Predicting the Impact of Alcohol Taxation Increases on Mortality—A Comparison of Different Estimation Techniques

Abstract

Aims

To examine how standard analytical approaches to model mortality outcomes of alcohol use compare to the true results using the impact of the March 2017 alcohol taxation increase in Lithuania on all-cause mortality as an example.

Methods

Four methodologies were used: two direct methodologies: (a) interrupted time-series on mortality and (b) comparing predictions based on time-series modeling with the real number of deaths for the year following the implementation of the tax increase; and two indirect methodologies: (c) combining a regression-based estimate for the impact of taxation on alcohol consumption with attributable-fraction methodology and (d) using price elasticities from meta-analyses to estimate the impact on alcohol consumption before applying attributable-fraction methodology.

Results and Conclusions

While all methodologies estimated reductions in all-cause mortality, especially for men, there was substantial variability in the level of mortality reductions predicted. The indirect methodologies had lower predictions as the meta-analyses on elasticities and risk relations seem to underestimate the true values for Lithuania. Directly estimated effects of taxation based on the actual mortalities seem to best represent the true reductions in alcohol-attributable mortality. A significant increase in alcohol excise taxation had a marked impact on all-cause mortality in Lithuania.

Short Summary: Four different analytic approaches were used to estimate the impact of a recent alcohol excise tax increase in Lithuania on all-cause mortality. While all resulted in decreases of mortality, the true effect was best estimated via a comparison between predictions derived from time-series methodology compared with actual number of deaths. The indirect analytic approaches based on alcohol-attributable fraction methodology seem to underestimate the true effect.

BACKGROUND

Alcohol use is a major risk factor for mortality and burden of disease (Rehm and Imtiaz, 2016; Shield et al., 2020) and incurs substantial costs to societies (Manthey et al., 2021). Increasing alcohol taxation as a means to increase the prices of alcoholic beverages is one of the most effective and cost-effective strategies to reduce alcohol use and attributable harm (Wagenaar et al., 2010; Chisholm et al., 2018; OECD, 2021). To plan and implement taxation policies for alcohol, governments often need to predict the effects of taxation increases, both with respect to health and to economic outcomes. This contribution examines how standard techniques to model health outcomes compare to the true results, using a recent example of a substantial alcohol taxation increase in Lithuania that was enacted on 1 March 2017 (Miščikienė et al., 2020). Specifically, there was an excise tax increase of 112% for beer, 111% for wine and 23% for ethyl alcohol (which impacted on spirits prices). This taxation increase resulted in alcohol becoming less affordable during a period of stable economic growth, with affordability being defined in terms of the price of alcohol increasing at a higher rate than average disposable income (Rehm et al., 2021b). We chose to use this taxation increase as an example given that, even though it was relatively recent, enough time had lapsed since it was implemented for the potential consequences to be measured, i.e. it was possible to conduct interrupted time-series analyses with sufficient time points available post-implementation (Beard et al., 2019).

Štelemėkas et al. (2021) showed that the increase in alcohol taxation in 2017 was associated with a marked decrease in all-cause mortality rates. Thus, we selected all-cause mortality as the health outcome to be measured. All-cause mortality is the most reliable mortality outcome, as all specific cause-of-death categories can be impacted by misclassification. This is particularly true for outcomes that are 100% attributable to alcohol, as they have been shown to be severely underestimated (Puffer and Griffith, 1967; Rehm et al., 2017b). Specifically for Lithuania, the coding rules for the most important 100% alcohol-attributable disease condition—alcoholic liver disease—changed several times, creating issues with comparability (Cicėnienė et al., 2019).

We compared the following analytic approaches, based on the theoretical considerations of how taxation may impact alcohol-attributable mortality (see Fig. 1):

Fig. 1.

Flow diagram on alcohol use, taxation and alcohol-attributable deaths.

• (1) An interrupted time-series analysis (Beard et al., 2019) of the direct effect of taxation on all-cause mortality, assuming an abrupt (i.e. immediate) and permanent effect (Štelemėkas et al., 2021)

• (2) A comparison between the predicted values based on a time-series analysis before the event, and the actual values (Jiang et al., 2021)

• (3) Regression-based analysis of the effect of taxation on alcohol consumption (on total alcohol per capita consumption, and separately by beverage type), combined with using attributable-fraction methodology (Rockhill et al., 1998; Rehm et al., 2010) to predict the impact of change in consumption on all-cause mortality

• (4) Modeling the potential effects of taxation increases on alcohol consumption via price elasticities (Sornpaisarn et al., 2017), and then using the predicted changes in alcohol consumption to predict changes in all-cause mortality via attributable-fraction methodology (as in 3)

While Fig. 1 explicitly pertains to alcohol-attributable mortality, the same mechanisms apply for other health indicators that are causally impacted by alcohol (such as all-cause mortality, or alcohol-attributable hospitalizations). In countries such as Lithuania, alcohol-attributable mortality is at such a high level (e.g. (Rehm et al., 2007)) that changes in alcohol-attributable mortality are reflected in the all-cause mortality trends as well.

MATERIALS AND METHODS

Interrupted time-series methodology

The model of the interrupted time-series methodology is described in detail in Appendix Text A1 (see also (Štelemėkas et al., 2021) for further details). In short, an interrupted time-series of monthly all-cause mortality data was estimated, using generalized additive models (GAMs (Hastie and Tibshirani, 1990, Beard et al., 2019)), which were adjusted for seasonality by smoothing splines. The alcohol control policy intervention was tested using a dummy variable to model an abrupt and permanent effect. Potential confounding factors included centered monthly gross domestic product (GDP), centered unemployment rate and a second-order polynomial of the time to account for the overall curvilinear trend in all-cause mortality. In this model, which will be referred to herein as the direct estimation approach, the intervention effect equals the beta weight as derived from the interrupted time-series model.

Monthly data on all-cause mortality were provided to the project from the Lithuanian Institute of Hygiene, age-standardized to the Lithuanian population of February 2017, the month before the taxation increase (Lithuanian Department of Statistics, 2021c).

Methodology to compare predicted continuation of trends prior to taxation versus actual mortality

For the prediction model, we fitted a time-series model as described above to estimate the trend before the intervention and subsequently used this model to predict the 12 months after the intervention. In this prediction approach, the intervention effect is the difference between the predicted and the actual numbers of death.

Estimated effect of taxation on consumption and subsequent mortality using alcohol-attributable fractions

The third indirect two-step approach is a regression-based approach used to predict changes in recorded adult per capita consumption (APC), thus estimating the impact of the excise taxation on APC. Given the relatively short time series of APC (yearly data between 2001 and 2019, n = 18), we performed an ordinary least-squares regression on alcohol consumption per capita, with the taxation policy in 2017 as a dummy-coded variable. We included two additional economic indicators (gross domestic product (GDP), and consumer price index (CPI)), which were yearly time-series data covering the same period (from 2001 to 2019), as well as quadratic and linear time-trend covariates. In addition, we performed a regression using a different estimator, a ridge regression, which accounts for potential multicollinearity between independent predictors (Hoerl and Kennard, 1970). For the full methods and results of the regressions used to estimate the impact on per capita consumption, refer to Appendix Text A2.

Consumption was then related to alcohol-attributable mortality using the following formula for alcohol-attributable fractions (AAFs; see (Rehm et al., 2010)), where P_i and RR_i denote the prevalence of consumption and the relative risk, respectively, for each consumption group—lifetime abstainers (‘abs’), former drinkers (‘form’) and current drinkers (‘CD’):

(1)

Note that this formula was applied to each disease category causally related and partially attributable to alcohol use ((Rehm et al., 2017a); for an overview of categories used, see Appendix Table A1), except for cancer where the impact of alcohol only manifests itself after a lag time. For cancer, we would thus not expect an effect within 12 months of the implementation of the policy.

The disease-specific risk functions (taken from (Shield et al., 2020), where applicable, separated by sex and age (Rehm et al., 2021a)), only depend on the level of alcohol ingested, as derived from reported APC (Lithuanian Department of Statistics, 2021b). The distribution of alcohol consumption by beverage type, sex and age was taken from Manthey and colleagues (Manthey et al., 2019; Shield et al., 2020). Finally, the data on the total number of deaths per year were obtained from the Institute of Health Metrics and Evaluation (Global Health Data Exchange, 2020).

However, not all the diseases causally impacted by alcohol use can be modeled using AAFs. For totally alcohol-attributable diseases, the AAF by definition is 1, independent of the level of consumption. To estimate change in these disease categories (e.g. alcohol use disorders), we followed the methodology used by Churchill and colleagues (Churchill et al., 2020); see Formula 2, where N is the number of deaths, S is the population size, p(x;k,t) is the risk function and F(x;μ) is the gamma distribution for the mean consumption μ.

(2)

The risk function only affects heavy drinkers, defined as women consuming 40 g or more ethanol per day, and men consuming 60 g or more ethanol per day.

Estimated effect of taxation on mortality based on alcohol price elasticity, consumption and subsequent mortality

Increasing taxes on alcoholic beverages means that their prices will increase, assuming the producers pass the cost entirely on to the consumer, which has generally been found to be the case (Nelson and Moran, 2020; Shang et al., 2020). The increase in product price can then be translated into a change in consumption if the price elasticity is known. Price elasticity is a measurement of the change in consumption of a product in relation to a change in its price (see Formula 3, where ∆Q is the percent change in consumption and ∆P is the percent change in price).

(3)

Price elasticity tends to depend on beverage type, and which beverages are preferred in different cultures but seem to be stable across countries (Fogarty, 2010). As in prior publications regarding European countries (e.g. (Rovira et al., 2021a, Rovira et al., 2021b)), we used elasticities based on published meta-analyses (Ornstein and Levy, 1983; Wagenaar et al., 2009; Fogarty, 2010; Sornpaisarn et al., 2013). That is, −0.36 (95% CI: −0.48, −0.24) for the most-preferred beverages, which for Lithuania in 2016 were spirits (Manthey et al., 2019); −1.20 (95% CI: −1.44, −0.96) for the least-preferred beverages, which were wines; and an intermediate value of −0.60 (95% CI: −0.72, −0.48) for beer. The negative values indicate the proportional decreases in the consumption of the respective beverages when the price is increased by 1 unit (Sornpaisarn et al., 2017).

For heavy drinkers, we modified the above-specified elasticities estimated based on the reported APC level, drinking status information (Manthey et al., 2019) and assuming a gamma distribution in the level of consumption (Rehm et al., 2010; Kehoe et al., 2012)). The distinction between heavy versus non-heavy drinkers was made based on a meta-analysis that showed that heavy drinkers or people with alcohol use disorders (for further discussion, see (Rehm et al., 2013)) tend to be less price-elastic (Wagenaar et al., 2009). In fact, continuation of alcohol consumption despite having good reasons to drink less or quit drinking altogether is one of the diagnostic criteria of alcohol use disorders (Carvalho et al., 2019). Based on the above-mentioned meta-analysis, the heavy drinkers, as defined above, were assigned an elasticity of −0.28 (95% CI: −0.37, −0.19), irrespective of alcoholic beverage type consumed.

As described above, the methodologies for partially and totally alcohol-attributable diseases can be applied to recorded consumption values before and after the taxation implementation to estimate the number of averted deaths.

Summary of assumptions

In Table 1, the main assumptions about each of the modeling approaches are given.

Table 1.

Main assumptions of the different modeling approaches used

Analytical approach	Assumptions regarding effect of taxation on alcohol use	Assumptions regarding risk relations alcohol use and causes of death	Assumptions regarding lag time between alcohol use and mortality	Assumptions regarding lag time between taxation and mortality	Confounders	Comments
(1) Interrupted time-series (ITS) of all-cause mortality data (see Appendix Text A1; (Jiang et al., 2021, Štelemėkas et al., 2021) Direct estimation approach	None	None.	None.	Abrupt and immediate effect	Assumes that all relevant confounders were included in the ITS model, and no other event at the same time could have caused the effect; effect depends on age-standardization used.	Modeling with different assumptions possible; if assumption of abrupt and immediate effect is not correct, impact estimate will be biased. Stockpiling may distort the results.
(2) Comparison of predicted (based on time-series) and actual all-cause mortality (see Appendix Text A1) Prediction approach	None	None.	None.	Abrupt and immediate effect	Assumes that time-series model before the taxation intervention is correctly identified and all relevant confounders/predictors are identified; effect depends on age-standardization used.	Modeling with different assumptions possible; if assumption of abrupt and immediate is not correct, impact estimate will be biased. Stockpiling may distort the results.
(3) Two-step approach: (1) Regression-based estimate of changes in alcohol use (see Appendix Text A2); (2) alcohol-attributable fraction approach to model effect of alcohol on mortality Indirect approach 1	No stockpiling.	(1) All COD that alcohol use impacts upon immediately are identified; (2) the risk relations between level of drinking and COD are correctly identified from meta-analyses.	Abrupt and immediate effect was modeled over one year.	None	Assumes that all relevant confounders were included in the model, and no other event at the same time could have caused the effect.	Modeling with different assumptions possible; if assumption of abrupt and immediate impact of alcohol on COD is wrong, this will not be detected.
(4) Three-step approach: (1) measuring the impact of taxation on price; (2) estimating the impact of price on consumption (via price elasticities); (3) using attributable-fraction approach to model effect of consumption on mortality Indirect approach 2	The price elasticities for different beverage types are correctly identified from meta-analyses, and there are no major cross-elasticities. The elasticities of non-drinkers do not play a role in taxation increases.	(1) All COD that alcohol use impacts upon immediately are identified; (2) the risk relations between level of drinking and COD are correctly identified from meta-analyses.	Abrupt and immediate effect was modeled over one year.	None	Does not assume other direct impact on use than price as all other impacts are only indirectly included in the meta-analyses to derive price elasticities.	If assumption of abrupt and immediate impact of alcohol on COD is wrong, this will not be detected. Stockpiling does not change the results of the modeling.

COD: cause of death

Validation of results

To validate the results, we compared empirical elasticities with the ones derived from meta-analyses, and we looked at how all-cause mortality and causes of death were impacted by alcohol from 2015 to 2019 (see Appendix Tables A1 and A2 for a list of causes of death).

RESULTS

While all methodologies indicated a decrease in mortality due to the implementation of the 2017 tax policy, point estimates of the number of deaths averted varied substantially (Table 2; see Appendix Texts A1 and A2 for details).

Table 2.

Results from four different modeling approaches for the 12 months following the taxation increases (March 2017–February 2018)

Analytic approach	Number of deaths averted in the 12-month period following the implementation of increased excise taxation
	Women	Men	Totala
Interrupted time-series of mortality rates (Jiang et al., 2021)	433 95% CI: −560, 1427	1718 95% CI: 203, 3232	2151 95% CI: 340, 3962
Comparison between predictions based on time-series and actual deaths (Jiang et al., 2021)	176 95% CI: −51, 404	979 95% CI: 618, 1340	1155 95% CI: 729, 1582
Two-step approach: (1) Regression-based estimate of changes in alcohol use (2) Alcohol-attributable fraction approach to model the effect of alcohol on mortality	77 95% CI: 54, 268	141 95% CI: 96, 210	217 95% CI: 171, 425
Three-step approach: (1) Measuring the impact of taxation on price (2) Estimating the impact of price on consumption (3) Attributable-fraction approach to model effect of alcohol on mortality	42 95% CI: 36, 169	65 95% CI: 53, 95	107 95% CI: 93, 244

^aTotal effects are based on population-adjusted addition of sex-specific effects.

The two direct methodologies using interrupted time-series or predictions based on time series yielded much higher estimates than the indirect methodologies. But which methodologies are more valid?

Table 3 gives the observed changes in all-cause mortality and causes of death, which are impacted by alcohol use for the years 2015–2019. There is not a lot of variation for 2015 to 2016, or 2017 to 2019, and indeed the big changes are seen between 2016 and 2017 (see last row in the Table), consistent with an impact from changes in taxation. Moreover, the changes are mainly in men (and, also, therefore, when both sexes are combined), and the level of change in causes of death impacted by alcohol accounts for most of the decrease in all-cause mortality (for women: 179 of 220; for men: 1199 of 1787; total: 1377 of 2007; see Table 3). Given these numbers, the prediction approach comes closest to the data.

Table 3.

Number of deaths by all causes and by causes of death causally impacted by alcohol in Lithuania 2015–2019

	Causes of death impacted by alcohol (as per Appendix Tables A1 and A2)			All-cause deaths
	Women	Men	Total	Women	Men	Total
2015	13,741	12,578	26,319	20,632	20,074	40,706
2016	13,547	12,552	26,098	20,422	20,114	40,536
2017	13,368	11,353	24,721	20,202	18,326	38,529
2018	13,482	11,357	24,840	20,306	18,377	38,683
2019	13,637	11,081	24,718	20,501	18,001	38,501
2017–2016	−179	−1199	−1377	−220	−1788	−2007

Source: Global Health Data Exchange (2020); for a graphical display, see Appendix Figure A2.

The italic line at the end refers to the mortality reductions between the years 2016 and 2017.

Thus, based on the values in Table 3 and the above considerations, it seems likely that the estimated effect of taxation from the interrupted time-series analyses (the direct estimation approach) was an overestimate of the true effect, and the results of the two indirect analytic approaches were underestimates. The results of the three-step estimation based on elasticities underestimates the true elasticities as calculated based on the empirical data (see Appendix Table A3), especially for spirits.

DISCUSSION

Overall, as previously shown, the implementation of a higher tax on alcohol on 1 March 2017 in Lithuania was associated with a marked decrease in mortality, which was to be expected since it decreased the affordability of alcohol (Štelemėkas et al., 2021; Rehm et al., 2021b). Different methodologies were used to estimate this impact and yielded widely varying results. A validation suggested that indirect methodologies likely underestimated the impact and that using a direct methodology comparing the predicted deaths based on times-series methodology might have yielded a more accurate estimate. One major reason for this may be that the indirect method often fails to account for the internal structure of the data—for example, autocorrelation, trend or seasonal variation.

However, our study included the following limitations. The meta-analyses for the risk relationship between level of alcohol use and different causes of death may underestimate the mortality pattern in a country such as Lithuania. Indeed, in the comparative risk analyses, alcohol-attributable mortality in Lithuania has traditionally been modeled based on Russian studies, with a higher relative risk based on its more irregular heavy drinking pattern of predominantly spirits (Shield and Rehm, 2015; Shield et al., 2020); for a review on differential health risks for spirits, see (Rehm and Hasan, 2020). Even though drinking patterns have changed in Lithuania in 2019, and beer is now more commonly consumed than spirits (Lithuanian Department of Statistics, 2021a), the values arising from the meta-analyses used for non-Russian countries may still be underestimates.

The current situation in Lithuania, where two beverage types each account for more than 40% of alcohol consumption, may pose a problem when distinguishing between the least- and most-preferred beverages to determine price elasticities (based on (Ornstein and Levy, 1983)). Also, the price elasticities used in the fourth scenario based on meta-analyses clearly did not capture the actual situation in Lithuania (see Appendix Table A3). However, alternative price elasticities, such as the one derived by Nelson (Nelson, 2013), or not applying the reduction for heavier drinking (by using the overall beverage-specific elasticities of (Wagenaar et al., 2009)), would not have changed the general conclusion that all indirect methods underestimated the actual changes in mortality.

Another limitation is that within one year of the tax increase, at the beginning of 2018, other policies were implemented that may have affected the outcomes (i.e. reductions in hours of sale and the introduction of a minimum legal drinking age (Miščikienė et al., 2020)). However, a closer inspection of the results of the prediction approach shows that after the implementation of these alcohol control measures in January of 2018, the deaths averted actually decreased, indicating that these other policies could not be responsible for the main result.

Lastly, all analyses were ecological and, while we used causal language based on the postulated mechanisms, causality could not be firmly established based on the data. While we are not aware of alternative events beginning in early 2017, which could explain the abrupt decrease in alcohol consumption and alcohol-attributable deaths, we cannot fully exclude such events.

There are three main conclusions that can be drawn from this study. First, direct methodology should be used whenever feasible. Time-series methodologies (Beard et al., 2019) with monthly data make much fewer assumptions than indirect methods and can measure direct associations for an event happening in a specific month without the use of assumptions regarding price elasticity or risk relations between alcohol use and mortality (see Table 1). Obviously, the models need to be correctly specified to arrive at good estimates, and by using two different models—such as interrupted time-series analysis and prediction models—it can help identify the potential biases. Second, when a three-step methodology is used, at least some checks on the plausibility of meta-analyses for price elasticities should be performed. Since some time has passed since the publication of the meta-analyses on price elasticities we used (Ornstein and Levy, 1983; Wagenaar et al., 2009; Fogarty, 2010), and many more empirical research papers on the topic have been published in the intervening years, it may be time to conduct a new meta-analysis using meta-regressions for region, beverage type and beverage preference. Third, the implementation of alcohol excise taxation increases in Lithuania was clearly associated with important public health effects. Tax increases leading to reduced affordability (Rabinovich et al., 2009) continue to be an effective alcohol control policy— indeed, perhaps the most effective alcohol control policy (Babor et al., 2022 (forthcoming)).

In this sense, we would clearly expect that all excise taxation increases that reduce affordability of alcoholic beverages would result—at least in the short run—in reductions of consumption and mortality (Rehm and Purshouse, 2021). Regarding the level of the effects found for Lithuania, potential transferability is harder to assess. As alcohol-attributable mortality is impacted both by level of consumption and by patterns of drinking (Rehm et al., 2017a), we would expect a similar impact for countries with the same level of consumption and patterns of drinking.

FUNDING

Research reported in this publication was also supported by the (U.S.) National Institute on Alcohol Abuse and Alcoholism of the National Institutes of Health (NIAAA), grant number 1R01AA028224. This research was conducted as part of the project ‘Evaluation of the impact of alcohol control policies on morbidity and mortality in Lithuania and other Baltic states’ and we would like to thank the whole team for their input to wider discussions in generating the research reported in this paper. Content is the responsibility of the authors and does not reflect official positions of NIAAA or the National Institutes of Health.

CONFLICTS OF INTEREST STATEMENT

The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses or interpretation of data; in the writing of the manuscript or in the decision to publish the results.