Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
15,033 characters · 6 sections · 15 citation commands
Working women and caste in India: A study of social disadvantage using feature attribution
Working outside the house has historically been considered a social-stigma or a low-status activity for women in India eswaran2013status. As a result, only the poorest women are working out of necessity in blue-collar jobs and once their family income increases, they withdraw from the workforce rao2010gender. Women tend to re-enter the work force only at high education levels that allow access to un-stigmatized white-collar jobs klasen2012push. In this cultural context, caste is an important determinant of a woman's work-status. Since scheduled castes and scheduled tribes (Sc/St) are the socially and economically disadvantaged caste groups in India, Sc/St women have always had higher workforce participation, mainly in blue-collar jobs. Upper-caste (or general caste) women have historically been discouraged from working in order to maintain a higher social-status eswaran2013status.\footnote{ bayly2001caste provides a comprehensive discussion of the caste system in India.}
We study whether the ability to infer a woman's work-status based on her caste is changing over generations. We also study the change in the effect of caste on work-status over generations. Adopting strategy similar to bertrand2018coming, we use a nationally representative dataset from the National Family Health Survey nfhs4 for training ensemble Gradient Boosting Decision Tree models to predict women's work-status. Our models uncover non-linear temporal patterns between caste and women's work-status using the SHAP (SHapley Additive exPlanation) feature attribution framework lundberg2017unified.
SHAP values of a feature (such as belonging to the Sc/St caste) measure how important that feature is in predicting the outcome of a model (woman's work-status). Using individual-level SHAP values, we find an upward trend across generations of working women, where caste is a less important determinant of work-status for younger women. Further, we unpack the impact of caste on blue-collar and white-collar jobs, finding that younger Sc/St women are moving out of blue-collar and into white-collar jobs in India.
Our paper is related to recent studies on machine learning for development de2017discovery,de2018machine,mullainathan2017machine. We demonstrate a novel usage of the SHAP framework to study social disadvantage and how it affects women's work in India. We also contribute to the aforementioned literature on caste and work by studying for the first time whether the younger generation of women are still facing the same level of disadvantage as older women.
We use a nationally representative dataset from the National Family Health Survey conducted in 2015-16 (NFHS-4) consisting of 699,686 women aged 15-49 years across all 29 states and 7 union territories of India nfhs4. The survey has information on caste, work-status, and occupation types for a representative sub-sample of 111,398 women. We consider women over 21 years of age, leaving us with 81,816 women for whom we have detailed socio-economic information for our analysis.
We classify working women into those who have blue-collar type jobs (agriculture, skilled and unskilled manual labour, and domestic services) and those who have white-collar type jobs (professional, technical, managerial, clerical, and sales). We use sixteen socio-economic features for training machine learning models to predict work-status. Appendix (ref) presents summary statistics of the dataset and the features used.
We design three binary classification experiments to predict a woman's work: (1) having a job or not (work-status), (2) having a blue collar job or not (blue-collar), and (3) having a white collar job or not (white-collar). Using the sampling weights provided in NHFS-4, we create a test set with 5% of the data and use the remaining 95% for training our models in each experiment. We use stratified sampling to ensure that positive-to-negative class balance remains the same across training and test sets.
For each experiment, we train an ensemble Gradient Boosting Decision Tree (GBDT) model using LightGBM ke2017lightgbm. We chose LightGBM due to generally strong empirical performance, fast training time, and easy compatibility with the SHAP framework. Optimal model hyperparameters are found using 5-fold cross-validated grid search over the training set. We use F1-score as the metric for choosing the best hyperparameters since it incorporates class imbalance. These hyperparameters are used to re-train each of the models on the entire training set.
Next, we interpret the trained models using the SHAP feature attribution framework for tree ensembles. Unlike traditional feature importance methods, SHAP allows for fast computation of complex tree-based, non-linear models. We compute SHAP values for our entire dataset, following lundberg2018consistent as a guide to using and interpreting the explanations obtained. These explanations allow us to understand how a single feature affects the model's output, summarize relative feature importance over the entire dataset, and analyze higher order interactions among feature pairs.
Since we use a nationally representative dataset, results from our experiments represent actual country-level patterns. However, it is important to note that the accuracy of feature attribution methods is proportional to model performance. See Appendix (ref) for a detailed discussion.
\paragraph{Global Feature Importance} The SHAP summary plot for work-status experiment in Figure (ref) shows the relative importance of features, the distribution of impacts of features on the model's prediction, as well as how the feature's value (Low to High) relates to its impact. Each dot represents a woman in the dataset and the x-axis position of the dot is the impact of that feature on the model's prediction. Dots that do not fit on the row pile up to show density. Since the GBDT model has a logistic loss, the x-axis has units of log-odds. Summary plots for blue-collar and white-collar experiments can be found in Appendix (ref).
Belonging to the Sc/St caste is the sixth most important feature in predicting work-status. State, household wealth index, age, years of education, and number of children below 5 years of age, are the five most important predictors of work-status. The observed patterns are similar to those found in existing literature: women from wealthier households are less likely to work rao2010gender, more educated women are more likely to work bhalla2011labour, and Sc/St women are more likely to work srivastava2010women.
\paragraph{Importance of caste over generations} For each of the three experiments, we plot the mean of the magnitude of SHAP values of the Sc/St Caste feature for discrete ages, along with the 99% confidence interval, to analyze whether the importance of caste has changed over generations (Figure (ref)). We find that caste is more important in predicting work-status of older women than of younger women. This pattern is also observed for blue-collar jobs. For white collar jobs, there isn't a clear monotonic pattern, but overall, caste is more important for women younger than 35 years of age.
\paragraph{Dependence and main effects of caste} The SHAP dependence plot in Figure (ref) plots the SHAP values of the Sc/St Caste feature from the work-status experiment and colours them by age to visualize underlying interaction effects. We can see the attributed importance of caste change as its value varies. Higher SHAP values represent higher probability of working: Sc/St women are more likely to be working than women of other castes. The dependence plot also captures vertical dispersion at a single value of the Sc/St Caste feature due to interaction effects with other features in the model. Colouring each dot by age (as the interacting feature) we find that older Sc/St women are most likely to be working.
Next, we analyze the main effects plot for the Sc/St Caste feature by removing all the interaction effects with other features in the work-status experiment (Figure (ref)). We find that Sc/St women are highly likely to be working while women of other castes are not likely to be working.
\paragraph{Interactions between caste and age} SHAP interaction plots capture all of the vertical dispersion that was present in the SHAP dependence plot (Figure (ref)) but was missing from the main effects plot (Figure (ref)). We only focus on interaction plots between the Sc/St Caste feature and age (Figure (ref)). For younger Sc/St women, SHAP interaction values are larger negatives, implying a high prediction for not-working. For older Sc/St women, SHAP interaction values are larger positives, implying a high prediction for working. Plotting the SHAP interaction values using age on the x-axis instead of caste, we get a sharper visualization of the same pattern (Figure (ref)).
Note that SHAP interaction values are expensive to compute. Hence, main effect and interaction plots use a random sub-sample of 10,000 women sampled using the weights provided in NFHS-4.
\paragraph{Interactions for job types} For fine-grained analysis, we plot the age and Sc/St Caste interaction plots for both blue-collar and white-collar experiments. We find that the pattern observed in the work-status experiment is only relevant for blue-collar jobs (Figure (ref)). We observe an opposite pattern for white-collar jobs (Figure (ref)). Younger Sc/St women are more likely to be working in white-collar jobs while older Sc/St women are less likely to be working in white-collar jobs. This pattern is prominent mainly for women younger than 26 and older than 45 years, respectively.
We train machine learning models to predict women's work-status in India and analyze the relationship between work-status and caste using the SHAP feature attribution framework. We find that, over generations, caste has become a less important determinant of younger women's work-status, especially their participation in blue-collar jobs. Next, we unpack the impact of caste on work-status using SHAP interaction effects, finding that younger women of scheduled castes and scheduled tribes are less likely to be working in blue collar-jobs, and more likely to be working in white-collar jobs.
For nearly 70 years, the government of India has implemented caste-based quotas in education and government jobs to ensure equal economic opportunity. Our work can be used as a cost-effective tool to monitor the impact of the existing caste-quota policy. In future work, we will look deeper into geographical variations to identify regions where younger women are lagging, and where job creation can be targeted. Our methodology can similarly be used to study the persistence of, as well as the nuanced patterns underlying other types of social disadvantage and bias in both developing and developed country contexts.
We would like to thank Maria De-Arteaga, Scott Lundberg, Sean Saito and the anonymous reviewers for helpful feedback and discussions that have been included in the Addendum (Appendix (ref)).