Fast-food jobs after New Jersey's 1992 minimum-wage rise: methods

Generated from minimum-wage-did.yaml. Status: provisional. Owner: data-visual-explainer worked example (public data). Updated 2026-10-07.

Question. When New Jersey raised its minimum wage in April 1992, did fast-food restaurants there cut jobs compared with similar restaurants just across the border in Pennsylvania, where the wage did not change?

Definition. We counted staff at fast-food restaurants in early 1992 and again late in the year. New Jersey raised its minimum wage on 1 April; nearby Pennsylvania did not. We compare how the average staff per restaurant changed in each state.

Result. More full-time-equivalent workers per new jersey restaurant than the pennsylvania trend predicts: +2.75 (plausible range +0.09 to +5.42). New Jersey restaurants averaged 20.4 full-time-equivalent workers before the rise, so this is about 13% of a typical store.

Population

The unit of analysis is the restaurant (One row per store in public.dat (410 rows); CHAIN 1 = Burger King, 2 = KFC, 3 = Roy Rogers, 4 = Wendy's. SHEET is documented as the unique store id, but sheet 407 is used by two different stores; rows, not SHEET values, are the unit.). Eligibility rules, in the order applied:

Fast-food restaurants in New Jersey and eastern Pennsylvania from four chains that took part in the first telephone survey. Implementation: public.dat: 410 stores, STATE = 1 (NJ, 331) or 0 (PA, 79). (n = 410) [S1, S3]
Reported staff numbers in that survey round. Each round's average uses every restaurant with a count for that round, so the two rounds are not exactly the same restaurants. Implementation: Wave-1 mean over stores with non-missing fte1 (PA 77, NJ 321); wave-2 mean over stores with non-missing fte2 (PA 77, NJ 319). "All available observations" (Table 3 rows 1–2). [S1, S5]
Restaurants that closed for good by the second round count as having zero staff. The four that were closed only temporarily are left out of the second round. Implementation: STATUS2 = 3 (6 stores): wave-2 employment fields are already 0 in public.dat. STATUS2 in (2, 4, 5) (4 stores, all NJ): wave-2 fields missing, so excluded from the wave-2 mean. [S1, S3, S6]

Time zero and windows

Time zero is the wage rise: New Jersey's minimum wage went from $4.25 to $5.05 an hour on 1 April 1992. Implementation: 1992-04-01 (paper, introduction and section I). Pennsylvania stayed at the federal $4.25.

Window	Role	Offsets (start inclusive, end exclusive)	Definition
First survey round	observation	[-46, -27) days	Wave 1: February 15 – March 4, 1992 (Table 1).
Second survey round	outcome	[218, 275) days	Wave 2: November 5 – December 31, 1992 (Table 1); DATE2 in the file spans 110592–123192.

Measures

Staff before (value, wave1): Full-time workers, including managers, plus half the number of part-time workers. Implementation: EMPTOT = EMPPT*.5 + EMPFT + NMGRS (check.sas line 54).
Staff after (headline) (value, wave2): The same count, taken in the second survey round. Implementation: EMPTOT2 = EMPPT2*.5 + EMPFT2 + NMGRS2 (check.sas line 55).

Aggregation and comparison

Average the staff count over restaurants in each state and survey round. PROC MEANS of EMPTOT and EMPTOT2 BY NJ (check.sas lines 128–131, 'PART OF TABLE 3').

Group	Definition	Value (other)	n
Pennsylvania, Feb–Mar	STATE = 0, non-missing EMPTOT (n = 77).	23.3312	77
Pennsylvania, Nov–Dec	STATE = 0, non-missing EMPTOT2 (n = 77).	21.1656	77
New Jersey, Feb–Mar	STATE = 1, non-missing EMPTOT (n = 321).	20.4394	321
New Jersey, Nov–Dec	STATE = 1, non-missing EMPTOT2 (n = 319).	21.0274	319

Take New Jersey's change in average staff between the two rounds. Subtract Pennsylvania's change over the same months. What is left is the change Pennsylvania's path does not explain. (NJ wave 2 − NJ wave 1) − (PA wave 2 − PA wave 1), FTE per store (Table 3, row 3, column iii). Target: ATT (effect among those who actually did it (not everyone)).

Recomputed from the group values: (21.0274 − 20.4394) − (21.1656 − 23.3312) = 2.754.

Uncertainty

The paper reports a t statistic of 2.03 for this number. We turned that into a 95% range of 0.09 to 5.42 workers. The paper does not say how it computed the spread. SE = 2.76 / 2.03 ≈ 1.36 (paper, section III.A); interval = 2.754 ± 1.96 × 1.36. Treating the four state-by-round means as independent gives SE 1.80 (interval −0.78 to 6.28). The balanced-sample change (row 4) gives SE 1.34. Not covered: The calculation behind the published spread is not stated, and check.sas does not compute it. Restaurants appear in both rounds, so the four averages are not independent; the simplest formula that ignores this gives a range that includes zero. Nothing here covers the assumption that both states would have moved alike.

Estimand

Attribute	Specification
Population	Burger King, KFC, Wendy's and Roy Rogers restaurants in New Jersey and eastern Pennsylvania in the 1992 sample frame that responded in wave 1 and reported employment in the given wave.
Exposure	Being in New Jersey when its minimum wage rose from $4.25 to $5.05 on 1992-04-01.
Time zero	1992-04-01.
Variable	FTE employment per store = full-time + managers + 0.5 × part-time, in wave 2 (Nov 5 – Dec 31, 1992) minus wave 1 (Feb 15 – Mar 4, 1992).
Intercurrent events	Permanent closures count as zero employment; temporary closures (4) are missing in wave 2.
Summary	Difference in state mean changes (unbalanced means), FTE workers per store; interval from the published t statistic.

Sensitivity

Specification	Estimate	Interval	Note
Same restaurants in both rounds (row 4)	2.75	[0.12, 5.38]	384 restaurants with a count in both rounds; SE 1.34.
All restaurants with a count (row 3, headline) (headline)	2.75	[0.09, 5.42]
Temporarily closed counted as zero (row 5)	2.51	[-0.13, 5.15]	Paper: 'only slightly smaller'.
Regression, Table 4 sample	2.33	[0.0, 4.66]	357 restaurants with wage data in both rounds; coefficient 2.33 (SE 1.19).

Limitations

Pennsylvania fell more than New Jersey rose (limits what the number can be used to conclude; high). New Jersey's average went up by 0.59 workers while Pennsylvania's fell by 2.17. Most of the gap is Pennsylvania shrinking. The paper attributes this to a worsening regional economy, which the design assumes hit both states equally.
The spread depends on an unstated calculation (the true uncertainty is larger than shown; medium). The published t statistic implies a 95% range of about 0.1 to 5.4 workers. A simple calculation that treats the two rounds as separate samples gives −0.8 to 6.3, which includes no effect. Counting only restaurants seen in both rounds gives 0.1 to 5.4 again.
Self-reported counts from a phone survey (the true uncertainty is larger than shown; medium). Managers reported staff numbers by phone. The paper estimates the reliability of this measure at 0.70 by reinterviewing a subset of stores, so some changes are noise.
One region, one industry, one rise (limits what the number can be used to conclude; medium). This is four fast-food chains in New Jersey and eastern Pennsylvania in 1992, after an 80-cent rise. It says nothing directly about other industries, larger rises or other places.

Description vs implementation

Store identifier. Described: Codebook: SHEET is the 'sheet number (unique store id)'. Implemented: Sheet 407 appears twice in public.dat: a Pennsylvania KFC and a New Jersey Wendy's. Impact: Anyone deduplicating or joining on SHEET loses or mixes one store. The paper's averages, which treat rows as stores, are unaffected.
Where closed stores become zero. Described: Table 3 note: 'Employment at six closed stores is set to zero.' Implemented: check.sas never sets anything to zero; the six STATUS2 = 3 rows already carry 0 in the wave-2 employment fields of public.dat. Impact: None for the published numbers. A re-analysis that recodes STATUS2 itself, or treats 0 as missing, will silently change the headline.
Headline value. Described: Text: 'The relative gain … is 2.76 FTE employees (or 13 percent), with a t statistic of 2.03.' Implemented: The public data give 2.754, which rounds to 2.75. Rows 4 and 5 and Table 4 reproduce to the published decimals. Impact: Rounding only. It is likely the text was computed from rounded table entries.
How the headline's uncertainty was computed. Described: A t statistic of 2.03; no method given. Implemented: check.sas computes only per-state means (PROC MEANS), not the standard error of the difference. The independent-samples formula gives t = 1.53; the balanced-sample formula gives t = 2.05. Impact: Unknown. It decides whether the headline is distinguishable from zero.
Restaurants without a starting wage. Described: Table 3 notes give New Jersey wage-group counts of 101, 140 and 73. Implemented: Those sum to 314 of 331 New Jersey stores. The 17 with no wave-1 starting wage go to check.sas's 'NJ STORE, BAD WAGE' group and are not mentioned in the text. Impact: Affects only the within-New Jersey columns (iv)–(viii), not the headline.

Reproduction

python examples/fetch_data.py minimum-wage (downloads and checks the data and paper).
python examples/minimum-wage/analysis.py (writes results/table3.csv and results/checks.csv).
Receipt: 2026-10-07: PA 23.33 → 21.17, NJ 20.44 → 21.03; DiD 2.754. Row 4 2.750 (SE 1.342), row 5 2.509.
Receipt: 2026-10-07: Table 4 column (i) 2.326 (SE 1.192), n = 357, mean change −0.238, SD 8.825; the paper prints 2.33 (1.19), −0.237, 8.825.

Completeness. The latest wage rise is 01 Apr 1992; its second survey round ends 31 Dec 1992. Data are complete through 31 Dec 1992, so every window is fully observed. This number should not move as more data arrive.

Sources

[S1] examples/minimum-wage/data/njmin-aer.pdf. Card & Krueger (1994), AER 84(4): intro, section I, Table 1, section III.A, Tables 3–4 notes
[S2] examples/minimum-wage/data/check.sas, lines 54-59, 128-139, 150-163
[S3] examples/minimum-wage/data/codebook. Column definitions; STATUS2 codes
[S4] examples/minimum-wage/data/read.me. Says the files support Myth and Measurement, chapter 2
[S5] examples/minimum-wage/analysis.py, lines 30-103. Outputs in examples/minimum-wage/results/
[S6] examples/minimum-wage/data/public.dat. 410 rows; downloaded from davidcard.berkeley.edu/data_sets.html