Minimum wage and the 'no disemployment' result

Minimum wages and job loss: what the "no disemployment" fight is actually about

The short version

The modern "no job loss" finding comes mainly from Cengiz, Dube, Lindner and Zipperer (CDLZ, 2019). They used public Current Population Survey wage microdata and found that state minimum-wage increases did not change the total number of low-wage jobs. The critics, centred on David Neumark and co-authors, do not say the data are wrong. They say the answer depends on which comparison you build:

• which places stand in for "what would have happened without the increase";
• which years of data are allowed to inform that comparison; and
• which estimator combines them.

The clearest documented case where one choice flips the sign is not CDLZ itself. It is the border-county branch of the same research programme:

• Dube, Lester and Reich (2010) found no effect.
• Jha, Neumark and Rodriguez-Lopez (2024) found negative effects after swapping county pairs for multi-state commuting zones.
• Dube, Reich, Bhatt and Sosinskiy (2024) replied that the negative result comes entirely from pre-2000 data, which a two-way-fixed-effects estimator mixes in.

Each side used the other's replication files, so the switch can be checked in principle. Details follow.

I did not find a peer-reviewed or working-paper reanalysis that takes CDLZ's own bunching estimate and reverses its sign by changing one choice. The one source I found claiming something like this, a blog post, was outside this environment's network allowlist, so I could not read it and do not rely on it (see Notes).

1. What the "no disemployment" studies claim, and on what data

CDLZ (2019), "The Effect of Minimum Wages on Low-Wage Jobs: Evidence from the United States Using a Bunching Estimator." Doruk Cengiz, Arindrajit Dube, Attila Lindner, Ben Zipperer. NBER Working Paper 25434, January 2019 (published in the Quarterly Journal of Economics, 2019). https://www.nber.org/system/files/working_papers/w25434/w25434.pdf

• Design. They count "missing jobs" just below a new minimum wage and "excess jobs" at or just above it, across 138 state increases. Abstract (PDF p. 2): "we estimate the effect of the minimum wage on the frequency distribution of hourly wages using 138 prominent state-level minimum wage changes between 1979 and 2016."
• Headline claim. Abstract: "We find that the overall number of low-wage jobs remained essentially unchanged over five years following the increase."
• Data. The NBER Merged Outgoing Rotation Group files of the CPS, 1979–2016, benchmarked to the Quarterly Census of Employment and Wages (QCEW) (Section 3, PDF p. 13). State minimum-wage histories come from a public GitHub series (fn. 15, PDF p. 14: https://github.com/benzipperer/historicalminwage/).
• Numbers. Employment elasticity with respect to the minimum wage of 0.024 (s.e. 0.025). Own-wage elasticity of 0.411 (s.e. 0.430) (PDF pp. 16–17).
• Stacked design. A robustness exercise builds a separate dataset for each event, containing "the treated state and all clean control states" (PDF p. 24 and appendix p. 80). This is the "stacked, clean-controls" event study that later work builds on.

The same camp's border-county work, as described by its critics. Jha, Neumark and Rodriguez-Lopez (2024, abstract, PDF p. 2) summarise Dube, Lester and Reich (2010): "Using minimum wage variation within contiguous county pairs sharing a state border, they find no relationship between minimum wages and employment in the U.S. restaurant industry." I did not open DLR (2010) itself.

The literature summary from the same camp. Dube and Zipperer, "Own-Wage Elasticity: Quantifying the Impact of Minimum Wages on Employment," NBER WP 32925, September 2024. https://www.nber.org/system/files/working_papers/w32925/w32925.pdf. Abstract (PDF p. 2): "the median OWE estimate of 72 studies published in academic journals is -0.13, which suggests that only around 13 percent of the potential earnings gains from minimum wage increases are offset due to associated job losses." They add that estimates published since 2010 "tend to be closer to zero."

2. The specific choice the critique says drives the result

The critique is not about one bug. It is about three related construction choices. For each, here is what each side says.

(a) Which places count as the control group: border counties or commuting zones

This is the sharpest, best-documented flip.

Jha, Neumark and Rodriguez-Lopez (JNR), "What's Across the Border? Re-Evaluating the Cross-Border Evidence on Minimum Wage Effects," NBER WP 32901, September 2024. https://www.nber.org/system/files/working_papers/w32901/w32901.pdf

• Position (abstract, PDF p. 2): "Using the same research design, we show that this result is overturned if we use instead multi-state commuting zones, which provide superior definitions of local economic areas."
• Mechanism.
  • With all cross-border county pairs inside multi-state commuting zones, the minimum-wage elasticity of employment is −0.244 (main estimate −0.242).
  • Using "only pairs of contiguous counties that belong to different commuting zones," the estimate is a small, non-significant −0.047 (Introduction, PDF p. 6).
  • Their explanation (PDF p. 6): counties "across the border, but not in the same commuting zone, do not provide good controls for common shocks potentially correlated with minimum wage changes."
  • Put plainly, their claim is that DLR's null result comes from pairing counties that touch but do not share a labour market, which biases the estimate toward zero.

(b) Which years and which estimator: two-way fixed effects over long panels, or event studies with clean windows

Dube, Reich, Bhatt and Sosinskiy (DRBS), "Restaurant Employment, Minimum Wages, and Border Discontinuities," NBER WP 32902, September 2024. https://www.nber.org/system/files/working_papers/w32902/revisions/w32902.rev0.pdf

• Position (abstract, PDF p. 2): "We show that JNR's results are confounded by parallel trends violations in the 1990s, when minimum wage events were rare and small in magnitude; JNR's outmoded two-way-fixed-effects model amplifies the biases introduced by these violations."
• Mechanism (PDF p. 9): "The larger negative estimate comes entirely from using a 1990 baseline instead of a 2013 baseline for evaluating minimum wage changes occurring after 2013."
  • A TWFE regression implicitly averages every pairwise before/after comparison in the panel, so a post-2013 increase is partly judged against 1990s employment.
  • With post-2000 data only, or with their event-study design that "analyzes only data that fall clearly within an event's window," they find no negative effect "whether we compare across all states, look within commuting zones or within border county pairs" (abstract).

So the dispute narrows to a concrete question: should the 1990s count as valid pre-period information? JNR's specification keeps those years and gets job losses. DRBS's specifications drop them, or wall them off by event window, and do not.

CDLZ make the same argument in reverse against an earlier negative finding. On Meer and West (2016), whose two-way-fixed-effects regressions on aggregate state employment find losses, CDLZ write (Introduction, PDF p. 6): "We show that the large negative effects on employment is driven entirely by inclusion of the 1980s in the sample—a period with very few minimum wage changes." They also claim the TWFE losses show up as "an unrealistically large drop in the number of jobs at the upper tail of the wage distribution" (PDF p. 6), meaning among workers too well paid to be directly affected. I did not open Meer and West (2016). CDLZ report Meer and West's baseline aggregate elasticity as −0.074 (PDF p. 16). That figure is CDLZ's citation of Meer and West, not my reading of Meer and West.

(c) Close or regional controls and state trends, against national comparisons (the older round)

Neumark, Salas and Wascher (NSW), "Revisiting the Minimum Wage-Employment Debate: Throwing Out the Baby with the Bathwater?" NBER WP 18681, January 2013 (published in ILR Review, 2014). https://www.nber.org/system/files/working_papers/w18681/w18681.pdf

• Target. The "close controls" designs of Allegretto, Dube and Reich (2011) and Dube, Lester and Reich (2010).
• Position (abstract, PDF p. 2): "new evidence based on methods that let the data identify the appropriate control groups leads to stronger evidence of disemployment effects, with teen employment elasticities near -0.3."
• Mechanism in Neumark's words. Neumark (2019, below) states it (Section III, PDF p. 8): the finding that effects are "less negative (and insignificant) when using close controls implies that minimum wages tend to be increased where there are negative shocks to low-skill labor markets." He then questions whether that is likely. The disagreement is whether nearby states and regional time effects are better counterfactuals, or whether they throw away the identifying variation.
• Responses I could not open.
  • Allegretto, Dube, Reich and Zipperer, "Credible Research Designs for Minimum Wage Studies: A Response to Neumark, Salas, and Wascher" (ILR Review 70(3), 2017).
  • Neumark and Wascher's reply in the same issue.
  • Both host links were outside the network allowlist (Notes). I know these papers exist from search listings only, so I quote neither.

(d) Is the bunching estimator itself immune to spurious effects?

David Neumark, "The Econometrics and Economics of the Employment Effects of Minimum Wages: Getting from Known Unknowns to Known Knowns," NBER WP 25043 (2018). Published in German Economic Review 20(3), 2019. https://www.nber.org/system/files/working_papers/w25043/revisions/w25043.rev0.pdf

• Fn. 11 (PDF p. 7), on the CDLZ bunching approach: "One curious finding, though – a very large positive implied employment elasticity for high school dropouts – suggests that the method may not be immune from spurious effects."
• How far it goes. This is a flag, not a reanalysis: "I leave it to future work to consider this approach more fully." Nothing I read follows it up with a specification that turns CDLZ's estimate negative.

Where each camp says the literature as a whole lands

• Neumark and Peter Shirley, "Myth or Measurement: What Does the New Minimum Wage Research Say about Minimum Wages and Job Loss in the United States?" NBER WP 28388, January 2021, revised March 2022 (published in Industrial Relations 61(4), 2022). https://www.nber.org/system/files/working_papers/w28388/w28388.pdf. Abstract (PDF p. 2): "there is a clear preponderance of negative estimates in the literature." They also say the evidence is stronger for teens, young adults, the less-educated and directly affected workers, and "less one-sided" for low-wage industries.
• Dube and Zipperer (2024) reply on method (PDF p. 3): such aggregations, naming "Neumark and Shirley 2022," "do not provide a very clear guide to the magnitude of effects." Their own reading is the median OWE of −0.13 quoted above.

The two summaries are not strictly in factual conflict. One counts signs of estimates; the other takes the median of a standardised elasticity. They answer different questions about the same body of studies, and each side says the other's metric is the less informative one.

3. Who is on each side, and their stake

Affiliations below are as printed on the papers I opened. "Funding" is listed only where a paper discloses it.

Name collision: the Employment Policies Institute, which funded part of NSW, and the Economic Policy Institute, Zipperer's employer, are different organisations. The papers abbreviate both as "EPI." I read no further disclosures from either side and make no claim beyond what is printed above.

4. How long this has gone back and forth

• Early 1990s onward. Neumark and Shirley (2022, abstract) date the "new minimum wage research," based on subnational variation, to "the early 1990s." That literature includes Card (1992), Neumark and Wascher (1992), and Card and Krueger (1994), all cited there.
• 2010–2017, close controls. Dube, Lester and Reich (2010) and Allegretto, Dube and Reich (2011) were followed by Neumark, Salas and Wascher (NBER WP January 2013; ILR Review 2014). Then came the 2017 ILR Review exchange (response and reply), which I could not open.
• 2016–2019, aggregate and bunching. Meer and West (2016) found negative aggregate effects. CDLZ (2019) attributed those to the inclusion of 1980s data. Neumark (2018/2019) flagged the bunching estimator's dropout result.
• 2021–2022, literature counts. Neumark and Shirley (2021 WP, 2022 published).
• September 2024, border-county round. JNR (NBER WP 32901) and DRBS (NBER WP 32902) appeared as consecutive working papers, alongside Dube and Zipperer's OWE repository (WP 32925).

The specific question of how to build comparison groups for US state minimum-wage variation has therefore been argued, paper against paper, for at least 14 years (2010–2024). The underlying question is older. NSW's abstract calls it "as old as the Department of Labor."

5. Could an outsider redo the comparison and see which choice flips the sign?

Yes in principle, and the border-county round is the place to do it. It is not a weekend job.

• The data are public.
  • CDLZ use the NBER CPS Merged Outgoing Rotation Group files and QCEW benchmarks (PDF p. 13), with a public minimum-wage series on GitHub (PDF p. 14).
  • JNR and DRBS use the QCEW and County Business Patterns. DRBS also use the ACS (DRBS abstract).
• The two sides have already passed code back and forth.
  • JNR say: "DLR's replication package is available at this link" (PDF p. 6), and they start from DLR's exact specification.
  • DRBS build their samples starting "with JNR's replication files (i.e. their pair definitions)" (Appendix, PDF p. 38).
  • So there is a chain of replication packages: DLR, then JNR, then DRBS.
• What the experiment would be. Take JNR's commuting-zone specification and toggle two things:
  1. sample start year: 1990s included vs. post-2000 only;
  2. estimator: TWFE vs. event-window ("stacked") difference-in-differences.

DRBS state that the negative sign depends on the first toggle (PDF p. 9). JNR state that the sign depends on commuting zones vs. cross-CZ county pairs (PDF p. 6). These are separate claims about separate switches, and both can be run on the same files.

• What checking would not settle. Running the code shows which switch moves the estimate. It does not settle which choice is correct. That depends on whether 1990s pre-trends are contamination (DRBS) or legitimate identifying variation (JNR's position, implied by their specification). The facts are checkable; the judgment is not mechanical.
• CDLZ's bunching estimate specifically is checkable in the same sense, since its inputs are public CPS data. However, I did not find a replication package link in the working paper text I read, apart from the minimum-wage series. I also found no third-party reanalysis that flips it. Neumark's only stated objection to the method (fn. 11) points to a suspicious positive result for dropouts, not a sign reversal of the headline.