55 years (reference: 15- < 25 years) | − 0.540 | − 0.403 | − 0.538 | − 0.551 |
| (0.03)** | (0.04)** | (0.03)** | (0.16) |
55 years+ (reference: 15- < 25 years) | −0.838 | −0.710 | − 0.823 | −1.007 |
| (0.00)*** | (0.00)*** | (0.00)*** | (0.02)** |
Currently married (reference: never married) | 0.495 | 0.458 | 0.467 | 0.749 |
| (0.04)** | (0.03)** | (0.06)* | (0.09)* |
Widowed, Divorced, Separated | 0.198 | 0.254 | 0.142 | 0.765 |
(reference: never married) | (0.53) | (0.39) | (0.66) | (0.22) |
Other backward castes (reference: upper castes) | −0.084 | −0.291 | −0.011 | −0.561 |
| (0.79) | (0.30) | (0.97) | (0.21) |
Scheduled tribes (ST) (reference: upper castes) | −0.339 | −0.646 | − 0.207 | −1.783 |
| (0.36) | (0.06)* | (0.57) | (0.03)** |
Scheduled castes (SC) (reference: upper castes) | −0.104 | −0.255 | − 0.044 | −0.431 |
| (0.75) | (0.40) | (0.89) | (0.39) |
Other castes or Muslims/Christians (ref.: upper castes) | −0.725 | − 0.917 | −0.632 | −1.433 |
| (0.06)* | (0.01)** | (0.10)* | (0.10)* |
Socio-economic variables |
Personal income: not able to spend 420 INR or more individually per month in 2011 (1=Yes) | −0.256 | −0.210 | −0.255 | − 0.277 |
| (0.03)** | (0.05)** | (0.03)** | (0.19) |
Sometimes or often debt problems (1 = Yes) | −0.347 | −0.244 | − 0.360 | −0.208 |
| (0.01)*** | (0.04)** | (0.01)*** | (0.40) |
Livestock or land (1 = Yes, ref.: livestock and land) | −0.110 | − 0.101 | −0.109 | − 0.116 |
| (0.47) | (0.48) | (0.47) | (0.68) |
No Livestock and no land (1 = Yes) | −0.606 | −0.596 | − 0.576 | −0.942 |
| (0.00)*** | (0.00)*** | (0.00)*** | (0.01)*** |
No work in 2011 (1 = Yes) | −0.387 | − 0.303 | −0.410 | − 0.166 |
| (0.03)** | (0.09)* | (0.03)** | (0.64) |
Illiteracy in 2011 (1 = Yes) | −0.302 | −0.323 | − 0.269 | −0.636 |
| (0.02)** | (0.01)*** | (0.04)** | (0.01)*** |
Person has at least one illness often in 2011 | −0.181 | − 0.241 | −0.146 | − 0.564 |
| (0.18) | (0.05)* | (0.29) | (0.04)** |
Constant(s) | 1.328 | −1.391 | 1.185 | −0.917 |
| (0.00)*** | (0.00)*** | (0.00)*** | (0.15) |
| | 2.297 | | |
| | (0.00)*** | | |
Number of observations | 1812 | 1812 | 1812 | 1812 |
Wald test for model | 128.17 | 160.14 | 173.01 | |
| (0.000)* | (0.000)*** | (0.000)*** | |
Pseudo R2 | 0.0792 | 0.0665 | 0.0744 | |
We start the following interpretation of the results for the two core hypotheses before we refer to the estimation results for the control variables.
Starting with hypothesis 1, all estimation results strongly support the hypothesis that – other things equal – reported well-being changes of MVP participants in the model villages are significantly better than those of villagers in the control villages. According to the simple binary logit regression MVP participants in Kadivala have a c. p. 24.3 percentage points higher probability to report that their well-being has a bit or very much improved than villagers in the control villages.
10 In Mangalagudda the corresponding marginal effect is also highly significant and reaches 20.4 percentage points. Both the ordered logit regression and the multinomial logit model strongly support this conclusion. The probability to report no change or even worsening of well-being in Kadivala is 27.4 percentage points lower than in the control villages for ordered logit regression and 23.4 percentage points lower for the multinomial logit estimation. For Mangalagudda the corresponding marginal effects are − 17.5 percentage points and − 20.4 percentage points. Moreover, multinomial logit results illustrate that compared to control villages in both model villages the significant impact of MVP participation both holds for smaller and for more substantial well-being improvements.
Also, hypothesis 2 is strongly supported both for Kadivala and for Mangalagudda and for all types of estimation methods. Table
3 presents the
p-values for Wald tests for equality of coefficients for non-participants and participants in the model villages: within Kadivala, the estimation results show that participants of the MVP activities report – other things equal – a significantly better well-being change than non-participants. A test on identity of the two estimated coefficients illustrates that the difference is significant at a 5%-level of significance for all types of estimations, only for small improvements in the multinomial logit model the p-value is 0.06. Within Mangalagudda, the corresponding differences are even significant at a 1%-level of significance, i.e. also MVP participants in Mangalagudda have a significantly better reported well-being change than non-participants.
Table 3
Wald tests for equality of coefficients for non-participants and participants in the model villages
…participants and non-participants in Kadivala | 0.026** | 0.011** | 0.060* | 0.026** |
…participants and non-participants in Mangalagudda | 0.001*** | 0.002*** | 0.002*** | 0.007*** |
Moreover, the estimation results reveal that c. p. even non-participants of the MVP in Mangalagudda report weakly significant better well-being development than people in the control villages. This might be explained by “positive spill-overs” within the model village to non-participants. For example, spill-over effects may occur if parents with small children appreciate additional corporate educational activities, even if their children are still not enrolled and do not participate and benefit from these activities. For non-participants in Kadivala, the corresponding estimated coefficient is also positive, but not significantly different from zero.
With respect to the socio-demographic control variables the estimation results indicate that reported well-being development is the better the younger the villagers are: young villagers aged 15–24 years have the highest probability of well-being improvements across the different estimation models.
11 According to logit regression the marginal effect for well-being improvements is about 16 percentage points lower for 55+ years old villagers than for the young villagers. Multinomial logit regression estimates show that this marginal effect is mainly driven by a 13.4 percentage points lower probability for small improvements.
For men and women, the results do not give hints for significant gender differences. Even if the estimated coefficients seem to indicate that women have a higher probability of reporting well-being improvements than men, this difference is never significant at a 10%-level of significance across estimation models.
Villagers who are currently married report better well-being developments than villagers who were never married. The estimated coefficients are significantly different from zero at a 5%-level of significance for the logit and the ordered logit model, and weakly significant at 10% for both reported outcomes of multinomial logit estimation.
With respect to castes there are only a few estimated coefficients that are significantly different from zero. Even if the results show a certain tendency that reported well-being development in SC, ST, in other backward castes and for Muslims/Christians/other castes has been a bit worse at least than in upper castes, significance is mostly not given at the 10%-level of significance. Only Muslims, Christians (including other castes) have a significantly lower probability of reported well-being improvements than upper caste villagers in all models.
12
Socio-economic control variables are taking into account that the initial economic situation of the villagers can have been very different thereby impacting the opportunity of future well-being changes.
Villagers who were not able to individually spend more than 420 INR per month in 2011 are shown to have a worse reported well-being development until 2014 than those who were. The marginal effects are rather low with – 4 to 5 percentage points. The multinomial logit model shows that the difference mainly depends on differences in small well-being improvements. Correspondingly, villagers living in a household which had sometimes or even often debt problems in 2011 show a worse reported well-being change than households without. The marginal effect is − 6.6 percentage points according to logit regression.
Persons from households without any land or livestock in 2011 report significantly lower well-being developments than persons from households with land and livestock. The marginal effect for persons in households without any land or livestock compared to persons from households with land and livestock is − 11.6 percentage points according to logit estimation and − 8.3 and − 3.3 percentage points for “a bit improved” and “very much improved” according to the multinomial logit model.
Villagers without work in 2011 have a lower probability of reported well-being improvements than villagers with work in 2011. C. p., the marginal effect is − 7.4 percentage points in logit regression and multinomial logit regression demonstrates that it mainly comes from the outcome “a bit improved”.
Illiterate people c. p. have a significantly lower probability of reporting well-being improvements compared to people who are able to read and to write. The estimated coefficients are highly significant in all three types of regressions. Moreover, there are significant differences between well-being changes of literate and illiterate people both for small and for substantial well-being improvements.
Persons with health problems in 2011 (measured as suffering from at least one out of 17 listed possible symptoms often in the last one year) seem to report worse well-being developments until 2014. The effects, however, are only significant for ordered logit estimation and for the outcome “very much improved” in multinomial logit.
To summarize, our estimation results for the socio-economic variables demonstrate that a better initial endowment seems to correlate with better reported well-being developments afterwards.