DONALD S. KENKEL
PennsylvaniaState University
DAVID C. RIBAR
PennsylvaniaState University
Alcohol Consumptionand Young
Adults' Socioeconomic Status
IN THE PAST FIFTFEEN years the consumptionand abuseof
alcoholbecame
mattersof intense public concern. By the end of the 1980s, the federal
governmentand many state governmentshad increasedexcise taxes on
to place
alcohol. The federalgovernmentbeganto requiremanufacturers
warninglabels on alcoholic beveragecontainers;the state governments
(with some federalpressure)set their legal minimumdrinkingages at a
uniformtwenty-oneyears. This surgeof governmentactivitymay not be
over. Policy initiatives still under considerationinclude furthertax increases,requirementsfor sternerwarninglabels, mandatedinsurancecoverage of alcoholism treatment,and restrictionson alcohol advertising,
especiallyadvertisingtargetedat young drinkers.
Advocates of new policies often point to estimatesof the economic
costs of alcohol abuse;in 1990 alcoholabuseanddependenceimposedan
estimated$98.6 billion in costs on the U.S. economy.' Key components
of the economiccost resultshave been a series of controversialestimates
thatalcoholabusesubstantiallyreducesproductivityandhenceearnings.2
This paperwas presentedat the December 1993 MicroeconomicsPanel Meetingfor
the BrookingsPapers on Economic Activity. The authorsthank the discussants, Phil
Cook and Sam Peltzman,for theirhelpful comments.The authorsalso thankEric Bond,
Keith Crocker, Peter Reiss, Mark Roberts, David Shapiro, Mark Wilhelm, Clifford
Winston,and the conferenceparticipants.
1. See Rice (1993), who updatedthe estimate for 1985 by Rice and others (1990).
2. Some of the controversy stems from a study by Harwood and others (1984)
showing a negative effect of alcohol abuse on household income. Disturbingly,using
the same data from the 1979 National Alcohol Survey, Heien and Pittman(1989) were
unable to replicate these results. Rice and others (1990, pp. 193-95) revisited the
controversyand attemptedto replicate the Harwood model using the 1984 National
119
120
Brookings Papers: Microeconomics1994
In the most recent economic cost study, estimatedlost productivityof
workerssufferingfromalcoholproblems(morbidity)accountedfor $36.6
billion in 1990, morethana thirdof the total costs of $98.6 billion.3The
size of this number,and its frequentuse in policy debates,motivateus to
examinemoreclosely the socioeconomicconsequencesof alcoholuse. In
additionto the policy interest, drinkingbehaviorrepresentsa situation
where individuals'consumptionchoices intersectwith their health and
socioeconomicstatus, posing intriguingquestionsin severalareasof applied microeconomics.
This paperuses unique data on young adultsto explore how alcohol
consumptionaffects earnings,laborsupply, and marriage.Datafromthe
NationalLongitudinalSurveyof Youth(NLSY)enableus to isolateunobserved family- and individual-specificeffects not capturedin previous
studies. We focus on these effects because alcoholismresearchshows
theremay be importantfamily backgroundand personalityvariablesthat
jointly cause alcohol problems and socioeconomic status. Recent economicresearchsuggests anotherline of inquiry:alcoholconsumptionand
earningsmay be simultaneouslydetermined.To addressthis possibility,
we develop instrumentalvariablesto explain alcohol consumption.Our
primarygoal is to see whetherthe new informationwe develop changes
or improveseconomic estimatescited in policy debates.
When we estimate benchmarkmodels similar to earlierstudies, we
replicatebothsome standardfindingsandpuzzles. Forexample,estimates
from ordinaryleast squares(OLS) regressionmodels that includeonly a
standardset of explanatoryvariablesimply that men with alcohol abuse
or dependenceearn about 5 to 10 percentless thanmen withoutalcohol
problems. However, the correspondingresults for women show a 10
percentearningsincrease associatedwith alcohol abuse or dependence.
Next, we estimatea series of modelsthataddressseveralof the empirical
problemssurroundingthe relationshipbetween alcohol use and socioeconomic status. These models producea range of sometimesconflicting
results, an outcome that is not altogethersurprisinggiven the strengths
andweaknessesof the variousestimationmethods.The effects of alcohol
Alcohol Survey. They were also unableto find a statisticallysignificantnegative effect
of alcohol abuse.
3. See Rice (1993). A much earlier reference is Fisher (1926, pp. 157-61). He
estimatedthat Prohibition,if effective, would increasenationalproductivityby at least
$3.3 billion (about $27 billion when expressedin currentdollars).
Donald S. Kenkel and David C. Ribar
121
consumptionon earningsremaindifficultto pin down. The resultsessentially rule out a large negative effect of alcohol use on labor supply for
young men. The relationshipbetweenalcohol consumptionandthe labor
supplyof youngwomenis harderto characterize.In fact, theconsequences
of alcohol consumptionappearto be differentfor men and women. Finally, estimatesfromthe differentempiricalapproachesconsistentlyshow
a negativeeffect of alcohol problemson maritalstatusfor both men and
women.
The remainderof this paperis organizedas follows. After providing
some backgroundto illustratethe currentscope of alcohol problems,we
review the general frameworkused to explain how these problemsmay
affect socioeconomic status. Based on medical and economics research
findings,we then identifyseveralimportanteconometricissues thatcomplicateestimation,andwe illustratetheseproblemsusinga simplestylized
model. Alternativeremedialeconometricapproaches,and some potential
pitfalls, are also discussed. Following a descriptionof the dataand variables used in the analysis, the resultsare presented.The last two sections
of the paperdiscuss policy implicationsand conclusions.
Background
Recent national estimates show that both alcohol consumptionand
alcohol problemsare common in the United States. Roughlytwo-thirds
of the adult population in this country consume some alcohol on a
regularor at least occasional basis. While most are light or moderate
drinkers, as many as 10 percent of adults may be problem drinkers,
depending upon the exact definition used.4 The third edition of the
American Psychiatric Association's Diagnostic and Statistical Manual
of MentalDisorders (DSM-I1) definestwo categoriesof problemdrinking: alcohol abuse and alcohol dependence. Using these definitions,the
4. See USDHHS (1990, chap. 2) for a descriptionof patternsand trendsin alcohol
use and abuse. In the literaturethere is no generally agreed upon precise definitionof
the terms problem drinking or even alcoholism. Arguing against the classic disease
concept of alcoholism, Fingarette(1988) proposed calling this level of consumption
simply heavy drinking. Following convention, we use the terms problem drinking,
alcoholism, and heavy drinking as they have been used in previous research, without
trying to precisely define them. In the empiricalanalysis we precisely define our measures of alcohol abuse, dependence, and heavy drinking.
122
BrookingsPapers: Microeconomics1994
nationalprevalence estimate for 1988 is that 8.63 percentof adults, or
over 15 million Americans, have diagnoses of either alcohol abuse or
alcohol abuse is definedas a maladaptive
dependence.5In the DSM-LLL
patternof alcohol use, and it is detected for the national prevalence
estimates by self-reportsof continueddrinkingdespite a persistentsocial, occupational,psychological, or physical problemrelatedto drinking or by drinkingin situationswheredrinkingis physically hazardous.6
A diagnosis of alcohol dependence requires self-reportsof symptoms
criteria:tolerance;
that meet at least three of the following DSM-LLL
withdrawal;drinkingto relieve or avoid withdrawalsymptoms;drinking
larger amounts than intended;spending a great deal of time drinking;
continued drinkingdespite problems;neglected responsibilitiesor obligations; or impaired control. The more serious diagnosis of dependence is actually more prevalent as defined in this way (only 2.38
percent are diagnosed with abuse, compared with 6.25 percent with
dependence). Alcohol abuse and dependenceare much more prevalent
among men (13.35 percent) than women (4.36 percent). For both genders the prevalence is highest among people eighteen to thirty years
old, the age group representedin the NLSY. In this age group almost
a quarterof men and about 10 percent of women suffer from alcohol
abuse or dependence.
Linking alcohol problems and socioeconomic outcomes is a logical
extension of human capital models of the determinantsof earnings.7
Given a well-functioning labor market, more productiveworkerswill
earn more. Alcohol problems may have both short-runand long-run
effects on productivity. In the short run alcohol problems may cause
increased absenteeism and lower productivity on the job due to the
aftereffects of heavy drinking. In the long run alcohol problems may
reduce productivityand wages indirectly throughthe worker's health
capital, schooling capital, and labormarketexperience. Most previous
analyses implicitly focused on the short-runconsequences.
Early research on economic costs treated alcohol abuse akin to a
disease randomlystrikingsome portionof the population.These studies
estimated models in which currentearnings or income were specified
5. See Grantand others (1991, p. 95).
6. Self-reporteddrunkdrivingon morethanone occasion is sufficientfor a diagnosis
of alcohol abuse.
7. Some standardreferencesare Mincer(1974) and Becker (1975).
Donald S. Kenkel and David C. Ribar
123
to be a function of exogenous currentdrinking. Consider a stylized
empiricalmodel of the form
(1)
Yi = ot +
PAi + ui,
where Yirepresentsthe naturallogarithmof individuali's earnings, Ai
is a measureof alcohol problems,anduirepresentsunobservedvariation
in the determinantsof Yi. Studies usually have included other characteristics of the individual as explanatoryrighthandside variables;for
purposes of exposition, let these terms be subsumedinto the constant
term o(.
In specification 1 the coefficient on the alcohol variable gives the
percentagechange in earnings associated with a change in alcohol status. Relying on ordinaryleast squaresanalyses of equationssimilar to
equation 1, researchers have interpretednegative estimates of i as
evidence that alcohol abuse causes lower earnings. Studies of the economic costs of alcohol abuse based on the OLS methodology have
generallyconcluded thatproblemdrinkingcauses earningslosses in the
range of 10 to 20 percent.8
There is no strong consensus in the existing empirical literatureon
the statistical significance or sign of P. While the failure to find statistically significant negative effects is important,new results reportedly
show that users of alcohol and illicit drugs earn more than nonusers.9
Positive effects are not implausible:recent medical and epidemiologic
evidence suggests that moderateuse of alcohol may improvehealth, so
moderate use may also increase productivity.10 The different results
may reflect the different phenomenaexamined by researchers-moderate versus problem drinking. It has been difficult to resolve the controversybecause few data sets contain informationon both alcohol use
and labor marketoutcomes. Several of these data sets were collected
8. See Berry and Boland (1977); Harwoodand others (1984); and Rice and others
(1990). Becauseof dataavailability, the firsttwo of these studiesestimatedthe relationship betweenalcohol problemsand household income, not earnings.
9. The switching regressions model by Berger and Leigh (1988) suggested that
moderatedrinkinghad positive economic effects. Cook (1991) re-examinedthese data
and concludedthat Bergerand Leigh's results were robust,but he did not find evidence
that even heavy drinking has adverse effects on earnings. Kaestner(1991) reported
resultssuggestingthat users of illicit drugsearn more than nonusers.
10. For a recentreview see the editorialby Shaper(1993) and relatedarticles in the
June 1993 American Journal of Public Health.
124
Brookings Papers: Microeconomics1994
to study alcohol problems, and they contain only very limited labor
marketmeasures."
Empiricalresearchoriginally emphasizedthe productivityeffects of
problemdrinking,as evidenced by lower earningsor income. The scope
of the inquirycan be broadenedto addressa numberof relatedissues.
First, since earnings depend on the wage rate and labor supply, the
coefficient a from equation 1 is the sum of the two corresponding
elasticities. The effect of alcohol on labor supply, in turn, can be
decomposed into its effect on labor force participationand its effect on
hours worked conditional on participation.Studies that have isolated
the effect of alcohol consumptionon participationhave yielded mixed
results.'2 Second, there is evidence that problemdrinkerssuffer additional socioeconomic consequences. Several studiesfindthatalcoholics
attainabout one and a half years less schooling thando people without
alcohol problems.13 In one study alcoholics appearedto be 8 percentage
points more likely to be separated,widowed, or divorced.14Of concern
by themselves, these outcomes provide additionalchannels by which
problem drinkingmay reduce earnings. There is a growing consensus
that these indirect channels may be nearly as importantas the direct
link between problemdrinkingandearnings;indeed, the indirecteffects
may account for almost one-half of the total effect of alcoholism on
earnings.'5 While not all of the indirect effects can be explored using
the NLSY data, below we estimate the effects of alcohol problemson
labor supply and marriage,as well as on earnings.16
11. For excellent critical reviews of the empiricalliterature,see Heien and Pittman
(1989); Cook (1991); and Mullahy(1993). It is importantto recognizethata numberof
recentstudies (Mullahyand Sindelar 1989, 1991, 1993; and Rice andothers 1990) have
analyzed the Epidemiologic CatchmentArea data, but it is not known yet if the same
patternswill be found in other data sets.
12. The regression results by Benham and Benham(1982) showed no statistically
significantrelationshipbetween alcoholism and employment, while Mullahyand Sindelar (1993) found that alcohol abuse or dependencewas negatively and statistically
significantlyrelatedto full-time work propensity.
13. See Mullahyand Sindelar(1989) and Benhamand Benham(1982).
14. See Benhamand Benham(1982, p. 210).
15. See Mullahyand Sindelar(1993).
16. It is difficult to study the effects of alcohol on educationalattainmentbecause
manyof the respondentsto the NLSY hadcompletedschoolingpriorto the surveypanels
that included questions on alcohol. Using the NLSY, Cook and Moore (1993) found
evidence that alcohol problemsdecrease attainmentof postsecondaryschooling.
Donald S. Kenkel and David C. Ribar
125
Problems, Solutions, and Potential Pitfalls
In the spirit of Zvi Griliches, consider an extension of equation I in
which earningsare a function not only of alcohol consumptionbut also
of some unobservedor partially observed personal attribute,Ci, such
that
(2)
Yi =a+
A + yCi + ui,
where Yi,Ai, and ui are defined as before.'7 Consideras well a standard
model drawn from the economic and medical research literaturein
which Ci is also a determinantof alcohol consumption.Specifically, let
alcohol consumptionbe a function of observablebackgroundand contextual characteristics,Bi, the personal attribute,Ci, and other unobserved determinants,ei, such that
(3)
Ai = rrBi+ XCi + ei.
The key feature of this model is that Ci is a common determinantof
both earningsand alcohol consumption.This specificationis motivated
by the observationthat many personal and backgroundfactors associated with the development of alcohol problems plausibly have direct
effects on productivityand earnings. For example, a descriptionof the
"pre-alcoholic home" is based on aspects of childhood environment
related to subsequent adult alcoholism. 18 Studies have consistently
found that alcoholics are more likely than nonalcoholics to have been
raisedin a home with heightenedmaritalconflict, inadequateparenting,
and lack of child-parentcontact. The parents of future alcoholics are
also more likely to have been alcoholic, antisocial, or sexually deviant.
In addition, depression and antisocial personalityare so strongly associated with alcoholism that alcoholism has been considereda manifestation of these disorders ratherthan as a separatedisorder.'9 It seems
reasonablethatpeople with deficientfamily backgroundsor personality
disorders may suffer socioeconomic disadvantage independent of
whetherthey develop alcohol problems.20
17. See Griliches (1979).
18. See Zuckerand Gomberg(1986).
19. See Hesselbrock(1986, pp. 44-49).
20. Relatedto this point, Benhamand Benham(1982), BartelandTaubman(1986),
and Frankand Gertler(1991) have estimatedthat mentalillness significantlydecreases
income or earnings.
BrookingsPapers: Microeconomics 1994
126
These considerationshave clear econometricimplications. An OLS
regression of equation 2 that omits or fails to completely control for
personal attributesyields a biased estimate of the effect of drinking.
Under the assumptionthat ui and ei are uncorrelated,the relationship
between the biased estimate, OLS, and the true effect, 1, can be expressed
2
(4)
(IOLS)
=
i
+ X
2
UA
where denotes the varianceof C1and oA denotes the varianceof
o'Ai.
Continuingwith the illustrationof C, as representinga deficient family
backgroundor a personalitydisorder,we hypothesizethatthe independent effects of C1on earningsand alcohol consumptionare negative and
positive respectively (that is, y < 0 and X > 0). The uncorrectedOLS
estimate of X would thereforebe hypothesizedto be negatively biased.
Other possible interpretationsare that C, representsunmeasuredproductivity or aspects of permanentincome. In these cases hypotheses
regardingthe likely biases are less clear cut.
Models similar to equations 2 and 3 describe the effect of drinking
on the other socioeconomic outcomes consideredin this paper-hours
worked and marital status. We again hypothesize that omitting the
personal attributemeasure C1results in an estimate of 1 that is biased
towardfinding a more negative effect of drinkingon hoursworkedand
maritalstatus.2'
A numberof potentialmethodologicalremediesaddressomittedvariables bias. The most obvious remedy is to include a sufficiently rich
set of personalattributesthateither directly capturesor indirectlyproxies the effects of C1. Ourempiricalanalysis begins by examiningresults
froma series of OLS regressionsbasedon equations 1 and2 thatinclude
increasingly detailed sets of personal controls. Although potentially
biased, the OLS estimates are easily computedand comparableto previous research. Further,underplausible assumptionswe can postulate
signs for the likely biases ex ante. The practicalityof this approachis
limited by our difficulty in obtainingprecise descriptionsof personality
problems and the quality of individuals' upbringing. Nevertheless,
21. Similar to our hypothesis, Zucker and Gomberg(1986) suggested that the parentalconflict in the pre-alcoholichome may be partof the cause of the maritalinstability
associatedwith adult alcoholism.
Donald S. Kenkel and David C. Ribar
127
comparisonsof models that do and do not incorporatevariouspersonal
characteristicsmay provide a rough indicationof the potentialseverity
of the bias.
The second, less direct, approachfollowed below comparessiblings.
If individuals' unobservedproductivecharacteristicsare entirely family-specific and identical across brothersand sisters, differencingequation 2 across siblings eliminates these factors and permits unbiased
estimationof the effect of alcohol on socioeconomic status. This strategy is theoretically attractiveand has clear relevance as a solution to
controlling for upbringing in a "pre-alcoholic home." However, its
usefulness is sensitive to-the assumption of identical effects across
family members. To the extent that siblings' upbringingand other relevant personal characteristicsare not entirely identical, biases will remain. Moreover, if family members are more alike in unobserveddeterminantsthat affect only alcohol consumptionthan in determinants
thataffect both alcohol consumptionandearnings,siblings' differences
exacerbatethe bias. Siblings' differences may also worsen biases associated with other econometric problems such as measurementerror
or endogeneity in A .22
A related corrective strategy is to exploit the longitudinalnatureof
the NLSY and estimate individualfixed-effects models of the effect of
drinking on socioeconomic status. As with the sibling difference
models, longitudinalfixed-effects models offer the promiseof removing
omitted variables bias associated with unobservedfamily-specific effects. In addition, the longitudinal models will remove bias from any
time-invariantperson-specificeffects. Unfortunately,caveatsmust still
be invoked. Longitudinaldifferencingwill not necessarilyremovefamily-specific effects, if these effects varyover time. Also, in the presence
of measurementerror or endogeneity, longitudinal fixed-effects estimators may again exacerbate, rather than reduce, statistical biases.
Finally, thereis a practicalproblem:only a subsetof the dataon alcohol
consumption and problem drinking is available longitudinally in the
NLSY.
As the precedingdiscussion suggests, the econometriccomplexities
in examiningthe socioeconomic consequencesof alcohol consumption
22. See Griliches (1979, pp. S40-S44) for a more complete discussion of these
issues.
128
BrookingsPapers: Microeconomics1994
are not necessarily limited to omitted variables. Considera case where
there is reciprocalcausality between drinkingand earnings. In particular, rewrite equation 3 as
(5)
Ai = 7rBi +
Yi + e.
For purposes of discussion, let us ignore problems associated with C,
and focus on the issues raised only by the simultaneityof YiandAi.
While equation 5 may appearto be ad hoc, it is in fact motivatedby
recentempirical and theoreticalstudies by economists. Simply appealing to demandtheory, it would appearthatoverall alcohol consumption
is likely to depend on income and earnings. The budgetconstraintmay
also influenceheavierdrinking,even if it is thoughtthatheavy drinkers
have little control over their behavior. Beyond this, the model of rational addiction suggests that it may be useful to view alcoholism as
subject to standardeconomic forces.23
Estimates of the effect of alcohol on socioeconomic outcomes that
neglect simultaneity are biased if the true model is as specified in
equations2 and 5. Assuming that y = 0 and ui and ei are uncorrelated,
the relationship between the naive estimate, IOLS, and the true X is
given by
(6)
(13OLS)
=
1
6
+
_13
2~
uv2
where U2ui.
Underthe assumptionthatthe true
denotes the varianceof
effect of alcohol consumptionon earningsis negative and that alcohol
consumption is a normal good (8 > 0), we would hypothesize that
simultaneity results in an upward bias in 1OLS and a corresponding
understatementof the socioeconomic consequences of drinking. As23. For empirical evidence that heavy drinkingand problem drinkingrespond to
standardeconomic forces, see Cook (198 1); Cook and Tauchen(1982); Kenkel (1991a,
1993); and Chaloupka,Saffer, and Grossman(1993). Grossman(1993) reviewed the
theoreticaland empirical issues regardingthe applicationof the theory of rationaladdiction to alcohol use. A complete rationaladdictionmodel of alcohol use and abuse
assumes that individualsbalance the relative costs and benefitsin deciding whetheror
not to consumealcohol. Currentandfuturesocioeconomicconsequencescan be included
as costs of consumption,providingan additionalreason why earningsmay be a determinantof drinkingbehavior.
Donald S. Kenkel and David C. Ribar
129
sumptionsdrawn from more intricaterationaladdiction models result
in more complicated predictions.
Models of the simultaneousdeterminationof alcohol use and hours
worked, and alcohol use and maritalstatus, are similar to the earnings
model describedby equations2 and5. While the hypothesizeddirection
of simultaneitybias in the model of hours workedis the same as in the
earnings model, income effects are less of an issue in explaining the
relationshipbetween alcohol consumptionand maritalstatus. Instead,
marital status may be a determinantof drinking, in addition to the
possibility that drinking reduces the probabilityof being (or staying)
married.24For this case our assumptionis that simultaneityresults in a
downwardbias, so thatsimple models overestimatethe negativeeffects
of drinkingon maritalstatus.
Anothereconometric difficulty arises if alcohol consumptionis imperfectly measured.This paperand most previous studies rely on selfreporteddrinkingbehaviorand drinkingproblems.The NLSY drinking
data used below come from thirty-dayretrospectivereportsand hence
are subject to recall error. The different periodicities of the drinking
data (monthly) and outcomes data (annual) also increase the chances
of measurementerror. In a standardOLS regression the presence of
randommeasurementerrorbiases the estimated effect of drinkingon
socioeconomic status toward zero. Such bias becomes an even more
severe problem in regressions that use sibling or longitudinaldifferences to control for fixed effects.25
One way of viewing all of the econometricproblemsdescribedabove
is that the measure of alcohol consumption in equations 1 and 2 is
effectively correlatedwith the errorterm, ui. Subjectto the availability
of appropriateidentifying measures, instrumentalvariables (IV) techniques can be used to obtain consistent estimates of the consequences
of drinking. This paper employs an IV methodology that has been
successfully applied in previous studies of the effects of youth problem
behavior on subsequent socioeconomic status.26Estimationof the si24. Miller-Tutzauer,Leonard, and Windle (1991) found evidence that entry into
marriageis associatedwith reductionsin drinkingand heavy drinking.
25. See Griliches (1979) and Griliches and Hausman(1986).
26. See Kaestner(1991); Mullahy and Sindelar (1992); Cook and Moore (1993);
and Ribar(1993).
130
Brookings Papers: Microeconomics1994
multaneousmodels is complicatedbecause several of our drinkingand
outcome measuresare qualitativeor limited dependentvariables. Consequently, we adopt a three-stageminimumdistance estimationprocedure.27
To sum up, each of the aforementionedproblems-omitted variables, simultaneity, and imperfectly measuredconsumptiondata-is a
relevantconcern. Therefore,we employ severalalternativeeconometric
techniques. The promise of sibling and longitudinaldifference models
lies in the possibility of eliminating, or at least reducing,bias associated
with omitted family- and individual-specificeffects. Unfortunately,to
the extent that simultaneity is a problem, the net effect may be that
differencing removes one bias just to uncover (or, worse yet, exacerbate) another. But if family- or individual-specificeffects account for
a large partof the unobservedvariationin socioeconomic status, sibling
and longitudinal comparisons may actually reduce biases associated
with both omitted variables and simultaneity.28Use of IV techniques
offers the promiseof consistent estimatesof the consequencesof drinking, but serious efficiency problems may arise if the instrumentshave
only modest explanatorypower.
Data
The primarydataused in this analysis come fromthe 1979-90 panels
of the National Longitudinal Survey of Youth. The NLSY contains
detailedeconomic anddemographicinformationfor 12,686 individuals
who were fourteen to twenty-one years old in 1979. Annual data on
work behavior, incomes, and maritalstatus are available for each individual in the survey. Retentionthroughthe 1990 panel is roughly 90
percent. Another important though often overlooked feature of the
NLSY is that interviews were conducted with all of the fourteen- to
27. The three-stageprocedureis similar to the methodologyoutlinedby Amemiya
(1978). The models are estimatedusing version 2.0 of the MECOSAprogramsystem;
see Schepers and Arminger(1992). Structuralestimatorsbased directlyon Amemiya's
procedurewere implemented by Mullahy and Sindelar (1992) in their study of the
employmentconsequencesof problemdrinking.
28. In particular,longitudinal fixed-effects estimatorsshould reduce simultaneity
bias implied by a rationaladdiction model of the joint determinationof alcohol abuse
and earnings.
Donald S. Kenkel and David C. Ribar
131
twenty-one-year-oldsliving in each of the surveyedhouseholdsin 1979.
Household and relationshipidentifiers from these data can be used to
form reasonably large data sets of siblings. Our analysis focuses on
same-gendersibling pairs; for each sex, the raw NLSY data contain
approximately 900 pairs.29
The NLSY also containsdetailed informationthatallows us to create
four alternativemeasures of alcohol consumptionand problemdrinking. Problem drinking is an inherently difficult concept to measure
empirically, and previous researchhas demonstratedthatthe estimated
impacts of alcohol consumption on earnings can be sensitive to the
alcohol measure used.30 The analysis in this paper concentrates on
drinkingbehaviorand socioeconomic outcomes in 1989, the only panel
in the NLSY that included questions correspondingto the DSM-III
diagnosticcriteriafor alcohol abuse anddependence.Ouranalysisuses
self-reportedanswers to these questions to form dummy variables indicatingwhetherindividualswere dependentuponor abusingalcohol.3'
These diagnostic measuresshould be useful indicatorsof two different
manifestations of problem drinking, but they do not measure other
potentially importantdifferences in alcohol consumption. To capture
these differences, we create two additionalmeasuresusing information
on individuals' drinkinghabits in the precedingthirtydays. The survey
containscategorical data on the numberof times duringthe past month
that individuals consumed six or more drinks on a single occasion
(possible responses are zero, one, two to three, four to five, six to
seven, eight to nine, or ten or more times). This categoricalvariableis
29. In households with three or more siblings of the same sex, our analysis selects
two siblings at random.
30. In an extremely useful illustrativeexercise, Sindelar(1993, pp. 212-18) estimatedthe effects of alcohol consumptionon income using ten alternativemeasuresof
alcohol use available in datafrom the 1988 NationalHealthInterviewSurvey. As might
be suspectedfrom the conflicting results of previous studies, the estimatedcoefficients
on the alcohol measuresvaried not only in magnitudebut in sign. As Sindelar(1993, p.
216) pointedout, "The conclusions one could reachon the effect of alcohol on income
do dependon the measureof alcohol used."
31. Answers to a similar set of questions were used by Grantand others (1991) to
estimatenationalprevalencerates and by Mullahyand Sindelar(1992) to examine the
economic consequences of problem drinking. As with the analyses by Mullahy and
Sindelar, our definitions of abuse and dependencedo not use informationspecifically
relatedto work/alcoholproblems. Data on alcohol-relatedemploymentproblemscould
introducespuriouscorrelationinto the analysis of drinkingand socioeconomicstatus.
132
Brookings Papers: Microeconomics1994
used as our measureof heavy drinking, and it will reflect consumption
that is potentially problematic whether or not the individual reports
alcohol problems.32 A continuousvariablerecordingthe numberof days
on which people had at least one drinkis used as a general measureof
alcohol consumption. Particularlywhen used in combinationwith one
of the otheralcohol measures, this measurehelps us separatethe effects
of moderate and problem drinking. Finally, because a young adult's
self-reportedalcohol use may not be reliable when others(for example,
parents)are present, we incorporatea variableindicatingthe presence
of people otherthanthe interviewerandrespondentduringthe interview
as a control for accuracy.33
The means and standarddeviations of the alcohol measuresand the
other variables used in the empirical analysis are reportedin appendix
A. The NLSY sample means for alcohol dependenceandabuseare very
close to previous nationalprevalenceestimatesfor the same age group:
20 percentof the men and 8 percent of the women are diagnosed with
alcohol dependence, while 15 percentof the men and 5 percentof the
women are diagnosed as alcohol abusers.34To show how the four alcohol measures compare, table I provides means for heavy drinking
and the numberof days of drinking by gender and problem drinking
status. As should be expected, the subsamples diagnosed as alcohol
dependentor alcohol abusersengage in more heavy drinkingand more
frequentdrinking. For example, men diagnosed as alcohol abusersengaged in almost four times as much heavy drinkingas nonabusers.Part
of the differencereflectsabstainers,buteven comparedwith nonabusers
who drankin the past month, abusers engaged in more than twice as
much heavy drinking. In terms of the amount of heavy drinking and
frequencyof drinking,respondentsdiagnosedas dependentappearsim32. When used as an explanatoryvariablein the OLS and fixed-effectsmodels, the
categoricalheavy drinkingmeasure is convertedto a continuousequivalentusing the
midpointsof the variable's scale. In the simultaneousequationsestimates, heavy drinking is modeled as an orderedcategoricalvariablewith known thresholds.
33. In an analysis of the 1984 and 1988 waves of the NLSY, Hoyt and Chaloupka
(1993) found that the presence of a parent is associated with statistically significant
reductionsin reporteduse and frequencyof alcohol, while the presence of a friend is
associatedwith higher reporteduse.
34. See Grantand others (1991). It should be noted that alcohol abuse and alcohol
dependenceare not mutuallyexclusive categoriesandthatthereis no necessaryordering.
That is, it is possible to be diagnosed as dependentbut not an abuser, or as an abuser
who is not dependent.
Donald S. Kenkel and David C. Ribar
133
Table 1. Mean Alcohol Consumptionby Gender and ProblemDrinking Status
Men
Category
N
Women
Heavy Days drinking
N
drinking per month
Heavy Days drinking
drinking per month
Full sample
4,604
1.98
(3.08)
6.35
(7.86)
4,910
0.63
(1.74)
2.79
(5.00)
Nonabusers
3,924
1.40
(2.56)
5.17
(6.98)
4,640
0.44
(1.38)
2.39
(4.47)
Nonabuserswho
2,616
drankin past month
2.10
(2.88)
7.75
(7.28)
2,335
0.88
(1.85)
4.75
(5.34)
680
5.36
(3.63)
13.18
(9.09)
270
3.77
(3.41)
9.61
(7.89)
3,689
1.16
(2.25)
4.64
(6.46)
4,499
0.36
(1.20)
2.14
(4.06)
Nondependentswho 2,318
drankin past month
1.79
(2.59)
7.19
(6.81)
2,194
0.73
(1.63)
4.38
(4.90)
915
5.32
(3.66)
13.26
(9.11)
411
3.57
(3.30)
9.91
(7.94)
Abusers
Nondependents
Dependents
Source: Figures based on 1989 data from the NLSY. Standard deviations appear in parentheses.
ilar to those diagnosed as abusive. While women engage in less heavy
drinkingand less frequentdrinkingthanmen, the relationshipsbetween
these measuresand the problemdrinkingindicatorsshow patternssimilar to the patternsfor men.
In additionto the consumptionandproblemdrinkingdata, the NLSY
contains several indicatorsof socioeconomic status. Our analysis considers three specific outcome measures-total annualindividualearnings, total annuallabor markethours, and maritalstatus. Annualearnings and laborsupply datafor calendaryear 1989 have been takenfrom
the 1990 survey; maritalstatus is recordedas of the 1989 survey date.
Observations with missing or censored outcomes data have been
droppedfrom the sample. In constructingeach of the dollar-denominated economic variables, nominal amounts have been converted to
constant 1982 dollars using the PersonalConsumptionExpendituredeflator.
A final reason for using the 1989 data is that by that year almost
every sample member had completed his or her schooling. All of our
empirical models exclude individuals who were currentlyenrolled in
134
Brookings Papers: Microeconomics1994
school. This is a less severe restrictionin 1989 thanin earlierpanels of
the NLSY.3s
Measuresof earnings, labor markethours, maritalstatus, interview
status, heavy drinking, and general alcohol consumptionsimilarto the
1989 variables can also be constructed from the 1982-85 and 1988
panels.36The paper's longitudinal fixed-effects models use outcomes
information from 1985 and 1988 along with the 1989 data. Where
possible, the robustnessof ourreportedresultshas been examinedusing
data from alternativetime periods.
Fromthe detailedinformationavailablein the NLSY, we constructed
a series of personaldescriptorsto be used as controlsfor contextualand
backgroundfactors. Year of birthis used as a controlfor age and cohort
effects. Indicatorvariables for African and Latino origin are incorporated as controls for differences in economic opportunitiesas well as
culturaldifferences in the determinantsof drinkingbehaviorand socioeconomic attainment.A variable indicating frequentattendanceat religious services as of 1979 capturesvariationin personalvalues.
Included among the productivityand humancapital controls in our
study are contemporaneous(1985, 1988, and 1989) dummyindicators
for healthproblemsthatlimited individuals'ability to work. Additional
humancapital differences are capturedby general measuresof educational attainment(dummy variables for high school and college completion) as well as by results from a standardizedintelligence test, the
ArmedForces QualificationTest (AFQT)portionof the ArmedServices
35. In 1989, 5.6 percent of the available NLSY sample were excluded from our
analysis because of their enrollmentstatus. For the longitudinalfixed-effects models,
we use datafrom as early as 1985. when roughly 11 percentof the samplewere enrolled
in school. All of the results reportedin this paper have been re-estimatedretaining
enrollees and found to be robustto the samplingrestrictions.These resultsare available
upon requestfrom the authors.
36. The 1982-85 panels also contain informationon consumptionof specific alcoholic beverages-beer, wine, and distilled spirits-in the past week. Research has
suggested that there are systematicdifferencesbetween "typical" drinkersof different
alcoholic beverages (Klein and Pittman, 1990). For example, beer drinkerstypically
drinkheavily, and wine drinkerstend to drink in moderation.Althoughdistinguishing
the type of alcoholic beverage consumed is potentiallyuseful, this same researchhas
found that drinkersin the NLSY age group mainly consume beer. This is borne out in
the NLYS datafor 1982-85. Of people who reporteddrinkingin the past week, roughly
three-quartersreporteddrinkingbeer, while only about one-quarterreporteddrinking
any wine, and somewhatunderhalf reporteddrinkingany distilled spirits.
Donald S. Kenkel and David C. Ribar
135
VocationalAptitudeBattery, which was administeredto survey participants in 198O.37
Several variables are used to account for variation in family background and upbringing. The paper includes two measures of family
structure-number of siblings and a dummy variablefor residing in a
nonintactfamily at age fourteen-as roughcontrolsfor householdeconomic circumstances, parentalsupervision, and family values. Information on parents' educationalattainmentis also used to capturevariation in family economic resources, values, and expectations;for each
parentthe education variables consist of continuousmeasuresof completed schooling and dummy indicators for missing data. A dummy
variableflagging whetherhousehold memberssubscribedto magazines
is takenas an additionalmeasureof family encouragementtowardreading and schooling and general household resources.
Retrospectivefamily alcoholismdataarealso availablein the NLSY.
Respondentsin the 1988 panel were askedto identify alcoholic relatives
andthe numberof years, if any, they residedwith those relatives. From
these data we constructedthree variables-dummies for an alcoholic
fatheror stepfather,alcoholicmotheror stepmother,or otherdistant(nonsibling, nonresident)alcoholic relative. Havingparentsor otherrelatives
who are alcoholics is taken to indicate a possible genetic or attitudinal
predispositiontowardproblemdrinking.Alcoholicparentsmayalso signal
socioeconomicdisadvantagein an individual'supbringing.
Longitudinalstate- andcounty-level economic andinstitutionalmeasureshave been collected separatelyand used to supplementthe NLSY
data. Local economic descriptorsinclude county-level measuresof the
unemploymentrate, employmentcomposition (percentageof all workers employed in manufacturing),and per capita total personalincome.
The unemployment data are available directly from the NLSY. The
other county-level economic variables have been obtained from the
Regional Economic InformationSystem of the Bureau of Economic
Analysis. The paper's policy measures include the price of beer (the
37. Roughly 94 percentof the NLSY sample completedthe standardizedtest. Individualswho were unavailablefor interview in 1980 or had been previouslyexposed to
the test account for most of the nonresponse(Center for HumanResource Research,
1990).
136
Brookings Papers: Microeconomics1994
beverage of choice in the NLSY age group) and the percentage of
individualsin the state who reside in dry counties.38
Results
The first step in the empirical analysis is to estimate models similar
to those used in earlier studies and attemptto replicate their results.
Benchmark Models
Table 2 reportsseparatelyfor men and women the estimatedeffects
of alternative measures of drinking on the three socioeconomic outcomes-earnings, laborsupply, andmarriage.In table 2 the continuous
earnings and hours outcomes are specified as naturallogarithms and
estimated using OLS. Marriageis modeled as a binary outcome and
estimated using maximumlikelihood probit.
The regressions for each drinking and outcome combinationhave
been estimated using alternative sets of control variables. The first
reportedspecification for each model uses a basic set of controls (age,
ethnicity, religiousness, local economic conditions, and health status)
that have been included in other studies. Estimates from the basic
specificationsappearin the firstrow undereach of the outcomeheadings
in table 2. Consistent with previous research, the overall results for
men point to significantconsequences from drinking.In particular,the
estimates indicate that alcohol dependencereduces men's annualearnings by 9.8 percent and that alcohol abuse reduces earnings by 6.3
percent. Each occurrenceof heavy drinkingin the basic model reduces
men's annual earnings by 1.7 percent. The impact of engaging in as
much heavy drinkingas the average alcohol dependentor abusingman
(from table 1, about 5.3 instances) is a 9.01 percentdecrease in earnings. While the measuresof problemandheavy drinkingappearto yield
very similar results, occurrencesof general alcohol consumptionhave
no noticeable effect on earnings. This is not surprising,since the measureof generalalcohol consumptionmakesless of a distinctionbetween
moderateand problemdrinking.
38. The alcohol policy variables were made available by Paul Gruenewaldof the
PreventionResearchCenter, Berkeley, California.
Donald S. Kenkel and David C. Ribar
137
In the basic model alcohol appearsto have less effect on men's hours
of work than on their earnings. While occurrencesof heavy drinking
are estimatedto reduce labor supply by a small but statistically significant amount (0.9 percent), the effects of alcohol dependence, abuse,
and general consumption are negligible. Strongernegative effects are
found for marriage.The calculated marginaleffects of alcohol dependence and abuse give a rough approximationof the magnitudeof the
relationships.39These calculationssuggest thatalcohol dependenceand
abuse reduce the likelihood of marriageby 15 and 12 percentrespectively. Occurrences of heavy drinking and general consumption are
estimatedto decrease the men's chances of being marriedby 2 and 0.6
percentrespectively.
In contrastto the seemingly reasonableresults for men, the results
for women are somewhatpuzzling and more suggestive of econometric
complications. Using the basic set of controls, we can see that alcohol
abuse and dependence each appear to increase women's earnings by
more than 10 percent. General alcohol consumptionis also estimated
to have positive earningseffects for women. The only resultto suggest
a negative effect is the coefficient on heavy drinking.Though statistically insignificant, it is nearly identical to the coefficient for men.
Similar to the men's results, alcohol consumptionhas little effect on
women's labor supply. With respect to marital status, the negative
effects of drinking for women are two to three times larger than the
effects for men.
Estimates from models that incorporatecontrols for family backgroundand ability are listed in the second and third rows under each
heading in table 2. Differences in estimates across specificationsindicate thatomitted variablesbias is presentin some of our initial models.
While most of the changes are consistent with expectations, the new
results do little to resolve the differences across gender. Specifically,
the additionof backgroundvariablesdiminishesthe estimatedearnings
effects for men and women. However, when ability variables are included, the negative earnings effects for men decrease in size, while
39. The marginaleffect of alcohol on marriageis calculatedas 4(rh),3,where 4(th)
is the standardnormaldensity functionevaluatedat the predictedlatentmarriageindex,
and 1 is the probitcoefficient on drinking.Cautionshouldbe appliedin interpretingthe
resultsince the marginaleffect in the probitmodel variesat differentpointsin the sample
(thatis, for differentgroups of individuals).
Table
Using
Alcohol
Alcohol and
Basic
Basic
2.
AddingAdding
AddingAdding
family
family
Drinking
OLS
abuse
Observations
Observations
background, controlsa
background, controlsz,
Alternative
controlscontrols
controlscontrols
specification
schooling
schooling
dependence
measurel
for
for
for
for
SetsAnalysis
ability, background'
ability, background'
of of
and
-0.041
-0.063
3,480 (0.036)
(0.038)
(0.038)
-0.066*
Explanatory
Relationship
3,480 (0.033)
(0.034)(0.034)
-0.058*
-0.092***
ln(earnings)
-0.098***
between
Variables
Alcohol
0.011
0.012
3,612 (0.024)
0.019(0.025)
(0.025)
Men
-0.011-0.016
3,612 (0.022)
ln(hours)
(0.022)(0.022)
-0.0003
Consumption
and
4,616 (0.051) (0.050)(0.049)
4,616 (0.057) (0.055)(0.055)
Marriage
-0.383***
-0.388***
-0.386***
-0.309***
-0.310***
-0.310***
Socioeconomic
0.
0.091
3,221 (0.073) (0.077)
(0.079)
132*
0.147**
3,221 (0.061) (0.064)
0.074
(0.065)
0.112*
Status,
0.120**
ln(earnings)
3,451 (0.057)
0.091(0.058)
0.085
0.068
(0.058)
Women
3,451 (0.047)
0.024
0.038 In(hours)
0.039(0.047)
(0.048)
-0.863
-0.784
4,917 (0.090) (0.087)
-0.781I
(0.070)
(0.070)
(0.086)
-0.799*
-0.857" 4,917 (0.072)
Marriage
-0.893**I-I:
I-I-I
b. a.
Days
Heavy
Basic
Basic
and
AddingAdding
AddingAdding
Measures
Measures
family
family
Significant
.Signilicant
at
forforS.Significant
drinking
at .
at
drinking
Observations
Observations
10 Regressions
background, controlsz,
background,
controlsz,
.01.05
controlscontrols past
controlscontrols
schooling
schooling
Source:
and
parents' level.
ethnicity.
level. based
level.
on
year
of
education. 1989
birth.
nonintact
for
for
ability, background'month
for
for
ability, background"
data
from
family
NLSY.
at religiousness. 3,477
0.001(0.002)(0.002)
(0.002)
0.00030.002
age
-0.005
3,478 (0.004)
(0.004)(0.004)
-0.015***
-0.017*s*
Standard
errors
fourteen,
unemployment
rate.
family
local
appear
in
per
-0.001 -0.001
-0.004
3,607 (0.001)
(0.001)(0.001)
(0.003)(0.003)
-0.0005 3,609 (0.003)
-0.008***
-0.009***
subscriptions
capita
parentheses.
to
income,
magazines
at local
age
4,608 (0.003) (0.002)(0.002)
4,610 (0.007) (0.007)(0.006)
-0.014***
-0.059***
-0.014***
-0.059***
-0.014***
-0.059***
fourteen.
manufacturing
number
of
employment,
0.011(0.004)(0.004)
-0.015
3,218 (0.004)
0.003(0.011)
3,219 (0.010)
(0.011)
-0.018*
*
0.019**s
0.013*5*
siblings,
andinterview
status,
alcoholic
and
health
relatives.
0.005
3,447 (0.003)
(0.003)
I
0.006
- (0.003)
3,448 (0.008) (0.008)
0.008
(0.008)
0.013* 0.006
0.008*1-iI
status.
0.
-0.048
-0.143
-0.049
4,913 (0.004)
144 (0.012)(0.012)
-0.046- 4,912 (0.013)
(0.004)
* (0.004)
'I
I
I
i
-0.143*I
I
140
Brookings Papers: Microeconomics 1994
the positive effects for women increase. Surprisingly, controls for background and ability have almost no effect on the estimated effects of
drinking on marriage.
Gender Differences: Additional Results
There are several reasons why estimates of the socioeconomic effects
of alcohol consumption might differ across men and women. An important explanation is that there are gender differences in drinking
behavior that are not captured by our measures of general and problem
consumption. From appendix A it is clear that women in our sample
drink less frequently than men. (Women report fewer than half the
number of days drinking reported by men.) On the days on which they
do drink, women appear to drink more moderately than men. (Twentytwo percent of women's drinking days and 31 percent of men's drinking
days involve instances of heavy consumption.) Alcohol abuse and dependence are also less common among women.
While these descriptive statistics provide evidence of substantial behavioral variation, they may not tell the entire story. For example, the
heavy drinking measure has been dichotomized at six drinks per occasion and does not describe whether even heavier drinking occurred. The
measures of alcohol abuse and dependence have been similarly dichotomized. Where it has been possible to examine the variables in detail,
we have found that women are much more likely than men to be at or
near the definitional margins.40 Unmeasured variation in the severity of
problem behavior may account for much of the difference in estimated
effects between men and women. Indeed, in table 2 and throughout
most of our empirical analysis, the variable with the most definitional
consistency across gender (heavy drinking) yields the most similar
results.41
40. Among the individualsin our sample for whom dependenceor abuse was diagnosed, men reportedmore instancesof symptomaticbehavioras well as more instances
of destructive behavior (for example, drunkdriving or hurtingthemselves or others)
thandid women.
41. More subtle differences may reinforcethe behavioraldistinctionsdrawnabove.
Alcohol researchershave found it useful to makea distinctionbetweenType 1 andType
2 alcoholism; see Cloninger(1987) for a thoroughdescription.Women predominantly
developType 1 alcoholism, which is associatedwith late onset (usuallyafterage twentyfive) and is found among individualswho are anxious, inhibited,and likely to feel guilt
over their dependence.Type 2 alcoholism, which typically involves early onset, a high
Donald S. Kenkel and David C. Ribar
141
Table 3 lists results from models that incorporatemeasuresof both
general consumptionand problem drinkingas well as all of the background, ability, and basic controls used in our previous specifications.
The resultingestimates are useful in distinguishingthe effects of moderate and excessive consumption. By providing a more complete descriptionof consumptionpatterns,the combined measuresalso reduce
some of the unmeasuredbehavioral variationacross gender. Not surprisingly, the results for men and women move closer together. For
men the earnings effects of dependence, abuse, and heavy drinking
become more negative when general consumptionis added, while the
effects of general consumption become slightly more positive. The
positive effects associated with dependence, abuse, and general consumptionfor women are smallerthanthe correspondingestimatesfrom
table 2. The earnings effects of heavy and general consumption for
women both increase in magnitude. Interestingly, the estimated net
effect of heavy drinking is almost the same for men's and women's
earnings. Similar to the earnings results, the effects of alcohol consumptionon marriagemove closer together when combined measures
are used. As in table 2 the labor supply effects for both sexes are
generally insignificant.
Beyond the differences in drinking behavior, there may be other
importantdifferences in the determinantsof socioeconomic status that
are sex specific. For example, men are more likely thanwomen to work
outside the home. Selectivity related to the labor force participation
decision may explain some of the gender difference in earnings and
hours effects; this will be especially true if alcohol consumptionhas
strong effects on individuals with marginalwork attachment.To test
for these effects, we estimatedselectivity-correctedversionsof each of
the specifications listed in tables 2 and 3. No evidence of selectivity
bias was found for the earnings equations. Selectivity was present in
the hours regressions; however, there was little difference in the correctedanduncorrectedalcohol coefficient estimates. As a consequence,
degree of risk taking, and violent and antisocial behavior, is more characteristicof
problemdrinkingby men. Given these differences, especially the persistenceand early
onset associated with Type 2 behavior, it is not unreasonablethat we might observe
moredrinking-relatedcomplicationsamongthe young men in our sample. Moreover,to
the extent thatenvironmentalfactorsare more likely to influenceType 1 alcoholism, we
shouldexpect simultaneitybias to be a more severe problemamong women.
a.
local
perSource:
fourteen.
Heavy
Days
R2/LLF Heavy
Days Alcohol
Days Alcohol
R2/LLF' Alcohol R2/LLFI Alcohol
capita
R-squared
Significant
Sifnilicant
at number
or Significant
abuse
at . of
at
Observations
Observations
Observations
log .0510
drinking
drinking
drinkingdrin1kin7g
abuse
Regressions
income,
.01
drinking
and
antd
per
level. local
based
level. siblinLs.
level.
on
likelihood
per
per
days
days
month
month
1989
alcoholic
data
function.
manufacturing
OLS
dependence
dependence Analysis
and
of
month
daYs
per
per
and
Specification
drinking
drinking
from
relatives,
Table
3.
drinking
Effects
of
NLSY.
3,475
0.233
employment,
(0.006)
(0.002)
1979
0.004*
0.233
0.003
-0.056
3,477
per
3,477
0.232
0.002
(0.002)
(0.036)
(0.039)
(0.002)
month
month
-0.083**
-0.011**
ln(earnings)
Combined
month
score Standard
on interview
errors
status,
Armed
appear
health
in
Forces 3,604
0.1020.001
Measures
-0.001
0.101
0.101-0.001
0.005 ln(hours)
Menof
0.026 3,607
(0.026)
(0.001)
(0.024)
(0.001)
(0.004)
-0.006*3,607(0.001)
status.
parentheses.
Qualification
education.
Test
Alcohol
Regressions
parents
-2720.25
(AFQT).
also
-2739.82
-2730.04
4,336(0.003)
(0.056) Marriage
(0.009) 4,340(0.003)
(0.061) 4,340 (0.003)
0.0002
-0.229**-0.007***
-0.011***
-0.327***
-0.060***
Consumption
education.
include
on
nonintact
measures
for
family
at
0.191
0.191
0.050
0.086 3,218
0.190
0.009
3,216
(0.004)
(0.067)
(0.004)
(0.077)
(0.012)
(0.004)
-0.021*3,218
R*
0.009**
ln(earnings)
0.015***
Socioeconomic
ageethnicity.
year
of
fourteen.
Status
birth.
fatily
-0.001
0.055
0.005
3,445
0.055
0.055
0.007 3,447
0.060 3,447
0.004
0.004
(0.003)
(0.061)
(0.003)
(0.003)
(0.009)
I (0.052) ln(hours)
Womeni
religiousness.
subscriptions
to
-2912.77
-2914.76
-2915.60
ttiagazines
unnieiploymicont
-0.548 Marr-iiage
-0.032
-0.097 4,696
6 (0.079)
-0.040
4,692(0.005)
-0.036
-0.637! 4,696 (0.005)
(0.094)
at
(0.014)
(0.004)
ager-ate.
1-
Donald S. Kenkel and David C. Ribar
143
we report only the uncorrected estimates. As part of the selectivity
corrections, we also estimated (but did not report) probit models of the
labor force participation decision. The estimated effects of the alcohol
measures on labor force participation are qualitatively similar to the
estimated effects on hours worked conditional on participation.
Factors associated with household structure may also account for
some of the differences in the market effects of alcohol across gender.
In particular, the labor market effects of marriage and children vary
greatly across men and women. While marriage and children are associated with positive labor supply and earnings outcomes for men, they
appear to have substantial negative effects for women.42 If we assume
that household structure is exogenous to the earnings and hours decisions but endogenous to alcohol consumption, then marital status and
the presence of children may represent an important source of omitted
variables bias in our regressions. To examine the effects that these
variables might have, we added marriage and children to the other
controls from table 2. The results from table 4 indicate that when
controls for household structure are included, there is no statistical
difference across gender in the economic effects of alcohol consumption.
Comparisons of Siblings
While the results in tables 3 and 4 are helpful in reconciling differences between men and women in our sample, they do not fully address
other econometric issues. To examine the problem of omitted personal
attributes more carefully, we make use of siblings comparisons in the
next set of results. An analysis of variance reveals that family effects
account for 50 to 60 percent of the total observed variation in the
socioeconomic outcomes and alcohol measures, suggesting that siblings
comparisons are a potentially valuable approach to reducing unobserved
heterogeneity.
Table 5 uses an estimator that is based on simple differences across
same-sex siblings to eliminate family-specific fixed effects. While the
specification is similar to those used previously, estimation of family
fixed-effects models requires a few modifications. The models include
42. Korenmanand Neumark(1991, 1992) are good examples of recent empirical
researchin this area.
144
Brookings Papers: Microeconomics 1994
Table4. OLS Analysis of Relationshipbetween Alcohol Consumptionand
SocioeconomicStatus, Controllingfor Marriage and Children
Men
Specification
Alcoholdependence
RI2
Observations
Alcohol abuse
R2
Observations
Heavydrinking
R2
Observations
Days drinkingper month
R2
Observations
Women
ln(earnings)
ln(hours)
ln(earnings)
ln(hours)
- 0.002
(0.032)
0.276
3,480
0.025
(0.022)
0.122
3,612
0.026
(0.060)
0.238
3,221
- 0.050
(0.046)
0.117
3,451
0.005
(0.035)
0.276
3,480
0.040
(0.024)
0.122
3,612
0.059
(0.072)
0.238
3,221
0.009
(0.056)
0.117
3,451
0.004
(0.004)
0.277
3,478
-0.001
(0.003)
0.122
3,609
-0.011
(0.010)
0.238
3,219
0.0001
(0.008)
0.117
3,448
0.003*
(0.002)
0.277
3,477
0.0004
(0.001)
0.122
3,607
0.003
(0.003)
0.238
3,218
- 0.001
(0.003)
0.117
3,447
Source: Regressions based on 1989 data from NLSY. Standard errors appear in parentheses. Regressions also include
measures for ethnicity, year of birth, religiousness, unemployment rate, local per capita income, local manufacturing
employment, interview status, health status, education, parents' education, nonintact family at age fourteen, family subscriptions to magazines at age fourteen, number of siblings, alcoholic relatives, and 1979 AFQT score, marital status, and
children.
* Significant at 10 level.
almost the same set of explanatorycontrol variablesused in the models
in tables 2 and 3, but measures that show no variationacross siblings
(for example, ethnicity, alcoholic relatives, and so on) cannot be included. Another modification requiredis the use of conditional logit
(insteadof probit) to estimate the determinantsof maritalstatus. Maximumlikelihood estimationof fixed-effects probitmodels yields inconsistent parameterestimates.43
The estimates from table 5 essentially confirm our earlier models
that used the full set of controls for family backgroundand personal
attributes.In principle, comparisonsof siblings have importantadvantages, but in this applicationthe alternativeapproachesof using either
explicit measures or family fixed effects to reduce omitted variables
43. See Maddala(1987) for a review of fixed-effectsmodels with qualitativedependent variables.
Donald S. Kenkel and David C. Ribar
145
bias appear to work equally well. At least in part this is due to the
detailed measures available in the NLSY.44 Taken together, the results
of the alternative approaches indicate that at least part of the negative
relationship between drinking and men's earnings reported in previous
studies is attributable to unobserved differences in family background.
Compared with the table 2 results, the results from the siblings model
suggest an even larger positive effect of alcohol problems on women's
earnings and hours. The point estimates imply that a diagnosis of alcohol dependence increases a woman's earnings by 49 percent and
increases hours worked by 36 percent; the point estimates for alcohol
abuse likewise imply large positive effects. Although there are plausibly
productive effects of moderate drinking, there is no indication in table
5 that this is the correct interpretation of the results. When entered by
itself, the measure of days drinking in the past month is not associated
with women's earnings. When the days drinking measure is entered in
combination with the problem drinking measures, the coefficient estimates on the problem drinking measures remain positive. Given the
lack of a plausible explanation for the positive results, the estimates
from table 5 suggest that simultaneity bias may be a substantial
problem.
The siblings results still show a strong negative effect of alcohol
consumption on the probability of marriage. Again using calculated
marginal effects to illustrate the magnitude, we can see that alcohol
dependence and abuse reduce the likelihood of men's marriage by 17
and 13 percent respectively, very close to the marginal effects implied
by table 2. As in table 2 the negative effects of drinking on the
likelihood of marriage for women are two to three times larger than for
men. The similarities between the results reported in tables 2 and 5
provide more evidence that omitted variables bias is not a significant
problem in the marriage results.
44. This contrastswith the results of Geronimusand Korenman(1992); they found
that siblings models yielded importantlydifferentestimates of the socioeconomic consequencesof teen childbearing.
45. The coefficients for the marriagemodels in table 5 cannotbe directlycompared
with the coefficients in table 2 because of the difference between probitand logit. For
the logit models reportedin table 5, the marginaleffect of alcohol on marriagecan be
)]2 3, where ,3 is the logit coefficienton drinking.
calculatedas exp(mh)/[l+ exp(rh
Days
Table
Heavy
Days Alcohol
Alcohol Alcohol
5.
R2/LLF R2/LLF R2/LLF R2/LLF
R2/LLF
Specification
drinking
drinking abuse
drinking
Siblings
per
dependence
per dependence
month
month
Fixed-Effects
Analysis
-0.035
-0.0830.104
-0.0840.106
-0.0040.105
0.0050.104
0.108
0.007
(0.084)
(0.090)
(0.012)
(0.098)
(0.005)
(0.005)
of
ln(earnings)
Alcohol
Men
-0.025
-0.0100.045
-0.0600.047
0.005
0.0090.044
0.0040.046
0.049
(0.009)
(0.061)
(0.071)
ln(hours)
(0.066)
(0.003)
(0.004)
Consumption
Effects
on
0.005
-148.879
-148.613
-146.108
-147.344
0.024
(0.036)
(0.315) -150.316
(0.014)
(0.016)
-0.066* (0.313)
-0.535* (0.281) Marriage
-0.675**
-0.884***
Socioeconomic
0.131
-0.003
0.137
0.128
0.149
0.012
0.149
0.004
(0.161)
(0.027)
(0.010)
(0.194)
(0.011)(0.185)
0.376*
0.493**
0.466***
ln(earnings)
Status
0.0110.050
0.001
0.368
0.0190.058
0.070
(0.123)
(0.138) 0.053
(0.022)
(0.145) 0.070
(0.008)
(0.008)
0.362*1ln(hours)
0.288**
Women
1I"I
-1.
-156.291
-158.225
-157.246
-0.033
-0.770
1.030
383
I
(0.359)
(0.402)
(0.022) -154.657
(0.076) -155.949
(0.445)
(0.024)
3"
Mcar-riage
-0.056i
-0.247:
I:-X
a.
capita
Days Heavy
Days Alcohol
R2/LLF'
R2/LLF'
inconie.
Significant
R-squarcd
Si-nificant Observations
at
or .Significant
abuse
at
local
at
drinking drinking
loe .05.10 Reeressions
drinking
Source:
.01
level.
level. based
level.
on
likelihood
manutacturing
1989
function.
per
per
month
month
data
from
eniployment,
NLSY.
-0.0180.110
-0.140
462 0.109
0.008
(0.104)
(0.015)
(0.005)
(0.006)
0.009*
interview
Standard
status.
errors
health
appear
status.
in 490
-0.043
0.003
0.0050.048
0.047
0.004
(0.011)
(0.004)
(0.075)
(0.004)
education,
parentheses.
and
1979
232 -146.668
Regressions
(0.047)
(0.018)
AFQT
0.035*
-148.180
0.018
(0.344)
(0.015)
-0.701***
-0.123*v*
also
score.
include
measures
for
347 0.132
0.005
-0.0200.138
0.340
0.016
(0.033)
(0.011)
(0.012)
(0.208)
year
oft
birth.
385 0.053
0.245
0.010
0.0040.059
0.006
(0.026)
(0.008)
(0.157)
(0.009)
religiousness.
unciiploymicent
rate.
loczl
pe-
242 -154.385
-0.019
-0.219
-0.032
-155.042
-1.136I
(0.085)
(0.479)
(0.025)
(0.024)
148
Brookings Papers: Microeconomics1994
Longitudinal Models
Table 6 reports the results of longitudinal fixed-effects analyses,
using bothone-yearandfour-yeardifferences.46 An analysisof variance
suggests thatlongitudinalcontrols, which explain from55 to 70 percent
of the total observed variation in the socioeconomic outcome and alcohol measures, also hold some promise in reducing unobservedheterogeneity. Because of data limitationsin the NLSY, these models can
only be estimatedusing the measuresof heavy drinkinganddays drinking per month. Estimationof longitudinalfixed-effects models requires
a few modificationssimilar to the modificationsnecessary for estimation of the siblings models: time invariantvariables(ethnicity, date of
birth, and so on) are not available, and conditionallogit is used for the
marriagemodels.
In the longitudinal results, positive earnings effects are found for
both men and women. While the benchmarkresults from table 2 imply
thateach occurrenceof heavy drinkingreducesearningsby 1.7 percent
for men and 1.5 percent for women, the longitudinalfixed-effects results imply that each occurrenceof heavy drinkingincreases earnings
by 1.3 percent for men and 1.5 percent for women. Again, the results
do not seem to be explained by a positive effect of moderatedrinking,
since a positive effect of heavy drinkingremainswhen both measures
of alcohol consumption are used. The one-year and four-yeardifferences show approximatelythe same picture.
It is useful to carefully compare the results reportedin table 6 with
the results from the OLS models and siblings comparisons. Consider
the estimatedeffects of alcohol consumptionon men's earnings. Starting with benchmarkresults showing substantialnegative effects, the
earlier approachesof using explicit controls and family fixed effects
yielded estimates which, while still negative, were muchcloser to zero.
In contrast, using individual fixed effects yields positive and statistically significant estimates. This suggests two plausible (and not mutually exclusive) interpretations.First, as a more complete control for
individual heterogeneity, the longitudinal fixed-effects models effectively eliminate omitted variablesbias but uncovera large positive bias
from contemporaneousbudget constraintsimultaneity. Alternatively,
46. Results similar to those reportedin table 5 obtain when we use meandifference
andrandom-growthlongitudinalfixed-effectsestimators.Resultsavailableuponrequest.
Donald S. Kenkel and David C. Ribar
149
the longitudinalmodels may simply exacerbatea small positive existing
simultaneitybias.
Once again, the patternsof results for women's earnings,hours, and
maritalstatus provide additionalclues. The individualfixed-effects estimatesshow smallerpositive impactsof alcohol consumptionon women's earnings and hours than did the previous methods. This is not
entirelyconsistent with the interpretationthatthe men's resultsindicate
the existence of a large positive simultaneity bias. Instead, a small
simultaneitybias may exist, or there may also be measurementerror
problemscreating a bias towardzero. The marriageresults continue to
be somewhat different and show that alcohol consumptionis still associated with a lower probability of being married. This is entirely
consistent with the earlier interpretationthat individual heterogeneity
was not an importantsource of bias in the marriagemodels.
Instrumental Variables Models
Table 7 reports results from the last approachused, instrumental
variables analysis. In theory, this approachprovides a solution to the
problems of omitted variables, simultaneity, and measurementerror.
Practical issues arise, however, in finding instrumentsthat are sufficiently strong predictorsof alcohol consumptionbut convincingly not
directly related to socioeconomic status. For the models reportedin
table7, the percentageof the state's populationresidingin drycounties,
the averagebeer price, parents'alcoholism, and otherdistantrelatives'
alcoholism are used as identifying variables.47Table 7 reportsthe coefficient estimates for the IV models estimated, as well as test statistics
for model fit and exogeneity. Particularlyfor the earnings models,
exogeneity of the alcohol measuresis usually rejected, providingsupport for the specification.
The IV results reported in table 7 provide the strongest indication
that alcohol problemshave importantsocioeconomic consequences. A
numberof featuresof the results are notable. First, in sharpcontrastto
the benchmarkOLS results, the results are much more similar for men
and women. In particular,the point estimates of the effects of alcohol
47. The resultsof alternativespecificationsusing all of the above IVs except parents'
alcoholismare similarto the resultsreportedin table 7. These resultsare availableupon
request.
Days
Table
Heavy
Days Heavy
6.
One-year
R2/LLP R2/LLF"
R2/LLFa
Specification
Observations
drinking
drinking
drinking
drinking
difference
per
Longitudinal
permonth
month
(1988-89)
Fixed-Effects
3,1780.007
(0.004)
(0.002) 0.006
(0.002)
(0.004) 0.005
0.003*
0.010** 0.005*** 0.013***
ln(earnings)
Analysis
of
Alcohol
0.001
-0.003
3,3980.006
0.006
(0.003)
(0.001) 0.004
(0.001)(0.003)
Men
ln(hours)
0.003**
0.004***
Consumption
Effects
-194.121
-0.011 (0.039)
315 -191.368
0.009
(0.043) (0.017) -191.472
(0.019)
-0.091**Marriageon
-0.100**
Socioeconomic
2,7190.006
0.009
0.006
(0.009)
(0.003) 0.005
(0.003)
(0.009)
0.015*
0.007*
0.008**
ln(earnings)
Status
3,0110.005
-0.003
0.005
0.003 0.003
(0.003)(0.007)
(0.003)
Women
0.006**(0.007) ln(hours)
0.006**
-0.093
315 -204.933
-205.087
-207.262
-0.036
-0.066
-0.072
(0.059)
(0.065)
(0.029)
(0.032)
Marriage
I:
a.
*'
*
local
Days
Heavy
Days Heavy
R2/LLF" R2/LLF Four-year
R2/LLF"
R-squared
Significant
Significant Observations
at
or Significant
drinking
at manufacturing
at
drinking
drinking
log .05 .10 Regressions
drinking
.01
per
per
difference
level. based
Source:
level.
level.
on
likelihood employment,
month
month
1985.
function.
(1985-89)
interview
1988.
and
status,
1989
and
2,233
data
health
from
0.0030.020
0.009
0.020
(0.005)
(0.002) 0.017
(0.002)(0.006)
R
0.006**
0.007***
status.
NLSY.
Standard
2,381
0.0010.010
-0.002
-0.004
0.002
0.010
0.010
(0.004)
(0.005)
(0.002)
(0.002)
errors
appear
in
-0.
-0.
parentheses.
0.006
946 -469.328
-471.451
(0.010)(0.028) -481.156
(0.009)
-0.016* (0.025)
124***
132***
Regressions
also
include
0.0010.004
0.001
0.0050.004
1,897 0.004
0.006
(0.014)
(0.005)
(0.005)
(0.013)
measures
for
-0.003
-0.013
0.003
0.005
2,096 0.006
(0.004)
(0.004)
(0.011)
(0.011)
0.009*
0.011P
I
unemploynient
rate,
local
per
-
capita
0.
1
1,015 -604.385
17* -609.430
(0.016)
(0.036)
(0.041)
(0.015) -612.079
-0.182I
0.059*
I
incomie.
-0.080,1-Ix
Days
Table
Heavy
7.
per
Model
ModelAlcohol ModelAlcohol
Model
fit
Specification
Exogeneity Exogeneity
Exogeneity
drinking
abuseExogeneity
month
drinking
fit4()
fit4()
fit4()
(X2)
(X2)
(x2)
(X2)
(X2)
dependence
Simultaneous
Equations
5.591
11.123
10.203
3.815
9.796
15.302
(0.044) 7.936
(0.100)
(0.110)
(0.041) 7.246
-0.107**
-0.310***
ln(earnings)
-0.301***
-0.118***
Analysis
of
Men
-0.011
-0.035
-0.042
-0.014
0.425
0.285
0.925
0.177
0.123
0.377
0.456
0.210
(0.061)
(0.018)
(0.019)
(0.050)
ln(hours)
Socioeconomic
Effects
of
2.941
0.541
0.021
-0.083
-0.073
-0.005
-0.017
2.908
3.108
0.318
0.760
3.234
(0.118)
(0.141)
(0.037)
(0.042)
Marriage
Alcohol
Consumption
-0.023
-0.109
0.0451.047
4.865
7.982
0.359
9.024
6.130
0.614
8.084
-0.282i
(0.163)
(0.120)
(0.075)
(0.032)
ln(earnings)
7.181
12.737
12.207
0.152
0.354
7.514
Women
3.719
5.089
i
(0.181) 3.834
(0.143)
(0.066)
(0.039) 10.527
0.442**'
* In(hou-s)
0.098''i
l'
I-0.
0.881
II
0.131
3.331
0.0971.513
-0.139
1.607
3.985
5.136
1.944
-0.232(0.151)
(0.061)
(0.140)
(0.054)
Marr-iatge
local
Heavy
perDays Alcohol
perDays
perDays Alcohol Model
Model
Model
capita
Significant
fit
fit
fit
Significant
Significant
at average Observations
Exogeneity
Exogeneity
Exogeneity
month
month drinking month abuse
at at .
(X2) drinking
10
(X3)
(x2) drinking
drinking
.05 beer Regressions
perSource:
counties,
.01
income,
(X)
(X2)
(X2)
dependence
level. based
level. price.
local
level.
on
parents'
1989
data
manufacturing
from
alcoholism,
-
0.931
-0.293
3,73116.603
-0.213
1.31216.687
2.915
0.077
0.274
2.047
0.15414.032
(0.287)
(1.182)
(1.412)
(0.364)
(0.163)
(0.144)
and NLSY.
employment,
other
Standard
interview
errors
nonsibling.
-
status,
appear
0.0680.454
-0.090
0.987
0.008
-0.022
0.097
0.139
0.012
0.176
0.014
in 3,8820.320
(0.076)
(0.047)
(0.257)
(0.066)
(0.182)
(0.046)
health
nonresident
status,
relatives'
parentheses.
education,
2.811
alcoholism
-0.360
-0.420
-0.142
1.917
1.785
2.033
0.109
0.070
0.114
0.654
4,3041.364
Regressions
(0.417)
(0.095)
(0.361)
(0.134)
(0.125)
(0.117)
family
are
also
used
as
include
background.
-
and
identifying
measures
1979
-0.247
0.025
-0.090
0.455
1.527
8.477
8.489
2.412
for 3,3575.630
(0.211)
(0.059)
(0.056)
(0.089)
(0.240)
(0.152)
0.446*
0.156*
0.106*
AFQT
variables.
ethnicity.
score.
year
ot
Percentage
birth,
of
state's
4.331
3,62112.968
-0.045
-0.254
-0.074
6.495
3.195
8.976
1.100
(0.215)
(0.857)
(0.051)
(0.230) 12.397
(0.076)
0.257!
(0.146)
0.504**
religiousness.
population
residing
unctiiploytiocnt
in
4,6572.263
1.444 -0.1380.0444.629
1.254
dryr-ate.
(0.090)(0.107)
0.131 0.104
0.1332.412
1,384
-0.136''
(0.106)(0.311)
(0.074)
(0.227)
154
BrookingsPapers: Microeconomics1994
abuse or heavy drinkingon men's and women's earnings are virtually
identical and suggest that alcohol problems cause about a 30 percent
drop in earnings. The results for the other alcohol measures are not
quite as similar. One anomalous patternremains: while there is little
effect of alcohol problems on men's hours, women with alcohol problems appearto work more hours.
Second, the IV models produceestimates of earningseffects that are
more consistently negative across alcohol measures;the estimates are
uniformlynegative for men. Thereis no longer any significantevidence
of positive productivity effects of moderate drinking. When entered
alone, the numberof days of drinkingin the past monthhas nearlythe
same negative effect on earnings as does the numberof days of heavy
drinkingin the men's equation. When entered in combination,the results do not provide any indications of positive effects.
Third, the relationship between alcohol consumption and marital
statusis usually negative. However, the coefficientestimatesaremostly
smaller in magnitude and less precisely estimated than were our previous results. The smaller point estimates are consistent with our hypothesis of reverse causality and a consequently reinforcingnegative
bias in the benchmarkestimates of the effects of drinkingon marital
status. However, exogeneity of the alcohol measures is formally rejected in only a handful of the marriagemodels.
Policy
Althoughestimates of negative effects of alcohol problemson earnings have attractedpolicy attention, estimates indicating no effect, or
even a positive effect, have appearedin the literature.In this paperwe
replicate and provide some methodological explanationfor this range
of results. On balance, we believe our results provide new evidence
that alcohol problems have a direct negative impact on earnings, althoughthis impact may be small. The negative effect of alcohol on the
probabilityof marriageappearsto be a more robustempiricalfinding.
While we believe our systematic investigation provides a foundation
for future consensus, it is clear that a consensus has yet to be forged.
How does evidence that problem drinkersmight suffer adverse socioeconomic consequences provide guidance for policy?
Donald S. Kenkel and David C. Ribar
155
To begin to answer this question, it is useful to consider how estimates of the socioeconomic consequences of alcohol consumptioncan
play a role in the policy process. In the past the estimates often were
used as a rhetoricaldevice to advance the argumentthat alcohol abuse
is a serious policy problem comparableto other health problems such
as cancer, heart disease, or AIDS. This use has led to the general
criticism that estimates of the total productivitylosses due to alcohol
problemsare not relevant for decisions about policies that will change
the consumptionof alcohol at the margin.48This criticismin turnpoints
to a more useful role for the earnings loss estimates, as an input when
evaluatingthe effects of specific policy changes (for example, the new
warninglabels on alcoholic beverages, advertisingbans, new alcohol
taxes, and so on). In fact, our analysis yielded additionalresults useful
for furthersteps in alcohol policy analysis. In the simultaneousequations models we estimated reduced-formequations showing alcohol
demandas a function of various exogenous influences, including alcohol prices. The estimated price coefficients from the reduced-formdemand equations (not reported)can be expressed as elasticities, taking
accountof the qualitativeand limited natureof the alcohol measures.49
The results imply that the price-elasticity of the probabilityof being
alcohol dependentis -0.45 for men and -0.41 for women. The correspondingprice-elasticities for the probabilityof abuse are - 1.2 and
-1 .6. The price-elasticityof the demandfor heavy drinkingis - 0.58
for men and -0.42 for women. The only anomalousresult is a positive
price-elasticity for the numberof days of drinkingfor women (0.29),
althoughit is negative for men (-0.24). These estimatescomparewell
to previous estimates in the literature;in particular,they supportprevious findingsthat even heavy drinkingrespondsto price.50Combined
with our estimates of earnings losses, these elasticity estimates imply
that increasing alcohol taxes could lead to increased productivityand
earnings.
Economistshave arguedthateven in the context of specific policies,
48. This pointdoes not applyto the analysisby Fisher(1926), since he was explicitly
consideringa policy, Prohibition,which if effective would eliminate all alcohol problems. However, in practice even Prohibitiondid not eliminate alcohol problems;see
Mironand Zweibel (1991).
49. See Maddala(1983).
50. Kenkel (1993) reviews estimates of price-elasticitiesfor young adults in this
range.
156
Brookings Papers. Microeconomics1994
estimatesof the lost earningsassociated with alcohol abuse offer insufficient and possibly misleading guidance. The fundamentalconceptual
problem is that the estimates fail to distinguish between costs to the
individual and costs the individual imposes on society (externalities).
The worker is the primaryloser when his or her earnings are reduced
by alcohol consumption.5'Using the standardof consumersovereignty,
even if estimates imply that increasing alcohol taxes can lead to increased earnings, it does not follow that the tax policy is justified.
However, if alcohol problemsare beyond the control of the individual,
it may make sense to view the entireearningslosses as relevantnumbers
to evaluate policy.52The earnings loss estimates also gain policy relevance to the extent that people are poorly informedabout the productivity effects of theirdrinking,in the same way they appearto be poorly
informedabout the health, safety, and legal consequences.53
Assuming that policymakers have decided that the socioeconomic
consequences of alcohol consumptionare of some relevance at a conceptual level, the practical problem remains: which set of empirical
estimates, if any, is correct? The lack of consensus on the correct
answersuggests caution in basing policy on the evidence we have now.
As futurestudiestry to remedythis situation,researchers,andthe policy
communityas eventual users of thatresearch,need to identify the most
importantempirical problems encounteredin estimatingthe effects of
alcohol problems on socioeconomic status. Our results suggest that
omission of family backgroundand personalitymeasurescreates some
bias towardfinding a negative relationshipbetween earningsand alcohol consumption. In our data family effects and individualeffects capture a significant amount of the observed variation in earnings and
51. See Manning and others (1991) for an in-depth discussion of the distinction
betweenthe internaland externalcosts of heavy drinking,smoking, and sedentarylifestyles. Heien and Pittman(1993) concludedthatearningslosses and otherinternalcosts
accountfor most (six-sevenths) of the total economic costs of alcohol abuse estimated
by Rice and others (1990).
52. In a similar vein Pogue and Sgontz (1989) arguedthat the optimaltax rate on
alcohol is much higher, perhaps as high as 306 percent of the net of tax price, if
alcoholism is considereda disease beyond the controlof the individual.
53. Kenkel (1991b) providedevidence that many consumersare unawareof at least
some of the health consequences of heavy drinking. Phelps (1987) found suggestive
evidence that young adults underestimatethe safety risks of driving after drinking.
Snortum,Berger, and Hauge (1989) found that manyAmericansare unawareof various
legal aspects of drinkingand driving.
Donald S. Kenkel and David C. Ribar
157
alcohol consumption,demonstratingthe potentialusefulnessof siblings
and longitudinaldata.54However, the more importantempiricalproblem appearsto be contemporaneoussimultaneity,wherehigherearnings
lead to more drinking.The ideal approachto this problemis a structural
dynamic model of alcohol and labor marketdecisions.
As more sophisticated econometric approachesare developed, researchersshould not lose sight of the inherentdifficulties in measurement. The researcherneeds to make the sometimes subtle distinction
between problem drinkingand other alcohol consumption. Moreover,
problemdrinkingis not a simple phenomenon.Researchcould usefully
address whether the different patternsof behavior that fall under the
general label of problem drinkinghave different impacts on socioeconomic outcomes. This is a particularlypromising approachto understandinggenderdifferences, since the "typical" male alcoholic appears
to be different from the "typical" female alcoholic.55The effects of
problemdrinkingon socioeconomic outcomes may also be quite subtle.
To shed more light it is probablynecessary to go beyond measuresof
wages and hours worked and examine other aspects of work performance, including both short-runoutcomes (such as absenteeism, tardiness, and on-the-job accidents) and long-runoutcomes (such as occupationalchoices and retirement).
Conclusion
Ourin-depthempiricalanalysis yielded manyestimatesof the effects
of alcohol on socioeconomic outcomes. We offer the following interpretationof the patternof results. The benchmarkOLS estimatesof the
effects of alcohol problems on socioeconomic outcomes may reflect
two sources of bias working in opposite directions. Because of omitted
variablesbias, the negative effects of alcohol problemsare overstated
in models that imperfectly control for deficiencies in childhood environmentand personalitydisorders.As alternativestrategiesare used to
controlfor the individualand family heterogeneity-including a richer
set of controls, siblings comparisons, and longitudinal fixed-effects
54. A promisingdata source is the LongitudinalAlcohol EpidemiologicSurvey by
the U.S. Departmentof Commerce, Bureauof the Census.
55. See footnote 41 above for a descriptionof differenttypes of alcoholism.
158
Brookings Papers: Microeconomics1994
models-another source of bias is uncoveredandpossibly exacerbated.
Contemporaneoussimultaneity(for example, from the income effect of
earnings on drinking problems) creates a positive bias. Once simultaneity is addressed using instrumentalvariables estimators, alcohol
problemsare estimated to lead to significantlyreducedearnings.
Ourinstrumentalvariablesestimatessuggest thatthe negativeimpact
of alcohol problems on earnings is somewhat larger than both our
benchmarkOLS estimates and many estimates from previous research
suggest. While we find little evidence that alcohol problems have a
negative impact on labor supply, there also appearsto be a negative
effect of problem drinkingon marriage.As we arguedin the last section, this type of evidence provides limited but useful guidancefor the
design of alcohol policies. It should be noted that the results in this
paper apply to a sample of young adults, none of whom is older than
thirty-one. Some other research suggests that older adults may suffer
largerearningslosses as the cumulativehealtheffects of chronic abuse
become apparent.56At the same time, the possible beneficial health
effects of moderate alcohol use may also only become apparentfor
older adults. Moderate alcohol use appears to protect against heart
disease, an effect that is unlikely to be relevant for people in the age
groupof the NLSY sample. Analysis of new datais neededto determine
if alcohol abuse is more harmfulor alcohol use is morehelpful for older
adults.
56. Mullahyand Sindelar(1993) and Rice and others(1990) find muchlargerlosses
for individualsaged thirtyto sixty than for young adults.
Donald S. Kenkel and David C. Ribar
159
Appendix
TableA-1. National LongitudinalSurvey of Youth(NLSY) VariableMeans:
Full Sample, Siblings Sample, and LongitudinalSample
Men
Variable
Full sample
1989 Outcomes
Log(earningsdividedby $1,000)
Log(hoursdividedby 100)
Mean
Women
Standard
deviation
Mean
Standard
deviation
9.48
7.57
(0.88)
(0.55)
8.99
7.25
Marriage
0.46
(0.50)
0.54
(0.50)
Alcoholdependence
Alcoholabuse
Heavydrinking
Days drinkingper month
0.20
0.15
2.01
6.44
(0.40)
(0.36)
(3.09)
(7.92)
0.08
0.05
0.63
2.81
(0.28)
(0.23)
(1.75)
(4.99)
0.27
0.16
60.65
0.41
0.12
0.04
0.80
0.16
183.45
1.04
(0.44)
(0.37)
(2.22)
(0.49)
(0.32)
(0.19)
(0.40)
(0.37)
(39.87)
(1.18)
0.26
0.16
60.51
0.49
0.06
0.05
0.84
0.17
184.44
1.47
(0.44)
(0.37)
(2.22)
(0.50)
(0.24)
(0.21)
(0.37)
(0.37)
(36.38)
(1.29)
9.99
0.07
9.25
0.15
0.32
3.88
0.55
0.04
0.18
0.23
(4.17)
(0.26)
(5.29)
(0.35)
(0.47)
(2.65)
(0.50)
(0.19)
(0.38)
(0.42)
10.17
0.05
9.19
0.15
0.33
3.96
0.54
0.05
0.23
0.25
(3.84)
(0.22)
(5.23)
(0.35)
(0.47)
(2.68)
(0.50)
(0.23)
(0.42)
(0.44)
5.50
1.31
0.15
2.59
(1.94)
(3.82)
(0.08)
(0.26)
5.55
1.37
0.15
2.59
(1.98)
(3.95)
(0.08)
(0.26)
1.89
(5.82)
2.24
(6.56)
Personalcharacteristics
Africanorigin
Latinoorigin
Yearof birth
Religious
Interviewstatus
Healthstatus
Completedhigh school
Completedcollege
AFQTscore
Children
Familybackground
Mother'seducation
Mother'seducationmissing
Father'seducation
Father'seducationmissing
Nonintactfamily at age fourteen
Siblings
Magazinesat age fourteen
Alcoholicmother/stepmother
Alcoholicfather/stepfather
Otheralcoholicrelative
Local variables
Unemploymentrate
Percapitaincome (dividedby $10,000)
Manufacturing
employment
Adjustedbeer price
Percentageof statepopulation
residingin dry counties
(1.07)
(0.79)
160
Brookings Papers: Microeconomics 1994
TableA-1. National LongitudinalSurvey of Youth(NLSY) VariableMeans:
Full Sample, Siblings Sample, and LongitudinalSample (Continued)
Women
Men
Variable
Siblingssample
1989Outcomes
Log(earningsdividedby $1,000)
Log(hoursdividedby 100)
Marriage
Alcoholdependence
Alcohol abuse
Heavydrinking
Days drinkingper month
Personalcharacteristics
Africanorigin
Latinoorigin
Yearof birth
Religious
Interviewstatus
Healthstatus
Completedhigh school
Completedcollege
AFQTscore
Children
Familybackground
Mother'seducation
Mother'seducationmissing
Father'seducation
Father'seducationmissing
Noniritactfamily at age fourteen
Siblings
Magazinesat age fourteen
Alcoholicmother/stepmother
Alcoholicfather/stepfather
Otheralcoholicrelative
Localvariables
Unemploymentrate
Percapitaincome (dividedby $10,000)
Manufacturing
employment
Adjustedbeer price
Percentageof statepopulation
residingin dry counties
Mean
Standard
deviation
Mean
Standard
deviation
9.41
7.55
0.44
0.20
0.15
1.90
6.14
(0.91)
(0.56)
(0.50)
(0.40)
(0.36)
(2.95)
(7.65)
9.04
7.25
0.51
0.08
0.06
0.64
2.83
(1.00)
(0.81)
(0.50)
(0.28)
(0.23)
(1.75)
(5.10)
0.30
0.18
60.99
0.43
0.12
0.04
0.78
0.14
176.69
0.95
(0.46)
(0.38)
(2.06)
(0.50)
(0.33)
(0.19)
(0.42)
(0.35)
(38.27)
(1.10)
0.30
0.15
61.03
0.53
0.06
0.04
0.85
0.17
181.88
1.43
(0.46)
(0.36)
(1.99)
(0.50)
(0.24)
(0.20)
(0.35)
(0.37)
(36.37)
(1.25)
9.78
0.08
8.79
0.16
0.31
4.48
0.51
0.03
0.17
0.24
(4.26)
(0.27)
(5.33)
(0.37)
(0.46)
(2.56)
(0.50)
(0.17)
(0.37)
(0.43)
10.24
0.04
9.22
0.13
0.30
4.59
0.52
0.04
0.23
0.26
(3.73)
(0.20)
(5.19)
(0.34)
(0.46)
(2.73)
(0.50)
(0.21)
(0.42)
(0.44)
5.53
1.18
0.16
2.59
(1.95)
(3.60)
(0.09)
(0.26)
5.52
1.17
0.15
2.59
(1.96)
(3.70)
(0.08)
(0.26)
2.10
(6.32)
2.30
(6.65)
Donald S. Kenkel and David C. Ribar
161
TableA-1. National LongitudinalSurvey of Youth(NLSY) VariableMeans:
Full Sample, Siblings Sample, and LongitudinalSample (Continued)
Men
Variable
Mean
Women
Standard
deviation
Mean
Standard
deviation
Longitudinalsample
1985 Outcomes
Log(earningsdividedby $1,000)
Log(hoursdividedby 100)
Marriage
Heavydrinking
Days drinkingper month
9.10
7.47
0.32
1.56
6.81
(1.04)
(0.65)
(0.47)
(2.94)
(8.03)
8.60
7.12
0.45
0.40
3.00
(1.21)
(0.89)
(0.50)
(1.48)
(5.13)
1989 Outcomes
Log(earningsdividedby $1,000)
Log(hoursdividedby 100)
Marriage
Heavydrinking
Days drinkingper month
9.44
7.55
0.47
2.09
6.48
(0.90)
(0.58)
(0.50)
(3.16)
(8.03)
8.91
7.21
0.55
0.64
2.72
(1.09)
(0.82)
(0.50)
(1.78)
(5.02)
1985 Personalcharacteristics
Interviewstatus
Healthstatus
Children
0.10
0.03
0.78
(0.29)
(0.18)
(0.97)
0.06
0.05
1.26
(0.23)
(0.22)
(1.13)
1989 Personalcharacteristics
Interviewstatus
Healthstatus
Children
0.13
0.04
1.08
(0.33)
(0.19)
(1.17)
0.06
0.05
1.56
(0.24)
(0.21)
(1.28)
1985 Local variables
Unemploymentrate
Percapitaincome (dividedby $10,000)
Manufacturing
employment
8.21
1.27
0.16
(3.00)
(3.68)
(0.09)
8.24
1.24
0.16
(3.05)
(3.55)
(0.09)
1989 Localvariables
Unemploymentrate
Percapitaincome(dividedby $10,000)
Manufacturing
employment
5.53
1.47
0.15
(1.99)
(4.07)
(0.09)
5.57
1.55
0.15
(1.99)
(4.16)
(0.08)
Source: Based on data from the NLSY.
Comments
and Discussion
Comment by Philip J. Cook: An early effort to estimate the productivity costs of drinking was made by Irving Fisher. "All of us know
thatindustrialefficiency was one of the chief reasonsfor Prohibition,"
he asserted in a book published in 1926.' In supportof his belief that
drinking slows down the "human machine," he cited an experiment
conducted in Germany. Four typesetterswere studied over a four-day
period; two of them were given drinks, and the other two served as a
control group. The experimenter'sconclusion was that drinkingthree
glasses of beer in a day reducedproductivityby about 10 percent.Fisher
made a breathtakingextrapolationfrom this result, estimatinga 5 percent increase in nationalproductivityas a result of Prohibition-induced
reductionin drinking.
More recent researchhas used largerdata bases and more sophisticated methods but has not necessarily reached more reliable conclusions.2The evidence utilized by Fisherwas derivedfroman experiment
and had a straightforwardinterpretation.Modernresearchin this area,
in contrast, has attemptedthe conceptually more difficult task of estimatingcausal effects from nonexperimentaldata. In theirpaper,Kenkel
andRibarconfrontthis challenge with an exceptionallyrichdataset(the
National Longitudinal Survey of Youth, or NLSY) and an arsenal of
econometric methods. Together with the recent paper of Mullahy and
Sindelar,3Kenkel and Ribarnow define the state of the art for this sort
of research.Yet we are left with considerableuncertaintyaboutwhether
1. Fisher(I1926,p. 158).
2. Cook (199 1).
3. Mullahyand Sindelar(1993).
162
Donald S. Kenkel and David C. Ribar
163
a successful national policy to reduce alcohol abuse (not necessarily
prohibition!)would increase productivity.
Causal Links
Kenkeland Ribarleave theirunderlyingmodel implicit in theireconometrics, but perhapsit is useful to make explicit the causal linkages
by which drinking is related to earnings and maritalstatus. Most obvious are links that, in the authors' framework,would be considered
"reverse" causation, and I begin with those.
Alcohol is an ingredientof beer, wine, and spirits, the consumption
of which is influencedby income. Researchon the effect of income on
alcoholic beverage expendituressuggests thatthe demandfor this commodity group is normal.4That is not the same thing as demonstrating
thatalcohol itself is a normalcommodity, since consumersmay increase
the quality of their drinks rather than the quantity as their incomes
increase. But it would be surprisingindeed if income played no role in
determiningalcohol consumption.
In a consumerchoice framework,it is standardto considerthe wage
rate rather than earnings as the relevant exogenous variable. To the
extent that drinking is complementaryto leisure activities, the wage
rate will affect drinkingboth throughthe income effect and the laborleisure choice.
"Reverse" causation is also likely to be importantin the case of
maritalstatus. A recent study, using the NLSY data, found thatpeople
tend to reduce their drinkingmarkedlythe year before getting married
and to sustain this reductionduringthe year after marriage.5It appears
that with love comes a tendency to substitutehome-orientedactivities
for barhoppingand the like.
The causal linkages that interest Kenkel and Ribar go in the other
direction, where earnings and marriageare viewed as consequences of
drinkingrather than the reverse. Such linkages are not plausible for
moderatedrinking, so the authorsutilize measures of alcohol dependence, abuse, and frequency of intoxication (six drinks or more in a
single occasion). "Dependence" and "abuse" bothserve as good prox4. Sammartino(1990).
5. Miller-Tutzauer,Leonard,and Windle (1991).
164
BrookingsPapers: Microeconomics1994
ies for frequent drinking and frequent intoxication (see table 1), and
their results may be interpretedaccordingly.
Alcohol abuse may reduce earnings(wage ratesor hoursworked)by
any of several mechanisms. In the near term, hangoversmay result in
absenteeismor impairedperformance,while drinkingduringthe working day may impair cognitive and physical capacity in various ways.
Over the longer term frequent intoxication may result in disabling injuries or health problemsand increase the chance of acquiringa police
record. These same mechanismsmay also be importantin the marriage
market, where an alcohol abuser will be a less attractivepartnerthan
an otherwise similar person who drinks moderatelyor abstains.
Adultdrinkingpatternsarestronglycorrelatedwith teenagedrinking,
suggesting a more subtle link between drinkingand earnings. In particular there is strong evidence that youthful drinkingimpairsthe acquisition of human capital by leading to early terminationof schooling.6
Althoughearly terminationof schooling and consequententry into the
labor market may initially enhance earnings, after a few years those
with more schooling will tend to earn more.
The Authors' Econometric Results
Kenkeland Ribarprovidean extensive arrayof coefficientestimates.
They use several alternativeestimation methods, definitionsof drinking, and sets of covariates. Most of these results characterizethe contemporaneous association between some measure of drinking (as a
causal agent) and maritalstatus, earnings, and hours worked (as consequences). This "shotgun" approachis consistent with best econometricpractice, given uncertaintyabout the appropriatespecification.
Kenkel and Ribar's results on marriageare quite robust. In nearly
every estimatethey find a negative relationshipbetweenheavy drinking
(however measured)and the likelihood of being married.In most specifications the effect is significant and quite large. There appearsto be
an importantcausal effect here. But in which direction?The fact that
the estimatedeffects are much smaller (andusually insignificant)in the
simultaneousequation estimates favors an interpretationthat empha6. Mullahyand Sindelar(1989); Cook and Moore (1993).
Donald S. Kenkel and David C. Ribar
165
sizes the reverse causal process. Love and marriage leads (at least
initially) to reduced alcohol consumptionfor both men and women.
The results for earnings and hours are more varied, and extractinga
coherentconclusion requiresstrongpriorsconcerningthe variousspecifications. A firstcut comes with a judgmentconcerningthe alternative
measuresof drinking.My preferenceis for the "heavy drinking" measure (which is the numberof occasions in the previous month where
the respondentconsumed six or more drinks), as opposed to the measures of dependence and abuse. Heavy drinkingis objective and meawhich drinking is
sures directly the mechanism-intoxication-by
likely to interferewith productivityfor youths.
The authors'estimatedeffects of heavy drinkingon earningsare for
the most part small and insignificant, a result that is compatible with
Mullahyand Sindelar's findingsfor this age group.7But in their simultaneousequationestimates, there is evidence that heavy drinkingdoes
indeed reduce productivity. The simultaneousequationestimates have
an edge over the other specificationsbecause, as explainedabove, there
is every reason to believe that earnings influence drinkingas well as
the reverse. What we find in these simultaneousequationestimates is
that heavy drinkingreduces earnings for both men and women and in
about the same proportion(although the effect is only significant for
men). Because hoursworkedare essentially unaffectedby heavy drinking for men and actually increase for women, it appearsthat the effect
on earningsis entirely the result of a reductionin the wage rate. Thus,
heavy drinkingby workers in their twenties appearsto limit access to
high-wagejobs but does not cause any reductionin hours worked.
The finding that women, but not men, work more hours when they
drinkheavily, is puzzling. Ratherthan stretchfor an ex post explanation, I would prefer to see more researchon this matter. The peculiar
findingmay resultfrom the selection process;only workersare included
in the data set used for these estimates.
Kenkel and Ribar's estimates may understatethe negative effect of
heavy drinking on earnings because their specifications incorporate
schooling as a covariate. That eliminates one potentially important
mechanismby which a tendency to drink heavily may influence productivity over the life course.8
7. Mullahyand Sindelar(1993).
8. Cook and Moore (1993).
166
Brookings Papers: Microeconomics 1994
Further Research
Kenkel and Ribar demonstratethe difficulty of extracting reliable
estimates of a complex causal process from natural (as opposed to
experimental) data. They try several stratagemsto remove potential
bias stemming from unobservedheterogeneity and from contemporaneous simultaneity. Because theirestimates are sensitive to the method
of estimation, readers' prior views concerning how drinking affects
earningsmay not be much revised by this exercise.
How could we hope to generatemore definitive results?Some guidance in this mattercomes from specifying the conceptual experiment
that underliesthis effort.9 To the extent that we, like IrvingFisher, are
interestedin what can be gained by restrictingthe availabilityof alcohol, then the desired experimentis something like this: Suppose that a
sample of adolescents is randomlyassigned to jurisdictionsthat differ
with respect to the stringency of alcohol control measures(tax, minimum drinking age, advertising) and other contextual determinantsof
drinking, and that this "treatment" is preserved through age thirty.
Will the subjectsin the morestringentjurisdictionsend up earningmore
on the average?
Several economists have producedestimatesof the effects of alcohol
controlmeasureson youthfuldrinkingand its consequenceswith something like this conceptual experimentin mind.10What carriesus from
the conceptual experiment to the analysis of nonexperimentaldata is
the assumptionthatthe observedgeographicvariationin alcohol control
measuresis independentof the unobservedheterogeneityamongyouthful residents. Perhaps the same approach will prove productive for
earningsas it has for alcohol abuse, schooling, and traffic accidents.
Alternatively, perhapswe should go back to plying typesetterswith
beer.
Comment by Sam Peltzman. The Kenkel-Ribarpapercan be viewed
as a description of behavior. It can also be viewed as part of a recipe
for policy. The paper is a solid contributionto knowledge, and I have
only minor reservations about it. My major reservation is about the
9. Mullahy(1993).
10. Cook and Tauchen (1982), Grossmanand others (1993), and Cook and Moore
(1993).
Donald S. Kenkel and David C. Ribar
167
potentialmisuse of the authors'results. Specifically, I believe it would
be incorrect to interpretthe structuralequations underlying some of
their findings as behavioral responses to variables amenableto policy
manipulation.
We learn much about the behavior of alcoholics in this paper. Perhaps the most interestingresults come from the authors'simultaneous
equation model. It tells us that the quantity (not just the quality) of
alcohol demanded increases with income and that more consumption
lowers income. My only reservationhere is the lack of muchdiscussion
of the context in which these results are obtained. Theirsis a sample of
young adults. As the authorspoint out, these are heavy drinkingyears.
Most of the heavy drinkersin their sample will soon be cut back substantially. This raises some obvious questions. Is the behaviorKenkel
and Ribardescribe permanentor temporary?Specifically, are the negative effects in income more than a matter of the timing of income
flows? It is impossible to answer such questionswith this database.But
the unrepresentativenatureof the samplehas to temperany conclusions.
The time period Kenkel and Ribar study is also unrepresentativein
that heavy drinking in general seemed to be going out of style in the
1980s. The decade witnessed a sharpdecline in per capita hardliquor
consumption, a more moderate decline in wine consumption, and a
flatteningof growth in beer consumption. Because the sample group
uses beer especially heavily, it was less affected by the trend toward
less alcohol than older devotees of hard liquor. Nevertheless, the fact
thatheavy drinking,like heavy smoking, was becomingless respectable
or acceptableshouldbe kept in mindin evaluatingthe results, especially
the longitudinalanalysis in table 6. For example, one resulthere is that
increasedheavy drinking is associated with increased (male) income.
This seems at odds with Kenkel and Ribar's other findings on the
income effects of drinking. But it may mean only that high income
males lagged in response to the pressureto reduce heavy drinking.
Kenkel and Ribar are appropriatelycautious aboutthe policy implications of their results. I want to add to that caution. They claim to
have found evidence that alcohol dependenceobeys the law of demand
while lowering income. If this model is taken as a behavioralsystem,
it will be temptingto reason as follows (using the authors'elasticities):
If the price of alcohol is raised, say, 20 percent, then alcohol abuse
will decline about 30 percent. This decline in abuse will raise average
168
BrookingsPapers. Microeconomics1994
labor incomes by about 1 percent." Because raising incomes is
'good," the policy implications seem clear-raise alcohol taxes.
This sort of inference would, I think, be wrong for reasons beyond
those emphasizedby Kenkel andRibar-namely, the economist's usual
plea to distinguish private from public gain. I agree with this, but the
political process has fewer hang-upsaboutpaternalismthaneconomists
do. The history of safety regulation, the currentantismokingcrusade,
andso forth, makeclear thatan antidrinkingcoalition will find "alcohol
taxes are good for you" an appealingslogan.
The more fundamentaldifficulty is with the behavioralinterpretation
of the model. Alcohol abuse can be viewed as an indicatorof a complex
set of endowments and lifestyle choices, which include dissipation of
earningcapacity. It also likely includes more present-orientation,less
happiness, less self-confidence, and less self-control than does nonabuse. These endowments and choices are likely reflectedin a myriad
of consumptionand investment decisions, all consistent with lowered
earning capacity. The abuser-type probably eats the wrong foods,
dresses poorly, has a bad temper, does not makes friends easily, etc,
as well as drinkstoo much. Only the latteris measuredin this data set.
The negative coefficient of abuse in the income regression inevitably
then summarizesall of the income-reducingbackgroundand decision
characteristics left out of the model. The coefficient almost surely
overstatesthe net contributionof alcohol abuse and thereforethe realworld effects of an alcohol tax. Indeed, the main real-worldeffects of
such a tax may well be simply to increaseconsumptionof good-substitutes for alcohol, for example, otherdrugsor more surlinesson thejob,
without any obvious enhancementof income.
It is easy to say that the problem I raise can be solved by a more
complete model. But no workable model is likely to begin to capture
the myriadof marginsimportantlyaffected by a change in the price of
alcohol. The kind of evidence we would need to get at the plausible
effects of an alcohol tax on income will not come from an elaboration
of the authors' model. It will come from observing if income really
rises after some jurisdiction substantiallyraises alcohol taxes.
11. Their result is that each abuser loses about -0.3 in In earningand about0.10
of the sample is abusers. Cuttingthe portionof abusersto 0.07 (thatis, by 30 percent)
of the sample raises the mean hourlyearningsin the sample by 0.09.
Donald S. Kenkel and David C. Ribar
169
General Discussion: The discussion was dominatedby data, measurement, and modeling issues. MarkPauly was concernedaboutthe variables used as exogenous measures of alcohol intake. The paper's alcohol abuse and dependence variables, he noted, correspondto DSMIII definitions, which are based on the consequences of alcohol consumption,such as its effects on an individual'swork. Thus, he asserted,
these variables appearto have potential endogeneity problems. Pauly
also argued that firms generally deal with workers who have serious
drinkingproblemsnot by reducingtheirwages, but by terminatingtheir
employment. As a result, he said, the labor marketmay not accurately
reflect the costs of alcohol abuse, because many abuserscould be unemployed. Some portion of lost output attributableto alcohol abuse,
therefore, will not be reflected in earningsdata.
Noting that the paper's results suggest that, on average, single
women drink more than marriedwomen, John Pencavel remindedthe
authorsthat, on average, single women work more hoursthan married
women. Based on the equation specifications used by the authors, he
suspectedthat the strongpositive effect of alcohol consumptionamong
women on hours workedreportedin the paperis merely a reflectionof
the absence from the equation of a marital status variable. Donald
Kenkelrespondedthat alcohol measuresare treatedendogenously, and
because marriageis not a regressorin the reducedform alcohol equation, alcohol consumption is not predicted based on marital status.
Pencavel replied that this did not deal with the problemhe raised.
FrankWolak suggested that firmswhich can identify their alcoholic
workersmight try to pay these employees less than other workers. He
arguedthatthe authorsshouldthereforeuse averagehourlywages rather
than earnings or hours in their regressions. Wolak also noted that
youngerworkerstend to move in and out of the laborforce, so that at
any given time many have zero earnings or zero hours of work. He
wonderedhow the authorstreat this problem. Kenkel replied that the
equationsinclude only those individualswho have eitherpositive hours
workedor positive earnings. Kenkel also said thathe and his co-author
hadestimatedversions of their models correctedfor selectivity bias and
had found no selection effects for the earnings equations. Although
there were some for the hours equations, correctingfor selection did
not change the alcohol coefficients, so these results are not reportedin
the paper.
170
BrookingsPapers: Microeconomics1994
RobertCrandallsuggested that it might be more socially acceptable
for single people than marriedpeople to admit to drinkingproblems
and argued that the resulting reporting biases may be affecting the
authors'results. Admitting that reportingbiases can cause a problem,
Kenkel said that although the authors could not control for the bias
Crandallraised, they did attemptto correctfor anotherbias problemby
using a control to indicate whethera thirdindividualwas presentwhen
a respondent was being interviewed. This control was used, Kenkel
said, because it is possible that some teenage respondentswould have
lied about their drinkinghabits if their parentswere presentduringthe
interview.
Given the potentialeffects of self-esteem on alcoholism rates, Patricia Anderson suggested that the authorsadd a self-esteem measureto
theirmodel. In response, David Ribarsaid thathe and Kenkelfirsttried
using both a self-esteem measure and a measure of internal versus
external locus of control, but these variables had predictableeffects
without changing any of the other results.
Suzanne Scotchmer argued that the differential effects of alcohol
consumptionon the marriagerates of men and women may result from
the fact that a higher proportionof men are alcoholics if it is also the
case the alcoholic men and women tend to marryeach other. If every
alcoholic woman marries an alcoholic man, she explained, then the
remainingalcoholic men must marrynonalcoholic women. If the marriages of double alcoholic families are more likely to dissolve than the
marriagesof single alcoholic ones, it will appearthat alcoholism has a
greaternegative effect on the marriagerate of women thanof men, she
said.
Noting thatthe negative effects of alcohol abuseon earningsrevealed
in the paperare not as strongas those shown by previousstudies, Martin
Baily speculatedthat certain common behavioralpatternsmay explain
these new findings. He suggested that people with a drinkingproblem,
especially those in high-paying stressfuljobs, tend to drinkin the evenings. They are still able to be at theirjobs the following morning,and
as a result, he said, their drinkingproblemsdo not have a large effect
on their employment situation. Instead, these problemsaffect the family, resulting, for example, in more conflict between husbandand wife
and, hence, in higher divorce rates.
William Niskanen assertedthat the effect of alcohol consumptionon
Donald S. Kenkel and David C. Ribar
171
productivityis not a sufficient basis for public policy action, which he
said should be based on marginalexternalbenefitsand costs. He noted
that the paperdoes not provide the sign of correlationbetween private
andsocial costs with respectto alcohol consumption.At best, Niskanen
said, the results provide guidelines for action within a paternalistic
environment,in which one makes suggestions to relatives and friends.
Ribardisagreed, arguingthat the transmissionof alcoholism from parents to children makes this issue an area of public policy concern.
172
Brookings Papers: Microeconomics1994
References
Amemiya, Takeshi. 1978. "The Estimationof a SimultaneousEquationGeneralized ProbitModel." Econometrica46 (September):1193-1205.
American Psychiatric Association. 1987. Diagnostic and Statistical Manual of
Mental Disorders: DSM-111-R. 3d ed. rev. Washington, D.C.
Bartel, Ann, and Paul Taubman. 1986. "Some Economic and Demographic
Consequences of Mental Illness." Journal of Labor Economics 4 (April):
243-56.
Becker, Gary S. 1975. Human Capital: A Theoretical and Empirical Analysis
with Special Reference to Education. 2d ed. University of Chicago Press.
Becker, GaryS., and Kevin M. Murphy. 1988. "A Theoryof RationalAddiction." Journal of Political Economy 96 (August): 675-700.
Benham, Lee, and AlexandraBenham. 1982. "Employment, Earnings, and
PsychiatricDiagnosis." In EconomicAspects of Health, edited by Victor R.
Fuchs, 203--20. Universityof Chicago Press.
Berger, Mark C., and J. Paul Leigh. 1988. "The Effect of Alcohol Use on
Wages." AppliedEconomics 20 (October):1343-51.
Berry, Ralph E., Jr., and James P. Boland. 1977. The Economic Cost of
Alcohol Abuse. The Free Press.
Centerfor Human Resource Research. 1990. NLS Handbook1990: The National Longitudinal Surveys of Labor Market Experience. Ohio State Uni-
versity.
Chaloupka, Frank, Henry Saffer, and Michael Grossman. 1993. "Alcohol
ControlPolicies and MotorVehicle Fatalities." Journal of Legal Studies 31
(January):161-86.
Cloninger, C. Robert. 1987. "Neurogenetic Adaptive Mechanismsin Alcoholism." Science, April 24, 1987, vol. 236, 410-16.
Cook, Philip J. 1981. "The Effect of LiquorTaxes on Drinking, Cirrhosis,
and Auto Fatalities." In Alcohol and Public Policy. Beyond the Shadow of
Prohibition, edited by Mark H. Moore and Dean R. Gerstein, 225-85.
Washington,D.C.: National Academy Press.
1991. "The Social Costs of Drinking." In Expert Meeting on the
Negative Social Consequences of Alcohol Use: Oslo, 27-31 August 1990,
49-74. Oslo: NorwegianMinistryof Healthand Social Affairs.
Cook, Philip J., and Michael J. Moore. 1993. "Drinking and Schooling."
Journal of Health Economics 12 (December): 411-29.
Cook, Philip J., and George Tauchen. 1982. "The Effect of LiquorTaxes on
Heavy Drinking." Bell Journal of Economics 13 (Autumn): 379-90.
Fingarette, Herbert. 1988. Heavy Drinking. The Myth of Alcoholism as a Dis-
ease. Universityof CaliforniaPress.
Fisher, Irving. 1926. Prohibition at its Worst. Macmillan.
Donald S. Kenkeland David C. Ribar
173
Frank,Richard,and Paul Gertler. 1991. "An Assessmentof the Measurement
ErrorBias for Estimatingthe Effect of MentalDistresson Income." Journal
of Human Resources 26 (Winter): 154-64.
Geronimus, Arline T., and Sanders Korenman. 1992. "The Socioeconomic
Consequencesof Teen ChildbearingReconsidered." QuarterlyJournal of
Economics 107 (November): 1187-1214.
Grant,BridgetF., andothers. 1991. "Prevalenceof DSM-III-RAlcohol Abuse
and Dependence." Alcohol Health and Research World 15 (1): 91-96.
Greenfield,Thomas K., KarenL. Graves, and Lee Ann Kaskutas. 1993. "Alcohol Warning Labels for Prevention." Alcohol Health and Research World
17 (1): 67-75.
Griliches, Zvi. 1979. "Sibling Models and Data in Economics:Beginningsof
a Survey." Journal of Political Economy 87 (October): S37-S64.
Griliches, Zvi, and Jerry A. Hausman. 1986. "Errorsin Variablesin Panel
Data." Journal of Econometrics 31 (February): 93-118.
Grossman,Michael. 1988. "Health Economics of Preventionof Alcohol-Related Problems." Paperpresentedat the Workshopon HealthEconomicsof
Preventionand Treatmentof Alcohol-RelatedProblems. National Institute
on Alcohol Abuse and Alcoholism, Washington,D.C., September29-30.
. 1993. "The Economic Analysis of Addictive Behavior." In Economics and the Prevention of Alcohol-Related Problems. Proceedings of a Workshop in Economic and Socioeconomic Issues in the Prevention of Alcohol-
RelatedProblems, edited by GregoryBloss and MichaelE. Hilton, 91-123.
Rockville, Md.: National Instituteon Alcohol Abuse and Alcoholism, U.S.
Departmentof Health and HumanServices.
Grossman, Michael, and others. 1993. "Effects of Alcohol Price Policy on
Youth." WorkingPaper4385. Cambridge,Mass.: NationalBureauof Economic Research. June.
Harwood, Henrick J., and others. 1984. Economic Costs to Society of Alcohol
and Drug Abuse and Mental Illness: 1980. Research Triangle Park, N.C.:
ResearchTriangleInstitute.
Heien, Dale M., and David J. Pittman. 1989. "The EconomicCosts of Alcohol
Abuse: An Assessment of CurrentMethods and Estimates." Journal of
Studies on Alcohol 50 (November): 567-79.
. 1993. "The ExternalCosts of Alcohol Abuse." Journal of Studieson
Alcohol 54 (May): 302-7.
Hesselbrock,Victor M. 1986. "Family History of Psychopathologyin Alcoholics: A Review and Issues. " In Psychopathology and Addictive Disorders,
edited by Roger E. Meyer, 41-56. New York:GuilfordPress.
Hoyt, Gail Mitchell, and FrankJ. Chaloupka.1993. "Self-ReportedSubstance
Use and Survey Conditions:An Examinationof the National Longitudinal
Surveyof Youth." Paperpresentedat the meetingsof the WesternEconomic
AssociationInternational,Lake Tahoe, Nev., June.
174
BrookingsPapers: Microeconomics1994
Kaestner, Robert. 1991. "The Effect of Illicit Drug Use on the Wages of
Young Adults." Journal of Labor Economics 9 (October): 381-412.
Kenkel, Donald S. 1991a. "Health Behavior, HealthKnowledge, and Schooling." Journal of Political Economy 99 (April): 287-305.
1991b. "What You Don't Know Really Won't HurtYou." Journal
of Policy Analysis and Management 10 (Spring): 304-9.
.1993.
"Drinking, Driving, and Deterrence:The Effectiveness and
Social Costs of Alternative Policies." Journal of Law and Economics 36
(October):877-913.
Klein, Hugh, and David J. Pittman, 1990. "DrinkerPrototypesin American
Society. " Journal of Substance Abuse 2: 299-316.
Korenman,Sanders,and David Neumark. 1991. "Does MarriageReally Make
Men More Productive?" Journal of Human Resources 26 (Spring): 282-
307.
1992. "Marriage, Motherhood,and Wages." Journal of HumanResources 27 (Spring):233-55.
Maddala, G.S. 1983. Limited-Dependent and Qualitative Variables in Econ-
ometrics. Cambridge:CambridgeUniversityPress.
. 1987. "LimitedDependentVariableModels Using PanelData." Journal of Human Resources 22 (Summer): 307-38.
Manning, Willard G., and others. 1991. The Costs of Poor Health Habits.
HarvardUniversity Press.
Miller-Tutzauer, Carol, Kenneth E. Leonard, and Michael Windle. 1991.
"Marriage and Alcohol Use: A LongitudinalStudy of 'MaturingOut'."
Journal of Studies on Alcohol 52 (September): 434-40.
Mincer, Jacob. 1974. Schooling, Experience, and Earnings. National Bureau
of Economic Research.
Miron, Jeffrey A., and Jeffrey Zwiebel. 1991. "Alcohol ConsumptionDuring
Prohibition."
American Economic Review Papers and Proceedings
81
(May): 242-47.
Mullahy, John. 1993. "Alcohol and the Labor Market." In Economics and
the Prevention of Alcohol-Related Problems: Proceedings of a Workshop in
Economic and Socioeconomic Issues in the Prevention of Alcohol-Related
Problems, edited by GregoryBloss and Michael E. Hilton, 141-74. Rockville, Md.: National Instituteon Alcohol Abuse and Alcoholism, U.S. Departmentof Health and HumanServices.
Mullahy, John, and Jody Sindelar. 1989. "Life-Cycle Effects of Alcoholism
on Education,Earnings,and Occupation."Inquiry26 (Summer):272-82.
. 1991. "GenderDifferences in LaborMarketEffects of Alcoholism."
American Economic Review Papers and Proceedings 81 (May): 161-65.
. 1992. "Employment, Unemployment,and ProblemDrinking." Departmentof Economics, TrinityCollege, Hartford,Conn. Mimeo.
Donald S. Kenkel and David C. Ribar
175
. 1993. "Alcoholism, Work, and Income." Journal of LaborEconomics 11 (July):494-520.
Phelps, CharlesE. 1987. "Risk and PerceivedRisk of DrunkDriving Among
Young Drivers." Journal of Policy Analysis and Management 6 (Summer):
708-14.
Pogue, Thomas F., and LarryG. Sgontz. 1989. "Taxing to Control Social
Costs:The Case of Alcohol." AmericanEconomicReview79 (March):23543.
Ribar,David. 1994. "Teenage FertilityandHigh School Completion."Review
of Economics and Statistics (forthcoming).
Rice, Dorothy. 1993. "The Economic Cost of Alcohol Abuse and Alcohol
Dependence, 1990." Alcohol Health and Research World 17 (1): 10-11.
Rice, Dorothy P., and others. 1990. The Economic Costs of Alcohol and Drug
Abuse and Mental Illness: 1985. DHHS Publication No. (ADM) 90-1694.
Alcohol, Drug Abuse, and Mental Health Administration,Public Health
Service, U.S. Departmentof Healthand HumanServices.
Sammartino, Frank. 1990. Federal Taxation of Tobacco, Alcoholic Beverages,
and MotorFuels. CongressionalBudget Office.
Schepers, Andreas, and Gerhard Arminger. 1992. MECOSAUser Guide.
Frauenfeld,Switzerland:Systeme fur Logistik und Informationstechnik.
Shaper,A. Gerald. 1993. "Editorial:Alcohol, the Heart, and Health." American Journal of Public Health 83 (June): 799-801.
Sindelar, Jody. 1993. "Measurement Issues in Alcohol Survey Data." In
Economics and the Prevention of Alcohol-Related Problems: Proceedings of
a Workshop in Economic and Socioeconomic Issues in the Prevention of
Alcohol-RelatedProblems, edited by GregoryBloss and MichaelE. Hilton,
201-28. Rockville, Md.: National Instituteon Alcohol Abuse and Alcoholism, U.S. Deparmentof Health and HumanServices.
Snortum,John R., Dale E. Berger, and RagnarHauge. 1989. "Legal Knowledge and Compliance:Drinking and Driving in Norway and in the United
States." Alcohol, Drugs, and Driving 4 (3-4): 251-63.
USDHHS (U.S. Departmentof Health and Human Services; Public Health
Service; Alcohol, Drug Abuse, and Mental Health Administration).1990.
Seventh Special Report to the U.S. Congress on Alcohol and Health from
the Secretary of Health and Human Services. DHHS Publication No. (ADM)
90-1656.
Zucker,RobertA., and Edith S. LisanskyGomberg. 1986. "Etiology of AlcoholismReconsidered:The Case for a BiopsychosocialProcess." American
Psychologist 41 (July): 783-93.