and Neumark 2011. In Table 4, Column 1 we replicate the coeffi cient estimates from Column 4 of Table 3, for comparison. In Column 2, we exclude respondents living in
commuting zones where the share of public employees is above the 90
th
percentile 21.6 percent and present the corresponding coeffi cient estimates. The same qualita-
tive conclusion is upheld: The estimated wage premium disappears once we control for cost of living and education. Therefore, we do not believe that our measure merely
proxies for public sector employment.
A second concern with the cost of living measure is that it might proxy for local school quality instead of local prices, because many of the areas with the highest
cost of living have poorly performing public schools for example, large cities. In response, we show that our main results hold when we also control for proxies of
school quality inputs to the respondents’ high school such as the share of teachers with higher degrees. These results are in Columns 3 and 4 of Table 4. In Column 3,
we present results for the sample of respondents with valid school input measures as a baseline note that this results in the loss of more than half the sample.
22
In Column 4, we include school input variables and fi nd the same pattern as our main result:
Controlling for cost of living and years of education erases the black- wage premium. Although we observe these measures of school quality, employers may not know the
quality of schools for workers who have relocated since high school. Therefore, in Column 5 we restrict the sample to workers who lived in the same CZ in 2006 and
age 14 since employers likely have a sense of school quality in their current location. Again, the qualitative conclusions are upheld, suggesting that our controls for cost of
living are not merely picking up school quality.
B. Alternative Controls for Cost of Living
Our preferred way to account for workers’ locations is to control directly for local costs of living but doing so requires geocoded data that are often available only
through restricted- use data licenses that are costly to obtain. For researchers estimat- ing wage gaps without such access, we examine the impact of using more widely
available controls for region and urban status. In Panel 1 of Appendix Table A2, we show baseline log wage regressions that control for quadratics in age and AFQT. Panel
2 adds a control for urban status, which reduces the estimated black wage premium substantially.
23
This is because the cost of living is higher in urban locations and black women are more likely to live in urban areas than white women. Adding region fi xed
effects without an urban indicator Panel 3 does not reduce the black wage premium relative to Panel 1. This is because black women are more likely to live in the South,
a region with lower costs of living. Including both a measure for urban status and
22. Even though the cost of living controls alone do not reduce the estimated premium in the school input re- sults shown, the main conclusion is upheld in this sample: Including controls for both cost of living and years
of education eliminates the estimates of a wage premium. Furthermore, in results not shown we fi nd that in the sample of respondents with school input information, cost-of-living controls reduce the black wage pre-
mium in every other specifi cation that is, OLS, median regressions without imputations, and median regres- sions when we only impute high potential wages for women whose spouse earns above the 90
th
percentile. 23. In the 2004 and 2006 NLSY79, respondents are considered to live in an urban location if they live in a
place with population of at least 2,500, or live in an urbanized area that is, population of 50,000 or more in “central core or city and adjacent, closely settled territory”.
Table 4 Median Regression Results from Different Samples, Impute Low Potential Wages and
Impute High Potential Wages if Spousal Earnings Above the 75
th
Percentile
1 2
3 4
5 Main
Results Table 3,
Column 4 Exclude
Commuting Zones with
High Share Public
Employment School
Input Sample
Control for School
Inputs Lived
in Same Commuting
Zone at Age 14
Panel 1 Control for age,
0.188 0.188
0.139 0.143
0.206 age
2
, AFQT, AFQT
2
0.026 0.029
0.042 0.061
0.035 Panel 2a
Controls in 1 + 0.126
0.142 0.112
0.119 0.133
housing- based cost of living
0.033 0.030
0.040 0.057
0.036 Panel 2b
Controls in 2a + –0.018
–0.005 –0.028
–0.028 –0.016
years of education
0.036 0.039
0.051 0.057
0.044 Panel 3a
Controls in 1 + 0.141
0.143 0.145
0.139 0.139
wage- based cost of living
0.030 0.031
0.042 0.058
0.038 Panel 3b
Controls in 3a + –0.006
0.001 –0.016
–0.021 –0.028
years of education
0.036 0.038
0.051 0.065
0.050 N
2,682 2,387
1,191 1,191
1,564
Notes: See notes to Table 3. In Column 2, we exclude commuting zones where the share of public employees is above the 90
th
percentile 21.6 percent. In Column 3, we preserve the individuals who have valid school input data and in Column 4 we use that sample and add controls for school inputs—that is, logenrollment,
log teachers, log guidance counselors, log library books, percentage of teachers with an MA or Ph.D., percentage of teachers that left during the year, and average teacher salary. In Column 5 we only
preserve those individuals who live in the same commuting zone CZ that they lived in at age 14.
region fi xed effects results in estimates of the black wage differential that are very similar to the corresponding results from Table 3 that control directly for local costs of
living.
24
We conclude that an indicator for urban status performs well as an alternative to direct cost of living controls in our context.
An alternative way of incorporating costs of living into a wage equation when geo- codes are available is to include fi xed effects that describe detailed locations. For
example, Black et al. 2012 estimate wage regressions with a separate intercept for each location defi ned as a Metropolitan Statistical Area or the balance of the state in
nonmetro areas. When we include a separate intercept for each commuting zone CZ in Panel 7, we obtain similar results to our direct controls for cost of living. Neverthe-
less, we believe that direct controls for cost of living are preferable when detailed geocode information is available. Because fi xed effects partial out all of the variation
across cities for example, Miami and Boston, and not just local prices, estimates of black- white wage gaps may be compromised if selective migration for example,
to large cities makes up an important part of productivity and wage differentials between black and white workers. No matter which alternative cost of living measure
we choose, we fi nd that once we control for years of education as well, the black wage premium is eliminated in most cases.
C. Testing for Presence of Black Wage Premium in Settings Most Likely to Show Premium