Poverty and Women Employment: Correlation, Causation, and What the Data Actually Says

Reader note: you can skip the hardcore maths — it’s not required. Every formula has a plain-English story right before it. Read the story, glance at the math only if you like the formal version. The tables and observations carry the argument on their own.

Personal note: this is opinion + reading notes, not research. I cite where I learned, I add where I wonder. Not advice, not a forecast.

Last winter I was on an overnight bus from Jaipur to Udaipur, stuck behind a group of women going for MGNREGA muster. One of them, probably in her late twenties, was telling her seat-mate she’d stopped going to the site after her daughter finished Class 12 — “ab padhai wali naukri dekh rahi hai.” Her mother, sitting next to her, laughed: “humne toh khet mein hi kaam kiya.”

That five-minute exchange has stayed with me more than any regression. Two women, same poverty history, different thresholds for what counts as “work worth doing.” The bus didn’t care about my $r(P,F)$. The papers did — and they helped me stop forcing poverty alone to explain what I was seeing.

The story I started with (and why it broke)

For a long time my mental model was tidy: poorer → more women work. Need pushes entry. It’s not wrong — it’s just incomplete in a way that makes the national numbers look contradictory.

India between 2017 and 2023 is the contradiction. Female labour force participation (PLFS, usual status, 15+) rose from about 23% to about 37% while poverty fell. If “poorer means more work” were the whole story, the line should have moved the other way. So I went back to four older papers and then to newer work that adds the missing hinge — caste — and to the latest survey rounds to see if the pattern held.

What I downloaded and re-read before rewriting this note: Goldin (1995) on the U-shape, Klasen & Lamanna (2009) on gender gaps and growth, Duflo (2012) on empowerment, plus for the caste dimension Thorat & Attewell (2007) on hiring discrimination, Deshpande (2011) The Grammar of Caste and Deshpande & Sharma (2016) on self-employment gaps, Afridi, Bishnu & Mahajan (2023) on FLPR and social identity, and Deshpande, Goel & Khanna (2024, CEEW) on care and transport. For data: PLFS 2017-18, 2019-20, 2022-23 and the new PLFS 2023-24 (released September 2024), NFHS-5 (2019-21), World Bank WDI Gender Data Portal 2024, ILO modelled estimates 2024, and the 2019 Time-Use Survey. All errors in summarising are mine.

1. Goldin’s U — why more development doesn’t always mean more women working (at first)

Claudia Goldin’s 1995 image still works as a map, not a prediction. At very low incomes, many women work — often unpaid on family farms. As work moves to factories and offices without childcare, transport, or social acceptance, participation dips. Later, with services, education, and norms shifting, it rises again. A U.

Rajasthan on that bus felt like the bottom of the U, just starting to climb. The national PLFS time series, kept small here on purpose, shows the same inflection — but only if you look at where the rise comes from.

* Poverty for 2017-24 is World Bank extrapolation from the 2011-12 Tendulkar baseline and later consumption surveys; PLFS and poverty come from different instruments, so I treat them as parallel proxies, not linked microdata.

Observation, not proof: rural FLPR gained almost 20 points while urban gained about 6. The PLFS micro-tables show that gain is overwhelmingly self-employment (own farm, household enterprise), not a regular wage. So the U here is not “more good jobs” yet — it’s any work reappearing at the bottom after a long dip.

If you like the formal version, Goldin writes the U as:

with turning point $y^{\star} = \exp(-\beta_1 / 2\beta_2)$. In words: the same income $y$ can sit on either side of the U depending on $X_i$ — norms, sector mix, care. That $X_i$ is where caste lives, which we’ll meet in a moment.

A point I keep underlined for myself: Poverty pushes entry to any work, but not to good work — the table tells the two stories at once, and the distinction matters more than the headline FLPR.

2. When the gap itself slows growth — Klasen and Lamanna, and why schooling flips the sign

Klasen and Lamanna (World Development, 2009) look at the reverse arrow. Across countries, larger gaps in education and employment between men and women predict slower subsequent growth, even after controlling for income and institutions. South Asia is one of the regions where the estimated cost is largest. I don’t read that as proof of causation, but as a reminder that the arrow runs both ways — low female employment can keep poverty sticky, not just respond to it.

That feedback is why a simple correlation $r(P,F)$ always feels muddied. In a cross-state view, schooling flips the sign.

Rank these four by poverty and the correlation with FLPR is weak, about $r \approx -0.15$. Rank by secondary completion and it jumps. That doesn’t prove schooling causes work, but it tells me my poverty-only story is missing the variable Klasen flags as costly to omit.

Formally, the trap is omitted-variable bias. The regression I mis-run in my head is $F_i = \beta_0 + \beta_1 P_i + \varepsilon_i$. What I should carry is:

where $E$ is girls’ schooling. Then $\mathbb{E}[\hat\beta_1^{\text{naive}}] = \beta_1 + \beta_2 \operatorname{Cov}(P,E)/\operatorname{Var}(P)$. In India $\operatorname{Cov}(P,E) < 0$ and $\beta_2 > 0$, so the bias is negative — even if the true distress effect $\beta_1$ is positive, my naive estimate looks more negative because I omitted schooling. With a district panel $F_{it} = \alpha_i + \delta_t + \beta_1 P_{it} + \beta_2 E_{it} + \varepsilon_{it}$, the $\alpha_i$ and $\delta_t$ sweep out norms and trends that my simple cross-section cannot. That’s why Duflo keeps coming up: context dominates a single $r$.

A second point I’ve started underlining: Schooling and norms flip the sign of the poverty–work correlation — the same poverty level can mean very different FLPR depending on who gets to finish school.

3. What counts as “employment” — Duflo, and why the salaried share matters

Esther Duflo’s 2012 JEL survey, “Women Empowerment and Economic Development,” is deliberately deflating. She walks through childcare subsidies, cash transfers, and role-model RCTs and finds effects that are real but heterogeneous: the same intervention works where care constraints bind and does little where norms or transport bind.

PLFS taught me the same lesson through a different door — quality of work. The rise in FLPR is not a rise in regular jobs.

If poverty were pulling women into good jobs, the salaried share would rise. It falls. So when I say “poverty correlates with women’s employment,” I now immediately ask: which $F$? For $F$ = any work, $r$ can be positive (distress, farm, own-account). For $F$ = regular wage, $r$ turns negative. That split reconciles Table 1 with this table without needing a new theory.

For an RCT like a crèche offer, the estimand is the average treatment effect:

where $Y_i(1)$ is work if offered childcare and $Y_i(0)$ if not. Duflo finds $\hat{\text{ATE}} > 0$ on average but larger where baseline $F$ is low due to care, near zero where norms or buses bind. I now ask “ATE for whom, where?” instead of “does childcare work?” — and I ask the same of any poverty–employment correlation. Time-use data makes the same point: women in poor households work more hours, just less counted, so FLPR understates work and overstates empowerment.

4. The hinge I was missing — caste makes the same poverty feel different

This is the dimension I had under-weighted in the first draft, and the newer papers force it.

Thorat and Attewell (2007) sent identical resumes with different caste names and found callback gaps; Deshpande (2011) shows how caste and gender compound — SC and ST women have higher FLPR than upper-caste women at the same poverty level, but lower earnings and a higher share of casual labour. Deshpande and Sharma (2016) find large gaps in self-employment earnings even after controlling for education and assets. More recently, Afridi, Bishnu and Mahajan (2023) and Deshpande, Goel and Khanna (2024, CEEW) show that FLPR differences shrink a lot once you condition on care burden, safety, and transport — and that caste and religion shift those constraints, not just income.

What that looks like in one more small table — PLFS 2022-23, all-India, usual status:

Two things stand out, and I’ve underlined the second because it reframes the bus conversation:

Poverty predicts entry for ST/SC women more strongly, but quality of entry is lower. At a given poverty level, an upper-caste woman in a village near a town with a bus may stay out of the labour force; an Adivasi woman ten kilometres further in, without that bus, is counted as “employed” on a family plot. Same $P$, different $F$, different meaning.

That line is why the hinge in the closing thought matters. Poverty down without moving the hinge — schooling, care, safety, title, transport — and the door barely opens. Move the hinge, and even modest income gains change who works and how. Caste, in the recent papers, is part of the hinge, not a separate story.

MGNREGA evaluations and the newer CEEW work point the same way: guaranteed work raises participation at the bottom for SC/ST women, but mostly in low-productivity, informal work. The distress entry is real; the good-work entry is not, unless something else moves with it.

Putting it together — my synthesis, after the evidence

If I have to compress Goldin, Klasen, Duflo, Thorat/Deshpande and the four tables into a few sentences for myself:

• Poverty correlates positively with any work (including unpaid/informal) and negatively with good work (formal, well-paid, safe). The sign flips with the definition of $F$.
• Norms, schooling, safety, care and caste flip the sign more than a one-point change in headcount poverty.
• Time-use undercount makes FLPR a proxy with a big asterisk — women in poor households often work more hours, just less counted.

A tiny math note I keep honest without making it the story: the headline $r(P,F)$ is not a number but a function $r(\text{definition of }F, \text{controls}, \text{aggregation})$. If I pool states, a state-level $r$ can hide district-level $r$’s of opposite sign (Simpson’s paradox), and if I omit schooling the poverty coefficient is biased by $\beta_2 \operatorname{Cov}(P,E)/\operatorname{Var}(P)$. That’s why I compute $r$ within-state and within caste-group before I trust a pooled $r$, and why I always split $F$ into $F^{\text{informal}}$ and $F^{\text{salaried}}$ before correlating.

What I’d test next, if this were more than a note

• In my district, does a new bus route or a crèche move FLPR more than an equal-cost cash transfer? Duflo would want it pre-registered, not story-told — and the 2024 CEEW findings suggest I should pre-stratify by caste and by baseline care burden.
• Does “poverty falls → women leave farms” show up as a welfare gain (fewer bad hours) rather than a loss? Goldin would remind me not to mourn strenuous farm self-employment — the 2023-24 PLFS suggests that’s still the margin.
• Do assets titled to women (land, house) shift the hinge more than household income alone? The micro literature, especially for SC/ST households, says yes, and PLFS can’t see title.

Closing thought

That bus from Jaipur taught me to listen for thresholds, not just headcounts. The mother stayed on the farm because that was work she was allowed to do; the daughter waited for work she wanted to do. Poverty was similar, thresholds were not.

The tables keep me from smoothing the story; the small, skippable math keeps me from overclaiming it. If you read the same papers — especially the newer caste and care work — and see the data differently, I’d like to know where your points sit, and which $F$ you’re counting.

References — what I (re)read, old and new

• Goldin, C. (1995). The U-Shaped Female Labor Force Function in Economic Development and Economic History. NBER Working Paper 4707.
• Klasen, S. & Lamanna, F. (2009). The Impact of Gender Inequality in Education and Employment on Economic Growth. World Development 37(10).
• Duflo, E. (2012). Women Empowerment and Economic Development. Journal of Economic Literature 50(4): 1051-1079.
• Thorat, S. & Attewell, P. (2007). The Legacy of Social Exclusion: A Correspondence Study of Job Discrimination in India. Economic and Political Weekly.
• Deshpande, A. (2011). The Grammar of Caste: Economic Discrimination in Contemporary India. Oxford UP.
• Deshpande, A. & Sharma, S. (2016). Disadvantage and Discrimination in Self-Employment: Caste Gaps in Earnings in Indian Small Business. Small Business Economics.
• Afridi, F., Bishnu, M. & Mahajan, K. (2023). Gender, Caste and Labour Force Participation. Working paper, ISI Delhi.
• Deshpande, A., Goel, D. & Khanna, S. (2024). Care, Transport and Social Identity: New Evidence on FLPR. CEEW Working Paper (cited for 2024 policy brief).
• Baird et al. + MGNREGA evaluations — survey readings on CCTs and guaranteed work.
• Data: PLFS 2017-18, 2019-20, 2022-23, 2023-24 (MOSPI, Sep 2024); NFHS-5 (2019-21); World Bank WDI / Gender Data Portal 2024; ILO modelled estimates 2024; Time-Use Survey 2019.

All errors in summarising are mine; the papers and datasets speak for themselves. Download links are open-access via NBER, World Bank, MOSPI, and author pages — I’ve kept the note to what’s in the public reports, not the paywalled PDFs.