Get documents about "Stop Smoking, Your Paycheck Will Thank You! Wage Effects from
W
Description
Stop Smoking, Your Paycheck Will Thank You! Wage Effects from
Document Sample
PRELIMINARY VERSION – PLEASE DO NOT QUOTE
Stop Smoking, Your Paycheck Will Thank You!
Wage Effects from Smoking Cessation
Silke Angera,* Michael Kvasnickab
a
German Institute for Economic Research (DIW Berlin), Mohrenstr. 58, 10117 Berlin, Germany,
sanger@diw.de, Tel: +49-30-89789-526, Fax: +49-30-89789-109
b
RWI Essen, Berlin Office, Hessische Str.10, 10115 Berlin, Germany
ABSTRACT
A growing body of literature has investigated the wage penalty attached to smoking.
Little research, in contrast, has been done on the wage effects of smoking cessation.
Using survey panel data from Germany, we study the relative earnings of smokers
and former smokers over an extended period of time. Our results from pooled OLS
regressions of wages on smoking status for ever smokers (smokers, former smoker)
point to a large wage gain from smoking cessation. This gain disappears once we
control for individual fixed effects, which suggests that the apparent wage gain
attached to smoking cessation is the result of a pure selection effect. However, when
adding controls for (past) smoking duration and time elapsed since quitting,
individuals who did not smoke very long or who quit many years ago do appear to
benefit in their earnings from quitting smoking. Both a selection effect and a causal
effect of smoking cessation hence seem to underlie the average wage gains attached
to quitting in our data. The prospect of a higher paycheck for short-time smokers and
long-term quitters provides an additional incentive to smokers to quit smoking, an
argument health authorities may want to utilize in public smoking cessation
campaigns.
JEL: J31, I19, C51
Keywords: Smoking cessation, smoking wage penalty, earnings regressions
Running Title: Wage Effects from Smoking Cessation
*
Corresponding author
INTRODUCTION
A small but growing body of literature has studied the consequences of
smoking for wages. Although covering different countries and time periods, and
making use of different empirical strategies and data sets, empirical studies tend to
concur in their findings of a significant wage penalty attached to smoking (see, for
example, Levine, Gustafson, and Velenchik, 1997; van Ours, 2004; Auld, 2005). In
contrast, relatively little is known on the effects that smoking cessation has on the
wages of smokers. Lack of research in this area is surprising. Wage gains, if indeed
materializing, would provide an additional incentive to smokers to quit and a
valuable argument policy makers and health authorities could fruitfully employ in
smoking cessation campaigns.
Existing research on the smoking wage penalty provides little if any
information on the effects of smoking cessation. In particular, estimates of the
smoking wage penalty may not be viewed as the mere negative of the wage gain that
is likely to accrue on average to smokers who quit. Most existing studies on the
smoking wage penalty sub-sum former smokers and never smokers in the group of
current non-smokers, which is used as the comparison group for current smokers in
the calculation of the smoking wage penalty. This shortcoming is likely to bias
causal inference for two reasons. First, former and never smokers might
systematically differ in productivity-relevant characteristics, that is self-select into
these two smoking states. Second, smoking may have adverse wage effects that
persist even when individuals quit, as caused for instance by irreversible health
damage or irreparable career setbacks. Treating former smokers who have been
exposed to smoking and never smokers who have not nevertheless alike, blurs the
very distinction between treatment and control group and makes it difficult to
2
interpret in any meaningful sense wage differentials calculated on this basis. It is for
these two reasons that existing estimates of the smoking wage penalty are likely to be
biased.1 For the same reasons, they are also inadequate as a measure from which to
infer the wage effects of smoking cessation.
In this paper, we study the wage effects of smoking cessation. Using panel
survey data from Germany, we analyze the relative earnings of smokers and former
smokers over an extended period of time. Restricting the analysis to ever smokers
throughout, that is excluding never smokers, provides for greater homogeneity in our
estimation sample and enables us to focus exclusively on the margin of direct
relevance to our research focus, the decision of whether or not to quit and the wage
effects of this change in smoking status. Results from pooled OLS wage regressions
point to a significant wage premium for past smokers relative to smokers. Fixed
effects panel regressions of wages on smoking status show that this premium
partially stems from a selection effect. However, additional analyses that control also
for (past) smoking duration and time elapsed since quitting produce evidence for a
positive effect of smoking cessation on wages for individuals who did not smoke
very long or who have quit long ago.
Our paper contributes to the literature on smoking and earnings in various
ways. To the best of our knowledge, our study is the first in-depth analysis of the
wage effects of smoking cessation. While few studies on the smoking wage penalty
provide some tentative co-findings on wage differentials for past smokers, none of
them addresses wages effects of smoking cessation and the underlying mechanisms
per se.
1
Anger and Kvasnicka (2009) show that the smoking wage penalty is reduced by as much as a third,
if past smoking of individuals is controlled for.
3
Second, in contrast to the previous literature on smoking and earnings we
restrict the estimation sample to ever smokers only. This restriction makes
individuals in the estimation sample more homogenous by providing for some
common initial condition (smoking initiation) and allows us to focus exclusively on
the effects of smoking cessation.
Third, we provide evidence on the short-term and long-term wage effects of
smoking and smoking cessation by considering both (past) smoking duration and
time elapsed since quitting as explanatory variables in the analysis. And finally, we
explicitly control for recall error, a measurement error known to be pervasive in
retrospective data on smoking that causes attenuation bias.
SMOKING, QUITTING AND EARNINGS
Several reasons are cited in the literature why smoking may adversely affect
earnings, including reduced individual productivity of smokers due to higher rates of
absenteeism and health problems, and potential discrimination of smokers by
employers and co-workers (see, for example, the discussion in Levine et al., 1997).
All of these channels may have persistent effects on earnings. Health damage
may be irreversible, career setbacks irreparable, and missing educational investments
not to be regained. If so, then smoking lowers both current and future earnings
capacity of individuals. Former smokers will be affected by their smoking history
even after quitting, and wage penalties for smoking will show no tendency of decline
even after years of smoking cessation. However, adverse health effects may also in
part subside with time and career setbacks be compensated to some degree or
4
altogether if individuals quit smoking. If so, then quitting smoking may well entail
significant wage gains for individuals.
Smoking cessation, however, is unlikely to be random. Smoking is strongly
addictive and successful smoking cessation hence very demanding in terms of drive
and will power, attributes that are also of great importance for worker productivity
and hence wages. Empirical studies on the wage effects of smoking cessation have to
control thoroughly for this potential self-selection of more productive workers into
quitting. Otherwise, wage differentials calculated between smokers and former
smokers is likely to suffer from selection bias.
PREVIOUS LITERATURE
Empirical studies on the earnings effects of tobacco use have found
significant wage penalties attached to smoking, ranging from 2% to 24% (Levine,
Gustafson, and Velenchik, 1997; Heineck and Schwarze, 2003; van Ours, 2004;
Auld, 2005). As noted, however, most existing analyses focus exclusively on the
current smoking status of individuals and calculate the smoking wage penalty as the
average difference in wages between smokers and current non-smokers.2
The group of past smokers has been long neglected as distinct group in
empirical analyses of smoking and wages, although it is the relevant comparison
group for a smoker that considers to quit. Levine et al. (1997) show, albeit
descriptively, that workers in the U.S. who quit smoking between 1984 and 1991
exhibit higher wage growth than workers who did not change their smoking status in
2
A notable exception is the study by Lee (1999) who, however, uses only very crude earnings
information: the average earnings in the occupational category of an individual.
5
this period. More recently, Anger and Kvasnicka (2009) first point to the general
failure in the literature to control for past smoking behavior of individuals, and show
that confounding past smokers and never smokers leads to biased estimates of the
smoking wage penalty. They also provide first evidence for a smoking wage
premium for past smokers, but did not control for unobserved heterogeneity in their
analysis which is based on cross-sectional data for Germany.
Most previous studies on the smoking wage penalty are based on cross-
sectional data or use only two or three years of panel data (e.g. Heineck and
Schwarze 2003, Levine et al. 1997). More recent studies by Brune (2007) and
Braakmann (2008) use longer-running panel data from the British Household Panel
Study (BHPS). Brune (2007) shows that the smoking wage penalty found in a cross-
sectional analysis is substantially reduced (to 2%) when using panel data methods
that control for unobserved individual heterogeneity. Using OLS regressions, he
provides indirect evidence on the wage premium for past smokers, as there are no
wage differentials discernable in his data between never smokers and quitters that
have stopped smoking more than 2 years ago. However, one of the shortcomings in
his study is that the identification of past smokers is based on retrospective data
without taking into account measurement error. Braakman (2008) uses annual data
from the BHPS, and finds likewise a substantial reduction in the smoking wage
penalty, as soon as unobserved heterogeneity is controlled for. Moreover, he finds
only small returns from starting or stopping to smoke within the observation period
relative to remaining smoker or non-smoker. However, he does not fully exploit the
smoking history of individuals, and treats past smokers who stopped smoking before
the sampling period as never smokers.
6
Previous studies on smoking and wages base their analysis unexceptionally
on all individuals, i.e. current smokers, former smokers, and never smokers despite
the differences in initial conditions for these groups. However, to analyze the wages
of smokers, the group of never smokers does not seem to be the appropriate
counterfactual, as current smokers merely have the choice between remaining a
current smoker and becoming a former smoker by quitting. In order to isolate wage
effects of smoking from selection effects into smoking the sample should be
restricted to ever smokers. Likewise, the analysis of wage effects of smoking
cessation should also be based on ever smokers only.
DATA AND SUMMARY STATISTICS
We use data from the German Socio-Economic Panel (SOEP), a
representative longitudinal survey of individuals in Germany conducted annually
since 1984 (see Haisken-DeNew and Frick, 2005 for a description of the dataset). In
the 2002 wave of the SOEP, extensive information on the current and past smoking
behavior of individuals was sampled, including whether or not individuals had
smoked regularly in the past. Unlike many previous studies, we may hence
distinguish between three groups of workers in our data: current smokers, past
smokers, and individuals that have never smoked. In addition, current and past
smokers are asked at what age they had started to smoke, and past smokers are asked
for the year when they had quit smoking. Hence, the data allows us to calculate the
duration of smoking for all ever smokers as well as the quitting duration for past
7
smokers. By matching the retrospective information on smoking behavior revealed in
2002 to the information on earnings and socio-economic characteristics in earlier
waves, we end up with an overall panel dataset for the period 1984-2002.
We restrict the estimation sample to workers who have ever regularly smoked
cigarettes in their life, i.e. who are current smokers or former smokers. We choose
this restriction in order to have a more homogeneous sample of individuals, since in
contrast to never smokers, all of them have once started smoking. Given the initial
condition of tobacco consumption at any point in life, one does not have to deal with
the problem of non-random selection into smoking, which would require to model
the starting decision.
Furthermore, we select only male workers that are of German nationality,
aged 27-55, work between 10 and 60 hours a week, earn a gross hourly wage of at
least €4, and live in West Germany. These restrictions are imposed to ensure
comparability with estimation samples used in the existing literature. Our results,
however, do not hinge on these restrictions. Furthermore, we drop cases with missing
information on the relevant variables. As we need the information on the starting age
for ever smokers and additionally the stopping year for past smokers to calculate the
length of the smoking duration and the time since they have quit, we drop all ever
smokers without information on the time when they started to smoke from the
sample. We also drop past smokers for whom the calendar year in which they
stopped smoking is not available. Hence, due to more missing information of past
smokers, they are slightly underrepresented in our sample, and the descriptive
statistics cannot be generalized to the whole population. However, since we assume
that the missing information on starting age and stopping year is random, we do not
expect a selection bias here.
8
A potential problem in the retrospective data on past smoking behavior
(starting age, stopping year) may be that the data suffer from recall error, since
survey respondents may misstate the true timing of an event (heaping). Lillard, Bar,
and Wang (2008) provide an extensive overview of measurement errors in
retrospective data, differentiating between calendar heaping, age heaping, and time
heaping. They show that a mismatch between the true and a reported date causes
attenuation bias. Figure 1 displays the age when started to smoke for all ever smokers
in our sample, as reported by the respondents. For the great majority of ever smokers
the start of tobacco consumption takes place during adolescents. While most of them
report a starting age of 16 or 18, it is clearly visible that respondents who started to
smoke later than 16 or 18 tend to report an age which is a multiple of five (20, 25,
30, 35, 40). Hence, the event of starting to smoke is rounded to multiples of age of
five, although the start of tobacco consumption should naturally occur over years.
Likewise, past smokers seem to round the reported calendar year, when asked for the
year when they stopped smoking. Figure 2 shows that the event of stopping to smoke
is “heaped” to multiples of 5, and even stronger to multiples of 10, as there are clear
peaks for the calendar years 1960, 1970, 1980, 1990, and 2000.3 To solve the
problem of measurement error from heaping, we follow Lillard, Bar, and Wang
(2008) who suggest to include parametric controls in the regressions to account for
the presence and for the type of heaping. Hence, we include a set of dummy variables
which indicate whether a respondent might possibly have rounded his reported age of
starting to smoke or year of quitting to units of 5 in age, time, or calendar year in all
our regressions.
3
The year 2001 was most frequently reported as stopping year, which can presumably be explained
by the survey on smoking behavior having been conducted in 2002. This phenomenon is not
necessarily due to recall error, as many past smokers might actually have stopped “last year”. We do
not have information on how many of them restarted smoking after the interview, but it is unlikely
that stopping was a definite decision for all of them.
9
– Figure 1 about here –
– Figure 2 about here –
Our dependent variable is the log of gross hourly wages (calculated from
gross monthly earnings and actual weekly hours of work). Covariates included in all
regressions are age, age squared, two sets of indicator variables for the respectively
highest schooling and professional degree obtained, and year dummy variables.
Variables that we use in our 2SLS regressions as instruments for past smokers
include a dummy for early age of smoking initiation (< 16 years), used also by van
Ours (2004), and dummies for co-residing with at least one no-time smoker (NS),
respectively one past smoker (PS). Current smoking status (either smoker or non-
smoker, i.e. past smoker in our sample) is determined by both past and current
factors influencing respectively the probabilities of smoking initiation and
(successful) smoking cessation. Our first instrument intends to capture systematic
differences between individuals at young age that affect their probability of early
smoking initiation when still residing with parents, such as parental smoking
behavior (see, for example, Bantle and Haisken-DeNew, 2002). The co-resident
variables, in turn, are used to capture potential influences on current smoking
behavior and (successful) smoking cessation probabilities of current co-residing
household members.4 Our instruments are strong (F-tests) and valid
(overidentification test), i.e. uncorrelated with the error term in the wage equations.
Summary statistics on workers in our estimation sample are provided in Table
1. As mentioned above, the share of current and past smokers is not representative
4
Clark and Etilé (2006) provide evidence on spousal correlation in smoking behavior.
10
for the whole population, as relatively more past smokers had to be dropped from the
sample due to missing information on the year when they had stopped smoking. As is
evident, past smokers in our sample differ markedly from current smokers in
productivity-related characteristics. While there are no significant differences in
educational degrees, past smokers in the sample are on average older and have higher
professional qualifications. Current and past smokers furthermore exhibit
substantially different likelihoods of co-residing with either a no-time smoker or a
past smoker, a feature we exploit in our 2SLS regressions. Finally, note that average
hourly wages of past smokers clearly exceed those workers who still smoke.
– Table 1 about here –
ESTIMATION METHODS
In the following, we analyze the wage effects from smoking cessation using a
standard Mincer-type specification of the earnings regression which is augmented by
measures of (past) smoking behavior. Let yit be the log of individual i’s hourly wage
at time t, xit individual characteristics, and sit the smoking status of the individual.
The estimated functions are based on the typical form:
yit = xit β + sitγ + α i + uit
'
where x is a vector of individual characteristics assumed to be related to
wages, s is the dummy variable for being a past smoker, β and γ are the
corresponding parameter vectors to be estimated, αi is the individual specific effect,
and uit denotes the idiosyncratic error term.
11
We include covariates as outlined above and estimate the wage of smoking
cessation using three different models. First, as a benchmark we use a pooled OLS
model which does not take into account unobserved worker heterogeneity, as it
assumes the unobserved individual specific effect αi to be identical for all persons.
Any violation of this assumption will lead to biased estimates. In addition to a pooled
OLS model, we will hence estimate a random effects model, which assumes the
individual specific effect αi to differ across individuals but to be constant over time.
The individual effect αi is assumed to be randomly distributed across individuals and
not to be correlated with the set of explanatory variables. To allow for correlation
between αi with the explanatory variables, a fixed effects model will be estimated.
Moreover, not only the individual specific effect αi but also the idiosyncratic
error term may be correlated with the smoking variable, as earnings and smoking
behaviour may be simultaneously determined. Another problem remains the
inaccurate measurement of the smoking status due to potential recall error with
respect to the retrospective questions on starting age and stopping year. Changes in
the smoking status over time may hence be significantly affected by measurement
error. To account for potential unobservable factors that might affect both smoking
behavior and individual earnings and to deal with the problem of measurement error,
we in addition instrument past smoking status by applying two stage least squares
regressions (2SLS). OLS and 2SLS are the two most commonly used methods in the
literature on smoking and wages (cf. van Ours, 2004; Heineck and Schwarze, 2003;
Levine et al. 1997). We use the instruments for past smoking status discussed above
12
which show to be sufficiently correlated with smoking cessation, but are uncorrelated
with the error term in the wage equation.5
REGRESSION RESULTS
Table 2 contains the results of our regression analyses from the pooled OLS
model (Model 1), the random effects model (Model 2) and the fixed effects model
(Model 3). In line with previous findings by Anger and Kvasnicka (2009), the simple
POLS regression shows that past smokers experience a sizable wage premium
relative to current smokers. The coefficient is highly statistically significant and
amounts to 5.4%, a wage mark-up that reduces to 2.4% if one controls for
unobserved heterogeneity by using a RE model. In the FE model, the coefficient on
past smoking changes sign, yet misses statistical significance. All of the applied
models above have been tested against each other. The Breusch and Pagan multiplier
test revealed for all estimates superiority of the random effects over the pooled OLS
model, whereas the result of the Hausman specification test was in favor of the fixed
effects model. This implies that the wage gains from smoking cessation found in the
POLS model can be explained by a pure selection on (time-invariant) unobserved
characteristics which are correlated with the explanatory variables.
– Table 2 about here –
5
As in any application using IV methods, the appropriateness of our instruments may or may not be
questioned. However, we use these instruments as they showed not to be endogenous in our data. The
tests of overidentifying restrictions have been carried out using the Stata ado file by Baum, Schaffer,
and Stillman (2003).
13
Instrumenting past smoking status to account for potential endogeneity and
measurement error confirms these findings. Table 3 shows that smoking cessation is
associated with a wage premium in the 2SLS model (Model 1), whereas estimating
IV panel regressions leads to a much smaller coefficient on past smoking (Model 2),
a change in sign (Model 3), and insignificant results in both cases. Again, the results
in Table 3 do not support the existence of a causal wage effect of smoking cessation.
Workers who stop smoking seem to have favorable unobserved (time-invariant)
characteristics, which drive higher earnings of past smokers.
– Table 3 about here –
If former smokers earn more due to non–random selection out of smoking,
they should have the same positive unobserved characteristics even shortly before
quitting. As an alternative identification strategy we therefore investigate whether
past smokers are remunerated for being a past smoker before they stop smoking, an
indicator that should capture productive unobserved worker characteristics. We
restrict our sample to current smokers and employ future quitting as control variable,
a dummy variable for quitting within the next 3 years. Table 4 shows results from
POLS, RE, and FE models with the same control variables as used in the regressions
above. The estimates from the POLS (Model 1) reveal that past smokers earn a wage
premium even before they stop smoking. The wage premium for future quitters is
about 3% and statistically significant at the 10% level. Although the wage mark-up is
clearly smaller and less significant than the wage effect of smoking cessation for past
smokers who have already quit (Table 2), it is evident that (future) quitting captures
unobserved individual characteristics which are remunerated well in the labor
14
market. The results of the RE and FE models (Models 2 and 3) support this finding,
as the future quitting coefficient becomes very small and statistically insignificant, as
soon as unobserved worker heterogeneity is controlled for. Overall, the results above
uniformly speak against any causal wage effects from smoking cessation, as
unobserved individual characteristics seem to drive earnings differences between
current and past smokers. The apparent wage gain attached to smoking cessation
seems to be the result of a pure selection effect.
– Table 4 about here –
Durations of smoking and quitting
So far we have looked at average wage effects for all past smokers who
turned out to possess individual productivity enhancing characteristics, which are on
average superior to those of workers who do not stop smoking. However, the group
of past smokers is heterogeneous itself, and while average wage gains attached to
quitting can be attributed to a selection effect, some worker groups might benefit
from a causal wage effect of smoking cessation. The group of past smokers consists
of individuals with both long and short smoking and quitting durations. Smoking for
a short period may have smaller adverse effects on health outcomes and career
opportunities. As a result, the earnings capacity of short-time smokers might be
reduced by less than those of workers who smoke for decades. Likewise, past
smokers who quit a long time ago might have enough time to recover in terms of
health, career, and wages from any negative effects of smoking. Of course, smoking
and quitting durations may again relate to productivity relevant individual
characteristics.
15
To investigate short-term versus long-term effects from smoking behavior we
first analyze whether wage effects from smoking and smoking cessation vary by
smoking duration. As shown in Table 2, smokers have a higher average smoking
duration (23.4 years) than past smokers (17.8 years). Controlling for smoking
duration in the regression models reveals that past smokers benefit from wage gains
even when (time-invariant) unobserved worker characteristics are taken into account
(Table 5, Models 2 and 3). The fixed effects estimation, which is our preferred
model, shows a wage premium for past smokers of 15.3% which is however reduced
by every additional year an individual has smoked in the past by 1.4%. The wage
penalty attached to smoking duration for former smokers implies that even past
smokers suffer a wage penalty if they have smoked for more than 11 years. This is in
contrast to current smokers for whom the length of tobacco consumption does not
seem to affect earnings if unobserved effects are controlled for, as the main effect of
smoking duration is very small and not statistically significant (Models 2 and 3). As
this wage impact for past smokers come into effect beyond any selection, the
duration of past smoking seems to causally affect earnings. Hence, negative wage
effects from smoking are persistent and have the power to harm earnings even after
quitting.
– Table 5 about here –
Second, we analyze whether wage effects from smoking cessation vary by the
time elapsed since a past smoker has quit smoking. The average quitting duration of
past smokers is about 11.7 years (Table 2), which would give the average past
smoker some time to recover from any causal wage penalty attached to smoking.
16
According to the results in Table 6, the wage premium for past smokers increases
with quitting duration, and is of similar size and statistical significance in all of the
models. The wage premium attached to an additional year of quitting is about 0.5%,
which amounts to a wage premium for the average former smoker of almost 6%.
– Table 6 about here –
Furthermore, we analyze non-linear wage effects from the length of quitting
in depth by differentiating between past smokers with different quitting durations.
This further allows us to analyze the persistence of wage effects from smoking and
smoking cessation. Hence, we split all former smokers according to their time since
quitting, and compare their wage differentials relative to current smokers. We use
quitters who have stopped smoking less than 5 years ago as reference category,
which represent about 20% of all former smokers, and include indicators for having
quit 5 to 10 years ago (25% of former smokers), 10 to 20 years ago (37% of former
smokers), and more than 20 years ago (18% of former smokers). Table 7 shows the
according estimates from the POLS, RE, and FE models. For past smokers who quit
smoking only recently (past smoker reference group), the POLS reveals a wage
premium of 3.3% (Model 1) which however vanishes completely when unobserved
worker characteristics are controlled for (Models 2 and 3). The results from the
preferred model, the FE estimates, reveal that there is no wage premium for recent
quitters. However, as is evident from the FE model, there are causal wage effects
from smoking cessation for past smokers who quit at least 5 years ago. The wage
premium attached to smoking cessation increases with the length of the quitting
period, being as high as 6% for former smokers who quit more than 20 years ago.
17
This finding again points to persistent causal wage effects of smoking which exist
even after quitting, and become weaker after time.
– Table 7 about here –
To sum up, the results of the above regressions point to an average wage
premium for past smokers compared to current smokers, which however vanishes
when controlling for (time-invariant) unobserved worker heterogeneity. Although the
first part of the analysis suggests that the apparent wage gain attached to smoking
cessation is the result of a pure selection effect, further in-depth investigation reveals
that causal effects come into play when the full (past) smoking history is taken into
account. The differentiation between past smokers according to their smoking
duration and quitting durations allows the identification of long-term and short-term
effects. The finding in the second part of the analysis that workers who did not
smoke for very long or who quit a long time ago benefit in their earnings from
quitting smoking point to causal effects of smoking and quitting on earnings. Both a
selection effect and a causal effect of smoking cessation hence seem to underlie the
average wage gains attached to quitting in our data.
As a robustness check, we include further control variables for marital status,
occupation, blue collar worker, tenure, public sector, firm size, region, and industry
dummies in our regressions, but did not find any significant changes in the results.
Our findings are also robust to various changes in the estimation sample, such as the
expansion of the age cohort to older workers, the omission of the minimum hourly
wage restriction, and the increase in the lower threshold for weekly hours.6
6
Results are available from the authors upon request.
18
DISCUSSION AND CONCLUSION
While numerous studies have investigated the consequences of smoking for
wages, there is a lack of evidence on the earnings effects of smoking cessation.
Quantifying the wage effects of smoking cessation is not trivial, as they are unlikely
to be the mere negative of the smoking wage penalty found in the literature. Most
existing studies on the smoking wage penalty do not differentiate between former
smokers and never smokers, which is a major shortcoming because former and never
smokers might systematically differ in unobserved productivity-relevant
characteristics. Furthermore, smoking may have adverse wage effects that persist
even when individuals quit.
In this paper, we study the wage effects of smoking cessation using panel data
from Germany. In order to achieve greater homogeneity in our estimation sample we
restrict the analysis to ever smokers, and analyze the earnings of former smokers
relative to current smokers. Results from pooled OLS wage regressions point to a
significant wage premium for past smokers relative to smokers. Fixed effects panel
regressions of wages on smoking status show that this premium partially stems from
a selection effect. However, additional analyses that control also for (past) smoking
duration and time elapsed since quitting produce evidence for a positive effect of
smoking cessation on wages for individuals who did not smoke very long or who
have quit long ago. We hence conclude that both a selection effect and a causal effect
of smoking cessation seem to underlie the average wage gains attached to quitting.
19
The findings in this paper are conditional on being a male full-time worker in
the sample. If smokers are more likely to suffer health problems which hinder them
to work (full-time), or if smokers are more strongly affected by sample attrition due
to unobserved individual characteristics, the wage premium attached to smoking
cessation will be underestimated. However, we leave the correction for these
selection effects for future research. Additional future research should be directed
towards the analysis of the unobserved factors which drive the selection out of
smoking to further illuminate the selection effect of smoking cessation.
Our findings of a causal effect of smoking cessation for short-time smokers
and long-term quitters provide an additional incentive to smokers to quit.
Furthermore, the revealed wage gains from smoking cessation provide a valuable
argument policy makers and health authorities could fruitfully employ in smoking
cessation campaigns. The simple message to smokers should be to stop as soon as
possible and to quit forever.
ACKNOWLEDGEMENTS
We are grateful to seminar participants at the University of Aberdeen and at the
University of Dundee. All remaining errors are our own.
20
REFERENCES
Anger, S., Kvasnicka, M (2009) Does smoking really harm your earnings so much?
Biases in current estimatesof the smoking wage penalty, Applied Economics
Letters, forthcoming.
Auld M. C. (2005) Smoking, drinking, and income, Journal of Human Resources,
40(2), 505-518.
Bantle C., Haisken-DeNew J. P. (2002) Smoke signals: The intergenerational
transmission of smoking behavior, DIW Discussion Paper, 277.
Baum C. F., Schaffer M. E., Stillman S. (2003) Instrumental variables and GMM:
Estimation and testing, Stata Journal, 3(1), 1-31.
Braakmann, N. (2008) The smoking wage penalty in the United Kingdom:
Regression and matching evidence from the British Household Panel Survey,
University of Lüneburg Working Paper in Economics, 96.
Brune, L.F. (2007) The Smoker’s Wage Penalty Puzzle – Evidence from Britain,
ISER Working Paper, 2007-31.
Clark A. E., Etilé F. (2006) Don’t give up on me baby: Spousal correlation in
smoking behaviour. Journal of Health Economics, 25, 958-978.
Haisken-DeNew J. P., Frick J. R. (2005) Desktop Companion to the German Socio-
Economic Panel Study. German Institute for Economic Research (DIW),
Berlin.
Heineck G., Schwarze J. (2003) Substance Use and Earnings: The Case of Smokers
in Germany, IZA Discussion Paper, 743.
Lee Y. (1999) Wage effects of drinking and smoking: An analysis using Australian
twins data, University of Western Australia Working Paper, 22.
21
Levine P. B., Gustafson T. A., Velenchik A. D. (1997) More bad news for smokers?
The effects of cigarette smoking on wages, Industrial and Labor Relations
Review, 50(3), 493-509.
Lillard, D. R., Bar, H., Wang, H. (2008) A Heap of Trouble? Accounting for
Mismatch Bias in Retrospectively Reported Data, Paper presented at the 8th
International German Socio-Economic Panel User Conference, July 9-11
2008, Berlin.
Van Ours J. C. (2004) A pint a day raises a man´s pay; but smoking blows that gain
away, Journal of Health Economics, 23(5), 863-886.
22
Figure 1: Age Heaping in Retrospective Measure of Age When Started Smoking
.8
.6
Density
.4 .2
0
10 20 30 40
Ag e W he n S ta rted To S m o ke
Source: SOEP 2002
Figure 2: Calendar Year Heaping in Retrospective Measure of Year When
Stopped Smoking
.25
.2
.15
Density
.1
.05
0
19 60 1 97 0 1 98 0 19 9 0 20 00
W h en G ave U p S m oki ng , Ye ar
Source: SOEP 2002
23
Table 1: Summary Statistics on Workers Who Have Ever Smoked
Current smokers Past smokers
Dependent variable
Hourly gross wage (in €) 13.0 (5.1) 14.8 (7.1)
Controls
Age (in years) 39.7 (8.4) 45.8 (8.7)
Education (share)
No school degree 0.55 0.55
Primary 0.22 0.20
Secondary 0.20 0.24
Higher 0.02 0.01
Professional degree (share)
No Professional degree 0.11 0.07
Vocational training 0.76 0.74
University 0.14 0.18
Smoking history
Smoking duration 23.4 (8.5) 17.8 (8.7)
Quitting duration - 11.7 (7.7)
Instruments
Co-residing with (share)
No-time smoker 0.33 0.49
Past smoker 0.08 0.20
Starting age less 16 (share) 0.24 0.23
N (Total = 9,766) 6,790 2,976
SOEP 1984-2002
24
Table 2: Regression Results for the Wages of Past Smokers
POLS RE FE
Model 1 Model 2 Model 3
Past smoker 0.0542*** 0.0243** -0.00160
(0.00834) (0.0115) (0.0142)
Age 0.0489*** 0.0484*** 0.0776***
(0.00307) (0.00255) (0.00265)
Age sqr -0.0005*** -0.0005*** -0.0005***
(0.0000) (0.0000) (0.0000)
Lower secondary school 0.0760*** 0.0496** -0.00953
(0.0184) (0.0231) (0.0304)
Intermediate secondary school 0.193*** 0.129*** -0.0106
(0.0190) (0.0245) (0.0333)
Higher secondary school 0.261*** 0.182*** -0.00419
(0.0203) (0.0261) (0.0359)
Vocational training 0.0727*** 0.0558*** 0.0364**
(0.00973) (0.0123) (0.0145)
University 0.279*** 0.221*** 0.0499*
(0.0146) (0.0203) (0.0294)
Constant 0.824*** 0.849*** 0.127**
(0.0669) (0.0580) (0.0648)
Year dummies + + +
Heaping controls + + +
Observations 9,766 9,766 9,766
SOEP 1984-2002
25
Table 3: Regression Results for the Wages of Past Smokers using Instrumental
Variables
IV IV RE IV FE
Model 1 Model 2 Model 3
Past smoker 0.155*** 0.0247 -0.223
(0.0448) (0.0878) (0.142)
Age 0.0491*** 0.0482*** 0.0787***
(0.00309) (0.00256) (0.00279)
Age sqr -0.0005*** -0.0005*** -0.0005***
(0.0000) (0.0000) (0.0000)
Lower secondary school 0.0689*** 0.0561** -0.0146
(0.0188) (0.0226) (0.0310)
Intermediate secondary school 0.184*** 0.141*** -0.0128
(0.0196) (0.0242) (0.0339)
Higher secondary school 0.247*** 0.195*** -0.00884
(0.0214) (0.0258) (0.0365)
Vocational training 0.0695*** 0.0582*** 0.0356**
(0.00990) (0.0122) (0.0147)
University 0.278*** 0.231*** 0.0470
(0.0147) (0.0203) (0.0299)
Constant 0.834*** 0.841*** 0.106
(0.0675) (0.0581) (0.0672)
Year dummies + + +
Heaping controls + + +
Observations 9,766 9,766 9,766
SOEP 1984-2002
26
Table 4: Regression Results for the Wages of Future Quitters
POLS RE FE
Model 1 Model 2 Model 3
Future Quitting within next 3 yrs 0.0288* 0.0216 0.00806
(0.0163) (0.0139) (0.0152)
Age 0.0497*** 0.0485*** 0.0765***
(0.00375) (0.00323) (0.00338)
Age sqr -0.0005*** -0.0005*** -0.0005***
(0.0000) (0.0000) (0.0000)
Constant 0.815*** 0.854*** 0.218***
(0.0793) (0.0697) (0.0770)
Schooling degrees + + +
Professional degrees + + +
Year dummies + + +
Heaping controls + + +
Observations 6,790 6,790 6,790
SOEP 1984-2002
27
Table 5: Regression Results for the Wages of Smokers and Past Smokers
According to Smoking Duration
POLS RE FE
Past smoker 0.112*** 0.0919** 0.153**
(0.0354) (0.0464) (0.0739)
Past smoker x smoking duration -0.00443 -0.00648 -0.0137**
(0.00319) (0.00414) (0.00646)
Past smoker x smoking duration 4.50e-05 0.000103 0.000253*
squ
(0.0000) (0.0000) (0.0000)
Smoking duration 0.00424* -0.000271 -0.00235
(0.00245) (0.00261) (0.00319)
Smoking duration squ -0.0001** -0.0001 -0.0001
(0.0000) (0.0000) (0.0000)
Age 0.0441*** 0.0497*** 0.0814***
(0.00398) (0.00388) (0.00449)
Age sqr -0.000417*** -0.000448*** -0.000483***
(0.0000) (0.0000) (0.0000)
Schooling degrees + + +
Professional degrees + + +
Year dummies + + +
Heaping controls + + +
Constant 0.868*** 0.814*** 0.0535
(0.0721) (0.0670) (0.0785)
Observations 9,766 9,766 9,766
SOEP 1984-2002
28
Table 6: Regression Results for the Wages of Past Smokers According to
Quitting Duration
POLS RE FE
Model 1 Model 2 Model 3
Quitting duration 0.00509*** 0.00484*** 0.00448**
(0.00131) (0.00158) (0.00188)
Quitting duration squ -0.00006 -0.00005 -0.00005
(0.0000) (0.0000) (0.0000)
Age 0.0499*** 0.0506*** 0.0799***
(0.00309) (0.00260) (0.00273)
Age sqr -0.0005*** -0.0005*** -0.0005***
(0.0000) (0.0000) (0.0000)
Constant 0.811*** 0.818*** 0.0962
(0.0672) (0.0586) (0.0655)
Schooling degrees + + +
Professional degrees + + +
Year dummies + + +
Heaping controls + + +
Observations 9,766 9,766 9,766
SOEP 1984-2002
29
Table 7: Regression Results for the Wages of Past Smokers According to
Quitting Duration Periods
POLS RE FE
Model 1 Model 2 Model 3
Past smoker 0.0332*** 0.0106 -0.00725
(0.0129) (0.0123) (0.0146)
Past smoker x quitting 5 to 10 yrs 0.00328 0.0267*** 0.0258**
(0.0150) (0.0102) (0.0107)
Past smoker x quitting 10 to 20 yrs 0.0370*** 0.0401*** 0.0327**
(0.0139) (0.0115) (0.0131)
Past smoker x quitting > 20 yrs 0.0384** 0.0625*** 0.0599***
(0.0167) (0.0162) (0.0191)
Age 0.0492*** 0.0494*** 0.0786***
(0.00308) (0.00258) (0.00268)
Age sqr -0.0005*** -0.0005*** -0.0005***
(0.0000) (0.0000) (0.0000)
Constant 0.823*** 0.837*** 0.116*
(0.0671) (0.0583) (0.0651)
Schooling degrees + + +
Professional degrees + + +
Year dummies + + +
Heaping controls + + +
Observations 9,766 9,766 9,766
SOEP 1984-2002
30
