Most of this book treats the wage as a price like any other, a number where a supply curve crosses a demand curve. That picture is not wrong, but it is incomplete in ways that matter for almost every policy fight you will ever read about. The labor market is where a person spends a third of their waking life, where the income distribution is mostly made, and where the cleanest predictions of competitive theory run into their most stubborn empirical resistance. This chapter takes the labor market on its own terms.
The front half rebuilds the standard apparatus. A worker chooses how many hours to sell, and whether to work at all, by trading consumption against leisure. A firm hires workers the way it buys any input, up to the point where the last worker's contribution to revenue equals their wage. Schooling and experience raise the wage a worker can command, and the Mincer earnings function turns that into a regression you can actually run. None of this departs from the consumer and producer theory of Chapter 5; it specializes it to the one factor that thinks back.
The back half is where labor economics becomes its own field. Real labor markets do not clear instantly. The search-and-matching model explains why a positive unemployment rate is a feature of the matching technology, not a failure. Employers often have wage-setting power, and a single buyer of labor, a monopsonist, marks the wage down below what the worker produces, which flips the textbook minimum-wage prediction on its head. That flip is the spine of the chapter: the minimum-wage debate is where the profession made a genuine within-mainstream call, and we make it here rather than waving at it. We close with the task framework's account of automation, with discrimination and its decomposition, and with the question of why the share of national income paid to labor has been falling for forty years.
Prerequisites: consumer and producer theory (Ch. 5), the utility-maximization and factor-demand machinery this chapter specializes; econometrics (Ch. 10), the difference-in-differences, instrumental-variables, and natural-experiment toolkit the minimum-wage and returns-to-schooling results rest on. Math: multivariable calculus and constrained optimization.
Wage labor in its modern form, a person selling hours to an employer for money, free to quit and free to be fired, is a historically recent arrangement; the economic-history book narrates how the Industrial Revolution turned scattered household and craft production into a market for labor time (Economic History, Ch. 7). We take that market as given and ask what a worker inside it actually decides.
The worker's problem is not "how hard should I work" but two separate questions: how many hours to sell, and whether to sell any at all. Start with the first. A day has a fixed number of hours, $T$. The worker splits them between leisure $\ell$ and work $h = T - \ell$. Every hour of work pays the wage $w$ and buys consumption; every hour of leisure is enjoyed directly. There may also be income $V$ that arrives regardless of work: a partner's earnings, a benefit, a trust fund. The budget constraint is the accounting of that trade.
Equation 21.1 is the consumption-leisure budget constraint. Its slope in $(\ell, C)$ space is $-w$: giving up one hour of leisure buys $w$ units of consumption. Non-labor income $V$ shifts the line up without changing its slope.
The worker maximizes utility over consumption and leisure subject to that constraint.
The first-order condition sets the marginal rate of substitution between leisure and consumption equal to the wage. Leisure is a good; the wage is its price. At the optimum the worker is indifferent between one more hour of leisure and the $w$ units of consumption that hour would have bought.
Why it matters: Forget the algebra for a moment. Every hour you do not work, you keep. Every hour you do work, you sell for the wage and spend the proceeds. You stop adding hours at the point where the next hour of free time is worth exactly as much to you as the stuff the wage would have bought. A higher wage makes that next hour of free time more expensive to keep, which pulls you toward working more. The whole subtlety of labor supply is what happens to that pull when the wage gets high enough that you are already rich from working.
A wage rise does two things at once, and they pull in opposite directions. The substitution effect: leisure is now more expensive (each hour off costs more forgone consumption), so the worker substitutes toward work. The income effect: the worker is richer at every hours level, and if leisure is a normal good, some of that wealth is spent on more leisure, that is, on fewer hours. Which effect wins is an empirical question, and the answer can change as the wage climbs.
At low wages the substitution effect typically dominates: raises buy more hours. At high wages the income effect can take over: a worker already earning a comfortable income responds to a further raise by taking more time off. When that happens, the individual labor-supply curve bends backward.
Everything so far concerns the first question, how many hours. The second question, whether to work at all, is governed by a single threshold.
The distinction is more than bookkeeping. Estimated intensive-margin elasticities are small: already-employed workers do not change their hours much when wages move. Extensive-margin elasticities are larger, because a wage or benefit change pushes people across the participation threshold, in and out of the labor force entirely. Policy that ignores the extensive margin, say a tax credit aimed at hours when the live response is participation, will mispredict its own effect.
Figure 21.1. Labor supply: income vs. substitution. Top panel: the consumption-leisure budget line pivots as the wage changes, and the optimal bundle traces the labor-supply locus. Bottom panel: the implied hours-vs-wage curve, with the backward bend highlighted once the income effect dominates. Drag the wage slider through the full range to see the bend.
Problem. A worker has $U = C^{0.5}\ell^{0.5}$, time endowment $T = 16$, non-labor income $V = 40$, wage $w = 10$. (a) Find optimal leisure and hours. (b) The wage rises to $w = 20$: does the worker supply more or fewer hours? (c) Find the reservation wage.
Solution.
(a) With Cobb-Douglas, $\ell^* = (1-a)(wT+V)/w = 0.5(160+40)/10 = 10$, so $h^* = 16 - 10 = 6$. Consumption $C^* = 10(6)+40 = 100$.
(b) At $w = 20$: $\ell^* = 0.5(320+40)/20 = 9$, so $h^* = 7$. Hours rise from 6 to 7; the substitution effect still dominates here. (With $V$ larger and $w$ higher, the spreadsheet of $\ell^* = 0.5(wT+V)/w = 8 + 0.5V/w$ shows leisure climbing back as $w$ keeps rising and $V$ matters less in relative terms, which is the backward bend.)
(c) The reservation wage solves $h^* = 0 \Rightarrow \ell^* = T \Rightarrow 0.5(wT+V)/w = 16$. Solving: $0.5 \cdot 16 + 0.5 V/w = 16 \Rightarrow 0.5V/w = 8 \Rightarrow w = V/16 = 40/16 = 2.5$. Below $w = 2.5$ the worker stays out of the labor force.
A firm does not want workers for their own sake. It wants what they produce, and it sells that output in a product market. So the demand for labor is not a primitive; it is read off the demand for the product. The marginal-productivity account of factor pricing, of which this is the labor case, originates with John Bates Clark in the 1890s.
A competitive, wage-taking firm hires up to the point where the last worker's marginal revenue product equals the wage. Hiring beyond that loses money on the marginal worker; stopping short leaves profitable hires on the table.
Why it matters: Picture a sandwich shop. The owner keeps hiring as long as each new worker brings in more sandwich revenue than they cost in wages. The first hire is worth a lot, because they can barely keep up alone. The fifth hire adds less, because the others are already covering the rush. The owner stops at the worker whose added sales just cover their pay. Now imagine sandwiches suddenly sell for double. Every worker's contribution to revenue doubles, so it pays to hire more at the same wage. That is the whole point of calling labor demand “derived”: a boom in the product market is, for workers, a hiring boom, and a bust in the product market is a layoff, even if nothing about the workers changed.
How much employment falls when the wage rises is governed by four forces, codified by Marshall and refined by Hicks. Labor demand is more elastic, meaning employment is more wage-sensitive, when (1) the product the labor makes has more elastic demand; (2) labor is more easily substituted by other inputs, such as machines; (3) the supply of those other inputs is more elastic; and (4) labor is a larger share of total cost. The fourth rule has a famous exception: labor demand is more inelastic when labor's cost share is large and output demand is inelastic, because then a wage rise that raises price barely dents output. These Marshall-Hicks rules are exactly what determines whether a minimum wage costs jobs, which is why they return in force at §21.6.
In the short run, with capital fixed, labor demand is just the downward-sloping $MRPL$ curve. In the long run, capital adjusts too: a wage rise prompts substitution toward capital (the scale effect and the substitution effect both operate), making long-run labor demand more elastic than short-run.
Figure 21.2. Labor demand from the marginal revenue product. Production is $Y = A K^{0.3} L^{0.7}$ with capital fixed, so $MPL = 0.7 A K^{0.3} L^{-0.3}$ and $MRPL = P \cdot MPL$. Employment is where $MRPL$ meets the prevailing wage. A higher output price or higher productivity shifts the curve out and raises employment. Drag the sliders.
The general theory of factor demand and cost minimization, covering isoquants, the conditional factor-demand functions, and the firm's output decision, is developed in Chapter 5 and Chapter 6; here it is specialized to labor.
Labor supply governs how much you work; labor demand governs whether a firm wants you. Neither yet explains why one worker commands triple another's wage. The bridge is human capital: the idea, due to Gary Becker and Theodore Schultz, that schooling and training are investments. You forgo income and pay tuition now to earn a higher wage later, exactly as a firm forgoes cash now to buy a machine that pays off over its life.
How much schooling? The investment logic gives a clean rule. Each extra year of school raises the wage by some proportion but costs a year of earnings forgone (plus tuition). The worker keeps investing until the marginal return on the last year of schooling falls to the discount rate, the return available on alternative investments.
Equation 21.6 is the schooling first-order condition: invest in education up to the year whose proportional wage return equals the interest rate $r$. If schooling returns 10% and money costs 6%, take the year; when the return falls to 6%, stop.
Jacob Mincer turned that investment logic into the single most-run regression in labor economics. Log earnings are linear in years of schooling and quadratic in experience.
Because experience enters quadratically with $\beta_2 < 0$, earnings rise with experience but at a decreasing rate, peaking at $x^* = -\beta_1 / (2\beta_2)$, typically in the late 40s to early 50s, and declining thereafter.
Why it matters: The Mincer equation is just the shape of a career, written down. Earnings jump for each year of schooling: that is the $\rho$ term, the schooling premium. Then, once you are working, pay climbs fast in your twenties and thirties as you learn the job, slows in your forties, and tips over before retirement as skills date and energy fades. That arc is the hump the quadratic in experience draws. The whole life-cycle of a paycheck is a straight line in schooling and an arch in experience.
There is a problem with reading $\rho$ as the causal return to education. The people who get more schooling are not a random sample: they tend to be more able, more motivated, from richer families. Some of the earnings gap that the regression attributes to schooling is really paying for ability that would have raised earnings anyway. This is ability bias, and it pushes the naive estimate up.
Separating schooling's true effect from ability's requires the credibility-revolution toolkit. The standard move is instrumental variables: find something that pushes people into more schooling but is otherwise unrelated to their earning power, such as compulsory-schooling laws and quarter of birth in the celebrated Angrist-Krueger design, and use only that variation to estimate the return. The instrument isolates a slice of schooling that is plausibly unrelated to ability, and the resulting IV estimate is the causal return for the people the instrument moved. The instrumental-variables machinery itself, including the exclusion restriction and what a local average treatment effect means, is developed in Chapter 10; the point here is only that it is what lets us read $\rho$ as a return rather than a correlation.
The signaling-versus-human-capital question is not fully settled, but the weight of the IV evidence is that schooling does build productive skill, with a causal return that is real and large even if somewhat below the naive OLS figure, rather than serving as pure certification.
Figure 21.3. Mincer earnings profiles. Log earnings plotted against experience for three schooling levels (12, 16, 20 years). The vertical gaps between curves are the schooling premium $\rho$; each curve is concave, peaking at $-\beta_1/(2\beta_2)$. The OLS/IV toggle shifts $\rho$ down by an illustrative ability-bias wedge to show the correction's direction (illustrative, not fitted). Drag the sliders.
Problem. An estimated Mincer regression is $\ln w = 1.5 + 0.10\,s + 0.04\,x - 0.0006\,x^2$. (a) What is the percentage earnings premium for a college graduate (16 years) versus a high-school graduate (12 years), at the same experience? (b) At what experience do earnings peak? (c) If ability bias inflates the OLS schooling coefficient by 0.025, what is the implied IV return, and the corrected college premium?
Solution.
(a) Four extra years at $\rho = 0.10$ raises log earnings by $0.40$, so the premium is $e^{0.40} - 1 = 0.492$, about 49%.
(b) The peak is $x^* = -\beta_1/(2\beta_2) = -0.04/(2 \times -0.0006) = 33.3$ years of experience, or around age 55 for someone who started work at 22.
(c) The IV return is $0.10 - 0.025 = 0.075$. The corrected college premium is $e^{0.075 \times 4} - 1 = e^{0.30} - 1 = 0.350$, about 35%, still large but a third smaller than the naive OLS figure once ability is netted out.
Everything so far has assumed the labor market clears: at the going wage, everyone who wants a job at that wage has one. But there is always unemployment, even in booms, and there are always job vacancies, even in slumps. A market that cleared would not have both at once. The search-and-matching theory of Diamond, Mortensen, and Pissarides, the work that won them the 2010 Nobel, explains why, and it does so without calling unemployment a market failure.
Why it matters: Think of the labor market like a dating market. There are single people who want partners and single people who want partners, so supply and demand are both there, and yet at any moment plenty of people are unattached. Not because the market “failed,” but because finding the right match takes time: you have to meet, screen, and agree. The same is true of jobs and workers. Unemployed people are looking, firms with vacancies are looking, and they are looking for each other, but the looking itself eats time. Some unemployment is just the snapshot of everyone mid-search. That is the whole idea: a frictional rate of unemployment is what a healthy market looks like, not a sign something is broken.
Dividing the matching function by $u$ gives the job-finding rate $f(\theta) = A\theta^{1-\alpha}$; dividing by $v$ gives the vacancy-filling rate $q(\theta) = A\theta^{-\alpha}$. The two move in opposite directions in tightness, which is the engine of the whole model: a hotter market is good for workers and slow for firms.
Unemployment is a stock fed by a flow in and drained by a flow out. Workers lose jobs at the separation rate $\delta$ (the inflow) and find jobs at rate $f(\theta)$ (the outflow). In steady state the two flows balance, which pins down the unemployment rate.
Setting inflow $\delta(1-u)$ equal to outflow $f(\theta)u$ and solving gives Eq. 21.9. A higher separation rate raises $u^*$; a higher job-finding rate (a hotter market or better matching) lowers it. The natural rate of unemployment is, in this model, just $u^*$ evaluated at the economy's structural $\delta$ and $A$.
Plot unemployment against vacancies over the business cycle and the points trace a downward-sloping curve: the Beveridge curve, the empirical signature of the matching model. Movements along it are the cycle. Shifts of it are the interesting events. An outward shift means the same number of vacancies now coexists with more unemployment, the diagnostic of a matching breakdown (skills mismatch, geographic immobility, a pandemic reshuffling who wants which job).
A matched worker and firm together create a surplus, the value of the filled job over the value of staying separate. Neither can produce it alone, so they bargain over the split. With worker bargaining power $\beta$, productivity $p$, the firm's recruiting cost $c$, and the flow value of non-work $b$, the bargained wage is a weighted average of the surplus the match creates and the worker's outside option.
The $\beta(p + c\theta)$ term is the worker's slice of the joint surplus, including the recruiting cost the firm saves by not having to search again; the $(1-\beta)b$ term is what the worker could get by walking away. Tighter markets raise $c\theta$ and so raise the bargained wage: labor's price rises when jobs are easier to find.
Where this frictional rate sits in the macro picture, including its relation to the natural rate, the Phillips curve, and the cyclical-versus-structural decomposition of unemployment, is taken up in Chapter 7 and Chapter 15.
Figure 21.4. Beveridge curve and DMP steady state. Top: the Beveridge curve in $(u,v)$ space (the locus of flow-steady-state points as tightness varies), with the current point marked; lowering matching efficiency $A$ shifts the whole curve outward. Bottom: the steady-state unemployment rate $u^* = \delta/(\delta + A\theta^{0.5})$ with the job-finding rate. Drag the sliders; watch the curve shift out as $A$ falls.
Problem. The separation rate is $\delta = 0.03$ (monthly), the matching function is $M = 0.5\,u^{0.5}v^{0.5}$, and tightness is $\theta = 1$. (a) Find the job-finding rate and the steady-state unemployment rate. (b) The separation rate doubles to $\delta = 0.06$ in a recession, tightness unchanged: what happens to $u^*$? (c) Describe the Beveridge-curve movement in each case.
Solution.
(a) $f(\theta) = A\theta^{0.5} = 0.5(1)^{0.5} = 0.5$. Then $u^* = 0.03/(0.03 + 0.5) = 0.03/0.53 = 0.0566$, about 5.7%.
(b) With $\delta = 0.06$: $u^* = 0.06/(0.06 + 0.5) = 0.06/0.56 = 0.107$, about 10.7%, nearly double.
(c) A pure separation-rate shock with unchanged matching efficiency moves the economy along a given Beveridge curve toward higher unemployment. If instead matching efficiency $A$ fell (skills mismatch), the curve itself would shift outward, the same vacancies coexisting with higher unemployment. Distinguishing the two is the central diagnostic of post-recession labor-market analysis.
The competitive model assumes a firm can hire all the workers it wants at the going wage: the firm is a wage-taker. But what if a firm is large in its local labor market, or workers cannot easily move between employers? Then the firm faces an upward-sloping labor supply curve: to hire more workers, it must raise the wage. That single fact changes everything.
Why it matters: Imagine the only hospital in a small town. If it wants one more nurse, it cannot just post a slightly higher wage for the new hire. It has to raise the pay of every nurse already there, because they will not stand for a newcomer earning more for the same work. So the true cost of that one extra nurse is not just her wage; it is her wage plus the raises for everyone else. That makes hiring expensive at the margin, so the hospital hires fewer nurses and pays them all less than they are worth. The nurse is not paid what she produces because the employer is the only game in town. That is monopsony: one buyer of labor, marking the wage down because it can.
The $w$ term is the new worker's wage; the $L\,(dw/dL)$ term is the raise paid to all $L$ existing workers because attracting the marginal hire bid the wage up. With linear inverse supply $w = a + bL$, the marginal cost of labor is $MCL = a + 2bL$, twice as steep in $L$ as supply, exactly as marginal revenue is twice as steep as a linear demand curve for a monopolist.
The monopsonist hires where $MCL = MRPL$, then reads the wage off the supply curve at that employment, which lies below $MRPL$. Both employment and the wage are below the competitive levels: $L_m < L_c$ and $w_m < w_c$. The worker is paid less than they produce, and fewer workers are employed than under competition.
Here is the result that makes monopsony matter for policy. Under competition, a binding minimum wage costs jobs, full stop. Under monopsony, a minimum wage set between $w_m$ and $w_c$ raises employment. The reason is the marginal-cost mechanism: a wage floor makes labor supply flat at the floor up to the supply curve, so hiring one more worker no longer requires raising everyone's wage. The marginal cost of labor drops to the floor itself. The monopsonist, no longer paying the markup penalty for expansion, hires more. Push the floor too high, above $w_c$, and the floor binds like the competitive case and employment falls. In between, the floor undoes the monopsony distortion and employment rises.
Figure 21.5. Monopsony. Inverse supply $w = 2 + bL$, marginal cost of labor $MCL = 2 + 2bL$, and $MRPL = 20 - 0.5L$. The monopsony point sits where $MCL = MRPL$, with the wage read off supply, below the competitive point. Drag the minimum-wage line up: employment rises while the floor sits between $w_m$ and $w_c$, then falls once it climbs above the competitive wage. Drag both sliders.
Joan Robinson coined "monopsony" in The Economics of Imperfect Competition (1933) and worked out the markdown geometry. For decades it was treated as a textbook curiosity: surely workers had many employers to choose from. The revival came with Alan Manning's Monopsony in Motion (2003), which made the case that monopsony power does not require a literal single employer. Search frictions are enough: if it is costly to find and switch jobs, every employer has a little wage-setting power, because workers cannot instantly flee a wage cut. This is the "new monopsony": pervasive, frictions-based, and a property of normal labor markets rather than company towns. Recent work measuring labor-market concentration directly (Azar, Marinescu, and Steinbaum) finds many local labor markets are concentrated enough for the effect to bite.
Monopsony as a market structure, the buyer-side mirror of monopoly, is introduced in Chapter 6; its labor-market application is the home built here.
Problem. Inverse labor supply is $w = 2 + 0.5L$ and $MRPL = 20 - 0.5L$. (a) Find the competitive wage and employment. (b) Find the monopsony wage and employment, and the markdown. (c) What minimum wage maximizes employment?
Solution.
(a) Competitive: $w = MRPL$ on supply, $2 + 0.5L = 20 - 0.5L \Rightarrow L = 18$, $w_c = 2 + 0.5(18) = 11$.
(b) $MCL = 2 + L$. Set $MCL = MRPL$: $2 + L = 20 - 0.5L \Rightarrow 1.5L = 18 \Rightarrow L_m = 12$. Wage off supply: $w_m = 2 + 0.5(12) = 8$. The $MRPL$ at $L_m = 12$ is $20 - 6 = 14$, so the markdown is $14 - 8 = 6$.
(c) A minimum wage flattens supply at the floor, so employment equals labor demanded at the floor up to the supply limit. Employment is maximized at the competitive wage: set $\bar w = w_c = 11$, giving $L = 18$. Any floor in the range $(8, 11]$ raises employment above the monopsony 12; the floor at exactly 11 hits the competitive employment of 18. Above 11, employment falls back along demand.
No question in labor economics has been fought harder, and on no question has the profession changed its mind more visibly. The competitive model gives a clean, confident prediction; a famous natural experiment appeared to refute it; the refutation was attacked on method; and out of three decades of that fight a within-mainstream settlement emerged. It is worth walking each step at full strength, because the settlement is only credible if you have felt the force of the position it moved away from.
Start with the textbook view, stated as strongly as it deserves. In a competitive labor market, the wage equals the marginal revenue product. A minimum wage set above that level prices out every worker whose product is below the floor: the firm will not pay someone twelve dollars to produce ten dollars of value. Employment falls, and the workers who lose their jobs are precisely the least-skilled and most vulnerable: teenagers, the inexperienced, the marginal. The policy meant to help the low-paid hurts a subset of them by pricing them out of work entirely. This is not a fringe position or a strawman; it follows rigorously from the model in §21.2, it was the consensus of the profession for most of the twentieth century, and the Marshall-Hicks elasticity rules tell you exactly how large the job loss should be. If labor demand is elastic, the disemployment is severe.
In April 1992 New Jersey raised its minimum wage from \$4.25 to \$5.05; neighboring Pennsylvania did not. David Card and Alan Krueger surveyed fast-food restaurants on both sides of the border before and after the increase and ran a difference-in-differences: the change in New Jersey employment minus the change in Pennsylvania employment over the same window. The competitive model predicted New Jersey employment should fall relative to Pennsylvania. It did not. If anything, employment rose slightly in New Jersey. The cleanest available test of the textbook prediction came back with the wrong sign.
The difference-in-differences design, using Pennsylvania as the counterfactual for what New Jersey would have done absent the increase on the parallel-trends assumption, is the credibility-revolution method developed in Chapter 10. The Card-Krueger study is one of the studies that made that method famous; what makes the result load-bearing is not a regression run here but the design's transparency about its counterfactual.
The result did not go unanswered, and the answer deserves its own voice. David Neumark and William Wascher argued that the survey data were unreliable: the phone-survey employment counts were noisy and possibly biased. Using administrative payroll records for the same establishments, they found that New Jersey employment did fall relative to Pennsylvania, reversing the headline. More broadly, they assembled a large body of studies and argued the preponderance still showed negative employment effects, especially for teenagers and the least-skilled, and that the Card-Krueger result was an artifact of one industry, one border, and questionable data. Their position is not obscurantism; it is a serious empirical case that the disemployment effect is real and that the cleaner the data, the more it shows up.
Three decades of replication, meta-analysis, and better data have produced something the profession can mostly stand behind, and it is not a draw. The competitive model's clean disemployment prediction does not survive contact with the moderate-hike evidence: across many studies and settings, minimum-wage increases in the range actually legislated produce employment effects clustered near zero, often statistically indistinguishable from no effect. Monopsony is why: search frictions give employers enough wage-setting power that a floor in the right range corrects a distortion rather than creating one, exactly the mechanism of §21.5. Card and Krueger have largely survived; the burden of the evidence sits on their side for moderate increases.
But this is a calibration, not a blank check. The honest reading is that the live disagreement is about magnitude and where the buffer runs out, not about which frame is right. A minimum wage corrects monopsony only up to the competitive wage; push it well past local productivity and the disemployment the textbook predicted reappears, because the buffer is finite and local. A floor that works in a high-wage city can cost jobs in a low-wage rural county where the same nominal number is a far larger share of the marginal product. So the profession's read is: monopsony is real and widespread enough that moderate minimum-wage increases do not cost the jobs the competitive model predicted, but the policy has to be calibrated to local labor-market conditions, because the room it exploits is bounded. That is a within-mainstream position taken with caveats, not a both-sides shrug, and not a claim that minimum wages are free.
Figure 21.6. The minimum-wage call, made visible. Predicted change in employment as the minimum wage rises, under a competitive labor market (monotone job loss) versus a monopsonistic one (gains, then losses past the competitive wage). The shaded horizontal band is the Card-Krueger / modern-meta employment estimate, near zero at moderate hikes, annotated from the literature and not fitted here. At moderate hikes the band sits on the monopsony side; at large hikes both models and the data agree on disemployment, where the buffer runs out. Toggle the regime and drag the floor.
Problem. Average full-time-equivalent employment per fast-food store (stylized, after Card-Krueger): New Jersey 20.4 before, 21.0 after; Pennsylvania 23.3 before, 21.2 after. (a) Compute the difference-in-differences estimate of the minimum wage's employment effect. (b) Interpret the sign against the competitive prediction.
Solution.
(a) New Jersey change: $21.0 - 20.4 = +0.6$. Pennsylvania change: $21.2 - 23.3 = -2.1$. Difference-in-differences: $(+0.6) - (-2.1) = +2.7$ FTE per store. The minimum-wage increase is associated with employment rising in New Jersey relative to the no-increase control.
(b) The competitive model predicts a negative number: the treated state should lose jobs relative to control. The estimate is positive. The sign is the wrong sign for the textbook prediction and the right sign for the monopsony mechanism, which is exactly why the result reframed the debate. (The DiD rests on parallel trends: that absent the increase, New Jersey would have tracked Pennsylvania, the identifying assumption examined in Chapter 10.)
This chapter is the apparatus home for the minimum-wage walkthrough. You now have the monopsony model and the Card-Krueger evidence; here is what they settle and what they leave open.
Two models, two predictions. The competitive market (§21.2) says any binding floor prices out the least productive workers: employment falls, and the fall is larger the more elastic labor demand. The monopsony market (§21.5) says the opposite over a range: because the single buyer marks the wage down below marginal product, a floor between the monopsony wage and the competitive wage corrects the distortion and raises employment. The whole policy question is which model describes a given low-wage labor market, and at what hike size the monopsony buffer is exhausted.
Neumark and Wascher's administrative-data critique is not dead; it is the reason the verdict is “calibrate” rather than “raise freely.” The honest disagreement is no longer competitive-versus-monopsony in the abstract; it is about the magnitude of monopsony power in different markets and the wage level at which the buffer runs out. A hike that is harmless in a high-wage metro can bind hard in a low-wage county where the same number is a much larger share of marginal product.
The within-mainstream call: moderate minimum-wage increases do not destroy the jobs the competitive model predicts, because monopsony power in low-wage labor markets is real and widespread. But the policy implication is calibration, not a free lunch: set the floor to local conditions, because the room monopsony provides is bounded. This is a position with caveats, taken on the evidence; the full controversy, with the politics and the \$15 fight, is the walkthrough's territory.
Both slogans assume their own model is the whole story. The competitive camp treats every floor as a job-killer; the activist camp treats monopsony as a blank check. The actual economics says the right number is local, and both confident soundbites are wrong about that.
AdvancedFor a long time the story about technology and wages was "skill-biased technical change": computers complement skilled workers, so they raise the skill premium, and inequality among workers tracks the spread of computing. The story fit the 1980s but broke on the 1990s and 2000s, when the middle of the wage distribution hollowed out while both the top and the bottom grew, a pattern pure skill bias cannot produce. The task framework, due to David Autor and to Daron Acemoglu and Pascual Restrepo, explains the hollowing.
Why it matters: The key move is to stop talking about jobs and start talking about tasks. Machines do not take jobs; they take tasks. A job is a bundle of tasks, and automation eats the ones that are routine: the codifiable, rule-following steps a payroll clerk or an assembly-line inspector does. What is left, and what gets created, sits at the two ends: abstract tasks that need judgment (the analyst, the manager) and manual tasks that need a human body in an unpredictable place (the nurse, the cleaner). So the middle empties out and the ends fill up. That is why automation does not simply mean “fewer jobs”; it means a different shape of jobs, with a hole in the middle.
Equation 21.13 is the Acemoglu-Restrepo task-content decomposition (schematic). Automation has three effects on labor demand: a displacement effect (negative, as machines take tasks), a reinstatement effect (positive, as new tasks emerge where labor has the edge), and a productivity effect (positive, as cheaper production raises output and thus demand for the tasks still done by labor). The net sign is an empirical, period-specific question, and the worry of the current frontier is that recent automation has been displacement-heavy and reinstatement-light.
Polarization is the framework's signature prediction and its strongest evidence. If technology simply favored skill, employment would shift smoothly up the skill ladder. Instead it hollows the middle, because the middle is where routine, codifiable tasks concentrate. The framework is the field's current frontier rather than a settled theory; Autor himself has revised his read of how fast and how far automation displaces, and the live debate over generative AI, whether it finally automates the abstract tasks that were supposed to be safe, is where the action is now.
The China trade shock of the 2000s is the other great middle-class-job displacement of the period, and the economic-history book narrates it: the globalization chapter covers the labor-market displacement of the import-competition era.
Figure 21.7. The task framework and polarization. Employment by task-content tercile: low (manual), middle (routine), high (abstract). Automation displaces middle-routine labor proportionally to intensity; reinstatement adds new high- and low-end tasks; the productivity effect scales all three. With reinstatement on, the middle hollows out while the ends grow into the polarization U. Displacement-only under-predicts the net change. Drag the intensity slider and toggle the channels.
The chapter closes with three topics that share a theme: how the labor market distributes income, not just how it sets the price of an hour. Each could fill a course; here they are the compact account of who gets what within the labor market and why the labor share of national income has been falling.
There are two leading economic models of labor-market discrimination, and they are best read as complementary lenses rather than rivals. Gary Becker's taste-based model treats discrimination as a preference: an employer who dislikes hiring a group acts as if their wage carried a surcharge. Becker's striking implication is that competition should erode it: a non-discriminating firm hires the cheaper-but-equally-productive workers the bigot passes over, and undercuts him. Taste-based discrimination is therefore hardest to sustain in competitive markets and most durable where firms have rents to spend on their tastes.
The Phelps-Arrow model of statistical discrimination needs no animus at all. Employers cannot observe a job applicant's true productivity, so they fall back on whatever they can observe, including group membership, as a noisy signal. A group with a lower average, or merely a noisier signal, is treated worse even by a perfectly unprejudiced employer. Worse, the belief can be self-fulfilling: if a group expects to be underpaid, it underinvests in the skills that would justify higher pay, confirming the statistic. The empirical anchor for both models is the audit or correspondence study, which sends otherwise-identical résumés that differ only in a name signaling race or gender and measures the callback gap. These studies consistently find gaps that cannot be explained by productivity, which is the evidence that some form of discrimination operates.
The explained term values the difference in average characteristics $(\bar X_A - \bar X_B)$ at a common return $\hat\beta$; the unexplained term values group B's characteristics at the difference in returns $(\hat\beta_A - \hat\beta_B)$. The unexplained part is consistent with both taste-based and statistical discrimination, and with unmeasured productivity, which is why the decomposition bounds rather than measures discrimination.
Problem. Group A out-earns group B by 20 log points (about 22%). Differences in observable characteristics, experience and hours, account for a gap of 12 log points when valued at common returns. (a) What is the explained share? (b) What is the unexplained residual, and how should it be interpreted?
Solution.
(a) Explained $= 12$ of $20$ log points $= 60\%$. Most of the raw gap reflects differences in measured experience and hours.
(b) Unexplained residual $= 20 - 12 = 8$ log points $= 40\%$ of the gap. This is consistent with discrimination, but also with unmeasured productivity differences and with characteristics the regression omits. It is an upper bound on discrimination, not a measurement of it. The figure below shows the split.
A union changes the wage-setting game from price-taking to bargaining. Two models bracket the outcomes. In the monopoly-union model, the union unilaterally sets the wage and the firm chooses employment off its labor-demand curve, so the union trades higher wages for fewer jobs, like a monopolist trading price for quantity. The right-to-manage model is more realistic: the union and firm bargain over the wage, then the firm sets employment. Either way the union raises the wage of its members above the competitive level.
The insider-outsider mechanism explains a puzzle the competitive model cannot: why high wages persist alongside unemployment. Insiders, already employed and costly to replace, bargain for wages that ignore the outsiders who would happily undercut them, because firing-and-rehiring is expensive enough to give insiders the leverage. The result is wage rigidity and a pool of outsiders who cannot bid their way in.
For most of the twentieth century the labor share looked like a constant, near two-thirds of national income, stable enough that Kaldor listed it among the "stylized facts" any growth model had to reproduce. Since around 1980 it has fallen, several percentage points across most advanced economies, and there is no consensus on why. Three mainstream accounts compete, each mapping to a different parameter moving.
The first is technology: capital-augmenting technical change and a fall in the price of capital goods (think automation, the task framework's displacement effect) make it cheaper to substitute capital for labor, shifting income toward capital. The second is rising market power, both product-market markups and labor-market monopsony power (the markdown of §21.5), which mechanically shrinks labor's slice as firms capture a larger surplus. The third is globalization: offshoring and the China trade shock put downward pressure on the wages and bargaining position of workers in import-competing sectors. The lean of the evidence is toward a combination of capital-biased technology and rising market power over pure globalization, since concentration and markups have risen in tandem with the labor-share fall and the timing fits. But this is a live question, not a verdict, and the three channels are not mutually exclusive.
Figure 21.8. The U.S. labor share of national income, schematic, with the ~1980 inflection marked. Roughly stable near two-thirds through the postwar decades, then declining several points after 1980. (Illustrative series, annotated from the literature; the inflection, not the exact level, is the point.)
The labor-share decline is one of the engines of the rise in top-end income inequality; the question of whether and how that inequality can be addressed is taken up at debate depth in the inequality walkthrough, which treats this chapter's labor apparatus as its source for the distribution mechanics.
A supporting stop. The labor-share decline and the wage-gap decomposition you just met are the labor-market mechanics behind the inequality debate.
A large part of rising inequality is made in the labor market, and §21.8 gives the mechanics: the falling labor share moves income from wages to capital (which is far more concentrated than labor income), and the polarization of §21.7 spreads the wage distribution itself. The Oaxaca-Blinder decomposition adds the within-group dimension: how much of a gap survives after observables. These are the pieces an inequality argument has to assemble before it can talk about remedies.
This chapter does not take the inequality question's position; that is the walkthrough's job. It supplies the labor-market apparatus the walkthrough reasons with: the three competing accounts of the labor-share fall, the polarization mechanism, and the monopsony markdown. The walkthrough carries it up to the efficiency-equity tradeoff and the policy fight.
| Label | Equation | Description |
|---|---|---|
| Eq. 21.1 | $C = w(T-\ell) + V$ | Consumption-leisure budget constraint |
| Eq. 21.2 | $\max U(C,\ell)$ s.t. $C = w(T-\ell)+V$ | Labor-supply problem |
| Eq. 21.3 | $MRS_{\ell,C} = w$ | Optimal-hours tangency |
| Eq. 21.4 | $MRPL = MPL \cdot MR\;(= MPL \cdot P)$ | Marginal revenue product of labor |
| Eq. 21.5 | $w = MRPL$ | Competitive hiring rule |
| Eq. 21.6 | $\frac{\partial w(s)/\partial s}{w(s)} = r$ | Schooling first-order condition |
| Eq. 21.7 | $\ln w = \ln w_0 + \rho s + \beta_1 x + \beta_2 x^2$ | Mincer earnings function |
| Eq. 21.8 | $M = A u^{\alpha} v^{1-\alpha}$ | Cobb-Douglas matching function |
| Eq. 21.9 | $u^* = \delta/(\delta + f(\theta))$ | Flow steady-state unemployment |
| Eq. 21.10 | $w = \beta(p + c\theta) + (1-\beta)b$ | DMP Nash-bargained wage |
| Eq. 21.11 | $MCL = w + L\,dw/dL$ | Marginal cost of labor (monopsony) |
| Eq. 21.12 | $MRPL = MCL > w_m$ | Monopsony equilibrium and markdown |
| Eq. 21.13 | $\Delta(\text{labor dem.}) = -\text{displ.} + \text{reinst.} + \text{prod.}$ | Acemoglu-Restrepo task decomposition |
| Eq. 21.14 | $\bar w_A - \bar w_B = (\bar X_A - \bar X_B)\hat\beta + \bar X_B(\hat\beta_A - \hat\beta_B)$ | Oaxaca-Blinder decomposition |
Becker (1964, 1971); Mincer (1974); Robinson (1933); Manning (2003); Azar, Marinescu & Steinbaum (2020); Diamond (1982); Mortensen & Pissarides (1994); Pissarides (2000); Card & Krueger (1994, 1995); Neumark & Wascher (2000, 2008); Autor, Levy & Murnane (2003); Autor & Dorn (2013); Acemoglu & Restrepo (2018, 2019); Angrist & Krueger (1991); Phelps (1972); Arrow (1973); Oaxaca (1973); Blinder (1973); Lindbeck & Snower (1988); Karabarbounis & Neiman (2014); Autor, Dorn, Katz, Patterson & Van Reenen (2020).