Beruflich Dokumente
Kultur Dokumente
Worked with:
Alexander Marcus and Alex Sloane
Background
In this data set we are considering the ability of a heavy drinker to hold
a job. Alcohol abuse is a serious problem that affects a number of
people in the United States. According to conventional wisdom this
abuse can also lead to problems holding a job as many customer
service jobs are directly dependent upon customer interaction and
being an alcohol abuser may make it harder to perform well at these
jobs. Thus, addressing this problem has importance because of not
only the effects on alcohol abusers but others who rely on service from
workers, which may include these alcohol abusers.
Data
The data used are from the 1988 Alcohol Supplement to the National
Health Interview Survey of the adult (age 18 and older) U.S.
population. There are a total of 9822 observations on an individual
level. The data are for men only. The data set contains information on
family characteristics, employment status, and socioeconomic
indicators. These data were linked to state-level data on alcohol and
cigarette taxes and to other pertinent labor market and health data.
The survey asks, among other things, about alcohol consumption in the
two weeks prior to the survey. The measure of heavy drinking used
here is that the respondent’s alcohol consumption falls at or above the
90th percentile of alcohol consumption, which is 18.0 ounces of
ethanol per two-week period (approximately 36 drinks in two weeks).
The survey asks about other health indicators and includes an overall
index of health status (excellent, very good, good, fair, poor),
constructed using the self-reported health status indicators.
Table 1
Estimates of the Effect of Alcohol Abuse
on Finding and Holding a Job, 1988
Regressors:
aabuse -.1490 -.0625 -.0167 -.0588 -.2542 -2.483*
+ (.100 (.1349) (.1371) (.2089) (1.261)
(.077 4)
4)
urate -.0577 -.0049** -.0050** -.0053 -.0496**
** (.0014) (.0014) ** (.0183)
(.017 (.0015)
0)
yrschool .0493 .0038** .0037** .0032* .0336*
** (.0007) (.0007 * (.0166)
(.008 ) (.0008)
6)
married .1631 . 0134* .0126* .0084 .0786
* (.0063) (.0063) (.0073) (.0833)
(.065
9)
white .3397 .0331** .0336** .0359* .3001**
** (.0076) (.0076) * (.0834)
(.064 (.0080)
3)
Q1 -.1871 -.0149** -.0147** -.0139 -.1391*
** (.0048) (.0048) ** (.0675)
(.055 (.0051)
6)
age .0050 .0004* .0004+ .0005* .0055**
+ (.0002) (.0002) (.0003) (.0025)
(.003
0)
goodhealth .1502 .0144 .0135 .0095 .0693
(.100 (.0112) (.0112) (.0121) (.1071)
6)
south .1127 .0074 .0033 -.0157 -.1478
+ (.0136) (.0138) (.0205) (.1579)
(.062
5)
alcoholsouth -.2095 -.0109 .0311 .2260 2.261+
(.164 (.1349) (.1371) (.2086) (1.326)
3)
J-test of 2.559
overidentifying (.112)
restrictions
Notes: Each row entry is the corresponding regression coefficient and,
in parentheses, its heteroskedasticity-robust standard error, estimated
using the regression method in the relevant column. The row labeled
“F-statistic testing coefficients on aabuse and aabuse*south (p-value)”
gives the F-statistic (and its p-value in parentheses) testing the
hypothesis that the coefficients on these variables are all zero. The
penultimate row presents the first-stage F-statistic testing the
hypothesis that the coefficients on the instruments are zero in the first-
stage regression, when TSLS and IV-probit are used, and the final row
presents the J-test of overidentifying restrictions.
+=sig. at 10% level
*=sig. at 5% level
**=sig. at 1% level
Methodology
We are trying to address the question whether problem drinkers have
trouble holding jobs. In this data set we can only look at the data for
those individuals who are in the labor force, meaning they are
employed are unemployed an cannot use data for those who reported
themselves as being outside of the labor force meaning those who
were not looking for a job. Since those who are out of the labor force
are not looking for a job we cannot include them in data that pertains
to whether they cannot hold a job. We will discuss this further in the
conclusion. However, to address the question of whether problem
drinkers have trouble holding jobs we used probit, IV and IV-probit
regressions while controlling for multiple variables in order to limit any
bias.
Regressions
We began by removing the data when lfstatus equals 1 because this
represents those individuals who are out of the labor force, meaning
they are not looking for jobs. We then recoded the variable lfstatus so
that 1 would now represent those who are employed and 0 would
represent those who are unemployed. We next decided on running a
probit regression with lfstatus as the dependent variable considering
that the data was now recoded to be binary, so it would be best to use
a probit model rather than a least squares regression. We first ran a
bare-bones probit model using only aabuse as the regressor, this
yielded a result in which aabuse had a coefficient of -.1490 and this
was significant at the 10% level. Meaning we initially found alcohol
abuse to decrease employment.
The next step was to add more regressors to the model in order to try
to see if the coefficient on aabuse was more significant than what we
saw in regression 1. To choose these new regressors we looked at
those variables that were either statistically significant or those
regressors whose omission would affect the coefficient on aabuse.
Thus we decided to use these regressors in the probit regression along
with aabuse to create regression 2. However, adding these additional
regressors led to a disappointing finding, which was the fact that these
made the coefficient on aabuse insignificant.
After pondering over the data set a little more we realized that there
was possible reverse causality as we were unsure whether problem
drinkers were losing jobs, or if people losing jobs were becoming
problem drinkers. Thus we decided on using IV regressions to tease out
this possible reverse causality. Regressions 3-5 are all TSLS regressions
that we used to control for possible reverse causality. We determined
livedwith and biomomordad were the best instruments after looking at
testing for relevance using the first stage F-statistic and by testing for
exogeneity using pairwise comparisons of the J-tests. Using regression
3 we determined that using two instruments was not effective; we
think this maybe due to multicolinearity issues in the data. Thus after
instrumenting for biomomordad in regression 4 and instrumenting for
livedwith in regression 5 we determined livedwith was the best
instrument.