Female-Targeted Hiring Subsidies, Firm Learning, and Women’s Employment

Mimosa Distefano, Lorenzo Incoronato and Anna Raute — “Female-Targeted Hiring Subsidies, Firm Learning, and Women’s Employment,” presented by Incoronato at the NBER Summer Institute, Gender in the Economy, on 29 July 2026 — the last paper of the meeting. No discussant; questions throughout, and several of them were sharper than the answers. Written from RFBerlin Discussion Paper 211/26.

Nearly everything we do about women’s employment after childbirth is a supply-side policy. Parental leave, childcare subsidies, flexible-hours mandates: all of them are about making it possible for a woman to come back. The evidence that they permanently return women to work is, at best, mixed.

One reason might be sitting on the other side of the market. Suppose a firm looks at a CV with a three-year gap and simply does not want it. Maybe it reads the gap as depreciated skill; maybe — and there is sociology on this — it reads a mother returning from leave as less committed. Either way, the firm is uncertain about how productive this kind of worker would be at this firm, and the uncertainty resolves only through hiring.

Which is a trap, and a very old-fashioned kind. To learn about a group you have to sample from it. If your prior about the group is low enough, you never sample. If you never sample, the prior never updates. This is exploration versus exploitation with the exploration arm switched off, and it can persist indefinitely without anyone doing anything a court would recognise as discrimination.

The paper’s question follows immediately: if a subsidy pays for the experiment, does the firm keep hiring from the group after the subsidy ends?

Bonus Donne

Italy introduced it in 2013: a 50 percent cut in employer social security contributions — roughly 10 percent of labor cost, comparable to hiring subsidies studied elsewhere — for firms hiring a woman who had been non-employed for at least 24 months. Twelve months of subsidy on a fixed-term hire, eighteen if the contract was permanent or was converted to one. The firm had to have shed no workers on net in the previous twelve months, which during an Italian recession was a real constraint.

The setting is well chosen: Italy’s gender gap is large and concentrated in employment rather than earnings, much of it attributable to the child penalty, and it runs a classic dual labor market where the protected insiders are men and the intermittent outsiders are disproportionately women.

The data are the Italian social security registers (INPS) covering the universe of private-sector workers and firms, with one feature that makes the paper possible: take-up is directly observable. You can see which hire was subsidized and which firm claimed it. That is rare, and it turns a policy-exposure design into a design around actual use.

Take-up was tiny. Between 2013 and 2019, only about 4 percent of eligible workers were hired through the subsidy — around 180,000 people. (The policy still exists and has since become much more generous, now covering essentially all employer payroll taxes, so a version of this paper written today would look different.)

The design, and its honest limitation

Treated firms are those that used Bonus Donne at least once between 2013 and 2019, with event time zero at first use. Control firms are propensity-score matched on prior wages, employment, female share and growth, and are additionally required to have hired a woman in the same year without the subsidy — not an innocent choice, as the presenter said, but necessary for some outcomes to be defined. Final sample: 38,270 firms, half treated.

The limitation was stated repeatedly and without prompting: this is selection on observables, and dynamic selection into treatment cannot be ruled out. Firms that reach for a hiring subsidy in a given year are firms doing something. The paper does not claim the firm-level effects are the causal effect of Bonus Donne, and I will not either. What it claims is narrower and more interesting — that a change in who a firm hired, induced by the subsidy, subsequently changes who the firm hires.

The spike at zero is mechanical — the law requires it. The point is what happens at event times 1 through 5. Treated firms keep hiring women with longer prior non-employment (about a quarter of a year more than controls, against a 1.5-year jump on impact), keep hiring more women out of long-term non-employment, and keep hiring about 2 percent more mothers. The persistence is the result.

Meanwhile firms grow: value added rises, and value added per worker does not move, so hiring long-non-employed women costs nothing in labor productivity. Male hiring, if anything, goes up too — which the presenter could not explain beyond “maybe they’re saving money and using it to expand,” and which is one more reason the causal claim is hedged.

Is it learning, or is it just cheap?

The obvious alternative is that treated firms keep hiring from the target group because they keep claiming the subsidy, which is a cost-saving story with no beliefs in it. The paper’s test is elegant, and it works by throwing away the control group entirely.

The model says persistence should depend on whether the firm was positively surprised by its first subsidized hire. To proxy for that, run a Mincer wage regression on all subsidized hires at treated firms in the 2013–2016 cohort, take the residual — the deviation of the actual hiring wage from the predicted one — and read it as the firm’s private signal from the interview, since the wage is set on that signal. Split treated firms at the median residual into good matches and bad matches, and compare them to each other, excluding the mechanical event-zero effect.

Good-match firms subsequently hire about 2 percent more workers through the subsidy than bad-match firms, hire more women out of long-term non-employment, and hire more mothers — and do not grow more, so this is composition, not expansion. Since both groups face identical financial incentives to keep claiming, the cost-saving story struggles to generate the gap. The paper also rules out the “firms learned to screen” mechanism that has been offered for the persistence of affirmative action: treated firms do not hire workers with higher Mincer residuals in their previous jobs, only in their current ones.

On the worker side — where the presenter is more comfortable claiming causality, since taking the subsidy is a firm decision — subsidized women are more likely to remain employed both at the hiring firm and anywhere, persistently. But there is no wage pass-through. The firm pockets the 10 percent, which is the standard finding for hiring subsidies and which one questioner found genuinely puzzling, since the whole logic of the policy is to compensate the firm for taking a risk, not to hand it a rent.

Where the room pushed

The most persistent objection was about the target group itself, and it was not answered. A senior questioner put it plainly: this is a strange category to discriminate against. Refusing to hire high-school dropouts, or people of a certain race or sex, is something you can do from a glance at an application. But “woman with at least 24 months of non-employment” requires you to have the CV and to be able to read a gap off it — and, as she noted, anyone who has read a lot of CVs knows you often cannot.

The follow-up was sharper still. The sociological evidence Incoronato cited is about firms believing mothers with young children are less committed. That is not the same group. If firms really held that belief about mothers, they would have fired the ones they already employed — “and of course we know there was a time even in this great country when that was the case.” She then named the case: Phillips v. Martin Marietta, the first Title VII sex-discrimination case to reach the Supreme Court, in which a company refused to accept applications from women with pre-school-age children. Incoronato’s answer was that the paper could characterize the group better, which is true and was conceded twice more.

Someone else offered a repair that is better than the paper’s own framing. Maybe the learning trap does not come from a low prior about the group at all. Maybe the interview is simply less informative about someone who has been out of the market for years — the signal is noisier, so the option value of interviewing is lower, so you don’t. That generates the same trap without requiring firms to hold a specific belief about mothers, and it is testable if you can find variation in signal precision. Incoronato wrote it down.

And the general-equilibrium question got the most honest answer of the session. If there are plenty of workers from other groups, why would a firm care about being stuck in a learning trap at all? “Nothing in my model” tells you it isn’t. Take-up was 4 percent; the presenter volunteered before being asked that the program is “tiny.”

There is also a data problem the authors did not choose. To separate “firms keep hiring from the group through the subsidy” from “firms keep hiring from the group, full stop” you need to see subsidy use in the later years, and INPS cut off access. It is top of their to-do list and they cannot do it.

(A hiring subsidy is normally justified as a wage subsidy: labor is too expensive relative to its product, so shave the payroll tax. This paper reframes it as buying information — you are not paying the firm to employ her, you are paying it to find out. Which, if it’s right, means the policy’s return is not the twelve months of employment it purchases but the posterior it leaves behind. And also that the case for such a policy rests on a fact that is difficult to establish and that this room, politely and repeatedly, declined to grant: that firms were wrong about this group to begin with.)