Explaining the Pay Gap

The gender pay gap is a number that gets quoted constantly and defined rarely. Depending on how it is calculated, the same firm can honestly report a gap of twenty-two per cent, five per cent, or one and a half.

All three figures are correct. They answer different questions. And the argument between people who quote different ones is, almost always, an argument about which adjustments are legitimate — which is 8.4.2's problem, and the same resolution applies.

One payroll, four numbers.

A firm of four thousand people publishes its figures.

Number one: the raw gap — 22 per cent. The difference between the average hourly pay of all men and all women employed there.

This is a real fact about the firm. It says that if you picked a woman and a man at random, the man would on average be paid substantially more. Nothing about it is misleading, and it is what determines the actual distribution of money between the firm's male and female employees.

Number two: adjusted for occupation and hours — 5 per cent. Compare people doing similar jobs at similar hours, and most of the gap disappears.

Number three: adjusted for occupation, hours, grade, tenure, region, qualifications and performance rating — 1.5 per cent. Almost nothing left.

Now watch what happened between numbers one and three.

The firm's senior management is 82 per cent male. Its administrative and customer-service staff are 74 per cent female. Those two facts are the raw gap. Controlling for occupation and grade does not explain the gap; it removes it from the calculation, by comparing people only with others in the same position (see 7.6.2).

And the adjusted number is genuinely informative about something else: whether two people in the same job at the same grade are paid differently. Which is a serious question with a serious answer, and it is not the question the raw number was asking.

Number four is the one nobody publishes, and it is the most revealing.

Take everyone the firm hired into the same graduate role in the same month, eight years ago. Their starting salaries were identical. Follow them.

Today the men in that cohort earn, on average, 31 per cent more. Not because anyone was paid differently for the same job — because of who moved to the roles with unpredictable hours, who took the client-facing account, who was available for the eighteen-month project abroad, who took nine months out, who went to four days, and who was promoted while doing so.

Number four is the gap opening in real time, and it is invisible in all three of the others.

The adjustment is the argument, not the answer.

Every control is a claim about what should be held constant — that is, a claim that this difference is legitimate and should not be counted as part of the gap.

Control for hours, and you have decided that working fewer hours should be paid proportionately less — which is a defensible position and also, in greedy jobs, factually not what happens (see 8.5.1). Hours are not paid linearly, so controlling for them does not remove a proportionate difference; it removes a disproportionate one.

Control for occupation, and you have decided that being in a lower-paid occupation is not part of the gender pay gap — which requires that occupational sorting be independent of gender, which it is not (see 8.5.1 on segregation and devaluation).

Control for the employing firm, and you have decided that working at a lower-paying firm is not part of it.

None of those decisions is obviously wrong. All of them are decisions, all of them are usually unstated, and the number falls monotonically as more of them are made.

Which yields the rule for reading any pay gap figure: what did they hold constant, and is it a cause or a channel? The same question as 8.4.2, on a different axis.

What the gap is made of, and how that has changed

The composition has changed completely over sixty years, and the residual has not.

In the mid-twentieth century, a large part of the gap was explained by measurable human capital: men had more education and more continuous experience. Both have gone. Women now exceed men in tertiary attainment in the great majority of rich countries, and the experience gap has narrowed sharply.

So the components that used to explain the gap now explain very little of it, and in some decompositions education contributes in the opposite direction — women's higher qualifications should, on their own, produce a gap in the other direction.

What now accounts for the measured, explained portion is occupation, industry and firm. Where people work, and for whom.

And the unexplained residual has proved stubborn. In careful American decompositions it has remained a substantial share of the total gap — of the order of a third or more — across decades in which everything measurable converged.

Which is the same shape of finding as 8.4.2's: the components that were expected to close the gap closed, and the gap did not close proportionately.

The lifecycle shape, which is the most informative single fact.

The gap at labour market entry is small. In several countries it is close to zero for graduates in the first year or two of work.

It opens sharply between the late twenties and the late thirties, and then flattens.

The best-documented case is a cohort of business school graduates followed for over a decade. At graduation, men's and women's earnings were nearly identical. Ten to fifteen years later the gap was very large — and it was accounted for almost entirely by three things: differences in training taken before the degree, career interruptions, and weekly hours. Modest differences in each, compounding through a convex pay-hours structure and through the promotion decisions that hours affect.

This is 8.5.1's mechanism observed in a payroll. The gap is not created at hiring. It is created by the interaction between an event — usually a first birth — and a job structure that penalises interruption and discontinuous availability at a steeply increasing rate.

And it explains why interventions at the point of hiring do so little. The gap does not open there.

Firms: sorting and bargaining.

A decomposition using matched employer-employee records — every worker linked to every firm, so that a firm's pay premium can be separated from a worker's characteristics — splits the gap into two components.

Sorting: women are employed disproportionately at firms that pay less to everyone.

Bargaining: within the same firm, women receive a smaller share of that firm's pay premium than comparable men do. When a firm does well and shares the gains, the gains are shared unequally.

Both components are substantial, with estimates indicating women receive roughly nine-tenths of the firm-specific premium men receive.

Why this matters for policy. A rule requiring equal pay for equal work within a firm addresses the bargaining component and does nothing about the sorting component — and sorting is the larger part in most estimates. Which is why equal-pay legislation, which is decades old in most rich countries, coexists with substantial gaps: it was aimed at one component.

Negotiation: the claim, the correction, and why the correction matters.

The popular claim is that women earn less partly because they do not ask — that the gap is, in part, a negotiation gap that could be closed by teaching women to negotiate.

The more careful evidence does not support it in that form.

Large representative data with direct measures of asking behaviour finds that women ask for pay rises at broadly similar rates to men. They receive them less often. The gap is in the response, not the request.

And experimental evidence identifies why the advice was bad advice. In studies where evaluators assessed candidates who did or did not attempt to negotiate, women who negotiated were penalised more than men who did the same thing — rated as less likeable and less desirable to work with, by evaluators of both sexes.

So the advice "ask more" was directed at people who were already asking, and into a situation where asking cost them more. This is worth dwelling on as a general pattern: an individualised explanation for a structured outcome, producing advice that shifts the burden onto the disadvantaged party and does not work (see 8.2.3 on individualisation, and 8.3.3 on aspirations, which is the same error in a different Topic).

Motherhood and fatherhood, measured experimentally.

An audit and laboratory study of the same question produces the clearest evidence available. Evaluators assessed otherwise identical applications that differed only in signals of parental status.

Mothers were rated as less competent and less committed, were held to higher standards, were recommended for hire less often, and were offered lower starting salaries than identically qualified women without children.

Fathers were not penalised. On several measures they were advantaged — rated as more committed, and offered higher salaries than identically qualified childless men.

And the field component found the same pattern in real callbacks.

Two implications. The penalty attaches to motherhood rather than to gender as such, which is why the gap opens when it does. And the same information about family produces opposite effects depending on which parent holds it, which no explanation appealing to hours or productivity can account for, because the applications were identical.

Two comparisons that mislead, and one that surprises.

Comparing raw gaps across countries is treacherous, because of selection into employment.

In a country where female labour force participation is low, the women who do work are a selected group — disproportionately highly educated, in professional occupations, from advantaged families. Comparing their pay to all men's produces a small measured gap.

As participation rises, less advantaged women enter the workforce and the measured gap widens — with no deterioration in anyone's position. Analyses correcting for this selection find that a meaningful part of cross-country variation in raw gaps is a participation artefact, and that some countries' apparently good performance is a composition effect.

This is 7.3.2's selection problem determining an international league table, and the league tables rarely mention it.

And the surprising comparison: gender discrimination in hiring is more heterogeneous than racial discrimination.

Audit and correspondence evidence on gender in hiring finds effects that vary substantially by occupation and context, with disadvantage concentrated in male-dominated manual, technical and senior roles, disadvantage attaching strongly to motherhood as above — and, in some experimental settings, including studies of academic hiring using written vignettes, a measured preference in favour of women.

Those last findings are vignette-based, context-specific and contested, and they should not be generalised into a claim that hiring discrimination against women no longer exists. What they do establish is that the picture is not uniform, which distinguishes it from the racial hiring evidence, where the direction is consistent across contexts (see 8.4.2).

Reporting both is what this course's standards require, and it is also more useful: an account that predicts disadvantage everywhere will not explain why the pay gap opens at thirty rather than at twenty-two.

Does pay transparency work? The evidence, which is modest and interesting.

Several countries have introduced requirements to disclose pay information or publish gender pay gaps.

A Danish law requiring firms above a size threshold to report gender-disaggregated pay statistics — with a threshold that permits a comparison of just-covered and just-uncovered firms (see 7.4.3) — was found to reduce the gender pay gap by a modest amount, of the order of two percentage points.

And the mechanism matters more than the magnitude. The reduction came mainly from slowing men's wage growth rather than raising women's. Firm wage bills and productivity effects were also detected.

British evidence on mandatory gender pay gap reporting finds a similar pattern: a modest narrowing, achieved substantially through slower male wage growth.

Two honest conclusions. Transparency does something, which is more than can be said for several better-publicised interventions. And what it does is smaller than advocates claim and operates in a way advocates rarely mention — by compressing pay at the top rather than lifting it at the bottom, which is a different political proposition and worth stating plainly.

Because this is the area where the same evidence is used to argue that the gap is entirely discrimination and that it does not exist, and both claims misuse the numbers.

Four questions.

Raw or adjusted? If adjusted, for what — and is each control a cause or a channel? Occupation, hours and firm are channels.

At what career stage? A gap measured at entry and a gap measured at forty are different quantities, and the second is the one the mechanism produces.

Within or between firms? Equal-pay rules address the smaller component.

And what is the comparison population? Cross-country raw gaps are confounded by who is in the workforce.

Two claims that do not survive contact with the evidence.

"The gap disappears once you control properly, so it is a myth." The controls remove the mechanism. A gap that vanishes on controlling for occupation is a gap that operates through occupation — which is a finding about how it works, not a refutation.

And "the gap is a measure of discrimination." The raw gap is a distributional fact with several sources, of which unequal treatment is one. Calling it a discrimination figure invites the first error as a response, and it is why the argument goes round in circles.

The formulation that survives: the pay gap is largely produced by the interaction of a household division of non-deferrable work with a labour market that pays convexly for uninterrupted availability, plus employer responses to parental status that are demonstrably not about productivity, plus sorting into lower-paying firms, plus a residual that has not closed as the measurable components did.

That sentence is long because the thing is. Any shorter version is a political position rather than a finding.

One payroll, four numbers: a raw gap of 22 per cent, 5 per cent adjusted for occupation and hours, 1.5 per cent adjusted for everything — and 31 per cent among a single cohort hired into the same job on the same day eight years earlier. The controls do not explain the gap; they remove the channels through which it operates.

Every control is an unstated claim that a difference is legitimate. Controlling for hours assumes linear pay, which greedy jobs contradict; controlling for occupation assumes sorting is gender-independent, which segregation and devaluation contradict.

The composition of the gap has changed completely and the residual has not. Education and experience once explained much of it and now explain almost nothing — women exceed men in tertiary attainment in most rich countries — while occupation, industry and firm carry the explained portion, and the unexplained share has remained around a third or more.

The lifecycle shape is the key fact: the gap is near zero at entry, opens sharply between the late twenties and late thirties, and flattens. A cohort of business graduates with identical starting earnings diverged enormously within fifteen years, accounted for by pre-degree training, career interruptions and weekly hours. The gap is not created at hiring, which is why hiring interventions do little.

Matched employer-employee data splits it into sorting — women at lower-paying firms — and bargaining — women receiving roughly nine-tenths of the firm-specific premium men receive. Equal-pay law addresses the smaller component, which is why it coexists with substantial gaps.

The negotiation claim does not survive: women ask at broadly similar rates and receive less, and experiments find women are penalised more for negotiating. An individualised explanation producing advice that does not work.

Motherhood is penalised and fatherhood is not — experimentally, on identical applications, with mothers rated less competent and offered less while fathers were advantaged.

Cross-country raw gaps are confounded by selection into employment: where participation is low, the employed women are highly selected and the measured gap is artificially small.

And gender hiring discrimination is heterogeneous — concentrated in male-dominated and senior roles and attached strongly to motherhood, with some vignette studies finding preference for women in academic hiring. Unlike the racial evidence, the direction is not uniform.

Pay transparency produces modest narrowing, mainly by slowing men's wage growth.

Raw (unadjusted) gap — the difference in average pay between all men and all women; a distributional fact.

Adjusted gap — the difference remaining after holding characteristics constant; the choice of characteristics is the argument.

Explained / unexplained portion — the decomposition's two parts; the second is not an estimate of discrimination.

Lifecycle profile of the gap — near zero at entry, opening between the late twenties and late thirties.

Sorting / bargaining components — employment at lower-paying firms; a smaller share of the firm's premium within them.

Motherhood penalty / fatherhood premium — the experimentally demonstrated opposite effects of identical parental information.

Negotiation penalty — the higher social cost imposed on women who negotiate.

Selection into employment — the composition effect making raw gaps artificially small where participation is low.

Pay transparency requirement — disclosure rules producing modest narrowing, mainly through slower male wage growth.

One — find your employer's four numbers. Where gender pay gap reporting is required, look up an organisation you know. Note whether it publishes anything resembling number four.

Two — classify the controls. Take any adjusted gap figure and mark each control as a confounder or a channel. Then decide what the adjusted number is actually measuring.

Three — check the age profile. Find a pay gap broken down by age band for your country. The shape will show you the mechanism without any further analysis.

Four — do the negotiation exercise. Before looking it up, predict whether women ask for pay rises less often. Then read the finding, and note what the advice industry built on the assumption.

Five — test the selection artefact. Compare the raw gender pay gap and the female employment rate across five countries. Look for countries with small gaps and low participation — and reconsider what their ranking means.