Mechanism · a sign that reverses, and the number that decides
The Wage Floor That Hires
Everyone knows what a price floor does to a market: it prices somebody out. Below is that market, with one switch. The switch does not touch the supply curve or the demand curve. It changes only what the employer was doing before the floor arrived. Flick it, and the same floor slider stops destroying jobs and starts adding them.
Wage on the vertical axis, jobs on the horizontal, both fixed in units of the wage-setting firm's own unregulated outcome (w_m and L_m) whichever way the switch is set. That is deliberate: the same absolute floor is applied in both modes, which is what makes the comparison fair. The price taker's unregulated point is the marked dot, and it is not at (1, 1).
Jobs at this floor
—
vs this mode's own unregulated level
Wage as a share of the last worker's value
—
before any floor
Marginal cost of one more worker
—
at the unregulated hire count
Nothing was swapped in. Both modes stand on the same two curves: a labour supply curve L = A·wε and a marginal-revenue-product curve MRP = B·L−1/η. What the switch changes is which curve the firm optimises against. A price taker hires until the value of the last worker equals the wage. A firm that sets the wage knows that paying one more dollar to attract the next worker means paying it to everyone already there, so the true cost of that next worker is not the wage but
and it stops short. That wedge, 1 + 1/ε, is the entire mechanism. It is also why this has nothing to do with there being only one employer in town: the wedge is finite whenever the supply of labour to a single firm is finite, which search costs, commuting, and simply preferring one job to another are all enough to produce. A market with a thousand employers can have a wedge; a company town with a perfectly elastic labour supply would not.
None of this is recent. The word monopsony enters economics in Joan Robinson's The Economics of Imperfect Competition (1933), and she credits the actual coining of it to the Cambridge classics scholar B. L. Hallward. The opposite prediction is the one that became conventional, and Card and Krueger state it in the opening paragraph of the paper you are about to recompute: "The prediction from conventional economic theory is unambiguous: a rise in the minimum wage leads perfectly competitive employers to cut employment (George J. Stigler, 1946)." So the vocabulary of the reversal is ninety-odd years old and the prediction it reverses was already the conventional one eighty years ago. Neither half of this argument is new, and neither was ever fringe. What has changed is that the parameter separating them is now something people measure.
Because the firm was standing short of the competitive hire count, a floor that pushes the wage up towards that count pushes employment up with it. That is not a story, it is an interval with two endpoints, and the model hands both over in closed form.
Four quantities, each computed two independent ways in your browser right now: once from the closed form, and once by a brute-force search over 20,000 points that knows nothing about the closed form. Three of them are read straight off a grid of floor levels. The fourth, the markdown, comes from a separate search that maximises the firm's profit over a grid of hire counts using only the two curves, so it never touches the marginal-cost formula the closed form is derived from. A search cannot resolve anything finer than its own step, so each row passes if the two columns agree to within the resolution of the search that produced it, printed beside every difference. If they ever disagreed by more than that, the closed form would be wrong. Move ε and η below and watch both columns move together.
| quantity | closed form | grid search | |difference| |
|---|
Both columns are computed live. The floor search maximises E(w̄) = min( A·w̄ε, (B/w̄)η ) over a uniform grid of floors and reads the peak and the crossing off the grid; its step at these settings is —. The markdown search maximises ∫MRP dL − ws(L)·L over a grid of hire counts, in two passes, and reads ws/MRP off the maximiser.
The markdowns the meta-analysis states in its own words
The wedge implies a wage markdown of 1/(1+ε), so a worker is paid ε/(1+ε) of what the last worker is worth. Sokolova and Sorensen's working paper states three of these in plain English. The right-hand column is our formula run on their ε; the middle column is the number they printed.
| their ε | their own words (published) | ε/(1+ε) computed here |
|---|
So the sophisticated dismissal, monopsony is a blackboard curiosity, not a Wendy's on Route 1, turns into a measurable question with a number attached. And the number has been measured. Sokolova and Sorensen collected 1,320 estimates of ε from 53 studies and found, in their own published abstract, "a prominent discrepancy between estimates of direct elasticity of labor supply to changes in wage (smaller) and the estimates converted from inverse elasticities (larger)". In their working-paper version the direct median is around 1.1, "implying that workers are payed 52% of their marginal product"; the inverse-method median corresponds to a supply elasticity around 4 and "a wage markdown of only 20%". Snap the dial to each. At 1.1 a floor can rise — before it does harm; at 4, —. The dismissal is answered by making it quantitative, and then the quantity turns out to be disputed inside a single meta-analysis. That is the honest state of the question.
Which brings us to the episode that made the question famous, and to the only part of this page where you can check the arithmetic against the original microdata.
The 410 restaurants
On 1 April 1992 New Jersey's minimum wage rose from $4.25 to $5.05 an hour, an increase of — computed live from those two figures. David Card and Alan Krueger telephoned 410 fast-food restaurants in New Jersey and eastern Pennsylvania before and after, and published the result in the American Economic Review in 1994. They also published the file. All 410 records of public.dat are embedded in this page; every number in the grid below is recomputed from them when you touch a control, and the value Card and Krueger printed is pinned beside it.
Card and Krueger's stated convention is 0.5. It is a convention, not a measurement: nothing in the file records hours.
Six stores closed permanently and enter wave 2 as zero throughout. Four closed temporarily; rows 3 and 4 treat them as missing, row 5 as zero.
| variable | stores by state | stores in New Jersey, by wave-1 starting wage | ||||
|---|---|---|---|---|---|---|
| PA | NJ | NJ − PA | $4.25 | $4.26–4.99 | ≥ $5.00 | |
Top line of each cell: recomputed here from public.dat, mean with standard error beneath. Second line, grey: the value printed in the AER. Green with ✓ = agrees to the printed precision, meaning our two decimals are theirs, for the mean and for the standard error alike. Amber with ✗ = does not, in at least one of the two.
Difference in differences
—
FTE employees per store
As a share of the NJ wave-1 base
—
the paper's "or 13 percent"
Stores in the wage groups
—
footnote a prints 101 / 140 / 73
The sign is robust. The famous percentage is not.
The difference in differences (green, left axis, FTE employees) and the same thing as a percentage of the New Jersey wave-1 base (amber, right axis) across every part-time weight from 0 to 1. Both curves are computed live from the file at 101 weights.
—
That is the honest robustness claim, and it is narrower than the one usually made. The sign of the New Jersey result survives every sample definition in the paper and every part-time weight between zero and one. The magnitude, the "or 13 percent" that everybody quotes, is a function of a counting convention: it is a different number at every weight, because the weight changes the denominator faster than it changes the numerator.
The Table 4 cell we cannot reproduce, and the Table 3 row whose standard errors we cannot
Table 4 of the paper regresses the change in employment on a New Jersey dummy, and alternatively on each store's initial wage gap. Four of its five models come back to the printed digits below. The fifth does not, and the two rows that report the fit for that same fifth model do.
Design matrix built and inverted in your browser by normal equations on the same n = — sample the paper describes: stores with usable employment and starting-wage data in both waves. The wage gap is (5.05 − W₁)/W₁ for New Jersey stores below the new minimum, and zero for Pennsylvania and for New Jersey stores already at or above $5.05.
| model | coefficient | std. error | s.e. of regression | p for controls | published |
|---|
All five rows recomputed live; ✓ marks agreement with the printed coefficient and standard error, ✗ disagreement. Published values from AER 84(4), Table 4, p. 779.
And there is a second, quieter thing the file will tell you. At Card and Krueger's own part-time weight, — of the thirty cells in the Table 3 grid further up print exactly the digits the AER printed, counted live with the same test that colours it. The ones that do not are of two kinds. Two are last-digit disagreements and nothing more: the balanced change at the $4.25 stores comes back — against a printed 1.21, and the row-2 difference column combines to — against a printed 1.07. The other six are all in row 3, and they fail in an unusually orderly way: every one of its standard errors is what you would print if you assumed a single fixed correlation between the two waves, rather than measuring the correlation the file actually contains. Row 3's point estimates do come back, to within six thousandths, though Pennsylvania's — rounds to one hundredth below the printed −2.16 and so is marked too.
| column | printed s.e. | correlation that printed s.e. implies | correlation in the file | s.e. from the file |
|---|
Every number in columns 3, 4 and 5 is computed live from public.dat. Column 2 is the printed value. Row 3 of the published table is the difference of two means from partly overlapping samples, so its standard error needs the wave-to-wave covariance.
This changes nothing about the sign and very little about the headline: on the file's own covariance the row-3 difference in differences is more precisely estimated, not less. We surface it because a page that recomputes a famous table and quietly drops the cells that do not match is not recomputing anything.
Heads, or hours
A year after Card and Krueger's paper, David Neumark and William Wascher re-ran the same policy change on a different kind of data: actual payroll records from 230 Burger King, KFC, Wendy's and Roy Rogers restaurants. Using Card and Krueger's own survey data their specification gave an employment increase of 17.6 percent in New Jersey relative to Pennsylvania. Using the payroll records it gave a 4.6 percent decrease, significant at the five percent level, an elasticity of employment with respect to the minimum wage of −0.24. Same state, same eight months, opposite sign.
The two files do not measure the same quantity. Card and Krueger's interviewers asked how many full-timers, part-timers and managers a store had: head counts, converted to a body-equivalent by that 0.5 weight. Payroll records count paid hours. A restaurant that trims everyone's shift by twenty minutes moves one series and not the other. Below is the accounting identity that links them, driven by the real head counts in the file. The hours per worker are yours to set, because public.dat contains no hours at all. Nothing in this instrument is a measurement of what happened.
These three sliders are assumptions you are choosing. They are not in the data and they are not in either paper.
Total paid hours per NJ store, wave 1
—
from the real head counts
The FTE weight your hours imply
—
part-time hours divided by full-time hours
The difference in differences at that weight
—
recomputed from the file
What would have to be true for both papers to be right
The head-count figure on the left of that identity is live: it follows your part-time weight in instrument 3. The −4.6 percent on the right is Neumark and Wascher's published payroll figure, not a computation. What the identity returns is the change in hours per worker that would reconcile them, which is an arithmetic consequence of the two published numbers and nothing more.
In 1998, in the working paper that became their 2000 Reply, Card and Krueger answered with a third kind of data: the Bureau of Labor Statistics ES-202 file, the administrative payroll record every employer files for unemployment insurance. Both a longitudinal and a repeated cross-section sample showed "similar or slightly faster employment growth in New Jersey relative to eastern Pennsylvania" after the rise. They also ran the natural reversal test. The 1996 federal increase raised the minimum wage in Pennsylvania but not New Jersey, pointing the experiment the other way; they found "no indication of relative employment losses in Pennsylvania". And on the payroll sample itself they reported that the difference from the administrative data was traceable to a small set of restaurants owned by a single franchisee, and that once the varying reporting intervals were controlled for, the combined sample "shows no difference in hours growth between New Jersey and Pennsylvania".
Neither side was refuted. Neumark and Wascher's payroll finding is real, published in the same issue of the same journal, and internally consistent. Card and Krueger's administrative-data reply is real too. The episode narrowed the question and did not close it, which is why the argument moved on to a different kind of estimator entirely.
The bunching ledger
The modern approach stops asking whether a group's employment went down and starts counting the wage distribution itself. Raise a minimum wage and jobs that used to pay below it must go somewhere: they either reappear just at or above the new floor, or they vanish. So count the missing jobs below the new minimum, count the excess jobs at and above it, and the net is the employment effect. Cengiz, Dube, Lindner and Zipperer did this for 138 state minimum-wage events between 1979 and 2016, across 847,314 wage-bin cells built from 4,694,104 workers, and found the two counts almost exactly cancel.
Almost exactly is the problem, and it is the frontier. In the benchmark column the answer is a difference of two numbers, one about — times its size and the other about — times, both computed live against the net that the column's own printed elasticity implies. Drag either one inside its own published standard error and watch the conclusion cross zero.
This instrument operates on the published estimates of NBER working paper 25434, Tables 1 and 2. It is not re-estimating anything: the underlying 847,314-cell regression cannot run in a browser. What is computed live is the estimator's arithmetic, using the four formulas quoted in the paper's own table note.
Net, Δa + Δb
—
share of pre-treatment total employment
Change in affected employment
—
net divided by jobs below the new minimum
Elasticity w.r.t. the minimum wage
—
published
Elasticity w.r.t. own wage
—
published
Below the resolution of the printed table
That is not an error in the paper. It is a property of what happens when the quantity in dispute is smaller than the last digit either of its ingredients is printed to. The envelope above is every value the four formulas can return given only the digits the table shows.
The specification is the result
The own-wage employment elasticity across the seven columns of the paper's own Table 1. Columns (1) to (6) differ only in bin-state trends and bin-division-period fixed effects; column (7) is not a trends variant at all but the paper's "simpler methodology", state-by-quarter data with state and year fixed effects, so it is a different estimator standing beside the other six. Bars drawn from the published Table 1; the whiskers are the published standard errors. Sign shown by the label as well as the colour.
—
The same paper also reports what a conventional two-way fixed-effects design on the log minimum wage returns for the overall state employment-to-population ratio: −0.089 (0.025), published, a large negative aggregate effect. Their bin-by-bin decomposition shows it is driven by employment shifts far above the minimum wage, where a minimum wage cannot bite. They also find one genuinely negative sector: tradables, and manufacturing in particular, with an own-wage elasticity "of around −1.4", "although the estimates are imprecise", against a positive 0.387 (s.e. 0.597) in the non-tradable sector where most American minimum-wage workers actually are.
The same shelf, read twice
Two careful surveys of an overlapping literature, both current, both by serious people, reach opposite headlines. Neumark and Shirley assembled every published US study using subnational minimum-wage variation since the early 1990s and asked the original authors which estimate they preferred, arriving at 130 preferred estimates. Dube, reviewing the international evidence for the UK Low Pay Commission, reported a median minimum-wage employment elasticity of about −0.05 across 439 estimates from 23 studies, and a median own-wage employment elasticity of −0.17 across the 36 US estimates for which it can be constructed, "around 1/6 as large" as the wage change.
Both statements are true. The disagreement is the reading rule. Set one.
Estimates counted
—
Negative but not distinguishable from zero
—
the gap between the two readings
Denominator
—
Where the denominators come from
The paper states one denominator: 130, for the whole set. The other nine rows print six percentages each and no count. For each row this page searches, live, for the smallest whole number of estimates that makes all six printed percentages round correctly, and shows the worst rounding residual it had to accept. Any multiple of that number would fit equally well, so these are the smallest consistent denominators and not the paper's own counts.
| slice | % negative | % neg, p<.10 | % neg, p<.05 | smallest N | worst residual | counts recovered |
|---|
Percentages are published (NBER WP 28388, Table 3). Smallest N, worst residual and counts recovered are searched and computed in your browser on load.
—
The coincidence you are about to notice
Run instrument 1 at the meta-analytic medians and the model predicts an employment gain in the neighbourhood of what New Jersey appeared to show. It is very tempting to call that a prediction confirmed. It is not, and the cleanest way to see why is to ask the model to hit the target on purpose. The answer is not a point. It is a curve.
Every (ε, η) pair on the curve produces exactly the target gain at the employment-maximising floor. Solved live by bisection on η at 240 values of ε.
—
Fitting one number with two free parameters is not evidence. The parameters were never estimated from New Jersey; the model is a static single-firm model with no product market, no entry, no capital and no hours margin; and the empirical estimate it is being matched to has a standard error roughly half its own size. What the model earns is the right to say the sign is not absurd, and to say exactly what would have to be true for it to hold. That is less than a prediction and considerably more than a story.
The check
What is computed in front of you. Every figure inside a bordered instrument is produced by JavaScript in your browser when the page loads and again on every input. Nothing in those panels is typed in. Live right now:
| quantity | value computed in this browser | what it is checked against |
|---|
What is published, and labelled published
These numbers are measurements or estimates from the sources below. They appear on the page as comparison values and are never presented as computed here: Card and Krueger's printed Table 3 and Table 4 (every grey "published" cell); Neumark and Wascher's 230 restaurants, +17.6%, −4.6%, elasticities +0.93 and −0.24, and the report that the standard deviation of employment change in the survey data is three times that in the payroll data; Card and Krueger's ES-202 findings, quoted from their January 1998 working paper, NBER WP 6386, which became their 2000 Reply; Cengiz and co-authors' Δa, Δb, wage effects, base shares, standard errors, the −0.089 (0.025) two-way fixed-effects elasticity, the tradable-sector −1.4 and the non-tradable +0.387 (0.597), and the counts 138 / 847,314 / 4,694,104; Neumark and Shirley's ten rows of percentages and their stated total of 130; Dube's −0.05 across 439 estimates from 23 studies and −0.17 across 36 US estimates; Sokolova and Sorensen's 1,320 estimates from 53 studies and their three stated markdowns; Lichter, Peichl and Siegloch's mean −0.508, median −0.386 and standard deviation 0.774 from 942 estimates in 105 studies; and New Jersey's $4.25 and $5.05.
Every free choice on this page
- ε and η have no true value here. The sliders open on ε = 1.1 and η = 0.386 because those are the medians two meta-analyses report, not because New Jersey was measured to have them. Nothing on this page estimates either parameter.
- The part-time weight of 0.5 is Card and Krueger's stated convention, not a measurement. The sweep exists precisely so you do not have to take it on trust.
- The hours sliders in instrument 4 are assumptions supplied by you. The file contains no hours. Anything that instrument returns is an accounting consequence of your inputs and two published numbers.
- The wage-gap variable uses (5.05 − W₁)/W₁, zero for Pennsylvania and for New Jersey stores already at or above $5.05, which is the paper's own definition. Stores with a missing wave-1 wage drop out of Table 4, which is how n = 357 arises.
- Which two New Jersey region dummies model (v) uses is not determined by the paper: it says only "two regions of New Jersey and two regions of eastern Pennsylvania", and the file carries four New Jersey location flags. We use two of the three mutually exclusive north/central/south flags (central and south); any two of the three give an identical fit, and all three at once are collinear with the intercept. This is the specification whose regression standard error and F p-value both reproduce the printed values exactly.
- The denominators in instrument 6 are inferred, not published. Only 130 is stated. The rest are the smallest whole numbers consistent with all six printed percentages; any multiple would fit. Four of them disagree by one with the category counts in the paper's own preceding table, and the instrument says so rather than picking a side.
- The rounding envelope in instrument 5 assumes each printed three-decimal input is the rounding of a value within ±0.0005. That is the standard reading of a printed table and it is an assumption.
- The monopsony model is iso-elastic in both supply and demand. That is a simplification chosen because it yields closed forms; a different functional form would move the endpoints of the interval, though not the existence of the hump.
Everything we could not verify, named
- Stigler 1946. We could not obtain the original. The citation (American Economic Review, June 1946, 36(3), pp. 358–65) is confirmed only from Card and Krueger's own reference list. Because we could not read it, this page attributes to Stigler only what Card and Krueger attribute to him in their opening paragraph, which is quoted verbatim above and is the sole mention of him in the body, and makes no claim about what else the paper argues. In particular we do not claim, as is often said, that Stigler also stated the monopsony reversal: that may well be in the paper, and we have not read it.
- Sokolova and Sorensen's published medians. The 1,320 estimates, the 53 studies and the direct-versus-inverted discrepancy are quoted from the published abstract. The specific medians used on this page (1.1, 3.036, 4) and the three markdowns beside them are read directly from the IZA working-paper version, which covers 801 estimates and reports different figures from the published article. The two versions are never mixed, and each number is tagged with the version it came from. We could not reach the published article's own table.
- Lichter, Peichl and Siegloch. The moments used here are read from the IZA working paper, which states 942 estimates from 105 studies. The published European Economic Review version may differ; we cite the working paper.
- Neumark and Shirley's percentages are taken from the NBER working paper (revised March 2022), which prints 79.2%. The published Industrial Relations article is a separate object and this page does not claim its numbers are identical.
- Table 4 model (v). We get —, read from the same fit the instrument above renders, where the paper prints 11.91 (7.39), on a specification whose regression standard error and F p-value both land on the printed digits. We cannot account for the gap and we do not hide it.
- Row 3 of Table 3. Its printed standard errors are not the ones the file's own covariance produces. Instrument 3c shows what they are consistent with instead. We do not know why, and we are not asserting an error.
What would falsify the central claim
The central claim is narrow and it is this: in the standard wage-setting model with iso-elastic labour supply and demand, a binding floor strictly between the unregulated wage and (1 + 1/ε) times it raises employment above the unregulated level, with the peak at (1 + 1/ε)η/(ε+η); and Card and Krueger's published difference in differences is reproducible from their own file.
- The model claim dies if the grid search in instrument 2 ever disagrees with the closed form by more than the resolution of the search itself, which that instrument prints beside every comparison, or if employment fails to rise anywhere on that interval. Both are tested live, at whatever ε and η you choose, and offline at six parameter pairs plus a sweep of the whole slider range.
- The empirical claim dies if the recomputed Table 3 stops landing on the printed values. It is checked cell by cell in front of you, on the strict rule that our two printed decimals must equal theirs, and the mismatches are flagged in amber rather than dropped. — of the thirty cells pass that rule, counted live; the rest are named in the paragraph under the grid and, six of them, taken apart in instrument 3c.
- The monopsony reading of a real labour market dies if employment falls monotonically from the very first cent of a binding floor, at floors far below (1 + 1/ε) times the going wage. That is a shape, and it is measurable.
- Either survey's headline dies to a preregistered set of elasticities with a preferred estimate named per study before its sign is known.
The verifier
Run node research/the-wage-floor-that-hires/verify-the-wage-floor-that-hires.mjs. It re-parses public.dat from a committed copy, asserts its SHA-256 is — and its second-interview status counts match the codebook exactly, then asserts: all twenty cells of Table 3 rows 1, 2, 4 and 5 against the printed values, to within 0.011 on each mean and 0.006 on each standard error; the row-3 point estimates and the fact that its printed standard errors are not the file's; the exact set of cells that agree with the AER at two decimals, which is the rule this page colours the grid with, so a change in either place breaks the other; the footnote-a group sizes 101 / 140 / 73; the difference in differences at every part-time weight from 0 to 1; the n = 357 sample and the dependent variable's mean and standard deviation; Table 4 models (i) to (v) with their regression standard errors and F p-values, recording the model (v) gap rather than suppressing it; the four monopsony closed forms against a 20,000-point grid search at six parameter pairs, and the resolution rule instrument 2 applies across a sweep of the whole slider range; the three published markdowns; the bunching estimator's arithmetic against a rounding envelope for all fourteen published columns; the seven own-wage elasticities and their standard errors, read out of this page's own source so the chart cannot drift from the paper; and every denominator in Neumark and Shirley's Table 3. It also asserts that the CSV payload embedded in this page is byte-identical to what parsing public.dat produces, so the page cannot drift from the file the script checks. It exits 0 on success.
Honest apparatus: what this page is not doing
It is not settling anything. "Economists now agree", in either direction, is false. Neumark and Shirley (2022) and Dube (2019) read overlapping literatures to opposite headlines and both are current. Instrument 6 exists so that you can produce both summaries from one table rather than be told which to believe.
It is not saying Card and Krueger were refuted, or that Neumark and Wascher were discredited. Both findings are real and published. They rest on data that count different things on partly different samples of restaurants.
Provenance, all sides or none. Neumark and Wascher's payroll data came from the Employment Policies Institute, and they disclosed it themselves on the working paper's title page: "We are grateful to Carlos Bonilla of the Employment Policies Institute (EPI), and to participating franchise owners and corporations, for providing us with the payroll data. The EPI is funded by business contributions and generally opposes minimum wage increases. However, the research described in this paper was conducted independently of the EPI, and neither author received any remuneration for conducting the research." That last sentence is part of the disclosure and quoting the first half without it would be a smear. On the other side: Ben Zipperer, a co-author of the bunching paper, gives his affiliation on its title page as the Economic Policy Institute; Dube's review was commissioned by the UK Low Pay Commission; Peter Shirley's affiliation is the West Virginia Legislature's Joint Committee on Government and Finance. The Employment Policies Institute and the Economic Policy Institute are different organisations with opposite alignments and the same initials. This page spells both out in full in its own prose and never writes the abbreviation alone outside the quotation above, where the authors' own words are kept exact.
Monopsony does not mean one employer. The wage-setting power in this model comes from the labour supply curve facing a single firm being finite, which search frictions, moving costs and idiosyncratic preferences over jobs all produce. Concentration is one route to it and not the only one.
What is not covered here, said out loud. Hours versus heads beyond the accounting identity; prices and pass-through to consumers; long-run substitution of capital for labour; who ultimately bears the incidence; non-US labour markets; and wage floors far above anything in the historical record, where every estimate on this page is out of sample.
No discovery is claimed anywhere on this page. The integer counts recovered in instrument 6 are an arithmetic inversion of printed percentages, not a finding; the paper states only the total of 130, so the remaining denominators are the smallest consistent whole numbers and any multiple would fit equally well.
Sources
- Card, D. & Krueger, A. B. (1994). Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania. American Economic Review 84(4), 772–793. Tables 3 and 4 read from the author's own PDF, p. 779; the opening sentence quoted above is from p. 772 of the same PDF.
- Replication file: njmin.zip (public.dat, 410 observations; codebook; read.me), davidcard.berkeley.edu/data_sets/njmin.zip. Embedded in this page and committed to research/the-wage-floor-that-hires/.
- Neumark, D. & Wascher, W. (1995). The Effect of New Jersey's Minimum Wage Increase on Fast-Food Employment: A Re-Evaluation Using Payroll Records. NBER Working Paper 5224. Abstract and title-page acknowledgment read directly. Published as American Economic Review 90(5), 1362–1396, doi:10.1257/aer.90.5.1362.
- Card, D. & Krueger, A. B. (2000). Reply. American Economic Review 90(5), 1397–1420, doi:10.1257/aer.90.5.1397. Working-paper version NBER WP 6386, whose abstract is quoted here.
- Ropponen, O. (2011). Reconciling the evidence of Card and Krueger (1994) and Neumark and Wascher (2000). Journal of Applied Econometrics 26(6), 1051–1057, doi:10.1002/jae.1258.
- Cengiz, D., Dube, A., Lindner, A. & Zipperer, B. (2019). The Effect of Minimum Wages on Low-Wage Jobs. Quarterly Journal of Economics 134(3), 1405–1454, doi:10.1093/qje/qjz014. Tables 1 and 2 read from NBER WP 25434, pp. 37–39.
- Neumark, D. & Shirley, P. (2022). Myth or Measurement. Industrial Relations 61(4), 384–417, doi:10.1111/irel.12306. Table 3 read from NBER WP 28388 (January 2021, revised March 2022).
- Dube, A. (2019). Impacts of Minimum Wages: Review of the International Evidence. Report for the UK Low Pay Commission. Paragraphs 4.13 and 4.17 read directly.
- Sokolova, A. & Sorensen, T. (2021). Monopsony in Labor Markets: A Meta-Analysis. ILR Review 74(1), 27–55, doi:10.1177/0019793920965562 (abstract, for the 1,320 estimates from 53 studies). Working paper IZA DP 11966, for the medians and markdowns used here.
- Lichter, A., Peichl, A. & Siegloch, S. (2015). The own-wage elasticity of labor demand: A meta-regression analysis. European Economic Review 80, 94–119, doi:10.1016/j.euroecorev.2015.08.007. Moments read from IZA DP 7958.
- Robinson, J. (1933). The Economics of Imperfect Competition. Macmillan (origin of the word "monopsony"; 1933 is the publisher's date in the library record, and Thornton's abstract calls it "her 1932 classic"); Thornton, R. J. (2004). Retrospectives: How Joan Robinson and B. L. Hallward Named Monopsony. Journal of Economic Perspectives 18(2), 257–262, doi:10.1257/0895330041371240. Bibliographic record confirmed from Crossref; the Hallward credit is from the article's own abstract, which reads "The term 'monopsony' was introduced by Joan Robinson in her 1932 classic The Economics of Imperfect Competition, although she gives credit to classics scholar B.L. Hallward of Cambridge for the actual coining of the term." The full text is paywalled and was not read.
- Stigler, G. J. (1946). The Economics of Minimum Wage Legislation. American Economic Review, June 1946, 36(3), pp. 358–65. Cited here as Card and Krueger cite it, from their reference list; we could not obtain the original and make no claim about its contents beyond theirs.
- Manning, A. (2003). Monopsony in Motion: Imperfect Competition in Labor Markets. Princeton University Press (standard treatment of the model used in instruments 1, 2 and 7).