Seeing beyond the Trees:

Using Machine Learning to Estimate

the Impact of Minimum Wages
on Labor Market Outcomes
Doruk Cengiz, OM Partners

Arindrajit Dube, University of Massachusetts Amherst, National

Bureau of Economic Research, and Institute of Labor Economics

Attila Lindner, University College London, Institute for Fiscal

Studies, Institute of Labor Economics, and Centre for Economic
and Regional Studies

David Zentler-Munro, University of Essex

We assess the effect of the minimum wage on labor market outcomes.

First, we apply modern machine learning tools to predict who is
affected by the policy. Second, we implement an event study using
172 prominent minimum wage increases between 1979 and 2019. We
find a clear increase in wages of affected workers and no change in
employment. Furthermore, minimum wage increases have no effect
on the unemployment rate, labor force participation, or labor market
transitions. Overall, these findings provide little evidence of changing
search effort in response to a minimum wage increase.

We thank Gábor Békés, Erin Conlon, Ezgi Cengiz, Ina Ganguli, Laura Giuliano,
Carl Nadler, Hasan Tekgüç, Michael Reich, Jesse Rothstein, and participants at the

Submitted December 12, 2020; Accepted December 9, 2021.

Journal of Labor Economics, volume 40, number S1, April 2022.
© 2022 The University of Chicago. All rights reserved. Published by The University of Chicago Press in
association with The Society of Labor Economists and The National Opinion Research Center. https://
[Link]/10.1086/718497

S203
S204 Cengiz et al.

I. Introduction
A long-standing question in economics centers around understanding how
minimum wages affect low-wage labor markets. A key challenge to convinc-
ingly answer this question comes from the difficulty in successfully identify-
ing most workers who are actually affected by the policy. While we can easily
locate workers who are currently earning the minimum wage, it is difficult to
identify all potential workers who also may have been working had the min-
imum wage been different. This difficulty has led many researchers to focus
on specific industries or demographic groups, such as teens (Card 1992; Neu-
mark and Wascher 1992; Giuliano 2013; Neumark, Salas, and Wascher 2014;
Allegretto et al. 2017; Totty 2017), younger workers with lower educational
credentials (Sabia, Burkhauser, and Hansen 2012; Manning 2016; Clemens
and Strain 2017; Clemens and Wither 2019), and individuals without a high
school degree (Addison and Blackburn 1999; Addison, Blackburn, and Cotti
2011). However, these groups constitute relatively small shares of all mini-
mum wage workers. As a result, there is a tension between what is often an-
alyzed (e.g., minimum wage effects on teens) and what is argued (effects of
the policy on affected workers largely composed of adults; Belman, Wolfson,
and Nawakitphaitoon 2015; Manning 2016).1
In this paper, we use machine learning (ML) tools to predict which individ-
uals were likely affected by minimum wage increases and then estimate the

Labor and Employment Relations Association 70th Annual Meeting, Institute for
Research on Labor and Employment Research Presentation seminar, 44th Eastern
Economic Association Conference, 2018 New School–University of Massachusetts
Economics Graduate Student Workshop, and the authors’ conference in honor of
Alan Krueger for very helpful comments. We are also grateful to Jon Piqueras for
outstanding research assistance. The previous version of the paper was circulated un-
der the title “Seeing Beyond the Trees: Using Machine Learning to Estimate the Im-
pact of Minimum Wages on Affected Individuals.” Attila Lindner acknowledges fi-
nancial support from the Economic and Social Research Council (new investigator
grant, ES/T008474/1) and from the European Research Council (ERC) under the
European Union’s Horizon 2020 Research and Innovation Programme (grant agree-
ment 949995). David Zentler-Munro acknowledges financial support from the Eco-
nomic and Social Research Council (ESRC) under grant ES/R005745/1. Contact the
corresponding author, Arindrajit Dube, at adube@[Link]. Information concern-
ing access to the data used in this paper is available as supplemental material online.
1
The discrepancy is particularly relevant when the measured outcome is the teen
employment rate, which has been the subject of extensive research in the United
States. Belman and Wolfson (2014) consider 30 studies that examined the employment
effects of the minimum wage on some demographic groups between 2001 and early
2013, and they find that 17 of them had teen employment as the dependent variable.
Neumark (2017) shows that 12 of 13 studies that examined minimum wage effects on
“lower-skilled” employment between 2010 and 2016 focused on teens (see his table 1).
However, teens are less likely to be in the affected group than nonteen adult minimum
wage workers (Lundstrom 2016). Compared with affected nonteens, only a relatively
small share of teens live in poverty. According to the 2016 American Community Sur-
vey, 18.4% of teens were in households with incomes under the poverty level.
Seeing beyond the Trees S205

impact of minimum wages on the individuals predicted to be exposed to the

minimum wage increase. Our prediction-based approach extends the classifi-
cation of low-wage workers that was developed by Card and Krueger (1995)
in Myth and Measurement (see p. 135) and has been undeservedly neglected
in the literature ever since.2
We construct various groups according to the predicted probabilities to
assess the impact of the policy on workers who are highly likely to be ex-
posed to the minimum wage (whom we refer to as the “high-probability”
group) and also a wider group where we can retrieve 75% of likely minimum
wage workers (whom we refer to as the “high-recall” group). We then study
the impact of the policy on various labor market outcomes, such as employ-
ment, unemployment, and labor force participation (LFP) of workers with
different exposure to the minimum wage. The impact of the minimum wage
on these latter outcomes has been extensively studied in the theoretical liter-
ature (e.g., Flinn 2011), despite the scant empirical evidence on them.
Our approach of using a prediction model to classify workers who are likely
to be exposed to a minimum wage treatment has several advantages. First, we
can assess the effect of the minimum wage on a large fraction of low-wage
workers and not just on some specific subgroups with high exposure, such as
teens or youth (see, e.g., Laws 2018). In that sense the spirit of our approach is
close to Cengiz et al. (2019), who assess the overall employment impact of the
policy using the frequency distribution of wages. Additionally, we can directly
study the impact of the policy on the affected nonteen (20–64) and prime
age (25–55) individuals who are more likely to live in low-income households
than teens and tend to be the intended beneficiaries of the policy.
Second, we can also study the impact of the policy on individuals with low
predicted probability of being minimum wage workers—that is, those who
should not be affected by the policy. Specifications that show an unrealisti-
cally large impact on workers not participating in the low-wage labor market
should be cautiously interpreted, as they raise questions about the credibility
of the particular research design. Using the impact on the “low-probability”
group as a falsification test is analogous to studying employment changes
in the upper tail of the wage distribution—a fruitful approach that success-
fully resolved some of the discrepancies in the minimum wage literature (see
Cengiz et al. 2019).
Finally, our prediction-based approach allows us to study the impact of the
policy on various labor market outcomes that would not be feasible with the

2
We are aware of only one previous publication that applied this method (Cengiz
et al. 2019), and the list of coauthors includes three of the authors of this paper. That
paper utilized the Card and Krueger prediction-based approach primarily to show
the differences between that method and the bunching method developed in Cengiz
et al. (2019).
S206 Cengiz et al.

distribution-based approach developed in Cengiz et al. (2019). This is a major

advantage, as it allows us to provide the first comprehensive picture on how
low-wage labor markets evolve in response to the minimum wage. The im-
pact of the policy on unemployment and participation rates, as well as flows
in and out of unemployment, are often mentioned in the policy discourse, but
evidence on these topics is limited and mainly available for narrow sub-
groups. Furthermore, as we will demonstrate later, understanding the impact
of the policy on outcomes besides employment is relevant, since it has welfare
implications in various noncompetitive models (e.g., Flinn 2011).
We implement our approach by using ML methods along with demo-
graphic information to predict which individuals are likely to be minimum
wage workers in the Current Population Survey (CPS) data between 1979
and 2019. In particular, we use individuals’ demographic characteristics to
predict their probability of having an hourly wage less than 125% of the stat-
utory minimum wage.3 We consider three tree-based learning tools in the
training data: decision trees, random forests, and gradient-boosting trees,
as well as the elastic net regularization of a logistic regression. A key advan-
tage of the ML tools over the Card and Krueger (1995) approach is that they
do not require the researcher to prespecify the functional form of the predic-
tion model, which is instead determined in a data-driven way. Then we com-
pare the performance of various prediction models on the test data.
The best-performing prediction comes from the gradient-boosting tree
model. The original linear prediction model proposed by Card and Krueger
(1995; with a judiciously chosen set of interactions) also performs relatively
well, although not quite as well as the state-of-the-art ML tools. When com-
pared with commonly used demographic groups in the literature (such as
teens or those under the age of 30 with high school or lower educational cre-
dentials), the boosting approach can form groups with a similar number of
(correctly classified) minimum wage workers while substantially reducing
the number of (misclassified) non-minimum-wage workers. The gains in
precision (i.e., the share of predicted minimum wage workers who are classi-
fied correctly) for a given level of recall (i.e., the share of true minimum wage

3
In the prediction exercise, we restrict the sample to states instituting prominent
minimum wage hikes and periods preceding those policy changes. The full set of
predictors and how they are coded are reported in app. B (apps. A–E are available
online). In the prediction model, we do not use variables related to past employ-
ment status or occupation/industry in the CPS Outgoing Rotation Group (CPS-
ORG). They are sometimes missing even if the individual is currently in the labor
force and looking for a job. We prefer to keep the observations with missing infor-
mation on these variables in the sample, as they potentially carry information about
labor supply effects of the policy. In our preferred prediction model, we do not use
state of residence or year information either. This choice is primarily to be able to
build samples that are comparable and consistent across time and space.
Seeing beyond the Trees S207

workers who are classified correctly) are sizeable when we limit attention to
nonteen workers—a group that is of particular interest to policy makers.
Armed with the prediction model, we implement an event study analysis
that exploits 172 prominent state-level minimum wage increases between
1979 and 2019. We assess the impact of the policy on various groups formed
on the basis of the predicted exposure probability. The high-probability
group comprises the 10% of the population with the highest likelihood of
being affected by the policy. We also study the impact of the policy on the
high-recall group, which captures 75% of all minimum wage workers.
For both groups, we find a considerable increase in wages after the policy
change; as expected, the wage increase is somewhat lower for the high-recall
group. At the same time, we detect a small, positive, statistically insignificant
effect on employment for both groups. The implied employment elasticity
with respect to own wage—the labor demand elasticity in the standard com-
petitive model of the labor market —is 0.29 (SE, 0.32) for the high-probability
group and 0.14 (SE, 0.25) for the high-recall group. The confidence bounds on
both of these estimates can rule out anything more than modest negative
disemployment effects at the conventional significance levels.
We find no evidence of substantial changes in the unemployment or par-
ticipation rates in response to the policy. We are also not able to detect any
economically meaningful (or statistically significant) changes in labor mar-
ket transitions between employment, unemployment, and nonparticipation.
This lack of response on the LFP margin provides new evidence that mini-
mum wages have a limited impact on search effort when we focus on indi-
viduals who are most likely to be affected by the policy.
Our results are robust to controlling for time-varying heterogeneity in a
wide variety of ways. Moreover, the increase in wages lines up well with the
timing of the minimum wage increases, and the effects emerge only in the group
of individuals likely to be exposed to the minimum wage increase. We find
no significant differences in labor market outcomes for the low-probability
group—suggesting that no unusual changes took place in the states’ labor
markets around the minimum wage increases we study here. All of these find-
ings reinforce the credibility of our research design.
Furthermore, we also study whether the responses to the policy vary
across demographic groups. Most importantly, we study whether differen-
tial responses can be detected on employment, unemployment, and partici-
pation margins for workers who are thought to have larger extensive margin
labor supply elasticities—such as teens, older workers, and single mothers.
In addition, we also assess the impact of the policy by the likelihood of mov-
ing into or out of the labor force. We use demographic information and ap-
ply ML tools again to classify workers as being more likely to move into and
out of the labor force. Even when we focus on the group of workers with the
highest predicted transition probabilities, we find no evidence of substantial
change in the unemployment or participation rates.
S208 Cengiz et al.

This paper contributes to several strands of the minimum wage literature.

Although this is a thick literature, there is only a handful of studies that ex-
amine broad segments of workers in the United States affected by the mini-
mum wage policy.4 Linneman (1982), Currie and Fallick (1996), and Clemens
and Wither (2019) study the short-term impact of minimum wages by exam-
ining the probability of remaining employed for workers earning below the
new minimum wage before the policy change. Assigning workers to groups
according to their baseline wages not only requires a richer panel data set than
the one used here but also misses the large fraction of minimum wage workers
who are entering the labor market in each year.5 This is especially problematic
when we are interested in understanding changes at the participation margin
and the impact on job flows. Moreover, it is difficult to study longer-term ef-
fects using this design, as the age composition of the cohort is changing along
with time elapsed since the policy change, and the share of the cohort earning
close to the minimum wage attenuates over time.
Only a few studies focus on the overall impact of the policy on low-wage
jobs. Cengiz et al. (2019) use the frequency distribution of wages to focus on
the number of low-wage jobs, while Meer and West (2015) simply consider
total state-level employment to study the overall impact of the policy. Our
estimates for the high-recall group complement the existing evidence by pro-
viding an alternative way of assessing the impact of the minimum wage on
overall employment.
Our paper also fills an important gap in the literature by going beyond
studying the wage and employment effects of the policy. While there has been
a long-standing interest in understanding the impact of minimum wages on
the unemployment and participation rates (see, e.g., Mincer 1976; Ragan
1977), there are only a handful of papers examining the impact on these out-
comes while applying credible difference-in-differences style estimators. The
few exceptions focus mainly on some specific subgroups, such as teens (see,
e.g., Wessels 2005; Laws 2018; Marimpi and Koning 2018), parents (see Go-
doy, Reich, and Allegretto 2020), and workers close to retirement (see, e.g.,
Borgschulte and Cho 2019). Adams, Meer, and Sloan (2018) study the impact
on aggregate job search. Similarly to this paper, they find no indication of sig-
nificant changes in job search activity. However, it is unclear whether they
have enough statistical power to detect significant changes by studying the
impact on all US workers, including those with high wages. Since a relatively

4
See Belman and Wolfson (2014) for a thorough literature review on the subject.
5
Assigning workers according to their baseline wages requires panel data. Panel data
sets, such as the National Longitudinal Survey of Youth (Currie and Fallick 1996) and
Survey of Income and Program Participation (Clemens and Wither 2019), are often
smaller than the CPS applied here and cover fewer years. The prediction probability ap-
proach can be applied to cross-sectional data, so we can include many more prominent
minimum wage changes in our analysis.
Seeing beyond the Trees S209

small fraction of the US workforce is typically affected by minimum wage

policies, positive wage effects cannot be detected without focusing on work-
ers earning close to the minimum wage (Cengiz et al. 2019). As a result, it may
not be surprising that Adams, Meer, and Sloan (2018) were not able to detect
significant changes in aggregate job search.
Methodologically, our use of a demographics-based predictive model for
minimum wage workers is inspired by Card and Krueger (1995), who exam-
ine the 1988 California minimum wage increase and use a linear prediction
model to sort individuals living in the state in 1987 according to their likeli-
hood of having hourly wages between the old minimum wage and the new
minimum wage ($3.35 and $4.25). The Card and Krueger (1995) model is
based on subjective judgments about predictors and the functional form,
which includes complicated multiway interactions. Even though this subjec-
tive assessment turns out to have been implemented incredibly well in Card
and Krueger (1995), the key advantage of the ML-based approach proposed
here is that we do not need to rely on such judgments. Instead, the ML tools
determine the prediction model in a data-driven way and can provide a guar-
antee against overfitting and specification hunting.
The remainder of the paper is organized as follows. In the next section we
briefly discuss the benefits of examining outcomes other than employment.
Section III describes the data sets used in our analysis. Section IV explains
how we apply various learning tools to predict exposure to the minimum
wage. Section V examines the empirical implementation and the key results.
Section VI concludes.

II. Participation, Unemployment, and the Welfare Impacts

of the Minimum Wage
The approach developed in this paper allows us to analyze the impact of
the minimum wage on a wide range of outcomes that go beyond the tradi-
tional employment and wage impacts emphasized in the literature. Studying
the impact on the unemployment and participation rates is particularly inter-
esting, since a large class of theoretical models have direct predictions on
these outcomes. For instance, efficiency wage models (Drazen 1986), models
with information asymmetry (Lang 1987), and search models (Swinnerton
1996) suggest that minimum wages will raise unemployment rates.
The participation margin is also of interest given the growing number of
empirical studies that find close to zero effect on employment. Flinn (2006)
and Ahn, Arcidiacono, and Wessels (2011) discuss in detail what mechanisms
are needed to find positive employment effects in a search model. They point
to endogenous participation rate or endogenous search effort as being essen-
tial. Flinn (2006) also points out that to simultaneously find an increase in the
employment rate and a decrease in the unemployment rate, one needs to in-
troduce two elements: an increase in both the participation rate and the search
S210 Cengiz et al.

effort. Therefore, by providing direct evidence on employment, unemploy-

ment, and participation rates simultaneously, we can assess the empirical va-
lidity of this search-based explanation.
In addition, the policy impact on labor market participation is potentially
informative about the welfare effects on some groups of minimum wage
workers. While the employment and unemployment responses often reflect
both supply and demand sides of the market, whether someone is participat-
ing in the labor market depends solely on the worker’s decision. As a result,
if some workers choose to participate in the labor market as a result of the
policy, they directly reveal that the payoff from searching and finding a job
makes them better off in expectation. Assuming that minimum wages do
not have an impact on the nonparticipation payoffs, the individuals who de-
cide to participate expect to be made better off by the minimum wage. Con-
versely, workers who leave the labor force in response to the minimum wage
expect to be made worse off.
Our emphasis on studying participation differs somewhat from the litera-
ture that interprets the presence of wage spillovers as an indicator of welfare
impacts (e.g., Flinn 2002). A key requirement for wage spillovers to be suffi-
cient for assessing welfare changes is the lack of nonwage amenities. If min-
imum wage policies alter job attributes, the positive (negative) wage spillovers
are neither a necessary condition nor a sufficient condition for a welfare gain
(loss). This is especially problematic, since some evidence suggests that min-
imum wages may affect job amenities. For instance, Dustmann et al. (2022)
find that the introduction of the German minimum wage led to an increase
in commuting times. Clemens, Kahn, and Meer (2018) find that minimum
wages alter the provision of the employer-provided health insurance.
The advantage of considering participation decisions is that they are driven
by workers’ overall assessment of the quality of jobs available (and the prob-
ability of getting those jobs). As a result, examining the impact on participation
can improve on a piecemeal approach of looking at wage spillovers or specific
measures of amenities. Of course, the change in participation is a discrete deci-
sion, so the welfare impact of the minimum wage is revealed only for workers
who are on the margin of the participation decision. As a result, we also study
the impact of the minimum wage on unemployment rates and transition prob-
abilities, which are affected by changes in search intensity. At the same time, we
note that these outcomes also reflect changes in labor demand (and are not just
workers’ decisions), so their relationship to welfare is more complicated.

III. Data
The primary data we use throughout the analysis come from the CPS.
We use the 1979–2019 CPS-ORG sample for the hourly wage and weekly
earnings variables. This is a subset of the Basic Monthly CPS (CPS-Basic), a
monthly survey of approximately 60,000 households in the United States. The
Seeing beyond the Trees S211

CPS-ORG includes only the fourth and eighth sample months, when usual
hourly wages, weekly earnings, and weekly hours worked are asked. These
variables are of primary importance for the prediction as well as for the esti-
mation, and thus we rely on their accuracy. For this reason, we exclude obser-
vations with imputed hourly wages, imputed weekly earnings, or imputed
hours worked. For hourly workers we use the reported hourly wage, and
for other workers we define the hourly wage to be their usual weekly earnings
divided by usual weekly hours. We also use a range of demographic variables
in the data set when predicting an individual’s likelihood of having a wage
close to the minimum. These variables indicate an individual’s age, race, His-
panic status, gender, education, veteran status, marital status, and rural status
of the residency (for the exact definitions, see app. B).
We use the 1979–2019 CPS-Basic files for the employment, unemployment,
and LFP variables as well as for a number of secondary variables describing the
nature of employment (part-time, overtime, and self-employment). Unlike the
CPS-ORG, CPS-Basic contains observations for every month that a respon-
dent is surveyed. Therefore, using the CPS-Basic to estimate employment, un-
employment, and LFP effects of the minimum wage results in greater precision
than using the CPS-ORG to estimate these effects. It also allows us to estimate
the impacts of the minimum wage on transitions between employment, unem-
ployment, and inactivity. We obtain the minimum wage data from Vaghul and
Zipperer (2016), which has been extended through 2019 by the authors.

IV. Predicting Who Is a Minimum Wage Worker

We build a prediction model to first explain the relationship between be-
ing a minimum wage worker (defined as having an hourly wage of less than
125% of the statutory minimum wage) and various demographic variables.6
Then we use the model to predict the likelihood of an individual being a min-
imum wage worker. As the model relies on demographic characteristics,
we can ascertain the likelihood of an individual being affected by the policy
even if that individual were currently not employed, or had no wage. As a
result, we can examine the effects of the policy not only on incumbent work-
ers but also on those who are currently nonemployed but are likely affected
by the policy.
We create the following data set to build the prediction model. First, we
select all workers in states and quarters that satisfy two criteria: (1) there had
not had been any prominent minimum wage events in the past 20 quarters

6
Our results are not sensitive to the definition of minimum wage workers based
on alternative cutoff values. Setting the thresholds to 3% above the minimum wage
or 200% above the minimum wage produces virtually the same ordering of obser-
vations according to predicted probabilities, suggesting that the specific definition
that we use has essentially no bearing on the conclusions.
S212 Cengiz et al.

and (2) there is a prominent minimum wage change in the next 12 quarters.
The former criterion ensures that we are not training the model using work-
ers in states/quarters where the wage distribution may not have stabilized
following a minimum wage event. The latter criterion ensures that we are
training the model on workers who will experience a minimum wage event
in the near future and are therefore pertinent to our analysis. There are
469,174 worker-level observations in the CPS that satisfy this screen be-
tween 1979 and 2019.
Second, we divide the 469,174 observations into two mutually exclusive
samples: a training sample and a test sample. To create the training sample
we randomly draw 150,000 observations. We apply various learning tools,
such as random forests, tree boosting, basic logistic, elastic net, and the linear
probability model along the lines of Card and Krueger (1995), and fit each
model on the training sample. In the next section we describe the key idea be-
hind each prediction algorithm. For further details on these prediction models,
we refer the reader to appendix C and Friedman, Hastie, and Tibshirani (2009).
The test sample is composed of the complement of the training sample.
We use the test sample to compare the performance of the prediction models
by plotting the precision-recall curves (explained below) along with other
descriptive statistics.
Once we have optimized over the prediction models, we use the preferred
(best-performing) model to calculate the predicted probability in the full
data set that includes all time periods and states between 1979 and 2019.
We use that full data set to investigate the causal effects of the minimum wage
on various predicted probability groups.

A. Prediction Algorithms
Decision Trees
A single decision tree lies at the heart of many learning techniques, including
random forests and gradient boosting. A decision tree recursively divides the
feature (predictor) space into two in a way that reduces the prespecified loss
function the most.7 More concretely, in the beginning the algorithm tries every
possible split to divide the entire sample space into two and picks the one that
diminishes the loss function the most. Subsequently, each subsample is treated
as the new sample, and the first step is repeated. Once the splitting is over, it
predicts the class of every observation according to the majority vote in the
subspace (terminal node) to which the observation belongs.
This procedure requires a decision on when to stop the splitting. In prin-
ciple, the splitting could continue until there is only one data point at each

7
The loss function is the deviance, defined as 22om ok nmk
logð p^mk Þ, where nmk
indicates the number of observations at terminal node m that belongs to class k and
^pmk is the share of observations at terminal node m that belongs to class k.
Seeing beyond the Trees S213

terminal node. Such a tree fits the training sample perfectly but would suffer
from overfitting. To overcome the problem, it is common to use cross val-
idation to determine the complexity of the tree. For a more accurate predic-
tion, we collapse some internal nodes (“prune the tree”) and decrease the
prediction variance at the expense of bias.
Decision trees are not among the most successful learners, yet they are
relatively easy to interpret. In figure 1 we plot a pruned decision tree pro-
duced to predict whether a worker has an hourly wage of less than 125% of
the statutory minimum wage using demographic and educational character-
istics. The tree predicts that the only group in the training sample with
hourly wages less than the threshold is the one with those who are 19 years
old or younger. The majority vote in all of the other terminal nodes is
“false,” indicating that nonteen observations are expected to work for hourly
wages higher than the threshold.
It is noteworthy that the recommendation based on a simple decision tree
is to proxy minimum wage workers with teens—which happens to be the
most common approach taken in the literature. However, as we show below,
it is possible to obtain much better predictions by combining multiple deci-
sion trees. The two most common ways to do so are the random forest by
Breiman (2001) and the gradient-boosting trees by Friedman (2001).

Random Forest
The random forest is a tree-based ensemble learning technique. It provides
a way to overcome the bias-variance trade-off of a single tree. It constructs a
multitude of fully grown decision trees formed using different training boot-
strap samples. Each tree produces unbiased predictions that have large var-
iances. We calculate the average of the predictions, thereby reducing the var-
iance. To further reduce the variance, we decrease the correlation among
trees by employing a randomly selected portion of the predictors at each
split. Although this results in the loss of the interpretability of individual
trees, it has no impact on the bias, since individual trees are still fully grown.8
Our fivefold cross validation finds that the optimum random forest is
achieved with 2,000 trees and only two predictors tried at each split.

Boosting
Boosting approaches the problem of how to combine multiple trees from
a different angle. Instead of producing many fully grown trees and averaging
them, the trees in this model are grown sequentially where subsequent trees
attempt to fix the errors of the preceding ones. As a result, while the first tree

8
Note that if trees are perfectly correlated, the reduction of the variance would
be nil. If they are independent, the variance of the final model would be j2 =B, where
j2 is the prediction variance of a single tree and B indicates the number of trees.
S214 Cengiz et al.

FIG. 1.—Minimum wage workers according to pruned trees. The figure plots a
pruned decision tree produced to predict whether an individual is a minimum wage
worker (has an hourly wage of less than 125% of the statutory minimum wage) using
demographic and educational characteristics. In the beginning the tree tries every pos-
sible split to divide the entire sample space into two and picks the one that diminishes
the loss function the most. Then each subspace is treated as the new feature space, and
the first step is repeated. “True” indicates that the tree predicts that workers in the ter-
minal node are minimum wage workers and “false” indicates otherwise. While the tree
explores characteristics such as gender, marital status, veteran status, and rural residency
status, it picks only age and education to make splits. LTHS 5 less than high school;
HSG 5 high school graduate; SC 5 some college; CG 5 college graduate.

in the boosting is interpretable, the subsequent trees are not independently

meaningful. Intuitively, with the boosted trees one starts with the lowest-
hanging fruit and, say, classifies teens as minimum wage workers and others
as not minimum wage workers. Then the second or subsequent tree builds
a model that focuses more on correctly classifying nonteen minimum wage
workers and teen non-minimum-wage workers, namely, the observations mis-
classified by the first tree. The change of focus is usually achieved by altering
Seeing beyond the Trees S215

the outcome variable (e.g., using the residual as the outcome variable) or slightly
changing the loss function (weighting the misclassified observations more
heavily). After building the subsequent tree, we combine the predictions of
all trees through a weighted majority vote. On the basis of our fivefold cross
validation, the optimum boosted tree model is obtained with the following
parameters: number of trees 5 4,000, shrinkage factor 5 0:005, depth of
tree 5 6, and minimum observations in a node 5 10.

Elastic Net
We use the elastic net regularization developed by Zou and Hastie (2005).
The underlying model is very similar to the logistic regression except that the
elastic net model penalizes model complexity. The penalty term is a linear
combination of the lasso and ridge methods; lasso tends to drop poor predic-
tors, while ridge tends to shrink their coefficients toward zero, so elastic net
combines both. As opposed to tree-based models, the elastic net regression re-
quires prespecification of the exact functional form for the predictors in the
prediction equation. Therefore, we purposefully build a fairly complex model,
where we include all of the predictors, their four-way interactions, and all of
the interactions with the quadratic, cubic, and quartic terms of the age variable.
We rely on the regularization to simplify the model and prevent overfitting.

Card and Krueger’s Linear Probability Model

This is a trial and error method employed by Card and Krueger (1995). We
follow the functional form proposed on page 135. They estimate a linear
probability model using the following independent variables: a set of three-
way interaction variables between teen, nonwhite, and gender indicators;
three-way interaction variables between young adult (aged 20–25), nonwhite,
and gender indicators; three-way interactions of age, categorical education,
and gender variables; quadratic and cubic terms of the age variable; and indi-
cator variables for Hispanic and nonwhite individuals. We can think of the
Card and Krueger model as a sort of lasso approach to predictor selection
but where the regularization is based on subjective judgment instead of a for-
mal learning algorithm.9

B. Precision-Recall Curves and Predicted Probabilities

To compare the models with each other, we employ two concepts from the
ML literature: precision and recall. “Precision” refers to the share of those we
classify as being in the predicted group who are true minimum wage workers.

9
We also tried to implement neural networks and support vector machines. While
the model constructed using the neural networks performs slightly worse than the
boosted tree model, the models using the support vector machines fail to provide a
well-performing prediction model.
S216 Cengiz et al.

“Recall” refers to the share of true minimum wage workers who we correctly
classify as being in the predicted group. For instance, if a predicted group has
only one observation and the observation is a true minimum wage worker, then
the precision is 1; however, here the recall is very small, as the sample will cover
only a minuscule fraction of the minimum wage workers in the population. On
the other hand, if the predicted group contains every observation in the popu-
lation, then the recall rate is 1, as the sample, by construction, includes all min-
imum wage workers in the population. However, here the precision is going to
be small, since the predicted group also includes all the non-minimum-wage
workers in the population. The ideal is to construct a predicted group that in-
cludes all the minimum wage workers and none of the non-minimum-wage
workers so that both the precision and the recall are 1. Generally, the higher
the precision for a given recall rate, the better the performance of the model.10
Figure 2A shows the precision-recall curves corresponding to the various
prediction algorithms. We also estimate and report the performance of a ba-
sic logistic model with age and the categorical education variables as predic-
tors for comparison. To plot the curve, we calculate the predicted probabil-
ities for each individual in the test sample. We then define the predicted
group for alternative probability thresholds, where all workers in the group
have a predicted probability greater than the threshold. We calculate the pre-
cision and the recall for each of these groups and obtain the curve. In other
words, each point on the curves corresponds to a separate predicted group.
When we raise the threshold, we expect the precision to increase but at the
cost of a reduced recall rate. How strong this trade-off is between the preci-
sion and recall rates for various prediction models is shown in the figure.
The figure shows that the boosted tree model (solid line) outperforms other
prediction models, since it provides the highest precision at almost all recall
levels. For comparison, in figure 2B we report the other prediction models rel-
ative to the boosted tree model. The boosted tree model (and also the other
prediction algorithms) improves precision considerably relative to the basic
logistic model. Nevertheless, the differences between the other prediction
models and the boosted tree model are relatively small, especially at higher
recall rate levels. The random forest model achieves almost the same result
as the boosted tree model. It is also notable that the Card and Krueger sub-
jective judgment approach does almost as well as the elastic regularization of
the logistic model, and the performance of their model is not far behind the
best-performing prediction model.11

10
Another approach commonly used to compare models is to plot the receiver operat-
ing characteristic (ROC) curve. The ROC curve plots the recall against the false positive
rate, the latter defined as the number of non-minimum-wage observations as a proportion
of the number of non-minimum-wage workers in the population. In our case, we reach the
same conclusion whether we use the ROC curve or the precision-recall curve.
11
An alternative way to assess model performance is to compare the fraction of true
minimum wage workers in each predicted probability decile. Table A.1 (tables A.1–A.5,
Seeing beyond the Trees S217

In figure 3 we compare the performance of the best prediction model, the

boosted tree model, relative to the strategy of choosing specific subgroups
to proxy minimum wage workers.12 First, in terms of precision, the teen
sample performs better than all of the other commonly used samples that
we compare (workers younger than 30 with less than high school education
[LTHS, age < 30], workers younger than 30 with high school or less educa-
tion [HSL, age < 30], and LTHS workers). In fact, when compared with
LTHS, age < 30, both recall value and precision values of the teen sample
are higher. This indicates that the teen sample includes more workers that
truly have hourly wages lower than 125% of the minimum, and it captures
them more accurately than the former samples. Second, the commonly
used samples that include nonteen workers tend to achieve a higher recall
than the teen sample. However, the rise in the recall is expensive in terms
of the lost precision. For instance, including all HSL workers younger than
30 increases the recall value by 0.124 compared with the teen sample, yet
the precision decreases by 0.256. It implies that many nonteen observations
in the HSL, age < 30 sample are actually non-minimum-wage workers.
Overall, it is clear that points on the curve are closer to the top-right corner
than the points corresponding to the commonly used samples. Furthermore,
the difference in the precision values between the commonly used samples
and the samples recommended by the tools increases as the recall increases.
For instance, the precision value that the learning tools achieve for the teen
sample’s recall value is only slightly greater than that of the teen sample
(the vertical distance between the black triangle and the curve). For the recall
value of the HSL, age < 30 sample, however, the learning tools achieve a sub-
stantially larger precision value than that of the former sample. Therefore, the
figure highlights that the learning tools improve the precision-recall trade-off
considerably, especially if the aim is to include nonteen observations.

C. Who Are the Minimum Wage Workers?

Before we study the impact of the minimum wage on labor market out-
comes, we examine who are the most likely minimum wage workers accord-
ing to our (best-performing) prediction model. To do this, we examine the

C.1, E.1 are available online) shows that the boosted tree model has a slightly higher
fraction of true minimum wage workers in the most likely predicted deciles and a
lower fraction in the least likely predicted deciles. This provides further support of
the slightly better performance of the boosted tree model.
12
Of course, it is possible that someone is directly interested in the impact of the
policy on the labor market outcomes of certain demographic groups or industries.
Nevertheless, in most cases researchers pick specific subgroups (e.g., teens) or sec-
tors (e.g., restaurants) not because they are the main subjects of interest but because
these are subgroups where the fraction of minimum wage workers is high. Further-
more, the prediction approach can also be applied if someone is specifically inter-
ested in the impact of the minimum wage on some subgroups (see table 5).
 FIG. 2.—Precision-recall curves. A plots the precision-recall curves for various
prediction models described in section IV.A and for a basic logistic model that we
estimate using (linear) age and categorical education variables. We obtain the precision-
recall curves in the following way: we use our prediction model to calculate the prob-
ability that someone is a minimum wage worker, and we assign all individuals to the
predicted group if that probability is above a certain threshold. The figure shows
the estimated precision and recall rates obtained when we vary the threshold value.
The horizontal dotted line shows the average share of minimum wage workers in the
sample. The areas below the precision-recall curves are 0.449 for boosted tree, 0.445 for
random forest, 0.443 for elastic net, 0.435 for Card and Krueger’s linear probability
model, 0.342 for basic logistic, and 0.269 for single tree. B shows the difference in pre-
cision rate between the best-performing model—the boosted tree model—and the
other models at each recall rate. A color version of this figure is available online.
Seeing beyond the Trees S219

FIG. 3.—Performance of the boosted tree model relative to demographic subgroups.

The figure compares the performance of the best-performing prediction model—the
boosted tree model—relative to the strategy of choosing specific subgroups to proxy
minimum wage workers. The solid line shows the precision-recall curve for the boosted
tree model. The black triangle shows the precision and recall level for teens, the black cir-
cle shows the level for individuals with less than high school (LTHS) education, the gray
circle shows the level for LTHS individuals younger than 30, and the gray triangle shows
the level for individuals with high school or less (HSL) education younger than 30. The
horizontal dashed line shows the average share of minimum wage workers in the sample.

main characteristics of individuals in various predicted probability deciles.

Table 1 shows the share of workers with various demographic characteristics
in each column, while predicted probabilities deciles are in descending order
(starting with the highest-probability group).
Examining the table, we observe that age is a highly important predictor.
Teens constitute 71.9% of the workers in the decile most likely to be exposed
to the minimum wage. This suggests that an analysis using only the high-
est decile would be very similar to an analysis based on just the teen sample.
However, the teen share drops to 4.7% in decile 9 and is virtually zero there-
after, and more than 34% of the observations in deciles 9 and 8 are adults
aged 20–30. This suggests that the importance of the age variable is not lim-
ited to determining whether an observation belongs to the highest decile.
The model also heavily employs educational attainment in determining the
predicted probabilities. Observations with less than high school education
are mostly in the highest deciles, and high school graduates are concentrated
on the top half of the exposed individuals. There is almost no observation
S220 Cengiz et al.

Table 1
Demographic Characteristics for Each Predicted Probability Decile
Black or
Teen 20 ≤ Age < 30 LTHS HSG Female White Hispanic
(1) (2) (3) (4) (5) (6) (7)
Most likely decile .719 .038 .752 .145 .592 .837 .244
Probability decile 9 .047 .405 .534 .238 .674 .847 .359
Probability decile 8 .004 .341 .344 .437 .594 .834 .243
Probability decile 7 .004 .298 .187 .575 .571 .833 .351
Probability decile 6 .000 .191 .085 .660 .673 .873 .150
Probability decile 5 .000 .187 .100 .475 .492 .784 .253
Probability decile 4 .000 .178 .067 .236 .512 .794 .237
Probability decile 3 .000 .162 .004 .297 .404 .865 .175
Probability decile 2 .000 .088 .000 .143 .385 .848 .122
Least likely decile .000 .015 .000 .039 .314 .741 .134
NOTE.—The table shows some demographic characteristics for each predicted probability decile. The pre-
dicted probability refers to the probability that an individual has an hourly wage lower than 125% of the min-
imum wage preceding the minimum wage hike and is calculated according to the best-performing prediction
model—the boosted tree model. Each row shows the average characteristics of individuals in the particular
predicted probability decile. The top (bottom) row shows the characteristics at the top (bottom) decile, which
consists of individuals that are most (least) likely exposed to the minimum wage according to our prediction
model. Each cell shows the share of the selected demographic group: col. 1, the share of teens (i.e., those youn-
ger than 20); col. 2, the share of individuals who are between 20 and 30 years of age; col. 3, the share of indi-
viduals with less than high school (LTHS) education; col. 4, the share of high school graduates (HSGs) with no
college education; col. 5, the share of females; col. 6, the share of white individuals; and col. 7, the share of
Black or Hispanic individuals.

without some college education in the lowest deciles. Another finding worth
noting is that the share of women workers is high in the top deciles (e.g.,
67.4% in decile 9) and is lower in the bottom deciles (31.4% in the least likely
decile), indicating that an individual’s gender also plays an important role.
Last, Black/Hispanic individuals’ share in the least likely decile is 13.4%,
whereas they make up at least 24% of the top two deciles.
An alternative way to examine who are the minimum wage workers is to
consider the relative importance of each predictor. In figure 4 we plot the “rel-
ative influences” of the variables calculated following Friedman (2001).13 The

The figure shows the reduction in the loss function caused by each variable
13

used in the nonterminal nodes. We normalize the relative influences so that they
sum up to 100. The average importance of each variable is
1 M 2
I 2l 5 o I l ðTm Þ,
M m51
where I l ðTm Þ is the reduction in the loss function due to the use of variable l in the
nonterminal nodes of tree m. However, we wish to caution against interpreting
them directly. First, the importance is in terms of prediction, not explanation. Sec-
ond, there are cases where one variable needs to be interacted with another one for
high predictive power. In those cases, only one of them is deemed to have a strong
influence, whereas both are essential.
Seeing beyond the Trees S221

FIG. 4.—Relative influences of the predictors in the boosted tree prediction model.
We plot relative influences of the variables in the best-performing prediction model—
the boosted tree model—calculated as in Friedman (2001; for details, see n. 13). The
bars, which indicate the decline in the loss function associated with the corresponding
variable, are normalized so that they sum up to 100.

figure largely confirms our previous observations. It finds age as the most im-
portant predictor in the sample with a very large margin. The variable for ed-
ucational credentials comes after age.14 Gender variables are also relatively im-
portant in the prediction. The indicator variables for Hispanic, rural, race, and
veteran status appear to have less influence on the prediction.

D. High-Probability and High-Recall Groups

While we show responses to the minimum wage change throughout the
whole predicted probability spectrum, in most of the paper we will focus on
two key subgroups in the baseline specifications. For the first group, we follow
Card and Krueger (1995) and define the high-probability group as consisting
of the top 10% of individuals in terms of predicted probability of being min-
imum wage workers. When we use the boosted tree prediction model, the
threshold probability that we need to apply to get this group is 35%—so all
individuals with predicted probabilities above this value are assigned to the
high-probability group. At this threshold, the precision rate is around 60%,
which means that around 60% of the individuals in the high-probability group

14
In fact, dropping teen observations from the sample decreases the relative im-
portance of the age variable substantially. It renders age to be the close second-most
important variable in the prediction. The educational credentials variable of the ob-
servation becomes the most important predictor.
S222 Cengiz et al.

are indeed minimum wage workers. The associated recall rate is 36%, which
means that this group covers around 36% of all minimum wage workers.
Since the high-probability group covers just over a third of all minimum
wage workers, we also study the impact of the minimum wage on a more
broadly defined group. In the high-recall group, we set a threshold probabil-
ity such that 75% of all minimum wage workers are captured. In practice,
this leads us to set the threshold at 12%; at that level we achieve a 35% pre-
cision rate. This high-recall group covers just more than 40% of all workers
in the data. To study the impact of the policy on workers unlikely to be af-
fected by the policy, we also define a group for whom the predicted proba-
bility is less than 12%. Throughout the paper, we refer to that group as the
low-probability group.

V. Impact of the Minimum Wage on Labor Market Outcomes

Identification Strategy
We estimate the effect of the minimum wage by implementing an event
study strategy similar to the baseline specification in Cengiz et al. (2019). We
exploit state-level variation in the minimum wage and compare labor market
outcomes in the treated states with the labor market outcomes in states without
any minimum wage hike. The event-based approach that we apply here is sim-
ilar to that of Autor, Donohue, and Schwab (2006).15 We focus on changes
within an 8-year window around 172 prominent state-level minimum wage
events instituted between 1979 and 2019. Prominent minimum wage changes
are those where the (real) minimum wages increased by more than $0.25 and
where at least 2% of the workforce earned between the new minimum wage
and the old minimum wage.16
Our basic regression specification is as follows:
4

o b treat t
g
Y st 5 t st 1 Qst 1 ms 1 rt 1 ust , (1)
t523

g
where Y st is the labor market outcome (e.g., employment rate, unemploy-
ment rate, participation rate) in state s and at quarter t for group g. Unlike
when estimating the prediction model, here we use all states and quarters
available in the CPS data. As we discussed above, we study the impact of
the minimum wage on various groups defined by the prediction model, such

15
Appendix G in Cengiz et al. (2019) shows how this event study approach is
related to alternative methods applied in the literature like the two-way fixed effects
estimator with log minimum wage.
16
We show the graphical distribution of prominent minimum wage changes and
the number of such changes in each year in fig. A.1 (figs. A.1–A.6, C.1, C.2 are avail-
able online). In fig. A.2 we plot the change in state-level log (real) minimum wage fol-
lowing a prominent minimum wage increase.
Seeing beyond the Trees S223

as the high-probability group and the high-recall group. Here, treattst is a bi-
nary variable that takes on the value of 1 if the minimum wage was increased
t years from date t in state s. This definition implies that t 5 0 represents the
first year following the minimum wage increase (i.e., the quarter of treatment
and the subsequent three quarters), and t 5 21 is the year (four quarters)
before treatment. Our benchmark specification controls for state and period
fixed effects, ms and rt, and we also include controls for small or federal in-
creases, Qst.17 We cluster our standard errors by state, which is the level at
which policy is assigned.
Our baseline approach uses staggered variation of minimum wage increase.
As shown in many recent papers, this can lead to negative weighting bias in
the presence of heterogeneous treatment effects (see, e.g., Sun and Abraham
2020). To alleviate these concerns, in table E.1 we present estimates with a
stacked regression approach following Cengiz et al. (2019). In this approach,
we align events by event time (and not calendar time) and use only within-
event variation (between the treated unit and clean control units), which is
equivalent to a setting where all of the events happened all at once and were
not staggered. Gardner (2021) derives the implicit weights for such a stacked re-
gression approach and shows that they do not suffer from negative weighting.

Main Results
Table 2 shows the estimated effects on the high-probability, the high-recall,
and the low-probability groups. We report 5-year averaged posttreatment
estimates for the key labor market outcomes (wages, employment, unem-
ployment, and LFP) relative to the t 5 21, formally ð1=5Þo4t50 ðbt 2 b21 Þ.
Columns 1 and 3 establish that the minimum wage has a significant positive
impact on wages for groups of workers predicted to be exposed to the min-
imum wage. In the high-probability group, wages increased by around 2.3%
(SE, 0.3%), while in the high-recall group—which captures 75% of the min-
imum wage workers—wage increase was a little smaller but still significant
(1.6%; SE, 0.3). In contrast, column 5 shows no indication of any significant
wage effects for the low-probability group. This confirms that the wage
growth occurred only for workers exposed to the minimum wages and not
for individuals unlikely to be directly exposed to the minimum wage shock.
Columns 2, 4, and 6 of table 2 show analogous estimates but classifying
workers according to the Card and Krueger linear probability model.18 It is

17
The variables we use to control for federal and small events are the same as the
ones employed in Cengiz et al. (2019). We collapse the windows for small and federal
events into three periods: EARLY, PRE, and POST. EARLY is for 3 and 2 years be-
fore, and PRE is for 1 year before the event. POST is for 0–4 years after the event.
18
Table A.2 shows that the boosted tree prediction model picks a sample that is a
bit older, more educated, more female, and less white than the sample selected by
the Card and Krueger model.
 Table 2
Impact of the Minimum Wage on Labor Market Outcomes

S224
(1) (2) (3) (4) (5) (6)
D wage (%) .023*** .025*** .016*** .014*** 2.001 2.001
(.003) (.004) (.003) (.003) (.003) (.003)
D employment (pp) .002 .001 .001 .001 .001 .000
(.002) (.002) (.001) (.001) (.001) (.001)
D unemployment (pp) 2.001 2.001 2.000 2.001* 2.000 .000
(.001) (.001) (.000) (.000) (.000) (.000)
D participation (pp) .002 .001 .000 .001 .001 .001
(.002) (.002) (.001) (.001) (.001) (.001)
Employment elasticity with respect
to minimum wage .071 .036 .020 .028 .012 .006
(.076) (.066) (.038) (.032) (.011) (.013)
Employment elasticity with respect
to wage .286 .138 .114 .191 NA NA
(.316) (.245) (.216) (.211)
 Table 2 (Continued)
(1) (2) (3) (4) (5) (6)
Number of events 172 172 172 172 172 172
Number of observations 7,854 7,854 7,854 7,854 7,854 7,854
Number of individuals in sample 6,639,492 5,812,367 20,917,455 24,737,455 29,370,470 25,511,485
Mean employment .338 .399 .415 .439 .741 .771
Mean unemployment .058 .064 .046 .043 .030 .030
Mean participation .395 .462 .460 .481 .771 .801
Group High probability High probability High recall High recall Low probability Low probability
Prediction model Boosted tree CK linear Boosted tree CK linear Boosted tree CK linear
NOTE.—The table reports the effects of the minimum wage on labor market outcomes according to the event study analysis (see eq. [1]) using 172 state-level minimum wage changes
between 1979 and 2019. The table reports 5-year averaged posttreatment estimates for each key labor market outcome: percent change in wages and the change in employment to pop-
ulation, unemployment to population, and labor force participation rate. We also report the employment elasticity with respect to the minimum wage and the employment elasticity with
respect to the wage, which is the ratio of the percent change in employment and wage. To calculate the percent change in employment, we divide the change in employment to population
by the mean employment to population rate preceding the minimum wage hikes (reported at the bottom of the table). “Number of observations” refers to the number of quarter-state cells
used for estimation, while “Number of individuals” refers to the underlying CPS sample used to calculate labor market outcomes in these cells. Columns 1 and 2 show estimates for the
high-probability group, which captures the 10% of the population with highest predicted probability. Columns 3 and 4 show estimates for the high-recall group, which consists of in-
dividuals whose predicted probability is above 12%—a threshold that leads to a 75% recall rate of minimum wage workers. Columns 5 and 6 show the estimates for workers whose
predicted probability is below 12%. Columns 1, 3, and 5 use the best-performing prediction model—the boosted tree model. Columns 2, 4, and 6 use the Card and Krueger (CK) linear
prediction model (here, the high-recall group is defined by having a predicted probability, using the linear prediction model, above 12%, which again generates a 75% recall rate). All

S225
regressions are weighted by state-quarter population. Robust standard errors in parentheses are clustered by state. pp 5 percentage point.
* p < .10.
*** p < .01.
S226 Cengiz et al.

worth noting that the wage effects are almost the same for the best-performing
prediction model and for the Card and Krueger approach and suggests that
the Card and Krueger model performs quite well in this setting.19
Table 2 also reports the effect of the minimum wage on employment. Con-
sidering the high-probability and high-recall groups, we find a small and statis-
tically insignificant positive effect on employment. The employment elasticity
with respect to minimum wage is around 0.07 (SE, 0.08) and 0.02 (SE, 0.04),
respectively. The 90% confidence intervals around these estimates can rule
out an elasticity of 20.1, the lower bound (in magnitude) of the range sug-
gested by Neumark and Wascher (2008).
Table 2 also shows that the employment effects are somewhat smaller for
the high-recall group than for the high-probability group, which is in line with
the wage effects. This leads to a similar elasticity of employment with respect
to own wage, which would be the labor demand elasticity in the neoclassical
model, in the two groups. When we calculate the employment elasticity with
respect to own wage, we obtain an elasticity of 0.29 (SE, 0.32) for the high-
probability group and 0.11 (SE, 0.22) for the high-recall group. The estimates
are quite precise, especially those from the high-recall group, which can rule
out all but a modest negative impact of the policy on employment.
A key advantage of the probability-based approach is that we can study
outcomes other than employment and wages. In table 2 we also report the ef-
fect of the minimum wage on unemployment rate and on participation rate.
For the high-probability group (col. 1), we find a slight decrease in unemploy-
ment and a slight increase in the participation rate. Importantly, however,
none of these changes in unemployment and participation rates are statistically
significant. The estimates for the high-recall group (col. 3) show no change in
either the unemployment rate or the participation rate.
The estimated slight decrease in unemployment or the slight increase in par-
ticipation (or just unchanged values of both) suggests that search effort is un-
likely to fall in response to the policy. This set of empirical findings is difficult
to reconcile with a Flinn-type search-and-matching model (Flinn 2006), where
adjustments on the participation margin play a vital role, or with models
predicting a considerable increase in the unemployment rate in response to
the policy (Drazen 1986; Lang 1987; Swinnerton 1996). Furthermore, as we
discussed in section II, the lack of a visible drop in participation implies that
the policy did not have a significant negative effect on the welfare of workers
at the margin of LFP, even as it raised wages for inframarginal workers.

19
However, the key advantage of applying ML tools is that someone can select
the predictors in a data-driven way without knowing much about the context. Even
if the functional form chosen by Card and Krueger (1995) performs very well, it is
unclear how someone with less knowledge about US labor markets could come up
with that functional form. Moreover, there is no guarantee that it would perform
well in all contexts.
Seeing beyond the Trees S227

Nonlinearity in the Extent of Exposure

Table 2 shows that the high-probability group (which captures the 10%
most exposed individuals) and the high-recall group (which covers a broader
group comprising 75% of the minimum wage workers) provide strikingly
similar estimates across the subgroups studied here. This suggests that the
impact of the policy does not seem to depend on nonlinearity in the extent
of exposure. Figure 5 explores this nonlinearity in more detail. We show the
estimated impact on the key labor market outcomes by applying various cut-
offs in the predicted probability. The solid lines show the 5-year averaged
posttreatment estimates for individuals whose predicted probability is above
the particular predicted probability threshold. As this threshold decreases,
we also naturally decrease the precision, as our group will include more non-
minimum-wage workers in the sample. At the same time, we also capture a
larger fraction of minimum wage workers, so we attain a higher recall rate.
Figure 5A shows that the wage effects tend to decline as we lower the min-
imum predicted probability. This is what we would expect given the declin-
ing precision rate: as we lower the threshold, our sample will cover more and
more non-minimum-wage workers whose wages are unaffected (directly)
by the policy. Figure 5B shows that employment responses, which start from
a small positive effect, decline as we lower the precision. As a result, when we
divide the employment effects by the wage effects, we obtain a stable employ-
ment elasticity with respect to own wage.
Furthermore, figure 5C shows that the unemployment rate declines at the
high-probability group but that the decline shrinks as we increase the recall
rate. The estimated change in participation rate, which is shown in figure 5D,
is close to zero and unrelated to the recall rate. Overall, the graphical evidence
in figure 5 finds no clear indication for nonlinearities in response to the policy.

Timing of the Impact

Figures 6 and 7 show the impact of the minimum wage on various labor
market outcomes over time for the high-recall and high-probability groups,
respectively. In the figures we plot bt expressed relative to the event date
t 5 21, or the year just before treatment. Panel A shows the evolution of
wages around the minimum wage increase. There is a clear increase in wages
in line with the timing of the minimum wage increase, and this is not driven by
preexisting trends. Over time the wage effects are somewhat attenuated, which
reflects that the most recent minimum wage changes tend to be larger.20

20
Since we do not fully observe the impact of the policy 5 years after the minimum
wage increase for the most recent events, those events impact the estimates only in
earlier posttreatment years. In figs. A.4 and A.5 we assess the timing of the policy
when we focus only on minimum wage changes where we see responses for the entire
event window. For these events with a balanced sample, which are not subject to the
composition effect, we find no decline in wage effects over time.
S228 Cengiz et al.

FIG. 5.—Impact of the minimum wage for alternative predicted probability thresh-
old values. The figure shows the main results from our event study analysis (see eq. [1])
using alternative predicted probability threshold values. We exploit 172 state-level min-
imum wage changes between 1979 and 2019. The figure shows the effect of a minimum
wage increase on wages (A), on employment to population (B), on unemployment to
population (C), and on labor force participation rate (D). In each panel the solid line
shows the 5-year averaged posttreatment estimates for individuals whose predicted
probability is above the minimum predicted probability threshold. On the x-axis we
also report the corresponding recall rate (the fraction of minimum wage workers re-
trieved by the prediction model if the particular minimum predicted probability thresh-
old is applied) and the precision rate (the fraction of minimum wage workers in the pre-
dicted group if the particular minimum predicted probability threshold is applied). We
also plot the thresholds corresponding to the high-probability group capturing the
10% of the population with the highest predicted probability and to the high-recall
group capturing 75% of all minimum wage workers. To calculate the predicted prob-
abilities, we use the best-performing prediction model—the boosted tree model. The
shaded areas show the 95% confidence interval based on standard errors that are
clustered at the state level. A color version of this figure is available online.

Panel B shows the impact of the policy on employment around the min-
imum wage hike. For both the high-recall group and the high-probability
group, we see a similar pattern: there is no clear evidence of preexisting trends
in employment, although there is a slight dip in employment 2 years before
the minimum wage increase when we look at the high-recall group (fig. 6).
Nevertheless, there is no unusual employment change if we look at the longer
Seeing beyond the Trees S229

FIG. 6.—Impact of the minimum wage over time, high-recall group. The figure
shows the main results from our event study analysis (see eq. [1]) using 172 state-
level minimum wage changes between 1979 and 2019. The figure shows the effect of
a minimum wage increase on wages (A), on employment to population (B), on un-
employment to population (C), and on labor force participation rate (D) for the high-
recall group. The high-recall group consists of all workers whose predicted probability
is above 12%—a threshold that corresponds to a 75% recall rate. To calculate the pre-
dicted probabilities, we use the best-performing prediction model—the boosted tree
model. We also show the 95% confidence interval based on standard errors that are clus-
tered at the state level. A color version of this figure is available online.

horizon between 1 and 3 years preceding the minimum wage increase. Fur-
thermore, the small drop in employment between 1 and 2 years preceding the
minimum wage increase would imply that the economy slightly deteriorates
before an average minimum wage hike, so we would expect a decrease in
employment rate after the policy change. In contrast, we see no clear break
in employment after the policy change—if anything, there is a small, statisti-
cally insignificant increase.
Panel C shows the impact on the unemployment rate. There is neither any
preexisting trend nor any break after the minimum wage increase. Panel D
shows the impact of the policy on participation rate, which also shows a flat
response following the policy change. Overall, the preexisting trends are re-
assuring, and we find no indication of any break in the participation rate at
the time of the minimum wage increases. Moreover, we do not find evidence
S230 Cengiz et al.

FIG. 7.—Impact of the minimum wage over time, high-probability group. The fig-
ure shows the main results from our event study analysis (see eq. [1]) using 172 state-
level minimum wage changes between 1979 and 2019. The figure shows the effect of a
minimum wage increase on wages (A), on employment to population (B), on unem-
ployment to population (C), and on labor force participation rate (D) for the high-
probability group. The high-probability group consists of the 10% of the population
with the highest likelihood of being affected by the policy. To calculate the predicted
probabilities, we use the boosted tree model. We also show the 95% confidence in-
terval based on standard errors that are clustered at the state level. A color version of
this figure is available online.

of lower employment (or higher unemployment) rates when we look up to

the fifth year following the policy change.

Impact by Predicted Probability Quintiles

Figure 8 shows the effect of the minimum wage separately for each pre-
dicted probability quintile. Figure 8A shows that there is a clear wage effect
among the individuals most exposed to the minimum wage. Wages also in-
crease for the second-highest quintile of predicted probabilities, but this in-
crease is statistically insignificant. Figure 8B–8D shows the effect of the policy
on labor market outcomes, such as employment, unemployment, and partic-
ipation. None of these outcomes shows any substantial change even at the
highest predicted probability quintile. Reassuringly, none of the outcomes are
Seeing beyond the Trees S231

FIG. 8.—Impact of the minimum wage by predicted probability quintiles. The fig-
ure shows the effect of the minimum wage separately for each predicted probability
quintile. The highest quintile comprises individuals with predicted probabilities (of
being minimum wage workers) in the top 20%. We estimate equation (1) for each
quintile separately and report the 5-year averaged posttreatment estimates. We use
172 state-level minimum wage changes between 1979 and 2019. The figure shows
the effect of a minimum wage increase on wages (A), on employment to population
ratio (B), on unemployment rate (C), and on labor force participation rate (D). We
also show the 95% confidence intervals based on standard errors that are clustered
at the state level. A color version of this figure is available online.

statistically distinguishable from zero in the bottom quintiles of predicted

probability, which provides support for the validity of the research design.

Robustness
In table 3 we assess the robustness of the main results shown in table 2 to the
inclusion of various versions of time-varying heterogeneity for the high-recall
group, while we report the same robustness checks for the high-probability
group in table A.3. In column 1 we report the estimates for the baseline spec-
ification shown in table 2. Column 2 allows the period effects to vary by the
nine census divisions. The results are similar to the baseline specification: we
find a positive and statistically significant wage effect and a positive employ-
ment effect—which comes from an increase in the participation rate.
 Table 3
Impact of the Minimum Wage on Labor Market Outcomes, Robustness to Alternative Specifications (High-Recall Group)
(1) (2) (3) (4) (5) (6) (7)
D wage (%) .016*** .016*** .018*** .012*** .016*** .016*** .016***
(.003) (.003) (.003) (.003) (.005) (.002) (.002)
D employment (pp) .001 .002 .001 2.001 .002 .000 .000
(.001) (.002) (.002) (.001) (.002) (.001) (.001)
D unemployment (pp) 2.000 .000 2.001 .000 2.001* 2.000 2.000

S232
(.000) (.000) (.000) (.001) (.000) (.000) (.000)
D participation (pp) .000 .002 .000 2.000 .001 .000 .000
(.001) (.002) (.002) (.001) (.002) (.001) (.001)
Employment elasticity with respect
to minimum wage .020 .047 .018 2.014 .033 .010 .007
(.038) (.038) (.040) (.037) (.052) (.021) (.035)
Employment elasticity with respect
to wage .114 .284 .096 2.109 .233 .071 .048
(.216) (.242) (.208) (.285) (.330) (.141) (.205)
Number of events 172 172 406 172 99 172 172
Number of observations 7,854 7,854 7,854 7,854 7,854 7,854 7,854
Number of individuals in sample 20,917,455 20,917,455 20,917,455 20,917,455 20,917,455 20,917,455 20,917,455
Mean employment .415 .415 .425 .419 .426 .415 .415
Mean unemployment .046 .046 .049 .045 .050 .046 .046
Mean participation .460 .460 .473 .464 .476 .460 .460
 Table 3 (Continued)
(1) (2) (3) (4) (5) (6) (7)
Controls:
State fixed effects Y Y Y Y Y Y Y
Quarter fixed effects Y Y Y Y Y Y Y
Division-quarter fixed effects Y
State federal events Y
Unweighted Y
Number of events after 2014:Q1 Y
State employment control: all Y
State unemployment control: all Y
State employment control: low-probability group Y
State unemployment control: low-probability group Y
NOTE.—The table reports the effects of the minimum wage on labor market outcomes according to the event study analysis (see eq. [1]) using 172 minimum wage changes between
1979 and 2019. We assess the impact of the minimum wage on the high-recall group. The high-recall group consists of individuals whose predicted probability is above 12%—a thresh-
old that leads to a 75% recall rate of minimum wage workers. The table reports 5-year averaged posttreatment estimates for each key labor market outcome: percent change in wages and
the change in employment to population, unemployment to population, and labor force participation rate. We also report the employment elasticity with respect to the minimum wage
and the employment elasticity with respect to the wage, which is the ratio of the percent change in employment and wage. To calculate the percent change in employment, we divide the
change in employment to population by the mean employment to population rate preceding the minimum wage hikes (reported at the bottom of the table). “Number of observations”

S233
refers to the number of quarter-state cells used for estimation, while “Number of individuals” refers to the underlying CPS sample used to calculate labor market outcomes in these cells.
In all of the regressions, we use the best-performing prediction model—the boosted tree model. Column 1 shows the preferred benchmark estimate reported in col. 3 of table 2. Column 2
augments the baseline model with division-by-quarter fixed effects. Column 3 reports estimates using 406 state or federal minimum wage increases. All regressions are weighted by state-
quarter population except col. 4, where we report unweighted estimates. Column 5 considers only minimum wage events that happened on or before 2014:Q1 to ensure a full 5-year post-
treatment period. Column 6 controls for state-level unemployment and employment rates (as a fraction of population), while col. 7 controls for the employment and unemployment rates of
individuals with low predicted probability of being a minimum wage worker (less than 12%). Robust standard errors in parentheses are clustered by state. pp 5 percentage point.
* p < .10.
*** p < .01.
S234 Cengiz et al.

So far, we have focused only on prominent state-level minimum wage

changes. In column 3 of table 3 we expand the event definition to include
(nontrivial) federal minimum wage increases. This leads us to 406 minimum
wage changes. Inclusion of these federal-level events produces similar results,
although the (small) reduction in the unemployment rate is now marginally
statistically significant. In column 4 we provide estimates without using pop-
ulation weights. The estimates are similar, albeit the employment and par-
ticipation estimates without population weight are somewhat closer to
zero and more noisily estimated. The implied employment elasticity with
respect to wage without population weights is 20.109 (SE, 0.29), which can
rule out only sizeable negative responses to the policy at the conventional sig-
nificance levels.
In column 5 we restrict our analysis to prominent state-level minimum
wage changes where we observe all of the 5 years after reform. In this anal-
ysis we include only events that took place on or before the first quarter of
2014. Restricting the sample to these events does not affect the key findings:
we obtain a slightly larger participation estimate (0.1 percentage points in-
stead of zero), but the estimates also become a little noisier. Finally, columns 6
and 7 control for the state unemployment and employment rates (as a fraction
of population), using all workers in the case of column 6 and workers in the
low-probability group (i.e., probability less than 12%—the high-recall cutoff)
in the case of column 7. Controlling for the state unemployment and employ-
ment rates—either using all workers or just workers with a low probability of
being a minimum wage worker—produces results that are very close the base-
line estimation. The ability to control for state-level labor market conditions of
the workers unlikely to be treated by the policy is another advantage of our
prediction-based approach. The findings confirm that our results are not driven
by unusual labor market conditions in treated states around the time of the
policy changes.

Effects on Labor Market Transitions

Table 4 shows the estimated impact on monthly labor market transition
rates between employment, unemployment, and participation. We find no
statistically or economically significant effects on the transition probabilities
in response to the policy change for the high-probability group (col. 1) or the
high-recall group (col. 2). This clarifies that the minimum wage increase did not
lengthen unemployment durations; if anything, the policy accelerated monthly
transitions from unemployment to employment by 0.006 (SE, 0.005) percentage
points for the high-probability group and by 0.004 (SE, 0.004) percentage
points for the high-recall group. Therefore, we find no indication that some in-
dividuals were pushed permanently into long-term unemployment or that the
labor markets became more sclerotic (with low job-finding rates) in response to
increases in the minimum wage.
Seeing beyond the Trees S235

Table 4
Impact of the Minimum Wage on Labor Market Transitions
(1) (2) (3)
D E-U flow as a share of employment (pp) .000 .000 .000
(.001) (.000) (.000)
D E-I flow as a share of employment (pp) 2.000 .001 .000
(.001) (.001) (.000)
D U-E flow as a share of unemployment (pp) .006 .004 .004
(.005) (.004) (.003)
D U-I flow as a share of unemployment (pp) .002 .003 .000
(.005) (.004) (.003)
D I-E flow as a share of inactivity (pp) .001 .000 .001
(.001) (.000) (.001)
D I-U flow as a share of inactivity (pp) .000 .000 2.000
(.001) (.000) (.000)
Number of events 172 172 172
Number of observations 7,854 7,854 7,854
Number of individuals in sample 4,701,665 14,856,017 21,156,039
Mean E-U flow as a share of employment (%) .028 .022 .009
Mean E-I flow as a share of employment (%) .107 .061 .020
Mean U-E flow as a share of unemployment (%) .230 .246 .258
Mean U-I flow as a share of unemployment (%) .369 .295 .190
Mean I-E as a share of inactivity (%) .057 .042 .053
Mean I-U as a share of inactivity (%) .036 .024 .023
Group High High recall Low
probability probability
Prediction model Boosted tree Boosted tree Boosted tree
NOTE.—The table reports the effects of the minimum wage on labor market transition rates according to
the event study analysis (see eq. [1]) using 172 state-level minimum wage changes between 1979 and 2019.
The table reports 5-year averaged posttreatment estimates for percentage point changes in the employment-to-
unemployment (E-U) transition rate as a share of employment (row 1), the employment-to-inactivity (E-I)
transition rate as a share of employment (row 2), the unemployment-to-employment (U-E) transition rate
as a share of unemployment (row 3), the unemployment-to-inactivity (U-I) transition rate as a share of un-
employment (row 4), the inactivity-to-employment (I-E) transition rate as a share of inactivity (row 5),
and the inactivity-to-unemployment (I-U) transition rate as a share of inactivity (row 6). We also report
the mean levels of each of these variables for the period preceding the minimum wage hikes. “Number of obser-
vations” refers to the number of quarter-state cells used for estimation, while “Number of individuals” refers to
the underlying CPS sample used to calculate labor market outcomes in these cells. Column 1 shows estimates for
the high-probability group, which captures the 10% of the population with highest predicted probability. Column 2
shows estimates for the high-recall group, which consists of individuals whose predicted probability is above
12%—a threshold that leads to a 75% recall rate of minimum wage workers. Column 3 shows the estimates
for workers whose predicted probability is below 12%. Columns 1–3 use the best-performing prediction
model—the boosted tree model. All regressions are weighted by state-quarter population. Robust standard
errors in parentheses are clustered by state. pp 5 percentage point.

Effects by Demographic Subgroups

In some cases, policy makers may be interested in the impact of minimum
wages on specific demographic groups. Moreover, individuals are more likely
to be on the participation margin in some demographic groups than in others.
As a result, the impact of the policy on participation decisions might vary
considerably across demographic groups. We study the heterogeneity in re-
sponse to the minimum wage in table 5. To make sure that changes in labor
market outcomes are driven by the minimum wage itself and not something
 Table 5
Impact of the Minimum Wage on Labor Market Outcomes by Demographic Group I

S236
(1) (2) (3) (4) (5) (6) (7) (8)
D wage (%) .016*** .012*** .015*** .001 .027*** .016** .015*** .014***
(.003) (.003) (.004) (.005) (.003) (.007) (.004) (.003)
D employment (pp) .001 2.003 .001 2.001 .004 2.002 .001 .000
(.001) (.003) (.001) (.002) (.003) (.004) (.002) (.001)
D unemployment (pp) 2.000 2.001 2.001* 2.000 2.001 .000 2.000 2.000
(.000) (.001) (.000) (.000) (.002) (.001) (.000) (.000)
D participation (pp) .000 2.004 .000 2.002 .003 2.001 .000 .000
(.001) (.002) (.001) (.002) (.002) (.004) (.002) (.001)
Employment elasticity with respect
to minimum wage .020 2.069 .020 2.044 .112 2.072 .019 .013
(.038) (.056) (.039) (.050) (.088) (.147) (.058) (.038)
Employment elasticity with respect
to wage .114 2.526 .128 NA .396 2.430 .120 .090
(.216) (.431) (.248) (.318) (.927) (.356) (.258)
 Table 5 (Continued)
(1) (2) (3) (4) (5) (6) (7) (8)
Number of events 172 172 172 172 172 172 172 172
Number of observations 7,854 7,841 7,854 7,854 7,854 7,854 7,854 7,854
Number of individuals in sample 20,917,455 5,680,302 13,357,475 4,883,626 3,763,811 2,163,408 9,118,096 16,755,128
Mean employment .415 .499 .385 .347 .333 .264 .361 .405
Mean unemployment .046 .061 .038 .024 .068 .015 .050 .049
Mean participation .460 .560 .423 .371 .401 .279 .412 .453
Group High recall High recall High recall High recall High recall High recall High recall High recall
Demographic group All Black or Hispanic Female Married female Teen Aged 60–70 LTHS HSL
NOTE.—The table reports the effects of the minimum wage on labor market outcomes by demographic group according to the event study analysis (see eq. [1]). We exploit
172 state-level minimum wage changes between 1979 and 2019. We assess the impact of the minimum wage on the high-recall group including individuals whose predicted probability
is above 12%—a threshold that leads to a 75% recall rate of minimum wage workers. The table reports 5-year averaged posttreatment estimates for each key labor market outcome:
percent change in wages and the change in employment to population, unemployment to population, and labor force participation rate. We also report the employment elasticity with
respect to the minimum wage and the employment elasticity with respect to the wage, which is the ratio of the percent change in employment and wage. To calculate the percent
change in employment, we divide the change in employment to population by the mean employment to population rate preceding the minimum wage hikes (reported at the bottom
of the table). “Number of observations” refers to the number of quarter-state cells used for estimation, while “Number of individuals” refers to the underlying CPS sample used to
calculate labor market outcomes in these cells. In all of the regressions, we use the best-performing prediction model—the boosted tree model. The demographic subgroups are Black
or Hispanic, woman, married woman, teen, aged 60 and older and less than 70, less than high school (LTHS) education, and high school or less (HSL) education. All estimates are
weighted by the corresponding subgroup’s population in the state. Robust standard errors in parentheses are clustered by state. pp 5 percentage point.

S237
* p < .10.
** p < .05.
*** p < .01.
S238 Cengiz et al.

else, in each subgroup, we focus on workers who are in the high-recall group.21
Restricting the sample to workers who are likely to be minimum wage workers
is also necessary for getting first-stage wage effects, which would not be pos-
sible if we had all workers in the sample (see table A.1 in the online appendix of
Cengiz et al. 2019).
For most subgroups in table 5, we find a clear and significant impact on
wages, which confirms a key advantage of restricting the sample to groups
that are likely to be exposed to the minimum wage. In column 1 we simply
reproduce the benchmark estimates for the overall high-recall group for
comparability. In column 2 we report the estimates for workers who are
Black or Hispanic.22 While the estimates are somewhat noisy, the point es-
timates indicate a nontrivial drop in employment and participation. This
suggests that in the Black or Hispanic group, some workers (who were at
the participation margin) may have been made worse off by the policy.
However, the standard errors are too large to draw a clear conclusion on this
question.
In columns 3 and 4 we examine the impact of the policy on all women and
on married women, respectively. Since the extensive margin labor supply elas-
ticities are often found to be larger for women than for men, it might be the
case that minimum wages lead to a greater increase in women’s participa-
tion. Furthermore, married women are also typically thought to have larger
responsiveness on the extensive margin. However, our estimated effects for
employment and participation outcomes for women and married women
are very similar to the benchmark specification; however, we note that there
is no statistically (or economically) significant wage effect for married women
in our sample, which makes it difficult to draw any strong conclusions. (We
find more informative evidence when we consider married mothers with
younger kids below.)
Columns 5 and 6 report the impact on teens and on older individuals (aged 60–
70), respectively. Both younger individuals and older individuals have lower
participation rates than prime-age individuals, and they are also thought to
have more elastic labor supply (Blundell, Bozio, and Laroque 2011). For
teens, we find larger wage effects and slightly larger employment increase

21
We use the predicted probabilities estimated on all workers. One could estimate
a separate prediction model for each subgroup and then use those predicted proba-
bilities. However, it is unlikely that there are substantial gains from estimating a sep-
arate prediction model for each subgroup. If there were substantial gains from using
some predictors differently within a particular subgroup, then the boosting algo-
rithm would tend to detect and incorporate this fact into the prediction model—even
if it were estimated using the full sample. We also explored separately estimating pre-
diction models for some specific demographic groups but found negligible changes in
the precision of our estimates.
22
We attempted to estimate the effect of the policy on Black and Hispanic indi-
viduals separately, but the estimates were too imprecise to be informative.
Seeing beyond the Trees S239

than for the overall sample. The increase in employment comes from the
changes in participation, which is in line with the idea that more teens are
on the participation margin and also with the evidence presented by Laws
(2018). For older individuals, the wage effects are imprecise (although the
point estimates are close to the overall sample). This makes interpreting the
results for older individuals difficult. Still, we document a slight positive (sta-
tistically insignificant) employment and participation effect, which is in line
with Borgschulte and Cho (2019), who document a similar-sized response
in employment for those aged between 62 and 70.
Columns 7 and 8 show the impact on those with lower educational cre-
dentials. The labor market impact of the policy on these education groups is
very similar to the impact of the policy on the overall sample. These findings
suggest that workers with lower education credentials seem to benefit from
the minimum wage policies.
Table 6 presents additional heterogeneity analysis. Columns 2–4 show the
estimates by the predicted probability of moving into and out of the labor
force. The goal of this exercise is to help us assess any heterogeneity in the
effects by how likely workers are to be at the margins of LFP. We predict
the probability that an individual changes their LFP status—either from non-
participation to participation or from participation to nonparticipation—by
applying the boosting tree ML method. For estimating this prediction model,
we use all of the demographic variables that we had used before to predict
minimum wage exposure but also add the number of children under the
age of 5, since it is likely to be an important predictor of LFP. The relative
importance of the predictors is shown in figure A.6. Similar to the prediction
model on minimum wage exposure, age, education, and gender are the three
most important predictors of changes in participation. In addition, race and
number of children under the age of 5 also substantially influence the predic-
tion model.
Since the number of children is coded consistently only since 1986 in the
CPS, we report estimates using the 1986–2018 period in table 6. Column 1
reports estimates for all workers using only that period, and they are very
similar to the estimates using the 1979–2019 sample. Column 2 shows the es-
timates for the group of workers that has a high predicted probability of
changing LFP status. The estimated impacts for this group are very similar
to those from the overall sample. Columns 3 and 4 show the estimates with
lower probability of switching. Again, we find responses similar to those
from the overall sample. They key takeaway is that even when we look at
individuals who are more likely to switch LFP, we find no meaningful dif-
ference in the causal effects of minimum wage policies. These findings cast
doubt that there was much of an impact from minimum wage changes in
our sample on job search and participation behavior, even among groups
that are likely to be at the margin of participation. In columns 5 and 6 we also
report estimates on single and married mothers, focusing on those with kids
 Table 6
Impact of the Minimum Wage on Labor Market Outcomes by Demographic Group II

S240
(1) (2) (3) (4) (5) (6)
D wage (%) .015*** .017*** .012*** .011 .025** .006
(.003) (.003) (.003) (.008) (.012) (.009)
D employment (pp) .001 .000 .002 2.001 2.000 2.005
(.001) (.002) (.002) (.001) (.006) (.004)
D unemployment (pp) 2.001 2.001 2.001 .001* 2.001 2.001
(.000) (.001) (.001) (.000) (.003) (.001)
D participation (pp) .000 2.001 .001 2.001 2.001 2.006
(.001) (.002) (.002) (.001) (.007) (.004)
Employment elasticity with respect
to minimum wage .017 .006 .035 2.084 2.005 2.115
(.035) (.035) (.038) (.071) (.113) (.097)
Employment elasticity with respect
to wage .107 .033 .276 NA 2.021 NA
(.217) (.193) (.339) (.442)
 Table 6 (Continued)
(1) (2) (3) (4) (5) (6)
Number of events 156 156 156 156 156 156
Number of observations 6,222 6,222 6,222 6,222 6,222 6,222
Number of individuals in sample 15,760,550 7,883,899 2,915,439 4,961,212 525,347 618,103
Mean employment .417 .512 .561 .148 .520 .417
Mean unemployment .047 .068 .042 .008 .091 .040
Mean participation .463 .581 .604 .155 .612 .457
Probability group High recall High recall High recall High recall High recall High recall
Demographic group All High LFP switch Medium LFP switch Low LFP switch Single mother Married mother
kids under 5 kids under 5
NOTE.—The table reports the effects of the minimum wage on labor market outcomes by demographic group according to the event study analysis (see eq. [1]). We exploit 156 state-
level minimum wage changes between 1986 and 2018. We assess the impact of the minimum wage on the high-recall group including individuals whose predicted probability is above 12%—
a threshold that leads to a 75% recall rate of minimum wage workers. The table reports 5-year averaged posttreatment estimates for each key labor market outcome: the change in em-
ployment to population, unemployment to population, and labor force participation rate. “Number of observations” refers to the number of quarter-state cells used for estimation, while
“Number of individuals” refers to the underlying CPS sample used to calculate labor market outcomes in these cells. In all of the regressions, we use the best-performing prediction model—
the boosted tree model. Column 1 shows the estimates on overall employment using the 1986–2018 periods. Columns 2–4 show the estimates for individuals with the different predicted
probability of moving into or out of the labor force. Column 5 shows the estimates for single mothers with children under the age of 5, while col. 6 shows the estimates for married mothers
with children under the age of 5. All estimates are weighted by the corresponding subgroup’s population in the state. Robust standard errors in parentheses are clustered by state. pp 5
percentage point.

S241
* p < .10.
** p < .05.
*** p < .01.
S242 Cengiz et al.

under 5. For single mothers, we find a very small increase in employment

and participation; for married mothers, we find a small reduction in employ-
ment and participation. Neither set of estimates are economically or statisti-
cally significant.23

Additional Labor Market Outcomes

Finally, in table 7 we study the impact on other labor market outcomes,
such as self-employment, part-time (working fewer than 30 hours per
week), and overtime (working more than 40 hours per week) status. Chang-
ing working hours and pushing workers to self-employment are often ar-
gued to be important margins of adjustment to the minimum wage. We find
little indication that the high-probability or the high-recall group experi-
enced any changes in self-employment. We do find a decline in the share
of employees in part-time jobs for the high-probability group, with close
to no change in the share of employees in overtime jobs. The former change
is statistically significant at the conventional levels. This implies that the
minimum wages increase the share of full-time jobs (without any overtime)
in the low-wage workforce.

VI. Conclusion
In this paper we estimate the impact of the minimum wage on various la-
bor market outcomes using 172 prominent minimum wage changes between
1979 and 2019. To capture the impact of the policy on a broad group of af-
fected workers, we utilize modern ML techniques to estimate the likelihood
that someone is a minimum wage worker. While the best-performing predic-
tion model does better than the linear prediction model of Card and Krueger
(1995), the gap is not large. One implication of these findings is that mini-
mum wage researchers who are not interested in investing in a ML approach
may do fairly well by simply applying the Card and Krueger linear proba-
bility prediction model. Of course, the advantage of the ML approach is that

23
We find an effect of minimum wages on employment for single mothers with
young children that is close to zero and statistically insignificant. This differs from
Godoy, Reich, and Allegretto (2020), who find a large, statistically significant in-
crease in employment for single mothers with kids under 5 in response to a mini-
mum wage increase. Table A.5 provides a reconciliation of the discrepancy between
these findings in greater detail. There are a number of differences between the two
papers, but the primary reason why our estimates differ from Godoy, Reich, and
Allegretto (2020) is that we use the CPS-Basic files while they use the (smaller)
CPS-ORG data. While use of the CPS-Basic versus CPS-ORG data by itself does
not make a difference in the overall sample (see cols. 1–5 in table A.5), the results are
more sensitive to the data sets for the single mother sample, which is much smaller
(see cols. 6–10 in table A.5).
Seeing beyond the Trees S243

Table 7
Impact of the Minimum Wage on Alternative Labor Market Outcomes
(1) (2) (3)
D self-employment as share of employment (pp) .001 2.000 2.001
(.001) (.001) (.001)
D part-time as share of employment (pp) 2.005** 2.001 .000
(.002) (.001) (.000)
D overtime as share of employment (pp) .001 .000 2.001
(.001) (.001) (.001)
Number of events 172 172 172
Number of observations 7,854 7,854 7,854
Number of individuals in sample 6,639,492 20,917,455 29,370,470
Mean self-employment .029 .066 .129
Mean part-time .414 .216 .080
Mean overtime .029 .067 .196
Group High High recall Low
probability probability
Prediction model Boosted tree Boosted tree Boosted tree
NOTE.—The table reports the effects of the minimum wage on alternative labor market outcomes according
to the event study analysis (see eq. [1]) using 172 state-level minimum wage changes between 1979 and 2019.
The table reports 5-year averaged posttreatment estimates for the following labor market outcomes: self-
employment, part-time (working fewer than 30 hours per week), and overtime (working more than 40 hours
per week). Each of these variables is expressed as a share of total employment. Column 1 shows estimates for
the high-probability group, which captures the 10% of the population with the highest predicted probability.
Column 2 shows estimates for the high-recall group, which consists of individuals whose predicted probabil-
ity is above 12%—a threshold that leads to a 75% recall rate of minimum wage workers. Column 3 shows the
estimates for workers whose predicted probability is below 12%. All columns use the best-performing pre-
diction model—the boosted tree model. All of the regressions are weighted by state-quarter population. Ro-
bust standard errors in parentheses are clustered by state. pp 5 percentage point.
** p < .05.

it will successfully locate minimum wage workers in environments different

from the one we study here, including in other data sets or time periods.
The prediction-based approach allows us to study the impact of the min-
imum wage not just on wages and employment but also on unemployment
and participation rates for groups that cover 75% of all minimum wage
workers. These groups include substantially more minimum wage workers
than the demographic-based and industry-based subsamples commonly
used in the literature. In line with much of the existing evidence in the liter-
ature, we find that the minimum wage has a positive and significant impact
on wages, while employment effects are modest in the US context. We also
show that the slight (statistically insignificant) employment increase comes
from a slight drop in unemployment and a slight increase in the participation
rate. These responses indicate that the minimum wage is unlikely to have a
negative impact on workers by discouraging them to search for new jobs.
We also find no significant heterogeneities in the responses to the mini-
mum wage. The most likely exposed group and a much boarder group that
covers 75% of all minimum wage workers responded very similarly to the
policy change. The only indication for a potential negative impact of the pol-
icy was on the Black or Hispanic subgroup, but those estimates are too noisy
S244 Cengiz et al.

to draw a definitive conclusion. Overall, our results underscore the positive

impact of the policy on key labor market outcomes.
Our findings suggest that the prediction-based approach of defining a
likely treatment group could be naturally applied to study the impact of the
policy on various other important outcomes measured in different data sources.
Since the most important predictors (age, education, gender) are available in
most data sources, it is straightforward to assess the impact of the minimum
wage hikes along the predicted probability spectrum (and show estimates like
those in fig. 5). A key advantage of doing so is to study directly whether there
is some nonlinearity in response to the minimum wage by the extent of expo-
sure. Furthermore, studying the impact of the policy on the low-probability
group can serve as an additional falsification test and can provide further sup-
port for the credibility of the research design. Given these advantages, the
prediction-based approach should be a part of the standard toolkit for ana-
lyzing the impact of minimum wages.

References
Adams, Camilla, Jonathan Meer, and CarlyWill Sloan. 2018. The minimum
wage and search effort. NBER Working Paper no. 25128, National Bu-
reau of Economic Research, Cambridge, MA.
Addison, John T., and McKinley L. Blackburn. 1999. Minimum wages and
poverty. Industrial and Labor Relations Review 52, no. 3:393–409.
Addison, John T., McKinley L. Blackburn, and Chad Cotti. 2011. Mini-
mum wage increases under straightened circumstances. IZA Discussion
Paper no. 6036, Institute of Labor Economics, Bonn.
Ahn, Tom, Peter Arcidiacono, and Walter Wessels. 2011. The distributional
impacts of minimum wage increases when both labor supply and labor
demand are endogenous. Journal of Business and Economic Statistics 29,
no. 1:12–23.
Allegretto, Sylvia, Arindrajit Dube, Michael Reich, and Ben Zipperer. 2017.
Credible research designs for minimum wage studies: A response to
Neumark, Salas, and Wascher. ILR Review 70, no. 3:559–92.
Autor, David H., John J. Donohue III, and Stewart J. Schwab. 2006. The
costs of wrongful-discharge laws. Review of Economics and Statistics
88, no. 2:211–31.
Belman, Dale, and Paul J. Wolfson. 2014. What does the minimum wage do?
Kalamazoo, MI: W.E. Upjohn Institute.
Belman, Dale, Paul Wolfson, and Kritkorn Nawakitphaitoon. 2015. Who is
affected by the minimum wage? Industrial Relations: A Journal of Econ-
omy and Society 54, no. 4:582–621.
Blundell, Richard, Antoine Bozio, and Guy Laroque. 2011. Labor supply
and the extensive margin. American Economic Review 101, no. 3:482–86.
[Link]
Seeing beyond the Trees S245

Borgschulte, Mark, and Heepyung Cho. 2019. Minimum wages and retire-
ment. ILR Review 73, no. 1:153–77.
Breiman, Leo. 2001. Random forests. Machine Learning 45, no. 1:5–32.
Card, David. 1992. Do minimum wages reduce employment? A case study
of California, 1987–89. ILR Review 46, no. 1:38–54.
Card, David, and Alan B. Krueger. 1995. Myth and measurement: The new
economics of the minimum wage. Princeton, NJ: Princeton University
Press.
Cengiz, Doruk, Arindrajit Dube, Attila Lindner, and Ben Zipperer. 2019.
The effect of minimum wages on low-wage jobs. Quarterly Journal of
Economics 134, no. 3:1405–54.
Clemens, Jeffrey, Lisa B. Kahn, and Jonathan Meer. 2018. The minimum wage,
fringe benefits, and worker welfare. NBER Working Paper no. 24635, Na-
tional Bureau of Economic Research, Cambridge, MA.
Clemens, Jeffrey, and Michael R. Strain. 2017. Estimating the employment
effects of recent minimum wage changes: Early evidence, an interpretative
framework, and a pre-commitment to future analysis. NBER Working
Paper no. 23084, National Bureau of Economic Research, Cambridge, MA.
Clemens, Jeffrey, and Michael Wither. 2019. The minimum wage and the
Great Recession: Evidence of effects on the employment and income tra-
jectories of low-skilled workers. Journal of Public Economics 170:53–67.
Currie, Janet, and Bruce C. Fallick. 1996. The minimum wage and the em-
ployment of youth evidence from the NLSY. Journal of Human Re-
sources 31, no. 2:404–28.
Drazen, Allan. 1986. Optimal minimum wage legislation. Economic Journal
96, no. 383:774–84. [Link]
Dustmann, Christian, Attila Lindner, Uta Schönberg, Matthias Umkehrer,
and Philipp vom Berge. 2022. Reallocation effects of the minimum wage.
Quarterly Journal of Economics 137, no. 1:267–328.
Flinn, Christopher J. 2002. Interpreting minimum wage effects on wage dis-
tributions: A cautionary tale. Annales d’Economie et de Statistique 67/
68:309–55. [Link]
———. 2006. Minimum wage effects on labor market outcomes under search,
matching, and endogenous contact rates. Econometrica 74, no. 4:1013–62.
[Link]
———. 2011. The minimum wage and labor market outcomes. Cambridge,
MA: MIT Press.
Friedman, Jerome H. 2001. Greedy function approximation: A gradient
boosting machine. Annals of Statistics 29, no. 5:1189–232.
Friedman, Jerome, Trevor Hastie, and Robert Tibshirani. 2009. The elements
of statistical learning: Data mining, inference, and prediction. New York:
Springer.
Gardner, John. 2021. Two-stage differences in differences. Working paper,
University of Mississippi.
S246 Cengiz et al.

Giuliano, Laura. 2013. Minimum wage effects on employment, substitu-

tion, and the teenage labor supply: Evidence from personnel data. Journal
of Labor Economics 31, no. 1:155–94.
Godoy, Anna, Michael Reich, and Sylvia A. Allegretto. 2020. Parental labor
supply: Evidence from minimum wage changes. IRLE Working Paper
no. 103-19, Institute for Research on Labor and Employment, University
of California, Berkeley.
Lang, Kevin. 1987. Pareto improving minimum wage laws. Economic Inquiry
25, no. 1:145–58. [Link]
Laws, Athene. 2018. Do minimum wages increase search effort? CWPE
Working Paper no. 1857. Cambridge Working Papers in Economics,
University of Cambridge.
Linneman, Peter. 1982. The economic impacts of minimum wage laws: A
new look at an old question. Journal of Political Economy 90, no. 3:443–
69.
Lundstrom, Samuel M. 2016. When is a good time to raise the minimum
wage? Contemporary Economic Policy 35, no. 1:29–52.
Manning, Alan. 2016. The elusive employment effect of the minimum wage.
Working paper, Washington Center for Equitable Growth, Washington, DC.
Marimpi, Maria, and Pierre Koning. 2018. Youth minimum wages and youth
employment. IZA Journal of Labor Policy 7:1–18.
Meer, Jonathan, and Jeremy West. 2015. Effects of the minimum wage on
employment dynamics. Journal of Human Resources 51, no. 2:500–522.
Mincer, Jacob. 1976. Unemployment effects of minimum wages. Journal of
Political Economy 84, no. 4:S87–S104.
Neumark, David. 2017. The employment effects of minimum wages: Some
questions we need to answer. NBER Working Paper no. 23584, National
Bureau of Economic Research, Cambridge, MA.
Neumark, David, J. M. Ian Salas, and William Wascher. 2014. Revisiting the
minimum wage-employment debate: Throwing out the baby with the
bathwater? ILR Review 67, suppl.: 608–48.
Neumark, David, and William Wascher. 1992. Employment effects of min-
imum and subminimum wages: Panel data on state minimum wage laws.
ILR Review 46, no. 1:55–81.
———. 2008. Minimum wages. Cambridge, MA: MIT Press. https://
[Link]/books/minimum-wages.
Ragan, James F. 1977. Minimum wages and the youth labor market. Review
of Economics and Statistics 59, no. 2:129–36.
Sabia, Joseph J., Richard V. Burkhauser, and Benjamin Hansen. 2012. Are
the effects of minimum wage increases always small? New evidence from
a case study of New York State. ILR Review 65, no. 2:350–76.
Sun, Liyang, and Sarah Abraham. 2020. Estimating dynamic treatment effects
in event studies with heterogeneous treatment effects. Journal of Econo-
metrics 225, no. 2:175–99.
Seeing beyond the Trees S247

Swinnerton, Kenneth. 1996. Minimum wages in an equilibrium search

model with diminishing returns to labor in production. Journal of Labor
Economics 14, no. 2:340–55.
Totty, Evan. 2017. The effect of minimum wages on employment: A factor
model approach. Economic Inquiry 55, no. 4:1712–37.
Vaghul, Kavya, and Ben Zipperer. 2016. Historical state and sub-state min-
imum wage data. Working paper, Washington Center for Equitable
Growth, Washington, DC. [Link]
/uploads/2016/09/02153029/[Link].
Wessels, Walter J. 2005. Does the minimum wage drive teenagers out of the
labor force? Journal of Labor Research 26:169–76.
Zou, Hui, and Trevor Hastie. 2005. Regularization and variable selection via
the elastic net. Journal of the Royal Statistical Society B 67, no. 2:301–20.