Objective. We evaluated the effects of tax increases on alcoholic beverages in 1983 and 2002 on alcohol-related disease mortality in Alaska.
Methods. We used a quasi-experimental design with quarterly measures of mortality from 1976 though 2004, and we included other states for comparison. Our statistical approach combined an autoregressive integrated moving average model with structural parameters in interrupted time-series models.
Results. We observed statistically significant reductions in the numbers and rates of deaths caused by alcohol-related disease beginning immediately after the 1983 and 2002 alcohol tax increases in Alaska. In terms of effect size, the reductions were –29% (Cohen's d = –0.57) and –11% (Cohen's d = –0.52) for the 2 tax increases. Statistical tests of temporary-effect models versus long-term-effect models showed little dissipation of the effect over time.
Conclusions. Increases in alcohol excise tax rates were associated with immediate and sustained reductions in alcohol-related disease mortality in Alaska. Reductions in mortality occurred after 2 tax increases almost 20 years apart. Taxing alcoholic beverages is an effective public health strategy for reducing the burden of alcohol-related disease.
Morbidity and mortality associated with consumption of alcoholic beverages constitute a substantial public health burden in the United States. An estimated 85 000 deaths per year are associated with drinking, including car crashes, other unintentional injuries, homicide, suicide, and a range of diseases, particularly those affecting the liver, pancreas, and heart.1–3 All states have taxes specific to alcoholic beverages; some of these taxes are imposed primarily to generate revenue,4 whereas others are ostensibly intended to promote public health and welfare by limiting alcohol consumption.5,6
More than 100 studies have examined the relationship between alcoholic beverage prices (or alcohol tax rates as a surrogate for prices) and various indices of sales or consumption of alcohol (see Babor et al.7 and Chaloupka et al.8 for recent reviews). With a few exceptions (e.g., Salomaa9), studies consistently find price or tax levels to be inversely related to sales or consumption of alcoholic beverages, with the magnitude of the effect in terms of elasticities ranging from –0.2 to –2.0, depending on population, methods, time period, and specific beverage.
Substantially fewer studies have examined the effects of alcohol prices or taxes on measures of alcohol-related morbidity or mortality, and of those that have, most have focused on injury rather than disease. Of 18 studies that examined price or tax effects on traffic crashes, 15 found that higher alcohol prices or taxes were associated with fewer crashes,10–24 but 3 found no effect.25–27 Sloan et al.25 found that alcohol prices were not related to falls, fires, or other unintentional injury rates, but Ohsfeldt and Morrisey28 found that higher beer taxes were related to lower rates of nonfatal industrial injuries. In addition, 8 studies (6 of which were conducted by 1 research team) found higher alcohol prices or taxes associated with lower morbidity or mortality from intentional injuries, including assault, homicide, suicide, child abuse, and spouse abuse.29–36 However, 1 of these studies found that higher beer taxes were not specifically related to levels of robbery or rape.31 Finally, a number of studies have found higher alcohol prices or taxes to be associated with lower rates of alcohol dependence37,38 and liver cirrhosis.23,39–42 In contrast, Schweitzer et al.43 did not find a significant relationship between alcohol prices and rates of alcohol dependence.
To elicit more information about the effect of alcoholic beverage taxes on disease mortality, we examined patterns of alcohol-related disease mortality in the state of Alaska over a 29-year period to determine whether 2 major increases in alcohol tax rates, 1 in 1983 and the other in 2002, affected alcohol-related mortality in the state.
The research design of the study can be represented as shown in Figure 1.
We utilized a time-series quasi-experimental design, which is similar to an experiment in that it is intended to identify the effect of an intervention while using comparison data series to rule out or control for possible alternative explanations for the effect.45 A good experimental design has a number of benefits: it can allow researchers to eliminate many confounding factors as a threat to a causal interpretation of an observed relationship without the need to identify, measure, and statistically control for all possible confounds; it can avert debates about which of the numerous specific control variables to include; and it can prevent researchers from introducing biases into the study by including only those control variables for which operational measures are available.
Given that Alaska is unusual compared with other states in terms of weather, population, economy, health, and many other factors, no other single state is optimal for comparison with Alaska. Therefore, we elected to use the aggregate of all other states as the comparison group. This comparison ensures that effects observed in Alaska are not caused by any changes in substantive factors (e.g., economy, society, policy) or measurement factors (e.g., diagnosis coding categories) that could affect observed alcohol-related mortality over time.
We further strengthened the study's design by examining 2 separate increases in tax rates occurring almost 20 years apart. This element of the study allowed us to test the replicability of observed effects, and it ensured that observed effects were not caused by unique circumstances not generalizable over time, thus ruling out the “contemporaneous history” threat to internal validity first articulated by Campbell and Stanley.46 Moreover, a simultaneous evaluation of the effects of the 2 tax increases allowed us to perform a long-term follow-up evaluation of the earlier tax increase.
Data on alcohol tax rates in Alaska were collected from a number of sources. The Alcohol Policy Information System provides summaries of US state and federal tax rates for multiple classes of alcoholic beverages, details on changes in rates, full legal citations, and text from the relevant codified statutes, starting with calendar year 2003.47–49 Data for the period prior to 2003 were collected by experienced research attorneys who used standard legal research methods to search the records of codified statutes in Westlaw databases (http://www.westlaw.com, fee-based membership) when available, and who searched law library hard-copy materials for earlier years.
Data on alcohol-related mortality outcomes were based on death-certificate data recorded by the National Vital Statistics System of the National Center for Health Statistics, which includes all deaths occurring within the United States. We obtained the complete annual data set containing 1 record on each deceased person for each year from 1976 through 2004 from the National Bureau of Economic Research.50 From these data we created quarterly counts of deaths stratified by underlying cause of death, first for Alaska and then for the other states as a group.
We cumulated counts by cause of death into 3 outcome variables. (1) Alcohol-caused mortality, which represents all deaths caused by diseases for which the alcohol-attributable fraction is 1.0 (e.g., alcoholic liver disease, alcohol-induced chronic pancreatitis, alcohol psychoses, alcohol abuse, alcohol dependence syndrome, alcoholic polyneuropathy, alcoholic cardiomyopathy, alcoholic gastritis, and acute alcohol poisoning). (2) Alcohol-related mortality, which represents all deaths caused by diseases with alcohol-attributable fractions 0.35 and higher but less than 1.0 (i.e., other cirrhosis; cholelithiasis; acute and chronic pancreatitis; malignant neoplasms of the mouth, pharynx, esophagus, liver, and breast; epilepsy; and cardiovascular diseases including hypertension, ischemia, arrhythmia, cerebrovascular disease, and ischemic and hemorrhagic stroke). (3) The third variable is the sum of the first 2 variables. Alcohol-attributable fractions for diseases, categorized by International Classification of Diseases (ICD)51 code, are based on Rehm et al.,52 English et al.,53 and Shultz et al.,54 which in turn were based on comprehensive reviews and meta-analyses of the literature (for the ICD codes used, see the table available as a supplement to the online version of the article at http://www.ajph.org). Finally, in addition to analyses of mortality frequency counts, we also used annual estimates of population from the US Census Bureau to calculate death rates per 100 000 population aged 15 years and older.55–58
Given the large number of repeated observations, we used an approach that combined a Box–Jenkins autoregressive integrated moving average (ARIMA) model with structural parameters,59,60
where Yi = 1 to Yi = 3 are the 3 outcome measures by quarter from t = 1 (first quarter 1976) through t = 116 (last quarter 2004); α is a constant; ω1 is the estimated effect of implementation of the 1983 alcohol tax increase; I1t is a step function equal to 0 before the 1983 tax change took effect and 1 after the change; ω2 is the estimated effect of implementation of the 2002 alcohol tax increase; I2t is a step function equal to 0 before the 2002 tax change took effect and 1 after the change; β is the estimated effect of Zt, the frequency (or rate) of alcohol-related disease in the comparison states; ψi is a vector of estimates that controls for outliers Xi; Θ is the first-order seasonal moving average parameter; ut is a random (white noise) error component; and B is the backshift operator such that B4(yt) equals yt–4. We used SAS version 9.1 Proc ARIMA (SAS Institute Inc, Cary, NC) to estimate all models, and we used a strict significance criterion of P at less than .001 to evaluate all models for effects of outliers. A maximum of 1 outlier was detected and controlled in any model.
For each Alaska outcome measure (alcohol-caused mortality, alcohol-related mortality, and the sum of both), 4 specific models were estimated: (1) frequency, (2) rate per 100 000 population aged 15 years and older, (3) rate per 100 000 population including the comparison states covariate, and (4) natural logs of rate per 100 000 population including the comparison states covariate. Frequencies were examined for initial evidence of effects. Rates per population were examined to assess whether observed differences were caused by population changes. Then the comparison states covariate was included to assess whether observed effects in Alaska may have been caused by other factors changing over time across states. The natural logs were modeled to ensure estimates were not affected by heteroscedasticity (changing variance in death rates over time). Finally, to statistically test whether the effects were permanent or temporary, we estimated an alternative specification of the time-series model that used a first-order transfer function on the differenced tax-change variable:
where ω is the estimated shift effect of It, representing the 1983 alcohol tax increase; δ is the estimated rate of decay of the initial effect; and B is the backshift operator, such that B(It) equals It-1. If the estimated value for δ in such a model is very close to unity, the effect is long-term and does not dissipate. If the estimate of δ is less than 1, the immediate effect decays over time. See chapter 3 of McCleary and Hay61 for a detailed description of transfer-function modeling that is accessible to nonstatisticians.
All of the models fit the data well and explained substantial proportions of the variance in deaths over time. Models of frequencies had R2 of 0.68 to 0.72 (after adjustment for number of degrees of freedom used in the model). Given that rates or ratios naturally have substantially higher measurement error do than the numerator frequencies alone, adjusted R2 for those models was somewhat lower, ranging from 0.13 to 0.33. Finally, parameter estimates in terms of change in frequency or population rate of death were transformed into 2 standardized metrics of effect: percentage change in the outcome based on the average of the 4 quarters immediately prior to a given tax rate change, and Cohen's d,62 the effect size in standard deviation units, calculated with the raw standard deviation of the outcome over the entire data series. Percentage change is a commonly used, easily understood metric of interest to prevention practitioners and policymakers, and the Cohen's d effect size permits comparison of effect sizes across a wide range of intervention and outcome domains.
Results show statistically significant reductions in the numbers of deaths caused by alcohol-related disease beginning immediately after the 1983 and 2002 alcohol tax increases in Alaska (Table 1). Estimated reductions in mortality are of clear substantive importance: the 1983 tax increase was followed by a –29% change in number of deaths (23 deaths averted per year), and the 2002 tax increase was followed by a –11% change in deaths (an additional 21 deaths averted per year). In terms of Cohen's d effect size, the reductions were –0.57 and –0.52 for the 2 tax increases, respectively, in chronological order. These effects are large enough to be clearly discernible from a plot of the data (Figure 2), especially for the 1983 policy change, for which a 19-year follow-up period is available.
|R2||Estimate (SE)||t||P||Percentage Change||Cohen's d|
|All alcohol-related and alcohol-caused mortality (AAF ≥ 0.35)|
|Effects of 1983 alcohol-tax increase|
|Rate per 100 000 population||0.33||–1.37 (0.50)||–2.77||.01||–22.5||–0.88|
|Rate per 100 000 population with other states covariate||0.33||–1.19 (0.52)||–2.28||.02||–19.6||–0.77|
|Log rate per 100 000 population with other states covariate||0.26||–0.19 (0.08)||–2.23||.03||–17.0||–0.79|
|Effects of 2002 alcohol-tax increase|
|Rate per 100 000 population||0.33||–1.23 (0.57)||–2.17||.03||–13.0||–0.79|
|Rate per 100 000 population with other states covariate||0.33||–1.36 (0.57)||–2.39||.02||–14.5||–0.88|
|Log rate per 100 000 population with other states covariate||0.26||–0.19 (0.09)||–2.08||.04||–17.1||–0.80|
|Subset analysis: alcohol-caused mortality (AAF = 1.0)|
|Effects of 1983 alcohol-tax increase|
|Rate per 100 000 population||0.30||–0.77 (0.38)||–2.05||.04||–20.8||–0.59|
|Rate per 100 000 population with other states covariate||0.31||–0.74 (0.38)||–1.96||.05||–19.9||–0.56|
|Log rate per 100 000 population with other states covariate||0.22||–0.13 (0.10)||–1.33||.19||–12.1||–0.40|
|Effects of 2002 alcohol-tax increase|
|Rate per 100 000 population||0.30||–0.59 (0.44)||–1.35||.18||–9.1||–0.45|
|Rate per 100 000 population with other states covariate||0.31||–0.55 (0.44)||–1.25||.21||–8.5||–0.42|
|Log rate per 100 000 population with other states covariate||0.22||–0.11 (0.11)||–1.00||.32||–10.8||–0.36|
|Subset analysis: alcohol-related mortality (AAF ≥ 0.35 < 1.0)|
|Effects of 1983 alcohol-tax increase|
|Rate per 100 000 population||–0.15||–0.44 (0.31)||–1.41||.16||–18.2||–0.53|
|Rate per 100 000 population with other states covariate||–0.13||–0.42 (0.36)||–1.19||.24||–17.7||–0.52|
|Log rate per 100 000 population with other states covariate||–0.04||–0.27 (0.13)||–2.21||.03||–24.0||–0.75|
|Effects of 2002 alcohol-tax increase|
|Rate per 100 000 population||–0.15||0.01 (0.36)||0.04||.97||0.4||0.02|
|Rate per 100 000 population with other states covariate||–0.13||–0.22 (0.37)||–0.58||.57||–7.4||–0.26|
|Log rate per 100 000 population with other states covariate||–0.04||–0.16 (0.13)||–1.20||.23||–14.9||–0.44|
Note. AAF = alcohol-attributable fraction. Results derived from autoregressive integrated moving average models combined with structural parameters. R2 adjusted for degrees of freedom used in model; calculated negative R2 in second subset analysis results from increased variance heterogeneity.
We analyzed the death rate to control for changes in size of the population over time, but this analysis did not appreciably change the estimated effects (Figure 3). Percentage reductions in alcohol-related mortality declined slightly to –23% for the 1983 tax increase and increased slightly to –13% for the 2002 increase. Estimated Cohen's d effect sizes for the tax increases were larger for the rate per 100 000 population measure than for the analyses of raw numbers of deaths (–0.88 and –0.79). These results confirm that observed effects of the tax policy changes are not attributable to overall changes in population.
We obtained a third set of effect estimates by adding the comparison states to the model, to determine whether observed alcohol-tax effects in Alaska reflected mortality reductions caused by any of a number of other possible factors operating across states. Again, there was no appreciable change in the estimates. The 1983 tax increase was associated with a –20% change in mortality (d = –0.77), and the 2002 increase was associated with a –15% change in mortality (d = –0.88). Thus, the comparison states did not experience the mortality declines that Alaska experienced, and the effects observed for Alaska cannot be attributed to broader trends or other factors experienced across the other states. Repeating this model on the natural logs of the death rate did not change the findings. Finally, the estimates and standard errors in Table 1 show that the magnitudes of the effects of the 2 tax changes were not significantly different from each other.
The second and third panels of Table 1 provide these same estimates of effect separately for 2 subsets of the overall outcome measure: (1) alcohol-caused mortality, for which the alcohol-attributable fraction is 1.0, and (2) alcohol-related mortality, for which the alcohol-attributable fraction ranges from 0.35 to 0.99. In the second subset, it is not known with certainty whether any individual death was caused by alcohol, but as the sample increases in size we are increasingly confident in the proportion of deaths that were caused by alcohol. We found no significant differences in the magnitudes of the estimated effects of the 2 tax changes across these 2 outcome subsets (Table 1). The sole obvious consequence of disaggregating the outcome is increased variability of the estimates (increased standard errors), an expected result of the smaller and more variable mortality counts in the subset analyses.
An examination of the actual death counts by quarter (Figure 2) could be interpreted to mean that the obvious sudden effect of the 1983 tax increase dissipated over time, because after 5 years the average numbers of deaths per quarter returned to the levels they were immediately before the policy change. Simple comparisons of the linear slopes of the pre-1983 and post-1983 periods show that they are similar (pre-1983 slope = 0.21 vs post-1983 slope = 0.32), but the slightly larger post-1983 slope might be interpreted as a dissipation of the effect seen in the sudden intercept difference in 1983. However, additional analyses that directly tested alternative hypotheses assuming temporary versus permanent effects demonstrated the effect did not dissipate over time. When we used the temporary-impact model specification, the estimate of ω remained essentially the same (–5.06, SE = 1.83), but the estimate of the decay parameter δ was 0.98 (SE = 0.03), very near unity, demonstrating that the effect did not decay over time.61 Moreover, the residual variance of the temporary-effect model was larger than that of the permanent-effect model, indicating poorer fit of the temporary-effect model.
When we repeated this procedure on the population-rate outcome variable, we again found no dissipation of effect over time. Similar statistical tests of permanent versus temporary effects for the 2002 tax change are not yet possible, because only 9 postchange data points are currently available.
We found that increases in alcoholic beverage tax rates were associated with significant and substantial reductions of alcohol-related disease mortality in Alaska. Reductions in mortality were observed for 2 tax increases 19 years apart, indicating that the observed effects cannot be attributed to a single historical period or event, atypical or otherwise. The long-term follow-up after the first tax increase allowed us to determine that the effect was not temporary, but was maintained over time. Our quasi-experimental research design included other states as a comparison group, demonstrating that the effects observed for mortality in Alaska were not caused by broader national trends, factors affecting mortality in common across states, or a sudden change experienced across states in 1983 or 2002.
Our results are consistent with basic economic theory and econometric analyses of the price elasticity of alcoholic beverages, which hold that an increased tax on beverage alcohol is presumed to raise the price to consumers, who respond by purchasing and drinking less alcohol. Lower alcohol consumption is then thought to reduce risk of death caused by a range of alcohol-related diseases, resulting in a decline in mortality counts and rates. Our study validated this account by demonstrating a clear link between the first and last factors in this presumed mechanism of effect (alcohol-tax increases and alcohol-related mortality). Complete and accurate quarterly measures of the 2 central intervening factors—retail prices and drinking behaviors—are not available for the 29-year period studied. However, the theory underlying this mechanism of effect is so well established that a lack of measures for each intervening effect does not reduce the plausibility of the findings.
A colleague and a reviewer both sought to explain the fact that the effect of the 1983 tax increase appears to be larger than the effect of the 2002 tax increase, as shown in the percentage-change column in Table 1. Several reasonable hypotheses quickly come to mind. For example, general price inflation means that any given alcohol tax increase in dollars per gallon (as alcohol taxes often are measured, rather than as a percentage of the sale price) would likely be a larger proportion of the total retail price of a product in 1983 than in 2002. However, close examination of the estimates and standard errors in Table 1 shows no significant differences between estimates of the effects of the 2 changes. Moreover, we only had access to 2.25 years of follow-up data for the 2002 tax change. Determination of whether the 2 tax changes had significantly different long-term magnitudes of effect, and the possible causes of such differences (if any), requires waiting for additional follow-up data to become available.
Some questions might be raised about the timing of observed effects of alcohol-tax increases. For instance, scientists and practitioners who focus on individual-level disease may find the claim that immediate population-level effects on mortality can be caused by modest changes in the social environment to be biologically implausible and “counterintuitive,” as a reviewer of this article said. A few of the specific causes of death included in our outcome measures are the result of the acute toxic effects of ethanol ingestion (e.g., poisoning), but most are chronic conditions that result from decades of high exposures to ethanol (e.g., cirrhosis or cancer). However, mortality caused by long-term, chronic alcohol use responds immediately to a change in drinking levels, because at any given time there is a reservoir of individuals in the population who are about to die from a chronic alcohol-related disease.63 Even modest reductions in current drinking immediately retards progression of alcohol-related disease for this population, resulting in an immediate reduction in the death rate, as found in the present study.
In this regard, alcohol-related disease is not unlike some other diseases, such as chronic obstructive pulmonary disease (COPD). Chronic, long-term exposure to air pollution is one cause of COPD.64 Development of COPD resulting from exposure to urban air pollution levels usually takes decades.64 Nevertheless, sudden reductions in ambient air pollution result in immediate reductions in COPD mortality. Pope et al. reviewed 18 “striking studies … that observed changes in daily death counts associated with short-term changes in particulate air pollution.”65(p475) This pattern of immediate population-level changes in mortality occurs for both COPD and alcohol-related disease, and for the same reason: at any given moment in time, there are many in the population whose cumulative exposure (to pollution or alcohol) is near the threshold that will cause death, and a reduction in current exposure delays death. A trajectory or cascade of delayed individual deaths results in population mortality declines that begin immediately when the exposure reductions occur.
Another question that could be raised about the time-ordered pattern of observed effects is related to the continual gradual reduction in real (inflation adjusted) alcohol taxes over time. Table 2 shows the Alaska tax rates immediately before and after each tax change in constant 2006 dollars (adjusted by the consumer price index for urban consumers in Anchorage66). For example, the beer tax increased from $0.46 per gallon in 1982 to $0.63 in 1983 at the time of the legislated increase, but then gradually declined to $0.40 by 2001 because of inflation, before increasing to $1.20 as a result of the 2002 law. Given such a pattern, one might have expected any effect of the 1983 tax increase on mortality to have dissipated after a decade and a half of inflation. However, we found little evidence of such dissipation of effect. Perhaps the sudden decline in drinking caused by increased prices established new normative and behavioral patterns that then were maintained in the face of a gradual elimination of the original price increase that stimulated the reduction in consumption. There is a need for additional study of the effects of sudden tax or price increases on drinking-related normative expectations and behavior patterns. The reduction of the real tax rate in Alaska also raises the plausible hypothesis that the substantial effects observed here would be significantly larger if such tax increases were maintained in real dollars via indexing to inflation.
|1982 (Preincrease)||1983 (Postincrease)||2001 (Preincrease)||2002 (Postincrease)|
Our study limited its focus to an evaluation of the effects of changing alcohol taxes on alcohol-related disease; we did not study mortality caused by the many categories of unintentional or intentional injury that have substantial fractions attributable to alcohol (e.g., traffic crashes, homicide, suicide). A replication of this research focused on injury morbidity and mortality is warranted. In addition, we limited our scope to mortality changes in Alaska; replication of the study for other states is also warranted.
In conclusion, Alaska's alcohol tax increases resulted in large (ranging from 11% to 29%) reductions in deaths caused by alcohol-related disease. The effect sizes are large compared with other efforts to prevent negative outcomes related to alcohol consumption. Three meta-analyses of efforts to prevent individual- and school-level alcohol problems show effect sizes ranging from d at –0.02 to d at –0.36,67–69 meaning that alcohol taxes had an effect 2 to 4 times larger than did other common prevention efforts. The size of the alcohol-tax effect is even more noteworthy given that state tax policy affects the entire population of a state, rather than the relatively small numbers of individuals affected by most other prevention programs. In addition, state alcohol taxation systems are already in place, so there are virtually no additional implementation costs associated with the large benefit to public health to be obtained by increases in alcohol taxes.
This study was funded by Robert Wood Johnson Foundation (grant 058005).
The authors appreciate the assistance of Linan Ma with data management and Amy L. Tobler with preparation of the article and are thankful for the helpful comments made by the reviewers and Frank Chaloupka.
Note. Findings and conclusions are solely the authors' and do not necessarily represent the views of the Robert Wood Johnson Foundation.
Human Participant Protection
The University of Florida institutional review board reviewed this study and deemed that its study protocol did not require approval because we had no contact with participants, data were based on public mortality files, and analyses involved aggregate mortality counts by state.