Abstract Grossman proposed an individual's health can be viewed as one aspect of their human capital. Following this line of thought a number of recent papers have reported a positive impact of alcohol consumption on earnings. The rationale for the existence of such a relationship is the positive impact of alcohol on physical and mental health. We conduct a meta-analysis to determine whether such factors as: the estimation technique, the presence of ex-drinkers in the sample, possible sample selection bias and publication bias may all contribute to these findings. An additional suggestion for the positive relationship between alcohol and wages is the presence of a common set of personality traits that determines drinking behaviour and also leads to higher earnings. We examine this relationship by reviewing the literature that investigates if the personality influencing aspects of alcohol consumption influences measures of human capital. We also survey the significant body of research that has examined how alcohol consumption has been found to influence educational outcomes and the work force participation of problem drinkers.

1. Introduction

The ‘measurement of human capital’ literature has heretofore focused primarily on attributes that directly impact a given individual's ability to have successful outcomes in a given labour market. There has been limited discussion in this broad literature on how ‘health effects’ impact human capital development and market outcomes (relatively speaking). One aspect of this body of work is how alcohol consumption impacts health, how health impacts the development of human capital and ultimately how these health effects impact labour market outcomes. There are various hypotheses on how alcohol consumption impacts a given individual's ultimate labour market outcomes, and this paper surveys that literature and attempts to survey those questions and bridge that survey evidence with the findings in other fields that have also explored this question. But beyond the survey aspects, we take the question further by reestimating some of the findings in this literature ourselves.

In this paper we examine the evidence for the existence of a positive relationship between moderate alcohol consumption and a person's labour market outcomes that becomes negative with higher levels of consumption. This process has often been characterized by an inverted U-shaped function between wages earned and the level of alcohol consumption. The main reasons for this relationship have been claimed to be the positive effects of moderate alcohol consumption on health and the other factors that influence the development of human capital and its efficient use. Previously, there has been little attempt to relate these results with the research from other fields concerning the implications of alcohol consumption. Here we standardize the results of previous studies of moderate alcohol consumption on earnings in order to make them comparable, and then we relate them to findings in other fields.

Standard human capital variables in specifications of wage equations have traditionally included education and on-the-job training. One mechanism by which alcohol consumption may influence human capital would be via an individual's health. The health of an individual can be viewed as one aspect of their human capital (Grossman, 1972; Wößmann, 2003; Becker, 2007). Grossman (1972, p. 224) observes that health capital differs from other forms of human capital in that while the stock of a person's knowledge affects their productivity, their health determines the total amount of time spent producing earnings and thus the utilization of their stock of human capital.

Several papers have estimated the simultaneous effects of health and wages based on the theoretical framework of health capital developed in Grossman (1972). Grossman and Benham (1974) use data on white males in the labour force from the 1963 health interview survey conducted by the National Opinion Research Centre and the Centre for Health Administration Studies of the University of Chicago. To account for the potential of simultaneity bias between health and earnings, they estimate a three-equation model for wages, hours worked and health status using two-stage least squares (2SLS). A significant negative relationship between ill-health and wages is found, although the ill-health qualitative variable is treated as a continuous variable.

Lee (1982) studies the relationship between health and wages using data from the National Longitudinal Study of Men aged between 45–59 for the 1966 survey year. His findings indicate that wages have a strong positive effect on the demand for good health and good health increases wages. Havemann et al. (1994) estimates a three-equation simultaneous model to examine the interrelations among work-time, wages and health. His data include 613 observations for white males from the Michigan Panel Study of Income Dynamics. The results of this analysis indicate that prior health limitations have a negative and significant effect on wages. Using data on males aged 25–64 from the third wave (2001) of the Household, Income and Labour Dynamics in Australia survey, Cai (2007) estimates a simultaneous equations model of health and wages accounting for the endogeneity of health. When measurement error and endogeneity of health are accounted for, good health is found to have a significant and positive effect on wages.

In the medical literature good health is considered to be a combination of factors beyond the individual's control including their genetic makeup and controllable factors such as drinking. The World Health Organization (2007) has estimated that in 2002 alcohol caused 3.7% of all deaths and 4.4% of the total burden of disease with a disproportionately higher impact on males. The health problems associated with the excessive intake of alcohol include liver damage, heart disease, ulcers, cancers of the head and throat, alcoholism and impairment of mental functions. Overuse of alcohol has also been shown to contribute to obesity and high blood pressure.

Multiple studies in medical research (e.g. Klatsky et al., 1990; Jackson et al., 1991; Rimm et al., 1991; DeLabry et al., 1992; Razay et al., 1992; Coate, 1993) have found a U-shaped or J-shaped relationship between alcohol consumption and risk of death from all causes. This implies that non-drinkers and heavy drinkers have an increased risk of death compared with moderate drinkers. This association is largely due to lower death rates from coronary heart disease among moderate drinkers. Numerous studies have indicated that drinking in small to moderate amounts decreases the risk of dying from coronary heart disease by almost one third. Although definitions of moderate drinking vary widely, a consensus in the medical literature defines moderate drinking as up to two standard size drinks a day for men, or about 25 g of alcohol, and for women this should be no more than one standard size drink (Klatsky, 2003).

In the economics literature the specific relationship between alcohol consumption and outcomes in the labour market has been examined. While it has been suggested that problem drinking may reduce productivity and wages, moderate alcohol consumption is thought to benefit health and increase productivity thus generating both greater promotional opportunities and wages.

In Section 2 of this paper, we conduct a review of those studies that have focused on the relationship between moderate alcohol consumption and earnings for those individuals with a relatively stable employment pattern. Section 3 discusses the evidence that has been put forward to establish the degree to which alcohol consumption may have an indirect influence on productivity via its effects on the acquisition of traditional human capital. We also review the research that focuses on the direct impact of problem drinking on employment and unemployment. Section 4 presents conclusions.

2. Moderate Alcohol Consumption and Labour Market Outcomes

Alcohol consumption is typically incorporated into the equation using a human capital framework for individual i:


where Wit is the log wage rate, Sit includes measures of the health components of human capital, Nit includes the non-health components of human capital such as school and experience, Xit includes demographic variables and vit is the error term. Specifying Sit = (Ait, Kit) where Ait is alcohol consumption and Kit is other health outcomes then (1) becomes


where γ1, γ2, γ3, and  β1 are vectors of parameters to be estimated.

The alcohol consumption measures used have been defined in different ways including a binary yes/no indicator of alcohol consumption, multiple binary indicators of alcohol consumption at different levels and frequency of consumption over a specified period. An example of the latter would be the number of drinks or the amount of ethanol consumed during the week.

The two main health related justifications for the existence of such a relationship between alcohol consumption and earnings are the impact on physical health and mental health. First, the impact of alcohol on a person's physical health may be due to lowering the incidence of illnesses such as heart disease and stroke. Consequently, this may lead to reduced absenteeism from the workplace and increased productivity which may generate greater promotional opportunities and hence higher wages (see for example MacDonald and Shields, 2001). And second, the impact on a person's mental health may be the increased psychosocial benefits of relaxation, stress reduction mood elevation and increased sociability obtained from moderate alcohol consumption (Peele and Brodsky, 2000). One aspect of sociability that may be important is that time spent consuming alcohol with work colleagues enables the development of informal mechanisms such as ‘networking’ through additional social time spent with colleagues and this may also act as a ‘signal’ to senior members of staff regarding motivation and commitment to job (MacDonald and Shields, 2001).

In the remainder of this section we examine the results of studies that have estimated wage equations that include indicators of alcohol consumption. In some cases the levels of consumption were not specified in detail while in others the units of measure of alcohol consumption may be highly varied. In order to compare and combine these results we need to employ a series of econometric techniques to allow comparisons across the various studies under consideration.

2.1 Single Indicator Variable Specification

Peters (2004), Peters and Stringham (2006), Tekin (2004), van Ours (2004) and Zarkin et al. (1998) all present results of estimated log-wage equations in which a single dummy variable has been included to distinguish between drinkers and abstainers in a regression on the log of wages. Table 1 presents a summary of the different samples these papers employ. Results are presented for both males and females using data from the USA, Canada, Russia and the Netherlands. Overall these papers use a consistent definition of alcohol consumption, although the proportion of the sample who are non-drinkers varies considerably across the surveys. Using ordinary least squares (OLS) estimates, the average of the coefficient on the dummy variable for alcohol consumption is around 0.9 for both the male and female results. However, this is reduced considerably when the wage equations can be estimated using fixed effects.

Table 1.  Summary of Papers in which a Single Overall Effect of Alcohol Consumption is Either Estimated or Calculated Based on Multiple Indicator Results.
PaperDummy for alcoholt-statDataSampleAbstainer%Abstainers in sampleAverage ageFixed effects
Peters (2004) Males0.0412.25National Longitudinal Survey of Youth 1979–1994Full-time workers 17–25 in 1982 and 29–37 in 1994Did not drink any alcohol in past 30 days 526.1 
Peters (2004) Males2.0e-40.01    526.1×
Peters (2004) Females0.0583.53   1026.1 
Peters (2004) Females0.0140.96   1026.1×
Peters and Stringham (2006) Males0.1043.59Canadian General Social SurveyFull-time workersAny occasion to use any alcoholic beverages? Abstainers answer no.2238.5 
Peters and Stringham (2006) Females0.1494.67   3238.6 
Tekin (2004) Males0.1051.12Russian Longitudinal Monitoring Survey Nov 1994 to Dec 2000EmployedDid not drink any alcohol in past 30 days2638.0 
Tekin (2004) Males0.0711.87   2638.0×
Tekin (2004) Females0.1642.19   4337.6 
Tekin (2004) Females0.1011.80   4337.6×
van Ours (2004) Males0.1282.62001 CentER data Netherlands26–55 years old who worked 10–60 hours per weekDid not drink any alcohol in past 30 days 741.4 
van Ours (2004) Females0.0370.9   1638.5 
Zarkin et al. (1998) Males0.0672.681991 and 1992 National Household Surveys on Drug Abuse30–59 years oldDid not drink any alcohol in past 30 days3440.5 
Zarkin et al. (1998) Females0.0401.38   4840.6 
Overall single effect based on multiple indicator results
Auld (2005) Males0.1013.18Canadian General Social Survey 1985 and 1991Employed 25–59 years oldLess than one a month during last year2037.9 
Barrett (2002) Males0.0642.311989/1990 Australian Health Survey25–59 employed full-timeNever drank or did not have a drink in the month prior to survey1438.4 
Hamilton and Hamilton (1997) Males0.082 1.96Canadian General Social Survey 1985Employed 25–59 years oldLess than one a month or not at all during last year1835.6 
Lee (2003) Males and females0.129 3.24Australian Twin Registry 1980/1982 and 1988/1989Employed 20–64 years oldPast 12 months did not drink at all or in typical week2435.4 
MacDonald and Shields (2001) Males0.072 3.391992–1996 cycles of Health Survey of EnglandEmployed 25–65 years oldNot had a drink in last 12 months1043.1 
MacDonald and Shields (2001) Females0.036 2.88 Employed 25–60 years old 2440.8 
Soydemir and Bastida, 2006 Males0.15314.821994–1999 Border Epidemiologic Study of AgingEmployed 37–62 years oldDid not drink socially or in company of others7260.8 

The papers by Auld (2005), Barrett (2002), Hamilton and Hamilton (1997), Lee (2003), MacDonald and Shields (2001) and Soydemir and Bastida (2006) all present results in which multiple discrete indicators of alcohol consumption are included in the estimated wage equation. A summary of these papers is reported in Table 2. Auld (2005), Barrett (2002) and Hamilton and Hamilton (1997) define the multiple indicators using similar definitions and obtain comparable results. Moderate drinkers were found to earn significantly more than abstainers, whereas heavy drinkers were not significantly different from moderate drinkers. By using data from the Australian Twin Registry Lee (2003) finds that light and moderate drinkers earn more than abstainers; when he accounts for the twin dimension in estimation some of this positive wage premium can be explained by genetic endowments and family background. Using six indicators of alcohol consumption to reflect the different categories of drinking intensity, MacDonald and Shields (2001) find evidence of an inverse U-shaped relationship between drinking and occupational attainment. Soydemir and Bastida (2006) also provide some evidence of an inverse U-shaped relationship for their sample of male Mexican Americans.

Table 2.  Summary of Papers that Use Multiple Indicators to Represent Alcohol Consumption.
Auld (2005)Barrett (2002)Hamilton and Hamilton (1997)Lee (2003)MacDonald and Shields (2001)bSoydemir and Bastida (2006)
  1. aAs not many individuals fell into the heavy category this group was excluded from estimation.

  2. bAlso present results using five indicator variables based on frequency of drinking, but similar results are obtained to the seven indicator variables.

Canadian General Social Survey 1985 and 19911989/1990 Australian Health SurveyCanadian General Social Survey 1985Australian Twin Registry 1980/1982 and 1988/19891992–1996 cycles of Health Survey of England1994–1999 Border Epidemiologic Study of Aging
25–59 employed men25–59 full-time employed men25–59 employed men20–64 employed males/females25–65 employed males25–60 employed females37–62 employed Mexican Americans
Less than one a month during last year
Not abstainer or heavy
At least one a week and had eight or more in one sitting last week
Never drank or did not have a drink in the month prior to survey
Not abstainer or heavy
Had eight or more in one sitting during reference week
Less than one a month or not at all during last year
Not abstainer or heavy
At least one a week and had eight or more in one sitting last week
Abstention Past 12 months did not drink at all or in typical week
Light Past 12 months 1–2/month or less and between 1–12/week on average
Moderate Not abstainer, light or heavy
Per week
Very heavy
Per week
Very heavy
Based on number of drinks when drinking with others
Four categories, including none, 1, 2–3 and >3
   Heavya In past 12 months drank more than 1/day or every day and had at least 25/week   
Abstainers earn 9% less than moderate, Heavy statistically insignificant from moderateAbstainers earn 5% less than moderate, Heavy statistically insignificant from moderateAbstainers earn 7.45% less than moderate, Heavy statistically insignificant from moderateLight and moderate earn 8.9% and 13.9% more than non-drinkers. These values are reduced when twin dimension is accounted for in estimationMagnitude of estimated coefficients increases with drinking intensity for both males and females. Males: peaks moderate. Females: peaks moderate/heavyNone of the female alcohol variables was significant. For men only 2–3 category is positively significant

For each of the results presented in Table 2, a single overall effect was obtained by weighting each of the coefficients by the proportion of response in each category. This estimate can be interpreted as equivalent to the value of a single dummy variable for drinking. The implied estimates and t-statistics for the single dummy variable are presented in columns 2 and 3 of Table 1 so that a comparison can be made with the single indicator results.

2.1.1 Meta-analysis on Single Indicator Models

In order to perform a meta-analysis of these results we estimate a weighted least squares regression in which the dependent variable is defined as the estimated coefficient on the alcohol consumption dummy variable (as reported in Table 1) and the explanatory variables are defined by the studies' properties (see for example Stanley and Jarrell, 2005). The different characteristics of the models included gender of sample, type of estimation, the percentage of abstainers in the sample and the average age of the sample. Following Stanley et al. (2008) we also include as an explanatory variable the standard error of each estimated coefficient on the alcohol consumption dummy variable as a measure of the estimate's precision to account for potential publication selection bias. In estimation the observations are weighted by the inverse of the variance of each estimate. Results of this estimation are reported in Table 3.

Table 3.  Regression Analysis of Characteristics of Studies on Dummy Coefficient of Alcohol Consumption.
  1. Notes: Dependent variable: dummy coefficient on alcohol consumption defined in Table 1. t-statistics are calculated from heteroskedastic-consistent standard errors.

Constant 0.02893 1.73670.102
%Abstainers2 3.46E-05 2.40660.029
Fixed effects−0.0395−4.04730.001
Standard error 1.6890 3.99480.001

In examining the results in Table 3 we found that the gender dummy variable was insignificant and was subsequently excluded. The average age was found to be highly positively correlated with the percentage of abstainers in the sample and thus it was not used for this model. The estimated conditional mean is approximately 0.03 (constant). The estimated coefficient on the publication selection explanatory variable (standard error) is positive and significant. The inclusion of fixed effects in the estimation appears to mitigate the magnitude of the dummy coefficient on alcohol consumption. This may be due to the fixed effects models' inclusion of individual specific effects which allow for time invariant heterogeneity. Thus some of the wage bonus from drinking may be accounted for by unobserved heterogeneity. The marginal effect and corresponding 95% confidence interval for %Abstainers is illustrated in Figure 1 (see Lye and Hirschberg (2009) for the interpretation of this type of figure). The results indicate that the marginal effect is 0 when %Abstainers in the sample is less than 28 and only becomes positive when %Abstainers is greater than 28.

Figure 1.

The Estimated Marginal Effect for the Percentage of the Sample Who Abstain from Alcohol Consumption with the 95% Confidence Interval.

The abstainers in the sample may be of two types, those that have never had a drink and those that are ex-drinkers. Ex-drinkers may have health problems, partly or wholly as a result of past drinking patterns and as a result have become abstainers (Heien, 1996a). Also as the sample becomes older, there may be an increase in ex-drinkers being defined as abstainers in the sample. In the medical literature, Fillmore et al. (2006) concluded that the cardiac protection associated with alcohol consumption may be over-estimated due to the inclusion of ex-drinkers in the abstainers. This may also be true in the wage models – the higher the proportion of abstainers in a sample indicates the higher the proportion of ex-drinkers that are now counted as abstainers. Possibly these individuals have stopped drinking due to negative impacts on their productivity and their potential for job mobility as modelled by Tchernis (2010). Thus the greater the number of ex-drinkers the greater the difference between the human capital of those in the still-drinking group and those that have had to stop drinking and thus there is a greater impact on the coefficient value.

2.2 The Wage Equation with a Quadratic Specification

When the actual amount of alcohol consumed was included in the survey a number of researchers have attempted to establish a continuous relationship between levels of alcohol consumption and wages earned. In many cases it has been proposed that the relationship between alcohol consumption and wages earned is non-linear in nature. This would follow if one expects that low to moderate levels of consumption are beneficial but that high levels would be detrimental. In order to allow for such an inverted U-shaped relationship between alcohol and wages the standard approach has been to estimate a model with a quadratic specification for the influence of alcohol consumption. Thus a typical specification is given by


These models have been estimated using a range of estimation techniques including OLS and instrumental variables (IV). In accordance with empirical evidence of the positive income elasticity for alcohol (e.g. see Fogarty, 2010), IV methods have been used to account for the possibility that the level of alcohol consumed may be considered endogenous in a wage equation.

Most papers that use cross-section data find support for the hypothesis of an inverse U-shaped relationship between alcohol consumption and earnings (see for example French and Zarkin, 1995; Heien, 1996a, b; MacDonald and Shields, 2001). Table 4 reports a summary of the papers that have used a quadratic functional form to define the relationship between earnings and alcohol consumption. In the majority of cases, the coefficient on the linear term is positive and the coefficient on the squared term is negative, thus implying that an inverse U-shape relationship exists. Although the absolute value of the t-statistics associated with the coefficient of the quadratic term range from 1.06 to 10.17, in all but one case one can reject the null hypothesis that the coefficient of the quadratic term is equal to 0 at the 5% level of significance.

Table 4.  Summary of Papers that Use a Quadratic Specification to Define Relationship between Alcohol Consumption and Earnings.
PapersDataEstimationEx-drinkerModel typeinline imageinline image
  1. Notes: t-statistics in parentheses. 3SLS, three-stage least squares.

French and Zarkin (1995)Uses a sample of randomly selected employees at four worksitesOLS estimation used ‘Full’– no human capital variables included in estimationIdentifiedFull effect – unbounded model5.44 × 10−5−4.22 × 10−8
  IdentifiedDirect effect-unbounded model6.71 × 10−5−3.99 × 10−8
Alcohol consumption equals total number of drinks during past year‘Bounded’ uses bounded influence estimation to account for outliersIdentifiedFull effect – bounded model1.61 × 10−4−1.31 × 10−7
  IdentifiedDirect effect-bounded model9.72 × 10−5−5.56 × 10−8
Heien (1996a, b)Uses data from the National Household Survey on Alcohol Use (NHSA) for 1979 and 1984 and data from Quality of Employment Survey (QES)OLS estimationIdentified1979 – NHSA70.62−0.454
 Non-linear 3SLS with religious preference variables as instrumentsIdentified1979 – NHSA332.9−3.1
 OLS estimationIdentified1984 – NHSA48.54−0.235
Alcohol consumption equals number of drinks per monthNon-linear 3SLS with religious preference variables as instrumentsIdentified1984 – NHSA128.2−0.61
 OLS estimationNot identifiedQES–75.4−0.55
MacDonald and Shields (2001)Uses samples of employees from the Health Survey for England between 1992 and 1996OLS estimationNot identifiedMales0.0033−0.000038
Alcohol consumption equals number of drinks per month. 1 unit = 8 g of alcoholIV – uses binary indicators for long-term non-acute illnesses as instrumentsNot identifiedMales0.0278−0.000519
 IV – uses binary indicators for whether or not the interviewee's mother or father smoked as instrumentsNot identifiedMales0.0107−0.00026
 IV – uses binary indicators based on individual's self-assessment on how much they drink as instrumentsNot identifiedMales0.0103−0.000144
Lye and Hirschberg (2004)Uses data from the 1995 Australian National Household
Alcohol consumption equals millilitres per week.
IV using country of birth, total exercise time, body mass index and state dummy variables as instrumentsNot identifiedMales – smokers−0.0011630.000003
  Not identifiedMales – non-smokers0.005659−0.000009

2.2.1 A Descriptive Comparison of Quadratic Specifications

For each specification that reported a negative sign on the quadratic term in Table 4 the estimated value of the extremum and corresponding 95% confidence intervals are reported in Table 5. To make a comparison across these papers and with the medical literature the turning points and confidence bands have been expressed as grams per day. The estimated turning points are found by setting the first derivative of the log wages with respect to alcohol consumption to zero and solving for the turning point in grams of alcohol (ω) as


where λ is defined as the conversion factor required to transform the units used in each study to grams per day; these are also reported in Table 5.

Table 5.  Extremum Estimates for Wage Equation Results (in Grams of Alcohol per Day).
Paper Modelλinline imageConfidence bounds – 95%
Lower boundUpper bound
  1. Notes: Both French and Zarkin (1995) and Heien (1996a, b) use samples that consist of both males and females. To make the conversion to grams per day the standard drink size per country used definitions from the International Centre for Alcohol Policies. If the original results were reported as per week values these were divided by 7, monthly by 28, yearly 365.

French and Zarkin (1995)1Full – UB0.0424.72−135.8369.77
2Direct – UB0.0432.25−70.6156.63
3Full – B0.0423.57−2.1849.24
4Direct – B0.0433.53−14.9972.18
Heien (1996a, b)51979 – OLS0.5038.9215.0095.36
61979 – IV0.5026.880.23134.67
71984 – OLS0.5051.6616.23199.45
81984 – IV0.5052.507.4210.99
MacDonald and Shields (2001)10OLS – males1.1449.6242.0459.16
11IV1 – males1.1430.617.0494.21
12IV2 – males1.1423.52−5.2378.45
13IV3 – males1.1440.8715.50114.88
15OLS – fem1.1445.7137.7158.67
16IV1 – fem1.1431.207.06119.73
17IV2 – fem1.1415.86−3.0070.56
18IV3 – fem1.1427.1814.9457.42
Lye and Hirschberg (2004)14Non-smokers – males0.1135.9329.0078.88

One can interpret the turning point as that level of alcohol consumption past which further consumption leads to negative impacts on wages and thus would be detrimental to the return on human capital. Up to the turning point there is a positive though diminishing benefit gained from additional consumption of alcohol; after that point the marginal effect is zero and ultimately may become negative. Although variations in the values of the estimated turning points have been found they are comparable. While the averages of the estimated turning points across countries in the studies are similar (USA = 35.37 g, UK = 33.07 g and Aus = 35.93 g), the turning points from studies estimated using OLS rather than IV are on average higher (OLS = 37.14 g, IV = 34.62 g) as are those that are based on samples of men only in comparison to samples of females only (men = 36.11 g, females = 29.99 g), whereas the average based on the samples which use both males and females is very similar to those based on males only (males and females = 35.37 g).

To determine the confidence bounds of the turning point defined by a ratio of estimated coefficients it is necessary to define the distribution of the ratio of normally distributed random variables. In the following analysis we follow the method proposed by Hinkley (1969) to define the cumulative density function (CDF) of the ratio of normally distributed random variables. Plassmann and Khanna (2007) demonstrate the use of this method for the determination of the turning point in a quadratic regression.

If inline image and inline image are normally distributed with mean β1 and β2 respectively and variance–covariance matrix


the bounds on the 95% confidence intervals are defined by the exact CDF of inline image defined by


where the components are defined as


This expression can be evaluated with widely available computer programs designed to compute the joint CDF of a bivariate normal distribution. In addition to the published results for each model we also estimated the corresponding covariance between the two parameters based on the information given in the paper including the reported delta standard errors (French and Zarkin, 1995; MacDonald and Shields, 2001), an F-statistic for a partial model (Heien, 1996a, b) and the actual data (Lye and Hirschberg, 2004). Note that the distribution of inline image is generally not symmetric. In addition, the ratio of the estimates is the median of inline image.

The 95% confidence bounds are reported in Table 5 and plotted in Figure 2. The confidence intervals are wide and indicate that there may be a positive effect of alcohol on wages. However, when the lower bound is less than zero, the relationship between wages and alcohol becomes insignificant. Conversely, when the level of consumption exceeds the upper bound the relationship between alcohol consumption and wages becomes negative. There are six cases for which the lower bound is negative and hence we cannot reject the hypothesis that there is no level of alcohol consumption for which there is a positive relationship. In all cases we find that the upper bounds fall well above recommended daily guidelines of about 28 g per day for men and 14 g per day for women (based on US guidelines, see Furthermore, they are generally higher than the reported maximum levels of alcohol consumption in the various studies. The higher the upper bound the greater the evidence that the relationship may be more of an inverted J-shape which would be present if the impact of alcohol consumption plateaus. This may well be due to the exclusion of very heavy drinkers in these samples.

Figure 2.

A Plot of the Turning Points inline image and the 95% Confidence Interval by Model Number (the Bounds are Restricted to the Range −20 to 100).

This conclusion is consistent with the argument that heavy or abusive drinkers engage in compensatory measures to conceal the problem so that problems at work may only be evident at the very late stages of alcohol abuse (Moore et al., 2000). It is also consistent with the finding of a negative relationship between problem drinking and labour force participation. Thus these studies may be considered as only partial analyses which do not account for the sample selection bias generated by the lack of labour force participation by those individuals who consume large amounts of alcohol.

2.2.2 Meta-analysis of the Quadratic Specification

In order to conduct a meta-analysis on these results we examine the lower bound of the 95% confidence interval for the turning point. This value can be interpreted as the limiting value of alcohol consumption that results in a positive impact of alcohol use on wages. The lower bound of the estimated turning point is the dependent variable in a weighted least squares regression with the characteristics of the studies, including estimation technique, gender and whether ex-drinkers were identified separately from abstainers, as explanatory variables. We also include the t-statistic on the quadratic term of each regression as a measure of precision to account for potential publication selection bias based on the common practice to take the significance of this t-statistic as evidence of the presence of an inverted U-shape relationship between earnings and alcohol consumption. The observations were weighted by a measure of the precision of the lower bound, defined as one fourth the inverse of the length of the 95% confidence interval – a value which would correspond to approximately the inverse of the standard deviation of a normally distributed random variable. The results of the estimation are given in Table 6.

Table 6.  Regression Analysis of Characteristics of Studies on Estimated Lower Bound of the Turning Point.
  1. Note: t-statistics are calculated from heteroskedastic-consistent standard errors.

Constant 17.1360 7.88070.049
IV estimation−16.8471 5.28250.007
Ex-drinker distinction used−30.337810.54890.013
Female-only sample  4.4249 2.99880.164
t-statistic on estimated quadratic coefficient −2.4819 0.78430.008

The results in Table 6 indicate that the models using IV to account for the potential simultaneity of alcohol consumption and wages result in lower bounds than the models using OLS. On average when IV estimation is used, the lower bound for the turning point is indistinguishable from zero and thus there is no level of alcohol consumption that has a positive impact on wages. Similarly, when ex-drinkers are accounted for we find that there is no level of consumption at which alcohol has a positive impact. The estimated coefficient for the explanatory variable t-statistic on estimated quadratic coefficient is negative and significant which indicates the potential presence of publication selection bias.

2.2.3 A Comparison with the Medical Evidence

In Table 7 we summarize the results of studies from the medical literature in which the daily intake of alcohol is related to health outcomes. From this table we can conclude that on average the maximum beneficial consumption level for health purposes is slightly less than the 15 g/day and is thus of similar magnitude to the estimate of the conditional mean (constant) in Table 6. We also find that when ex-drinkers are accounted for the evidence from both the economics literature and the medical literature indicates that alcohol consumption is not beneficial.

Table 7.  Level of Alcohol Consumption at which Mortality is Minimized (Grams of Alcohol per Day).
  1. aTakes into account misclassification of abstainers and/or occasional drinkers.

White (1999)– USA1950–19959.90 g/per day3.73 g/per day
White (1999)– UK1950–199516.59 g/per day 
Corrao et al. (2000)1966–199825 g/day10 g/day
Bagnardi et al. (2003)1966–20006–7 g/day5 g/day
Gmel et al. (2003)until 20009–15 g/day3–13 g/day
Fillmore et al. (2006)a1950s to mid-20042 g/day2 g/day

Again as we found in Section 2.1, the economics literature finds a positive impact of alcohol consumption on wages but that there may be mitigating factors to influence this relationship. One such factor is the identification of past-drinkers in the sample without which there will be a misclassification of abstainers. A second factor is the need for the appropriate accommodation of sample selection bias caused by the absence of heavy drinkers in the samples – this may be why the upper bounds for the turning points are so high. And third it has been shown that the endogeneity of alcohol consumption with respect to wages needs to be considered in the choice of the estimation procedure. Other authors have proposed that omitted variable bias may also be present in these models. In the next section we investigate some proposed solutions to this ‘puzzle’.

2.3 The Alcohol–Income Puzzle

Auld (2005, p. 505) refers to such finding as we have investigated above – that moderate drinking increases wages – as the ‘alcohol–income puzzle’. He concludes that health is not driving the wage penalty associated with abstention because there is very little change in the result when other health measures are omitted from the specification. Some authors have concluded that the puzzle is due to omitted variable bias. Thus it is not the drinking itself that influences wages but that there may be some attribute of individuals that is unmeasured and correlated with the decision to drink that has a positive influence on human capital and hence the wages earned (see for example Peters, 2004). It has also been suggested that this attribute is a personality trait that contributes to both drinking behaviour and higher earnings (see for example Auld, 2005, p. 514; Soydemir and Bastida, 2006, p. 426).

2.3.1 Personality Traits and Earnings

Using data from the Wisconsin Longitudinal Study, Mueller and Plug (2006) find that men who were antagonistic, open to experience and, to a lesser extent, emotionally stable received higher earnings than otherwise similar men. In the same sample it was found that women received a wage premium for being more conscientious and open to experience.

Groves (2005) examined longitudinal data on women from the National Child Development Study that follows the lives of all children born during week 1 of 1958 in England, Scotland and Wales. The two personality variables used describing aggression and withdrawal were found to have a negative and statistically significant influence on wages. Nyhus and Pons (2005) use data from the DNB Household Survey based on a large sample of the Dutch population. They find that emotional stability is positively associated with the wages of both men and women while agreeableness is significantly associated with lower wages with women.

2.3.2 Alcohol and Personality Traits

Lipton (1994) found that depression scores were higher for abstainers, light drinkers and heavy drinkers compared to moderate drinkers. Cook et al. (1998) using a sample of municipal employees in a rural area found that abstainers were more introverted than drinkers. They also found that alcohol consumed is correlated positively with sociability and extraversion but negatively with conscientiousness and willingness to conform. On the basis of two studies using participants who were undergraduate students at a large Midwestern university enrolled in a psychology course, Walton and Roberts (2004) found individuals who abstain from substance abuse to be highly conscientious and a bit inhibited. Using Wave 5 of the HILDA longitudinal survey of Australian households, Losoncz (2007) found non-drinkers scored lower on extraversion than moderate drinkers. Respondents who never consumed alcohol scored the lowest on openness to experience. High quantity drinkers also scored lower on this scale than moderate drinkers. Agreeableness, conscientiousness and emotional stability were highest among moderate drinkers and lowest among high quantity alcohol intake respondents. One can summarize these results to indicate that those on the tails of the distribution of alcohol consumption are less extraverted and less open to new experiences than those in the middle group. Thus if drinking behaviour is a marker for personality types then what can be said about the relationship between personality and human capital?

Because most research has either used alcohol consumption or personality traits in models of human capital, the measures of alcohol consumption may serve as a proxy for these personality traits. From the research findings where attempts have been made to measure personality traits, we may conclude that in both men and women the relationships between alcohol consumption and personality and between personality and wages suggest that alcohol use in the wage equation may be accounting for such personality factors as emotional stability and openness to experience. In addition, for women the alcohol consumption may be an indication of conscientiousness as well. The emotional stability component appears consistent with the logic that increased psychosocial benefits such as relaxation, stress reduction and mood elevation justify the inclusion of alcohol consumption in wage equations. It is also consistent with a report by the UK Mental Health Foundation (2006) which finds that daily drinkers cite the alleviation of anxiety and depression as a common reason for drinking. However, there are some personality traits that correspond to alcohol consumption that may not be related to improvements in human capital. For example, there was little evidence for extraversion as an important personality factor in the wage equation but it has been found to be related to alcohol consumption. Thus alcohol consumption may prove to be an imperfect proxy for all personality traits that have a positive influence on human capital.

Interestingly this imperfection in the use of alcohol consumption as a marker for human capital improving personality traits has been found in research conducted by French and Zarkin (1998). They examine the relationship between symptoms of emotional and psychological problems and earnings among employees at a large manufacturing worksite. They concluded that workers with three or more emotional/psychological symptoms had 13% lower earnings than workers without these symptoms, all else equal. They also included measures of alcohol use in the specification. It was found that the number of days drunk in the past year was significantly related to earnings even after controlling for emotional symptoms as was fair/poor health status. However, daily drinking, defined as drinking 20 or more days in the past 30, was not significant. Thus the measures of personality traits dominated the explanatory effects of moderate alcohol consumption.

In Sections 2.1 and 2.2, we found that moderate alcohol consumption appears to positively influence wages. In this section we have examined the literature that may provide clues to this link. First we discussed the findings that physical health and alcohol may have an independent impact on earnings since leaving them out of the model seems to have little effect on the factor left in. We then discussed the literature that finds that alcohol influences certain aspects of mental health but not all of these aspects are influential on earnings. Thus when measures of emotional stability are included along with alcohol consumption in a model it has been found that the alcohol consumption does not appear to influence earnings. This would indicate that moderate alcohol consumption is serving as a proxy for the trait of stability. This would also allow us to explain why the proportion of ex-drinkers in the sample is crucial to the findings. It has been documented that ex-drinkers are more likely to suffer from depression. Manninen et al. (2006) and Marlowe (2002) present results that suggest that depression has a major impact on workplace efficiency. If ex-drinkers are present in those classified as abstainers this may lower the average level of emotional stability of this group compared to moderate drinkers and thus appears to increase the influence of alcohol on wages.

3. Other Potential Effects of Alcohol Consumption on Human Capital

Although our main focus in this paper is on the contemporary influence of alcohol consumption on measured labour market outcomes there may be other influences of drinking behaviour that would not show up in the data used in the studies examined above. In particular, there may be long-term influences at work as well. It may be that alcohol consumption at one point in time may have long-term effects via the impact on the acquisition of human capital. In addition, the effects of alcohol abuse may be such that the studies we have examined above will never be applied to the population whose use of alcohol is abusive.

3.1 The Effects of Alcohol Consumption on Education

Benham and Benham (1982) and Yamada et al. (1996) report a negative effect of drinking on educational attainment. Cook and Moore (1993) use data from the 1979–1988 National Longitudinal Survey of Youth to estimate structural equations that relate schooling to drinking and drinking to alcohol price and availability. They conclude that heavy drinking in high school reduces the average number of years of schooling completed after high school. However, these results have been criticized on the basis that the IVs used to account for the endogeneity of the drinking decision rely solely on cross-state variation (Dee and Evans, 2003, p. 179).

More recently, Koch and Ribar (2001), using data from the 1979–1990 panels of the National Longitudinal Survey of Youth on same-sex sibling pairs, conclude that delaying drinking onset by a year increases schooling by no more than 0.47 years for men and 0.36 years for women. Dee and Evans (2003) use matched cohorts from the Monitoring the Future Surveys and 1960–1969 birth cohorts in the Public Use Microdata sample. Their results from two-sample IV estimation indicate that teen drinking has a statistically insignificant effect on college entrance, college completion and on the probability of completing high school. Renna (2007) uses the 1979 National Longitudinal Survey of Youth to examine the consequences of binge drinking during senior year of high school. A system of two probit equations is estimated for drinking behaviour and on-time educational attainment defined as completing 12 years of education before 19. The results indicate that bingeing decreases the probability of graduating on time for both men and women. Bray (2005) develops and estimates a theoretical model of wage determination that examines the relationship between alcohol use, human capital accumulation and wages using the 1979 cohort of the National Longitudinal Survey of Youth. His results suggest that (page 301) ‘… moderate alcohol use has a positive effect on the returns to education or experience, and therefore on human capital accumulation, but that heavier drinking reduces this gain slightly’.

Other studies find a significant negative relationship between drinking and measures of education that reflect the quality of human capital accumulation. Williams et al. (2003) use data from the 1993–1999 Harvard School of Public Health's College Alcohol Study to estimate a three-equation model to examine the relationship between study levels, drinking and grade point average (GPA) using 2SLS to allow for the endogeneity between drinking and hours spent studying. The direct effect of drinking on GPA is positive but outweighed by a negative indirect effect. Thus the combined direct and indirect effect of drinking on GPA is negative via a reduction in the hours spent studying. DeSimone and Wolaver (2005) use the 2001 and 2003 Youth Risk Behaviour Survey on high school students. Their results indicate that while binge drinking has a significant negative impact on GPA, there is little evidence to expect moderate drinking to have a detrimental effect on school performance. Wolaver (2007) simultaneously models drinking behaviour, GPA and study hours to determine the effect of drinking on study hours and grades using generalized methods of moments and the 1993 and 1997 Harvard College Alcohol Study. For both men and women, binge drinking is shown to lower predicted grades.

Thus while these results suggest little effect of drinking on the number of years of education, binge drinking may have a negative impact and there is also a growing body of literature that suggests there is a negative effect of drinking on school performance. However, for the most part the studies of labour market outcomes as a function of alcohol consumption are limited in that the time frame over which the level of alcohol consumption is measured is the current level of activity and little is collected on the past history of alcohol consumption that may have inhibited one's educational performance.

3.2 Problem Drinking and Labour Market Outcomes

Another area of research concerning labour market outcomes and alcohol relates to the consequences of problem drinking. These studies examine the relationship between alcohol abuse and labour market outcomes by including indicators of alcohol dependency or alcohol abuse in addition to other health components of human capital in order to determine labour market outcomes. These indicators are formed using surveys which are specifically designed to identify symptoms of alcohol dependency. Alternative definitions of problem drinking are based on measures of quantity and frequency of alcohol use, including binge drinking and definitions of at-risk drinking.

The usual form of these analyses are regressions in which the dependent variables are either a measure of income or an indicator variable to describe the outcome of participation in the labour market with the indicator of problem drinking used as a regressor. Findings from a range of papers and different datasets have consistently found a negative relationship between problem drinking and positive labour market outcomes (higher wages and less unemployment). Most of these analyses have given more emphasis to male workers than female ones on the basis that alcohol abuse is much more prevalent for males.

Using data on males aged 22–64 from the 1980–1981 Wave 1 of the New Haven Epidemiological Survey and treating alcoholism as an exogenous variable, Mullahy and Sindelar (1993) present evidence that alcoholism may affect income more by restricting labour market participation than by affecting the wages of workers. Mullahy and Sindelar (1996) examine the relationship between measures of problem drinking and employment and unemployment. The analysis is based on the 1988 Alcohol Supplement of the National Health Interview Survey. Using IVs to test the relationship between problem drinking and the propensity to be employed the authors find that, for both men and women, problem drinking results in reduced employment and increased unemployment. However, the results are not statistically significant. Terza (2002) using the same data as Mullahy and Sindelar (1996) for the male subpopulation estimate a multinomial logit model with an endogenous treatment effect. Problem drinking is found to have a positive effect on the probability of unemployment and a negative effect on the likelihood of being employed. The latter estimate is statistically significant. MacDonald and Shields (2004) using data from the Health Survey of England estimate a bivariate probit that accounts for the endogeneity of problem drinking. They conclude that being a problem drinker leads to a substantial reduction in the probability of working for males aged 22–64. The results are found to be robust to a number of definitions of problem drinking. Using data from a large Finnish health survey, Johansson et al. (2007) estimate a bi-probit model and find that both men and women have lower employment probabilities. By accounting for the endogeneity of problem drinking the measured negative effect of alcohol consumption is increased in these models.

Although two papers have found little evidence that alcohol problems have a negative impact on labour supply, both of these papers have some serious limitations. Kenkel and Ribar (1994) use the 1989 panel of the US National Longitudinal Survey of Youth and define problem drinking using the DSM-III diagnostic criterion for alcohol abuse and dependence as well as additional measures based on information on individuals' drinking habits. In estimation, simultaneity and heterogeneity are accounted for via IVs. However, labour supply is defined as hours worked and is thus not really a true consideration of labour participation. Furthermore, the sample used is relatively young with no respondents older than 31.

A more recent study by Feng et al. (2001) using data drawn from a sample of prime age men and women from six southern states in the USA found a similar conclusion. Problem drinking is defined using DSM-IV criteria and a range of potentially problematic drinking patterns. However, there are limitations to this analysis including a non-representative survey, the information from the survey is self-reported via telephone interview, and the standard IVs used for problem drinking could not be used because the data did not contain information such as a history of living with alcoholic parents.

In the analysis described above, Kenkel and Ribar (1994) do find evidence that alcohol problems have a negative effect on the earnings of men and women. Two recent papers have also examined the presence of a negative relationship between earnings and problem drinking. Using data on males and females from the 1989 and 1994 waves of the National Longitude Survey of Youth, Jones and Richmond (2006) estimate propensity scores that are sub-classified into quintiles. Earnings are compared between those who abuse and those who do not across the different quintile groups. They find alcoholism to have a negative impact on earnings that may become more pronounced over the life cycle. Keng and Huffman (2007) use data on males and females from the National Longitudinal Survey of Youth. Defining binge drinking as the number of occasions when an individual consumed six or more drinks in one session during the past 30 days they conclude that individuals younger than 40 who engage in binge drinking have significantly lower earnings.

4. Conclusions

In the labour market, traditional measures of human capital stock in earnings functions are completed schooling and work experience. However, since Grossman (1972) health can also be viewed as a form of human capital. There are now numerous papers that have examined the specific relationship between alcohol consumption and outcomes in the labour market. A rationale for these papers is the medical literature that has found positive influences from moderate alcohol consumption.

In the economics literature there are a number of papers in which a wage equation has been estimated with a single dummy variable indicating alcohol consumption or not or multiple indicators of alcohol consumption at different levels. The re-evaluation of these studies indicates that, while alcohol consumption has a positive influence, it could partially be explained by unobserved heterogeneity in the sample and by classifying ex-drinkers who may be past heavy drinkers as abstainers. This result mirrors the findings from the medical literature where alcohol consumption is treated as a dichotomous variable.

We have also investigated the nature of the studies which find an inverse U-shaped relationship between various levels of alcohol use and job performance. These results were re-evaluated by examining the estimated turning point and their corresponding confidence interval. We find that the lower bounds of the confidence interval tend to be consistent with the levels of alcohol consumption observed in the medical literature. However, there is not as much evidence for an inverse U-shape as there is for an inverted J-shape since in most cases the upper bound of the turning point is out of the range of consumption. Thus, although there is a positive relationship with alcohol little evidence is present to conclude that there is a downward portion in the relationship. Perhaps this is indicative that heavy or abusive drinkers conceal their problem at work. Or alternatively, these studies can only be interpreted as partial analyses in that they do not account for potential sample selection bias caused by the lack of labour force participation by the heavy or abusive drinker.

A common conclusion by many researchers is that the wage bonus associated with drinking may be due to omitted variable bias – that personality traits are missing from the analysis. A review of the literature that examines the relationship between alcohol consumption and personality and the literature that studies the impact of personality traits on earnings suggests that the drinking wage implied bonus may be accounting for such personality factors as emotional stability.

Another related area of research is the examination of the indirect effects of alcohol consumption on traditional human capital variables. A set of these studies focuses on problem drinkers and their labour force participation, and for those in the labour force, on their earnings. Another collection of studies has concentrated on the relationship between alcohol consumption and earnings for individuals with a relatively stable employment pattern. In addition, there is evidence from some studies that binge drinking may negatively impact on the number of years of education completed. Findings also indicate that drinking may affect the quality of school performance as measured by GPA scores. A consistent finding is that those individuals who are problem drinkers are less likely to be in the labour force and, for those that are, there is a negative impact on earnings that may become more pronounced over the life cycle.

In this survey we have demonstrated that a close examination of the literature on alcohol consumption and wages indicates that there may be cumulative impacts of alcohol consumption that are not accounted for in the use of current consumption patterns alone. This was shown in the health literature by the discovery that the proportion of ex-drinkers that are misclassified as non-drinkers is important, and also in the literature that identifies the poor education outcomes of individuals that drink heavily at an early age. Thus the impact of alcohol consumption on health could also be considered as an element of a person's intangible investment in human capital as discussed in Folloni and Vittadini (2010). Furthermore, the impact of a person's life history of such health related activities such as drinking, smoking, diet and exercise could be used to identify the health related aspects of their worklife expectancy as defined by Millimet et al. (2010). An alternative approach to the use of regression analysis for measuring the impact of alcohol would be the latent variable approach proposed by Lovaglio (2010) where the administrative data used in this research could be complemented by health records.


We thank the anonymous referees as well as Jan van Ours for comments on the earlier draft of this paper. We also wish to thank the Department of Economics and Finance of La Trobe University and the Faculty of Economics and Commerce of the University of Melbourne for partial support of this research.