SmartZena
dortyboy
sbh-reanalysis-revisedmay2014-2aogjgs.pdf
Are the Effects of Minimum Wage Increases Always Small?
A Re-Analysis of Sabia, Burkhauser, and Hansen
Saul D. Hoffman Department of Economics
University of Delaware
May 23, 2014
Abstract: In a recent article, Sabia, Burkhauser, and Hansen report very large negative employment effects of the 2004-2006 increase in the NY state minimum wage on young, less- educated workers. I re-examine their estimates using data from the full CPS, rather than the smaller CPS-MORG files they use, and find no evidence of a negative employment impact. In this case, the full CPS, which is the source of U.S. official labor market statistics, is certainly the more appropriate and reliable data. Furthermore, when I repeat their analysis using three states and the District of Columbia that also had a substantial increase in the state minimum wage in the same time period, I find evidence of a small positive employment effect. Together, the two findings are consistent with other more recent research that reports very weak or zero disemployment effects of the minimum wage.
Key Words: Minimum Wage
JEL Codes: J08, J21, J38
In a recent important contribution to the minimum wage literature, Sabia, Burkhauser,
and Hansen (2012) ask “Are the Effects of Minimum Wage Increases Always Small?” Using
evidence from the 2004-2006 increase in the New York state minimum wage from $5.15 to
$6.75, they answer emphatically “no.” They find that employment for less-educated workers
under age 30 fell by 20%, which yields an employment elasticity of approximately -0.7, far
larger than estimates found in most of the more recent empirical minimum wage literature.
Indeed, they conclude that “these findings provide plausible evidence that large state minimum
wage increases can have substantial adverse labor demand effects for younger, less-
experienced, less-educated individuals that are well outside the consensus range of –0.1 to –0.3
found in the literature” (p. 372). This result has been cited by some conservative think tanks
and on-line commentators as important evidence against an increase in the federal or state
minimum wages.1 This result also figures prominently in other analyses of the redistributive
and anti-poverty impact of the minimum wage by the same authors (Sabia and Burkhauser
2010).
Their analysis is based on employment data from the Current Population Survey-Merged
Outgoing Rotation Group (CPS-MORG) files for 2004 and 2006. They use a variety of difference
methods to compare employment changes in NY to the corresponding changes in either
neighboring states or a synthetic control group. The analysis is very capably executed, but it is
ultimately undermined by two factors. First, the data set they use yields estimates of the
employment rate in NY and the control group states that differ substantially from the
corresponding official rates derived from the full CPS sample. The MORG files used by Sabia,
1 See, for example, Hotz-Eakin (2013) in the Huffington Post and Employment Policies Institute (2012).
1
Burkhauser and Hansen (hereafter SBH) are a subset of the regular CPS that includes the one-
quarter of the CPS panel that is rotating out of the sample after either four or eight months in
the survey.2 The full CPS, not the one-quarter MORG subsample, is the source of official BLS
tabulations of employment and unemployment and it is clearly the preferred data source. SBH
used the MORG data because, unlike the regular CPS, the MORG includes information on wage
rates for workers paid by the hour and weekly earnings for other workers. This is essential
information for computing wage impacts of the minimum wage, but not for estimating
employment effects. As shown below, the employment rate effects computed from the full
CPS files for 2004 and 2006 yield a very different picture of the impact of the minimum wage
increase in NY. While the MORG files are, in principle, an appropriate data set to use, in
practice their representativeness may fail for relatively small state-by-age group samples, such
as are used in their analysis.
Second, NY appears to be a somewhat idiosyncratic treatment state. In natural
experiments like this one, it is always necessary to assume that the treatment and control
groups are similar except for the treatment itself, here, the minimum wage increase. If that
were true, then, by extension, states with minimum wage increases similar to that in NY would
be expected to have relatively similar responses. But that is not the case. Three other large
states (Illinois, Florida, and New Jersey) and the District of Columbia had minimum wage
increases at the same time that were quite substantial—an average increase of $1.03 or 18.7%.
The employment response to the minimum wage increase in these states is substantially
2 CPS sample members are interviewed for four months consecutively, leave the survey for eight months, and then return for another four months. They are part of the ORG files in both of their final months of interviewing.
2
different from that in NY. Indeed, the employment rate of the same group analyzed by SBH
increased in these states relative to states that had no increase in the minimum wage.
In this paper, I re-examine the NY minimum wage experiment analyzed by SBH using
both the CPS-MORG and the full CPS data for 2004 and 2006. I also apply the same methods to
examine the impact of the minimum wage increase in the other states with a sizeable increase.
The next section of the paper briefly reviews the analysis and findings of SBH and then focuses
on the NY experiment. The following section provides a parallel analysis of the impact of
minimum wages in the other states that also had substantial increases during the same time
period.
II. Employment Effects of the NY Minimum Wage Increase
Background. Between 2004 and 2006, the state minimum wage in New York was
increased in two steps from $5.15 to $6.75, while the federal minimum was unchanged at
$5.15. Three geographically-proximate states—New Hampshire, Pennsylvania, and Ohio—had
minimum wage rates of $5.15 throughout the period and are used by SBH as a control group.3
The use of geographically proximate areas with different minimum wages was first famously
used in a natural experiment context by Card and Krueger (1994) following the 1992 increase in
New Jersey’s minimum wage. Similar approaches have been used subsequently in research by
Dube, Naidu, and Reich (2007), who compared restaurant employment in San Francisco and
neighboring cities after a local increase in the minimum; Hoffman and Trace (2009), who
compared Pennsylvania and New Jersey after a federal minimum wage increase that affected
3 Four other neighboring states (Vermont, Massachusetts, Connecticut, and New Jersey) either had an increase in their state minimum or had a constant, but higher minimum.
3
only Pennsylvania; and Dube, Lester, and Reich (2010), who compared restaurant employment
in adjacent counties that are across state boundaries and are subject to different minimum
wages.
SBH use primarily difference-in-difference methods, with and without control for
covariates. As the group potentially most adversely affected by the minimum wage increase,
they focus on 16-29 year olds without a high school degree. They also use a difference-in-
difference-in-difference model to compare employment changes of the target group to the
employment changes for a putatively unaffected group across the two sets of states. Finally, in
addition to the three neighboring states, SBH also compare NY to a synthetic control group
using the methods of Abadie, Diamond, and Hainmueller (2010).
Their analysis uses data from the CPS-MORG files, which are the merged annual files for
the outgoing rotation groups of the regular CPS. Each month’s ORG file contains one-quarter of
the full CPS sample who are rotating out of the sample after four or eight months of interviews.
Thus, the annual MORG file contains three times the sample size of any single month’s CPS and
one-quarter the sample size of the full annual CPS. Monthly sample sizes for a sample that
includes just a few states and a restricted age and education range can be relatively small. For
16-29 year olds with less than a high school degree, the CPS-MORG annual file includes 989
persons in NY in 2004, 916 in 2006, and 1765 and 1499 for the control group. Monthly sample
sizes average about 75-80 for NY and 125-150 for the control group. Sample sizes for subgroups
by age bracket are obviously much smaller.
4
SBH use the MORG files because they first examine whether the minimum wage
increase affected the distribution of wage rates. Only the MORG file contains information on
wage rates. For this reason, the annual MORG files are the data source used in the annual BLS
reports on the characteristics of minimum wage workers (BLS 2013) and are occasionally also
used in analyses of wage inequality (Card and DiNardo 2002). While they are essential for that
purpose, they are not ideal for the analysis of employment rates, because of their smaller
sample size. Indeed, for employment analyses, they have no advantage whatsoever over the
full CPS sample.4 The full CPS sample is always the source for official tabulations of labor
market outcomes, including employment, labor force participation, and unemployment. In
many cases, the MORG files may be a suitable substitute for the full CPS; they are, after all, a
random part of a nationally-representative sample. But with smaller sample sizes, the
representativeness may not carry through.
Analysis. To re-examine the impact of the NY state increase in the minimum wage, I
downloaded the MORG files for 2004 and 2006 from the NBER website and the corresponding
monthly CPS files from the US Census site using Data Ferrett. Table 1 summarizes the age, race,
and education distribution of the CPS and MORG samples. The estimates shown utilize sample
weights and thus are population estimates. In terms of these observable characteristics, the
CPS and MORG files are very similar. Age, race, and the proportion male are virtually identical
and the education distributions differ only slightly. The only mean that is statistically different
across the data sets at the 10% level or more is the proportion with very low education in the
4 This point has been made previously by Addison, Blackburn, and Cotti (2013), who recommend that researchers use the MORG files only to examine wage effects and then use the full CPS to examine employment effects.
5
control states, where the MORG files has a higher proportion. In both data sets, the NY
samples have a much higher proportion of blacks from the control state samples, a lower
proportion with 10 years of education and a higher proportion with twelve years (but no
degree); these differences are statistically significant. The full annual CPS files provide samples
for the NY and control group states that are about four times as large as the MORG samples.
In my re-analysis, I focus on the comparison to the geographically-proximate states
rather than the synthetic comparison group. The results of the two analyses in SBH are virtually
identical.
In Table 2, I show the re-analysis of the NY v NH/PA/OH natural experiment separately
by state, year, and data source. For each age group, I show the MORG results from their Table
3 and my estimates from the full CPS. The huge adverse employment effect reported by SBH is
easily seen. The employment rate for the less-educated younger workers in NY plummeted
from .362 to .291 between 2004 and 2006, a 20% decline. Employment in the control group
states was essentially unchanged, yielding a difference-in-difference estimate of −0.076 that is
statistically significant at the 5% level. Since the wage increased 31%, the employment
elasticity is a very sizeable −0.63. My estimates from the MORG files are identical to theirs,
both for means and sample size, which confirms that their analysis does not involve any
idiosyncratic coding or sampling whatsoever. There is no question that the MORG files show a
very substantial adverse employment effect of the minimum wage increase in NY on this group
of relatively young, less-educated workers.
6
The estimates from the full CPS, presented in the second row, show a very different
picture, however. The 2004 employment rate for NY is a full 2.6 percentage points lower than
the MORG estimate, while the 2006 CPS estimate for NY is about 1.6 percentage points higher.
For the control states, the 2004 employment rate from the full CPS is very close to the MORG
estimate, but the 2006 rate is 1.7 percentage points lower than in the MORG file. The net effect
of all these adjustments is a DID estimate of the impact of the NY state minimum wage increase
of less than one percentage point (−0.008) that is statistically insignificant, compared to the
statistically significant -0.076 estimate from the MORG data. The two DID estimates are clearly
statistically different. The lower bound of the 95% confidence interval for the DID estimate
from the full CPS is -0.035, which is less than half the point estimate in SBH’s analysis.
A similar pattern is seen in the next two rows, which focus on the subset of 16-19 year
olds. The employment rates from the MORG files yield a DID estimate of 6.3 percentage points,
equivalent to an elasticity of −0.79, given the lower baseline employment rate. Again, the main
factor is a very sharp decline in the employment rate in NY, while the employment rate in the
control states is essentially unchanged. With the full CPS data, the 2004 NY employment rate
for this subgroup is 2.7 percentage points lower and the 2006 rate is 1.1 percentage points
higher than in the MORG files. Thus, the estimated decrease in the NY employment rate is 2.6
percentage points, less than half the decrease in the MORG data. At the same time, the control
state employment rate change moves in the opposite direction, from neutral in the MORG data
to a 2.1 percentage point decrease in the CPS data. None of the four sets of mean differences
are large enough relative to their standard error to be statistically significant, but the net result
is a DID estimate of less than half a percentage point, rather than the 6.3 percentage points
7
computed by SBH. Again, the lower bound of the 95% confidence interval for the DID estimate
from the CPS (−.033) is about half the MORG point estimate presented by SBH.
The remaining rows of the table show the employment rates for the other age
subgroups examined by SBH. The general pattern follows what has been seen in rows (1)-(4).
For 20-24 year olds, SBH find a difference-in-difference estimate of −12.4 percentage points,
primarily due to an enormous 10.7 percentage point drop in the NY employment rate,
equivalent to a 19.9% decrease. With the CPS, the difference-in-difference estimate is about
one-quarter as large (−3.7 percentage points) and is not statistically significant. The underlying
NY employment rate decline in the CPS is half as large as in the MORG and it is partly offset by a
two point decline in the control states, rather than the 3.5 percentage point increase in the
MORG data. The employment rates and sample sizes in the MORG imply that the NY sample
contained 94 employed 20-24 year olds without a high school degree in 2004 and 63 in 2006. If
the employed numbers had, instead, been 91 and 68—hardly large changes—the MORG
employment rates would have matched those from the full CPS.
For 25-29 year olds without a high school degree, the MORG difference-in-difference
estimate of the effect of the minimum wage on employment is −5.3 percentage points. In the
CPS, the estimate is actually positive (0.011), but not close to statistical significance. Again, the
very small sample size in the MORG, which ranges from 109 to 158, is potentially an issue.
Finally, in the last rows, which focuses on 20-29 year olds with at least a high school degree—a
group plausibly largely unaffected by the minimum wage—the two sets of estimates are very
similar. Note that this is the largest sample size in the MORG by a very substantial margin.
8
The final column shows difference-in-difference estimates adjusted via regression for
age, education, race, and gender.5 Again, I report the estimates from SBH for the MORG file
and my corresponding estimates from the full CPS files. The procedures I follow are similar to,
although probably not literally identical, to what SBH do.6 In their estimates, the adjusted DID
effects are very similar to and sometimes slightly greater in absolute value than the unadjusted
results. Thus, for example, the adjusted DID estimates are −0.073 for the full sample, −0.072 for
teens, and −0.141 for 20-24 year olds. All of these effects are statistically significant at the 95%
level. With the CPS data, adjustment for covariates makes the impacts a bit larger in absolute
value, but still quantitatively small and not statistically significant. For all young, less-educated
workers, the adjusted DID estimate is −0.018 with a standard error of −0.0137. The largest
adjusted effect and largest t-statistic is for 20-24 year olds, where the DID estimate is −0.054
and the t-statistic is 1.5.
Difference-in-difference-in-difference estimates are shown in Table 3. Here, the
comparison is between across-state employment rate change differences for an at-risk group
and one essentially unaffected by an increase in the minimum wage across the two sets of
states. The DIDID allows for further control for otherwise unmeasured factors that might differ
between the treatment and control states (Hoffman, 2014). As the unaffected group, SBH use
persons age 20-29 with a high school degree, whose employment rate changes were shown in
the bottom rows of Table 2. Their estimates are shown in their Table 4 for models including
covariates or, alternatively, can be computed from the figures presented in their Table 3
5 Full regression results are available on request. 6 They do not present the estimates or exact details of coding. 7 Treating the point estimate as if it were statistically significant yields an elasticity of −0.11, which is at the low end among previous studies that find negative employment effects.
9
without adjustment for covariates. The results are virtually identical, so to keep the analysis
simple, I focus on the DIDID without covariates.
All of the figures in Table 3 were previously presented in Table 2. The first panel shows
the DID estimate (2006−2004) of −0.008 for the affected group (age 16-29, not a high school
graduate) from the full CPS. The next panel shows the corresponding estimate for the
unaffected group (−0.0005). The DIDID estimate from the CPS is, therefore, −0.0075 with a
standard error about twice as large as that. The estimate from SBH using the MORG is shown in
the bottom row: it is −0.086 with a t-statistic of 2.6. The corresponding regression-adjusted
DIDID estimate from their Table 4 is −0.078 with a t-statistic of 1.70. I do not present the DIDID
estimates for the other subgroups shown in Table 2, but it is obvious that they will be very
similar to the DID estimates in that table, since the control group DID is itself very small,
implying that correction for other unmeasured effects is not quantitatively important.
Interpretation. What should we make of the differing CPS and MORG estimates? As I
have argued above, the CPS data is the source for official BLS employment estimates. With the
CPS data, for example, any researcher can exactly replicate national published BLS employment
estimates for teens or any other age group. This includes not only the annual average (not
seasonally-adjusted) employment and unemployment rate, but also the underlying monthly
rates and the corresponding number of persons employed and unemployed. I have done that
10
with the 2004 and 2006 CPS samples for teens.8 All estimates exactly match the BLS figures
available at http://data.bls.gov/pdq/querytool.jsp?survey=ln.
In contrast, estimates from the MORG do not replicate the official figures, although at
the national level, the differences in estimates from the two data sets are relatively small. The
official average teen employment rate in 2004 computed from the CPS and reported by the
Bureau of Labor Statistics was 36.4%, while in the MORG file, it is 35.7%, a difference of 0.7
percentage points (about 2%). In 2006, the difference is 0.3 percentage points. The 2004
difference is statistically significant at the 5% level, while the 2006 is not. In both cases, the
differences are quantitatively small.
At the state level, however, where sample sizes are much smaller, the differences are
often much larger. Consider 2004 when, as noted above, the difference in the average annual
employment rate for teens at the national level between the two data sets was 0.7 percentage
points. I calculated the state-level teen employment rates with both data sets and then
computed the difference. The average absolute value of the state differences was 1.4
percentage points and the median difference was 1.5 percentage points. In 15 states, the
difference was greater than two percentage points, with four states having a difference greater
than three percentage points.
For the analysis sample used by SBH (age 16-29, not high school graduate), the same
pattern holds. The national employment rates in 2004 are .388 in the CPS data and a virtually
identical .385 in the MORG data. But the average absolute value of the difference at the state
8 I use this group rather than the sample used by SBH because national published estimates are available for comparison. The teen employment data is for 16-19 year olds.
11
level is 1.4 percentage points, and 15 states have a difference of two percentage points or
more. Unfortunately for the SBH analysis, NY and PA are conspicuous outliers of opposite signs:
for NY, the MORG employment rate, as seen in Table 2, is 2.6 percentage points higher and in
PA, it is 1.4 percentage points lower (result not shown separately in Table 2). In OH and NH, the
estimated employment rates differ by just 0.2 percentage points in the two data sets.
Figures 1 and 2 provide some further insight into the differing NY employment rate
estimates. The employment rate in the CPS is a weighted average of the rates for the ORG and
non-ORG parts of the sample and it is easy, therefore, to back out the employment rate for the
non-ORG sub-sample. The two figures plot the employment rate by month for the two
subsamples. In 2004 (see Figure 1), the ORG series is clearly far more variable, which is not
surprising given its smaller sample size. The average month-to-month change in the
employment rates is 7.5 percentage points, compared to 2.9 points for the non-ORG sample.
The two series are within 1-3 percentage points in five months, and in another four months,
they differ by four to six percentage points, with the ORG higher in two and the non-ORG higher
in the other two. But in the remaining three months, they differ by 10-15 percentage points,
with the ORG estimates always higher, and in each case followed in the subsequent month by a
change in the ORG rate that eliminates most of the difference between the estimates. The
three outlier months account for almost all of the 3.5 percentage point difference in the annual
rates for the two subsamples. In the other nine months, the simple average difference is less
than one percentage point. The monthly differences are statistically significant at the 5% level
in two months and for the year as a whole.
12
In 2006 (see Figure 2), the ORG series is again much more variable from month to
month; the average month-to-month change is again .075, compared to .042 for the CPS.9
There is no clear pattern in the differences; the ORG rates are lower in eight months, higher in
three, and close to the non-ORG rate in only one. Again months where the two rates differ the
most are typically followed by months where the ORG rate moves toward the more stable non-
ORG rate. This is true for February, April, June, September, and November. Overall, because
the ORG employment rate is more often lower than the non-ORG rate, the average annual rate
from the MORG is 1.7 percentage points lower than the CPS (see Table 2, rows 1 and 2).
The same comparison for the control states also shows far more variability by month for
the ORG sample than the non-ORG10, but the difference in the rates is smaller. In 2004, for
example, in eight months the two series are very close, in three months the non-ORG rate is
higher, and in one month the ORG rate is higher. As previously seen in Table 2, the annual rates
differed by about three-quarters of a percentage point. In 2006, the two rates are similar in
eight months, but now the ORG rate is higher in three months and lower in one. On average,
the months with a positive ORG difference, which are as large as 9.8 percentage points, yield a
1.7 percentage point higher employment rate.
In both years, the higher variability in the employment rate from month-to-month in the
MORG than in the CPS is undoubtedly related to its smaller sample size. Why this translated
into a higher employment rate for NY in 2004 and a lower one in 2006 is a puzzle, but it is
genuine—and unfortunate for the SBH analysis. It is, I suspect, simply a small sample problem.
9 Some seasonal variability in the employment is expected for a population that includes many students. 10 In 2004, the average month-to-month change is .022 for the CPS and .049 for the MORG. In 2006, the corresponding averages are .019 and .047.
13
As I noted above, a relatively small change in the number of persons reported as employed in
the MORG is all that is necessary to yield employment rates similar to the CPS, especially for the
subgroups with smaller samples. One possibility is fluctuation in the proportion of the MORG
NY sample that is teenaged. In 2004, this proportion ranges from under 50% to almost 75%
with an average of 66.6%, while in 2006 it ranges from 63% to 79% with an average of 70%.
Monthly variation in the teen population share explains 55% of the variation in the difference
between the two sets of monthly employment rates in 2004, but almost none of the variation
in 2006.
Whatever the explanation, when estimates differ, as they do here, there is no option
but to accept those from the full CPS, which is four times larger and indisputably more fully
representative. On that basis, I conclude that the natural experiment created by the increase in
the minimum wage in NY shows a negligible impact on employment of persons age 16-29
without a high school degree.
III. How Representative is New York?
If the New York minimum wage natural experiment is to be of policy importance, it
ought to have some predictive value for other states with minimum wage increases. As a test,
one can conduct the same kind of natural experiment using other states to assess the validity of
using NY as a representative case. In fact, such a natural experiment can readily be done. Over
the same time period, New Jersey, Florida, Illinois and the District of Columbia increased their
minimum wage substantially, while another seven states increased their minimum wage by
much smaller amounts.
14
To examine this, I again use the full CPS data for 2004 and 2006. I focus on the three
states plus DC with the largest increases and compare them to all states that had no increase
over this time period. More refined comparisons could undoubtedly be made, but the results
of this exercise ought to be suggestive. The minimum wage increased in these states11 by an
average of $1.03, equivalent to an 18.7% increase.12 For the same sample restrictions as in SBH
(age 16-29, not a high school graduate), the CPS files include 8,000-9,000 observations for the
four states with an increase in the minimum wage and approximately 65,000 in the 39 states
with no increase.
Table 4 shows the employment rates in the two groups of states before and after the
minimum wage increase and the corresponding difference-in-difference estimate. In 2004, the
employment rate in DC, IL, FL, and NJ was 36.6%, while in 2006, after the increase, the
employment rate increased by 3.4 percentage points to 39.9%. In the states with no increase,
the employment rate increased 0.6 percentage points. This yields a difference-in-difference
estimate of 2.74 percentage points that is statistically significant at the 5% level. Panel B shows
the comparable information for 20-29 year olds with at least a high school education.
Employment rates in both years are very similar in the two groups of states, rising by about one
percentage point, presumably for reasons having nothing to do with any change in the
minimum wage and reflecting, instead, employment changes due to the overall state of the
economy. The difference-in-difference estimate is a miniscule −0.0007. Interpreted as an
indicator of the general state of the economy, this estimate suggests that overall conditions
11 For ease of exposition hereafter, I refer to DC as a state. 12 The increase in the other seven states ranged from $0.25 (Maine) to $0.55 (Wisconsin) with an average increase of 7.5%.
15
were quite similar in the two sets of states. Finally, Panel C combines the two sets of estimates
to compute the DIDID estimate of the impact of the minimum wage increase in DC, FL, IL, and
NJ. The DIDID estimate is 0.0281 and it is statistically significant at the 5% level or better. Thus,
this natural experiment suggests that the minimum wage increase in these states had a positive
effect on employment of young, less-educated workers.13
I also computed employment rate changes using the CPS-MORG sample to see whether
the same sample issues that affected the NY v control state comparison would arise here.
Sample sizes are about twice as large as the samples for NY and the control states used by SBH.
In this case, the DID estimates from the MORG files are essentially identical to those from the
CPS. The employment rate in the MORG is one percentage point lower than the CPS in both
years for the states with an increase in the minimum wage, while in the states with no increase
the MORG employment rate is about 0.2-0.3 percentage points lower than the CPS in both
years.14 Although the employment levels differ, the trend is identical, resulting in a DID
estimate for the employment change of .0275 with a standard error of .0154 and t-statistic of
1.79. The DIDID estimate from the MORG is lower than with the full CPS, because the MORG
files show a more positive employment rate change for 20-29 year olds with at least a high
school degree in the states with a minimum wage increase than in the states with no increase.
The DIDID estimate is 0.0105 but with a t-statistic barely greater than one.
13 I also estimated adjusted DID models with the same covariates used in Table 2. The estimated minimum wage DID effects for the two samples are .0206 (standard error =.013) for the age 16-29 year olds and .0070 (standard error =.009) for the more educated 20-29 year olds. Full regression results are available on request. 14 The estimated 2004 and 2006 MORG employment rates are 35.6% and 39.1% for the four states with a substantial minimum wage increase and 39.1% and 39.7% for the states with no increase.
16
III. Discussion and Conclusion
Sabia, Burkhauser, and Hansen asked “Are the Effects of Minimum Wage Increases
Always Small?” and answered emphatically “no” for the case of New York’s increase in the state
minimum wage between 2004 and 2006. The implied employment elasticities in their paper
are in the range of −0.6 to −0.8, well above the consensus estimate in the earlier minimum
wage literature of −0.1 to −0.3 (Brown, Gilroy, and Kohen 1983; Neumark and Wascher 2008)
and even further above some more recent estimates that show essentially no effects. The
employment rate changes they report are so large that any reasonable policy analyst would
have to question the wisdom of such a policy. They are also so large that labor economists
might well wonder about their accuracy.
My re-analysis of the SBH natural experiment yields results that are substantially
different than theirs. I find no evidence of a negative employment impact for young, less-
educated workers in NY following the minimum wage increase. The difference in results reflects
the different data sources used, rather than differences in method. SBH used the CPS-MORG
files, which are a one-quarter subsample of the full CPS, while I used the full CPS files. In this
case, the MORG files yield incorrect estimates of the employment rate changes in NY and in the
control states, substantially overstating the apparent impact of the minimum wage change. A
closer examination reveals very large month-to-month employment rate changes in the MORG
files, a result that is not terribly surprising in light of the small monthly sample sizes. For
example, the difference between the annual employment rates in the two data sets for NY in
2004 is fully accounted for by three outlier months, each of which is followed by a month that is
17
very close to the CPS estimate. The huge employment decline for 20-24 year olds in the MORG
data could be eliminated if 4-8 additional sample members reported a different employment
status.
I also presented evidence from another natural experiment involving an increase in
state minimum wages in Florida, Illinois, New Jersey, and the District of Columbia, all of whom
had quite large increases in their minimum wage. I compare employment rate changes in those
states to changes in the 39 states that had no increase. I find evidence of a positive
employment effect of 2.74 percentage points or 7.5%. Interestingly and perhaps reflecting the
larger sample sizes involved or just the laws of sampling variability, I find very similar results
using the MORG files.
SBH were not inherently wrong in using the MORG files and their analysis and methods
are appropriate. Rather, they were unlucky. The difference between employment rates at the
state level from the CPS and MORG is a cautionary tale for applied labor economists, especially
for analyses using a DID strategy with relatively small samples and a population mean that is
reasonably low. In that case, a small difference across data sets in the number of persons
employed can end up yielding very different estimates. It is true that previous minimum wage
employment analyses have relied on the MORG data, including Hoffman and Trace (2009) and
Allegretto, Dube, and Reich (2111). It may be prudent to revisit those studies to see whether
estimates based on the full CPS sample validate the MORG estimates. As shown in this paper, it
is an empirical issue: for the NY natural experiment the CPS and MORG estimates were quite
18
different, but for the other states with an increase in the minimum wage, the two sets of
estimates were quite similar.
My findings of employment effects that are either negligible, as in the case of New York,
or positive, as in the case of DC, FL, IL, and NJ, are largely consistent with the newer round of
minimum wage employment estimates. Dube, Lester, and Reich find no negative employment
effects comparing counties across state lines with different minimum wages and Hoffman finds
no negative effect of the 2009 federal minimum wage increase in a comparison of individuals in
states where the minimum did increase and those where the minimum did not increase
because the state minimum already exceeded the new federal standard. Belman and Wolfson’s
(Belman and Wolfson 2014) meta-analysis review similarly concludes that minimum wage
effects in the US are very small in magnitude and not statistically insignificant.
It is important to caution that the findings reported in this paper reflect the range of
minimum wage increases observed in the data. They support the idea that modest minimum
wage increases in the 10-20% range phased in over a two-year period may not be problematic
in terms of employment. But they are not informative about what the employment
consequences might be for much larger increases. At the current $7.25 level of the federal
minimum wage, a 20% increase would boost the minimum to $8.70. The recently proposed
increase to $10.10 amounts to a 39% increase and the often-discussed $15 per hour minimum
wage is far outside that range.
Finally, SBH find particularly large impacts on 20-24 year olds without a high school
degree and my estimates from the CPS, although considerably smaller than theirs and not
19
statistically significant, are also largest for that group. Unlike teens without a high school
degree, for most of the workers in this age group, their educational attainment is terminal.
Hoffman and Trace also found larger effects for workers in this age group. It may well be that
this group merits further attention in minimum wage analyses.
20
Bibliography
Abadie, Alberto, Alexis Diamond, and Jens Hainmueller. 2010. "Synthetic Control Methods for Comparative Case Studies of Aggregate Interventions: Estimating the Effect of California’s Tobacco Control Program." Journal of the American Statistical Association 105: 493-505.
Addison, John T., McKinley L. Blackburn, and Chad D. Cotti. 2013. "Minimum Wage Increases in a Recessionary Environment." Labour Economics 23: 30-39.
Allegretto, Sylvia A., Arindrajit Dube, and Michael Reich. 2111. "Do Minimum Wages really Reduce Teen Employment? Accounting for Heterogeneity and Selectivity in State Panel Data." Industrial and Labor Relations Review 50 (2): 205-239.
Belman, Dale and Paul J. Wolfson. 2014. What Does the Minimum Wage Do? Kalamazoo, MI: W.E. Upjohn Institute.
Brown, Charles, Curtis Gilroy, and Andrew Kohen. 1983. "Time-Series Evidence of the Effect of the Minimum Wage on Youth Employment and Unemployment." The Journal of Human Resources 18 (1): 3-31.
Bureau of Labor Statistics. 2013. Characteristics of Minimum Wage Workers: 2012, U.S. Department of Labor.
Card, David and John E. DiNardo. 2002. "Skill-Biased Technological Change and Rising Wage Inequality: Some Problems and Puzzles." Journal of Labor Economics 20 (4): 733-783.
Card, David and Alan B. Krueger. 1994. "Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania." American Economic Review 84 (4): 772-793.
Dube, Arindrajit, T. William Lester, and Michael Reich. 2010. "Minimum Wage Effects Across State Borders: Estimates Using Contiguous Counties." Review of Economics and Statistics 92 (4): 945-964.
Dube, Arindrajit, Suresh Naidu, and Michael Reich. 2007. "The Economic Effects of a Citywide Minimum Wage." Industrial and Labor Relations Review 60 (4): 522-543.
Employment Policies Research. 2012. "New Study: Past New York Wage Hike Caused Substantial Job Loss," http://www.epionline.org/release/o335/, accessed May 20, 2014.
Hoffman, Saul D. 2014. "Employment Effects of the 2009 Minimum Wage Increase: New Evidence from State Comparisons of Workers by Skill Level." The B.E. Journal of Economic Analysis and Policy: 1-27.
21
Hoffman, Saul D. and Diane M. Trace. 2009. "NJ and PA Once Again: What Happened to Employment When the PA–NJ Minimum Wage Differential Disappeared?" Eastern Economic Journal 35 (1): 115-128.
Holtz-Eakin, Douglas. 2013. "The Mythology of the Minimum Wage." Huffington Post, November 15, 2013, accessed May 20, 2014.
Neumark, David and WIlliam L. Wascher. 2008. Minimum Wages. Cambridge, MA: MIT Press.
Sabia, Joseph J. and Richard V. Burkhauser. 2010. "Minimum Wages and Poverty: Will a $9.50 Federal Minimum Wage Really Help the Working Poor?" Southern Economic Journal 76 (3): 592-623.
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." Industrial & Labor Relations Review 65 (2): 350-376.
22
Table 1. Sample Characteristics, NY and Control States, Workers Age 16-29, No High School Degree, by Data Set
New York Control States
CPS-MORG CPS CPS-MORG CPS
BLACK 0.249 0.246 0.135 0.136
MALE 0.531 0.532 0.522 0.526
AGE 19.199 19.165 18.649 18.626
EDUC <=9 0.272 0.257 0.260 0.244
EDUC = 10 0.285 0.290 0.328 0.333
EDUC = 11 0.339 0.339 0.353 0.361
EDUC=12 (No Degree)
0.105 0.114 0.060 0.062
Number of Observations
1905 7436 3265 12986
Note: Control states are New Hampshire, Pennsylvania, and Ohio. All estimates are population characteristics using sample weights.
23
Table 2. Employment Rate Effects of NY State Minimum Wage Increase, 2004-2006 (Standard Error in Parentheses; Sample Size in Brackets)
Group and Data Source
NY 2004 NY 2006 NH, PA, OH 2004
NH, PA, OH 2006
Diff-in-Diff Adjusted Diff-in-Diff
16-29, w/o HS degree
MORG
CPS
0.362 [989]
0.291 [916]
0.409 [1765]
0.414 [1499]
−0.076** (0.029)
−.073** (0.028)
0.336 [3854]
0.308 [3582]
0.417 [6909]
0.397 [6077]
−0.008 (0.014)
−0.018 (0.013)
16-19, w/o HS degree
MORG
CPS
0.260 [685]
0.196 [659]
0.357 [1383]
0.356 [1198]
−0.064** (0.032)
−0.072** (0.036)
0.233 [2698]
0.207 [2547]
0.366 [5406]
0.344 [4836]
−0.005 (0.015)
−0.016 (0.014)
20-24, w/o HS Degree
MORG
CPS
0.537 [176]
0.430 [148]
0.524 [224]
0.560 [170]
−0.124 (0.077)
−0.141** (0.071)
.515 [686]
0.458 [604]
0.549 [877]
0.528 [720]
−0.037 (0.038)
−0.054 (.036)
25-29, w/o HS Degree
MORG
CPS
.604 [128]
.620 [109]
.603 [158]
.671 [131]
−0.053 (0.034)
−0.070 (0.051)
0.593 [470]
0.631 [431]
0.599 [626]
0.627 [521]
0.011
(0.043)
−0.006 (.041)
20-29, >= HS Degree
MORG
CPS
0.694 [2082]
0.700 [1844]
0.759 [3422]
0.754 [3503]
0.010 (0.009)
0.005 (0.005)
0.695 [8197]
0.701 [7323]
0.755
[13612]
0.762
[13791]
−0.001 (0.009)
−0.003 (0.008)
MORG estimates from Sabia, Burkhauser, and Hansen, Table 3.
** Statistically significant at 5% level or better. * Statistically significant at 10% level or better.
24
Table 3. Difference-in-Difference-in-Difference Estimates of Employment Rate Effect of Minimum Wage Increase, NY v NH/OH/PA, 2004-2006
(Standard Error in Parentheses; Sample Size in Brackets)
Group NY NH/PA/OH
16-29, w/o HS degree
2004 0.336 [3854]
0.417 [6909]
2006 0.308 [3582]
0.397 [6077]
Difference-in-Difference -0.008 (0.014)
20-29, high school degree or more
2004 0.695 [8197]
0.755 [13612]
2006 0.701 [7323]
0.762 [13791]
Difference-in-Difference -0.0005 (0.009)
Diff-in-Diff-in-Diff (CPS) -0.0075 (0.017)
Diff-in-Diff-in-Diff (MORG) -0.086** (0.033)
Source: Current Population Survey, 2004 and 2006. MORG estimates from Sabia, Burkhauser, and Hansen, Table 3.
25
Table 4. Employment Rate Effects of State Minimum Wage Increase, 2004-2006, DC, FL, IL, and NJ and States with No Increase
(Standard Error in Parentheses; Sample Size in Brackets)
Group DC, FL, IL, NJ States with No MW Increase
A. Age 16-29, w/o HS degree
2004 0.3664 [9014]
0.3945 [66548]
2006 0.3999 [8673]
0.4006 [66534]
Difference 0.0335** (0.0073)
0.0061** (.0027)
Difference-in-Difference 0.0274** (.0078)
Adjusted Difference-in-Difference
B. Age 20-29, at least HS degree
2004 0.7416 [19017]
0.7464 [148336]
2006 0.7520 [18697]
0.7575 [131314]
Difference 0.0104** (0.0025)
0.0111** (.0016)
Difference-in-Difference -0.0007 (.0048)
C. Diff-in-Diff-in-Diff 0.0281** (0.0091)
Source: Current Population Survey, 2004 and 2006.
26
0%
10%
20%
30%
40%
50%
60%
J F M A M J J A S O N D
Em pl
oy m
en t
Ra te
Month
Fig. 1. Employment Rate by Month, Persons 16-29 without HS degree, NY, 2004, ORG and Non-ORG CPS Samples
Non-ORG
ORG
27
0%
10%
20%
30%
40%
50%
J F M A M J J A S O N D
Em pl
oy m
en t
Ra te
Month
Fig. 2. Employment Rate by Month, Persons 16-29 without HS degree, NY, 2006, ORG and Non-ORG CPS Samples
Non-ORG
ORG
28
