Examining the effects of beer excise taxation on cross-border sales in border regions of the Baltic States

ABSTRACT Cross-border alcohol shopping threatens health outcomes and tax revenue, yet offers potential jobs and investment for the Baltic border regions. Exploiting the COVID-19-induced border closures in 2020 with an event study, and several large tax changes using differences-in-differences, this paper measures beer purchases in border regions of the Baltic States. Using monthly data from the Carlsberg group, beer sales in lower tax border regions dropped sharply during the COVID-19 closures, while sales increased modestly in higher tax border areas. Evidence from tax changes in Lithuania and Estonia also suggests dramatic effects on sales in lower tax Latvian border districts.


INTRODUCTION
In fighting to reduce chronic levels of alcohol consumption, while also maintaining excise tax revenues, the Baltic States (Lithuania, Latvia, Estonia) present a unique opportunity to study the regional effects of excise tax competition.The uncoordinated and centralised nature of taxsetting, combined with the tax differentials in their Schengen Zone 1 neighbours (Poland and Finland), has led to the growth of a cross-border alcohol shopping industry in several countries.Finns are able to access cheap alcohol by taking a short ferry ride to Tallinn in Estonia, while Estonians can visit many of the specialised alcohol shops just across the border in Latvia.A similar situation has also been noted on the Polish side of the border with Lithuania.Understanding the degree to which such a cross-border trade in alcohol exists is essential for policymakers interested in reducing the loss of excise tax revenues and/or reducing problematic excessive alcohol consumption in their given countries and regions.
One challenge of studying the causal impact of excise taxation on alcohol sales lies in identifying the type of behavioural change exhibited by consumers.They may choose to decrease consumption, make their own illegal alcohol or opt for cross-border shopping (Devereux et al., 2007).Cross-border shopping is especially egregious for policymakers as it both lowers domestic tax revenue while also negating the health benefits of reduced alcohol consumption.Cross-border shopping is also difficult to stop because it is legal and requires no special knowledge, unlike smuggling or home production.This paper addresses the cross-border aspect of alcohol consumption, contributing to the broad literature on taxation and crossborder shopping for a wide variety of products encompassing: cigarettes (Chiou & Muehlegger, 2010;Nagelhout et al., 2014), marijuana (Salar et al., 2020), sugary drinks (Cawley et al., 2019), gasoline (Devereux et al., 2007) and others (Chandra et al., 2013).While not the main focus, the present paper also adds to the general literature on how excise taxation affects consumption (e.g., Chaloupka et al., 2002;Chetty et al., 2009;Colchero et al., 2016).
Studies on cross-border shopping for alcohol have often been observational in nature, relying on price changes and distances from the border to estimate the degree of cross-border shopping indirectly (e.g., Asplund et al., 2007;Beard et al., 1997;Leal et al., 2010).These studies generally find that distance from the border and lower price differentials decrease cross-border shopping.As an example, Stehr (2007) studies the effect of alcohol excise taxes and Sunday sales restrictions in American states and finds that excise taxes lessen alcohol sales, but that these effects are dampened when a large proportion of a state's population lives close to the border.In the Baltic region specifically, there have been attempts to study the impact of excise taxation on consumption, but these studies often rely on surveys (e.g., Karpuškienė, 2021).The most in-depth report on excise taxation in the Baltics to date presents a descriptive analysis of alcohol sale changes on the borders (Pluta et al., 2020), but lacks the data to perform a deeper analysis of the cross-border trade.
More recently, a small literature has developed incorporating quasi-experimental designs to study cross-border alcohol shopping.However, these studies are limited to single-event studies due to the infrequency of alcohol excise tax changes in most developed countries.In one such study on increased wine and spirit taxation in Illinois, Gehrsitz et al. (2021) find no effects of a tax change on cross-border shopping.The authors suggest that this result most likely stems from the small differences between state alcohol excise taxes in the United States.They also argue that consumers substituted between different alcohol types locally, rather than purchasing the same type of alcohol in another state.In a related study, using a differencein-differences (DiD) design to explore the cross-border effects of Washington's privatisation of its alcohol monopoly, Ye and Kerr (2016) find that liquor sales in Oregon increased by 10.1% and in Idaho by 8.2%.The present study contributes to the quasi-experimental literature in this area by applying a variety of state-of-the-art techniques to multiple tax change events and a full border closure precipitated by the COVID-19 crisis.The size and frequency of the tax changes in the Baltics, combined with the uniqueness of the border closure event, make the present study an important addition to the literature.
The vast majority of cross-border tax studies use US data due to the large variance in tax rates at the state and municipal levels.However, this fiscal federalism can encourage tax competition, which complicates the measurement of the effect of tax changes on sales by introducing endogeneity.Ideally, tax changes should come as exogenous shocks in order to measure the pure effect of the tax on beer sales.However, with tax competition, beer sales may drive tax rate changes.The three Baltic countries involved in the present study do not give municipalities the power to independently regulate excise tax rates, thus the problem of competition is restricted to the country leveland is, therefore, much less of an issue.A further major drawback of using US data for alcohol excise changes is that state tax rates rarely ever change.For example, Alabamawhich has one of the highest rates in the country (US$0.139/0.5 litre)has not changed its tax rate once in the period from 1982 to 2021 (Tax Policy Center, 2021).Finally, beer excise taxes in the United States are low when compared with the European Union (EU)in 2015, the average beer unit had an excise tax of only US$0.03 (Naimi et al., 2018).
Using an event study design robust to heterogeneous treatment effects, I exploit the COVID-19 border closures in early 2020 to measure the relative change in beer sales within border districts.I show that districts on the higher tax side of a border experience a modest relative increase in sales.This finding supports the idea that consumers who would normally cross the border to buy cheaper beer must now substitute towards more expensive local beer.In contrast, districts on the lower tax side of the border see a dramatic decrease in sales due to the complete stop of cross-border shopping.This suggests that cross-border shoppers are travelling not only from adjacent districts but also from farther afield.Next, I identify how beer excise tax changes affect domestic and neighbour beer sales by studying three of the largest tax rate changes in the Baltics using a DiD design.I establish that own-country tax changes have no statistically significant effect on one's own border relative to the interior.However, across the border, beer sales grow significantly relative to their neighbouring interior districts.These findings are strengthened by a host of robustness checks, including placebo tests, and provided evidence that the excise taxes were shifted to consumers through price rises.
The results of the present study highlight the inherent problem for the tax authorities when attempting to set optimal excise tax rates.They also raise issues concerning the welfare of those living in border districts.The rural border regions on the lower tax side of a border may benefit from the increased sales due to alcohol tourism, but are made vulnerable by the ease with which their tax advantage can be withdrawn.Along higher tax borders, the loss of tax revenue and public health concerns due to easy cross-border access pose additional problems.
The rest of the paper is organised as follows.Section 2 briefly provides the background for alcohol excise taxation and the cross-border beer trade in the Baltics, including the dataset.This is followed by the empirical part of the paper in section 3, which begins with an event study using the COVID-19 border closures, including its results.Next, a closely related DiD study on three of the largest excise tax changes is presented along with its results.Section 4 concludes.

Background
After regaining independence from the Soviet Union, excessive alcohol consumption has created major health and social problems for all three Baltic States.According to data from Eurostat (2021), Lithuania, Latvia and Estonia occupy the top three spots in the EU for alcohol consumption as a percentage of income.Efforts to curb alcohol consumption have been taken, albeit in an uncoordinated fashion.Estonia has implemented some of the most far reaching anti-alcohol policies, including limiting alcohol sales at night and bans on advertisements (Täht et al., 2020).Lithuania has implemented similar measures: in 2018, alcohol sales in Lithuania were restricted at night, the legal drinking age was raised to 20 and all advertisements for alcoholic beverages were outlawed.Furthermore, even announcing the reduction of prices (promotional sales) has been banned (Berdzuli et al., 2020).Latvia, while more liberal in its approach, has made steady progress in rising excise tax rates over the years.Similar trends in alcohol policy combined with their shared and intertwined history, both in terms of economic developments and political outlook, suggest that a parallel trends assumption is more likely to hold in the case of the Baltic States as compared with using a set of more heterogeneous countries.
The three Baltic countries (Figure 1) are small and centralised states that do not provide autonomy at the local level regarding excise taxation.The uncoordinated, centralised tax setting has arguably led to a specialised border retail alcohol industry in Latvia that focuses on cross-border sales.For example, SuperAlko has three shops in northern Latvian, all within walking distance of the border (for a map illustrating this point, see

Figure A1
in Appendix A in the supplemental data online).They also have a shop at the port of Tallinn that is purposefully within walking distance of the ferry coming from Finland.Many of these alcohol shops are understandably located along the major highways (and ports) entering each respective country, which suggests that the impact of excise tax changes on border districts will be heterogeneous, depending on whether the district lies along a major highway or shipping lane.

Data
The present study is based on wholesale sales data from the Carlsberg group that consists of monthly beer sales (0.5litre units) from January 2016 to December 2020.In addition, I have collected tax rate data for the three Baltic States plus Finland and Poland, which allows border tax differentials to be constructed.The panel data consist of 59 municipalities for Lithuania, 15 counties for Estonia and 28 districts in Latvia, 2 for a total of 102 panels.As Estonia is smaller and less densely populated than the other two countries, Estonian counties roughly correspond to the municipal/district size level observed in Lithuania and Latvia.For simplicity, the municipalities, counties and districts will all be referred to as districts in the rest of the paper.A map of the Baltic States at the district level (Figure 2) highlights the border districts to be used in the study.
The dataset comprises monthly wholesale sales figures for all alcoholic beer brands within the Carlsberg group, aggregated at the district level.While the Carlsberg group consists of 687 different beer brands worldwide, only a subset of these is available in the Baltic market.Certain international brands in the group (e.g., its namesake Carlsberg) are represented in all three markets, but a significant portion of sales in the region comes from local subsidiary breweries, such as Švyturys-Utenos in Lithuania, Aldaris in Latvia and Saku in Estonia.The total market share of the Carlsberg group is 40% in Lithuania, 26% in Latvia and 40% in Estonia. 3 The dataset therefore reflects a broad range of wholesale beer sales in the Baltic States across various brands and alcohol strengths.Sales from all the brands are aggregated together and reported at the district level in monthly intervals.
One major concern with the dataset is that it does not report final retail sales data.As a result, there may be some delay between the sales observed in the dataset and the ultimate sales to consumers.However, the extensive storage requirements and limited shelf life of beer suggests that retailers will not maintain large inventories, meaning that there should be a high correlation between wholesale and final consumption sales.Thus, wholesale sales data used in the study can serve as a reliable proxy for final retail sales.
The price data used in this study represent average national prices, taking into account the market mix of the Carlsberg group in each country.These final retail prices are sourced from Nielsen and do not represent the wholesale prices charged by the Carlsberg group.Since these data do not include a regional breakdown, they will only be used to measure the impact of excise tax changes on prices as part of the robustness checks.The lack of detailed price and promotion data are problematic as these factors are known to influence consumer purchasing behaviour.Identifying the pure effect of excise taxation becomes more difficult during the pandemic if shops in border regions opted for differing pricing or promotion strategies as compared with internal regions.Thus, for identification, I rely on the assumption that the pricing and promotion strategies are identical in the control and treatment groups.In addition, determining the exact tax amount per beer is problematic, as tax rates are set in alcohol percentage points/hectolitre in the Baltic States and Finland, while Poland uses the Plato gravity scale.In the Baltics, most of the major beers have 5% alcohol by volume.For example, Carlsberg (5%), Švyturys (5.2%), Utenos (5%) Aldaris (5%) and Saku (5.2%).Thus, for simplicity I will assume a 5% alcohol rate for calculating tax amounts per beer.Figure 3 illustrates the evolution of the tax rates on beer in the three Baltic States from 2016 onward using monthly intervals.
One limitation of using data on beer sales alone is that changes in the tax rates for wine or spirits has been shown to induce substitution effects between alcohol types (Gehrsitz et al., 2021).However, while there are periods when the tax rates of wines and spirits fluctuate without a corresponding change in the beer tax, a large proportion of beer, wine and spirit tax changes occur concurrently in the sample.Finally, value-added tax (VAT) will also likely affect beer sales, but during the time horizon of the study, there were no changes in VAT tax rates in any of the countries involved.
The beer sold by the Carlsberg group, as analysed in this study, constitutes a normal good within the beer market, in that sales should increase with incomes.It is important to note that the findings of this study may not be generalisable to other segments of the market, particularly those comprising of inferior goods such as cheap or discount beer brands.These brands are often sold in large volume containers, typically with higher alcohol strengths, 4 and are preferred by low-income and alcoholdependent individuals.Due to the income effects created by rising excise taxation on prices, these consumers may resort to increased consumption of such brands.As a result, a limitation of this study is that the findings do not necessarily pertain to overall beer consumption in the Baltic States.In addition, by using data solely from the Carlsberg group, the results of the study may inadvertently incorporate certain aspects of the group's marketing, pricing or distribution strategies which are not representative of the general industry.For example, the group may have responded to the COVID border closures differently than its rivals.Thus, the results of the present study can only be generalised to other beer sellers which behave similarly and share the same customer base with the Carlsberg group.
Table 1 summarises the relevant data.It begins with the first difference of the logarithm of beer sales, which approximates the growth rate.I purposely do not include data on the amount of beer sales used in the study, as they are sensitive to the firm.In order to gauge the price incentive for cross-border beer shopping, I construct a measure of the border tax differential which take the value of zero for interior districts and the difference of country A and country B's excise tax for border districts.Carlsberg wholesale price data are not included because they are sensitive to the firm, and are not particularly relevant as they are not retail prices.However, beer excise taxes in the EU have been shown, on average, to be almost fully shifted to prices (Ardalan & Kessing, 2021), meaning that changes in excise tax amount per unit should roughly equal price changes per unit.However, further tests of this assumption will be conducted in the robustness checks.Finally, excise tax data appear in Table 1 as euros per beer unit sold.Unfortunately, the inclusion of additional control variables was not possible due to lack of district level data at monthly intervals from the respective national statistics bureaus.
While the majority of the alcohol tax changes were implemented according to a pre-determined schedule, one notable exception that prompted the so-called 'Baltic booze war' (LRT, 2019) was reactionary in nature.The government of Estonia was concerned with losing tax revenues due to perceived cross-border shopping and it subsequently lowered its excise tax rates in July 2019 on all alcohol categories.Latvia responded by dropping excise tax rates on wine and spirits the following month, but kept the beer excise tax rates unchanged.Such endogenous tax changes are problematic in an econometric setting as they can lead to estimate bias (for a deeper discussion, see Romer & Romer, 2010).Although it is impossible to know whether the other excise tax changes in the sample were purely exogenous, they nevertheless produce steep changes in consumption that are most likely orthogonal to the unobserved determinants of beer sales.If it is assumed that beer buyers who engage in cross-border shopping are solely motivated by prices, there is little reason to believe that the political or revenue motivations behind excise tax movements would affect a buyer's decision-making process.

THE STUDY
I employ two distinct empirical strategies to measure the cross-border beer trade caused by excise taxation.First, I use the exogenous border closures between countries in response to the COVID-19 pandemic to identify the relative size of the beer trade, comparing border districts with interior districts in an event study using a weighted  624 Aras Zirgulis

REGIONAL STUDIES
DiD estimator.The Baltic State border closures all occurred at roughly the same time: from mid-March to mid-May 2020.This provides a quasi-natural experiment to test the degree to which such cross-border trade exists.Second, I exploit three of the largest excise tax changes in the sample to measure their effects on sales for either side of the given border using a classical DiD design.This approach allows for the estimation of the causal impact of excise taxes on cross-border trade in beer.

Study 1: COVID-19 border closures and cross-border trade
Under ideal experimental conditions, the border closures would have been implemented by surprise, independent of any other shocks.However, in the run up to the border closures, given the growing panic and unease amongst the general population, there may have been some COVID-19-related effect on alcohol sales.To mitigate this potential confounding effect, an event study design is employed.By differencing out any common trend in both border and interior districts, the estimator is able to isolate the policy effect of the border closures.In addition, an event study design enables the comparison of pre-treatment periods with the treatment event, tracking the dynamic evolution of the policy effect at each month.This is necessary because the border closures occurred only partially for March and May, and fully in April.
An event study often takes the form of the two-way fixed effects model (TWFE) because it incorporates time and location fixed effects.While the generalised TWFE method is popular amongst practitioners (de Chaisemartin & D'Haultfoeuille, 2020), a host of recent contributions to the literature have raised concerns with the validity of this approach (Athey & Imbens, 2018;Borusyak et al., 2021;Callaway & Sant'Anna, 2021;de Chaisemartin & D'Haultfoeuille, 2020;Goodman-Bacon, 2021;Sun & Abraham, 2021).It has been shown that estimating a TWFE model where the treatments have heterogeneous effects and/or occur at differing times may lead to ordinary least squares (OLS) assigning negative weights to unit-time treatment effects, creating potentially misleading inferences.
Despite border closures occurring concurrently, treatment effects are likely heterogeneous in multiple dimensions.The strength of the border closure effect should depend on the size of the border tax differential, which varies quite substantially between some of the border treatment districts (Table 2).In addition to the heterogeneity in the size of the treatments, the strength of treatment effect will vary over time.As the closures only began in mid-March and ended in mid-May, those two months only experience a partial treatment effect, while April receives a full-treatment effect.Finally, there are also likely to be heterogeneous effects on either side of the same border.As more specialised alcohol shops lie on the lower tax borders, the effects of such closures will presumably be more pronounced in those districts.To illustrate this point, Figure A1 in Appendix A in the supplemental data online shows the location of SuperAlko brand alcohol shops on the Latvian/Estonian border.
During the border closures, cross-border beer purchases factually dropped to zero, implying that the size of the trade can be inferred by the degree to which sales changed on either side of the border.However, the law of demand states that as prices increase, consumption will decrease.Thus, on the higher tax side of the border where beer is more expensive, the increase in consumption will unlikely correspond to the full level of cross-border beer purchased at a cheaper price.Therefore, the amount of cross-border trade is predicted to differ depending on the sign of the tax differentialthe lower tax side of the border should experience a drop in sales that is larger than the increase on the higher tax side of the border.
In the study, I run regressions where the treatment group consists only of higher tax border districts and a separate regression using only lower tax border districts as the treatment group.I exclude the Lithuanian/Latvian border as the tax differential during this time is close to zero (€0.017 per beer).The lack of a substantial border tax differential would mean that this border region is closer to the control group than to the treatment group.In an unreported robustness check, I also run the regressions with the Lithuanian/Latvian border districts classified as controls and the results are essentially the same.The higher tax border districts to be included in the study are along the southern Lithuanian border with Poland and the southern Estonian border with Latvia.The lower tax border districts are along the northern Latvian border with Estonia and the district of Tallinn, which is designated the border with neighbouring Finland due to the ferry connection.
An event study typically includes dummy variables for a set of time periods before and after an event, making it possible to assess the effects of the policy at each individual time period with regards to a reference point (one excluded time period).The control districts receive values of zero for all time periods and act as the units from which comparisons can be drawn.In the present study, using a simple dummy for all treatment districts would ignore the differences in the relative size of the border tax differentials.Therefore, I substitute the absolute value of the tax differentials in order to incorporate the heterogeneity of treatment sizes.A typical set-up for an event study, which has been tailored to the present case, is presented below Examining the effects of beer excise taxation on cross-border sales in border regions of the Baltic States 625 REGIONAL STUDIES (Clarke & Tapia Schythe, 2020): The first lead is omitted and acts as the baseline of comparison for all other b j coefficients.An important assumption of event study designs requires no anticipatory effects, meaning that there should be no treatment effects contaminating the pre-treatment period.The lockdown started in March 2020; however, it is important to consider the possibility that individuals may have been stockpiling beer in anticipation of the lockdown.Thus, I will designate the treatment as starting in February 2020 to incorporate any possible anticipatory effects in the treatment.
The main coefficient of interest, b j , will measure the effect of the border regions for the months before and after the treatment.The dependent variable (y i,t ) is measured as the log of beer unit sales in each district at each month.For this particular case, I limit the lags to six months and the leads to five months in order to avoid the inclusion of an excise tax change for Estonia in July 2019.In this set-up, m c represents the location effects, l t represents time effects and 1 i,t is the unobserved error term.
Some authors have developed robust estimators to deal with the bias TWFE estimator encounters in the face of heterogeneous treatment effects described above (e.g., Athey & Imbens, 2018;Callaway & Sant'Anna, 2021;de Chaisemartin & D'Haultfoeuille, 2022).Conveniently, de Chaisemartin and D'Haultfoeuille (2022) have developed a weighted DiD estimator that allows for dynamic estimation and is robust to heterogeneous treatment effects even when treatments are non-binary, which is the present case.The authors have written the did_multiplegt program in Stata, which I use to implement their robust estimator.
One of the major assumptions of any event study is that the pre-trends of the treatment and control groups should be parallel.This assumption is often tested by including pre-treatment time dummies and checking that the coefficients do not significantly differ from zero.Sun and Abraham (2021) show that these pre-treatment coefficients can become contaminated by the values of neighbouring coefficients when treatment effects are heterogeneous, calling into question the validity of such parallel trends tests.In response to this issue, the did_multiplegt program constructs placebo tests for the pre-treatment periods which are robust to the previously discussed contamination.
As emphasised by Bertrand et al. (2004), among others, DiD studies often suffer from downwardly biased standard errors due to serial correlation.One of the proposed solutions involves the use of clustered standard errors with a large sample that should preferably include 50 or more clusters.However, the present study only incorporates sales data from three countries, rendering clustered standard errors unfeasible.One proposed strategy to deal with small numbers of clusters is through the use of the wild cluster bootstrap, but this has been shown to massively over-reject when the null hypothesis is applied and massively under-reject in the unrestricted version (MacKinnon & Webb, 2018;Roodman et al., 2019).Another proposed strategy involves randomisation inference, which has been argued as a strategy to correct for serial correlation bias with a small number of treated clusters (Conley & Taber, 2011).Given the issues present in the present dataset, I will use robust standard errors and rely on comparisons with the placebo estimators generated in the pre-treatment periods for inference.

Results
In the event study results shown below, the border shutdown lasts from t ¼ 1 (March 2020) to t ¼ 3 (May 2020).As the borders were closed during April (t ¼ 2), I will focus on t ¼ 2 as the target treatment date.The results presented in Figure 4 and Table 3 show that beer sales increased in higher tax border districts during the lockdown and essentially returned to their pre-treatment levels after the travel restrictions were lifted.As compared with the baseline (t ¼ −1), beer sales were 14% higher in the border areas relative to the control non-border districts at the height of the border closure.It should be noted that this figure corresponds to a full-treatment effect, but the treatment weights for each border were calculated according to the border excise tax differentials, meaning the true effect will vary significantly by border.This finding strongly suggests that consumers who normally bought beer across the border in lower tax districts substituted their consumption for higher taxed beer at home.In the lower tax border districts sales decrease by 52% relative to the control districts (Figure 5 and Table 4), implying the existence of the cross-border trade being much larger on the low tax side of the border.In conclusion, the effect of the border lockdown was 3.7 times stronger on the lower tax side of the border as compared with the higher tax side.

Robustness checks using districts that border Russia and Belarus
As a counterfactual, I also test the effects of border closures on beer sales along the Russian and Belarusian borders.During the time period, Russia's excise tax rate was €0.1375 per beer and Belarus's rate was €0.0648 per beer (the excise tax rates in Russia and Belarus are levied on a per unit of beer basis, regardless of strength).These relatively low rates of tax would imply a large amount of cross-border shopping.However, these countries are not in the EU and, thus, are subject to import restrictions.Consumers may only bring 16 litres of beer into the EU when coming from abroad, while within EU countries, 110 litres of beer may be transported.This in addition to visa restrictions implies that cross-border shopping with Russia and Belarus should be severely limited as compared with within the Schengen Zone.Therefore, there should not be any large effect on sales due to border lockdowns.However, using the Belarusian and Russian border regions as a counterfactual may be less than ideal if illegal smuggling takes 626 Aras Zirgulis

REGIONAL STUDIES
place in large volumes.Instead of increasing domestic beer sales during the border closure, residents of these districts may instead resort to consumption of undocumented, contraband beer.While the smuggling of tobacco and hard liquor products across Schengen borders is a significant concern, there has not been one recorded incidence of beer smuggling from Russia or Belarus in recent memory according to Lithuanian customs. 5 To perform this robustness check, I use the same empirical method as in study 1, setting the treatment group as districts that border Russia and Belarus, while using the interior Baltic districts as the control group.I also eliminate all other border zones from the sample to prevent contamination.The results are presented in Figures A2 and A3 and Tables A1 and A2 in Appendix A in the supplemental data online.These results show that there were no statistically significant sales changes in the Russian and Belarusian bordering districts during the COVID-19 border closures.
I also run event studies at the country level for each individual border but do not report them as they are in line with the aggregate findings.The individual border results for either side of the Lithuanian/Latvian border fail the parallel trends assumption test, and are thus not reported.

Robustness checks using price data
In study 1, I rely on the assumption that price trends are parallel between interior and border regions during the COVID-19 border closures.If borders experienced larger  Examining the effects of beer excise taxation on cross-border sales in border regions of the Baltic States 627 REGIONAL STUDIES price changes than interior districts in response to the pandemic, then the interpretation of the results as representing the overall level of cross-border shopping would be misleading.In order to test the assumption that price movements were parallel, I compare interior versus border prices using an event study design as already described in study 1.For this test, I was only able to acquire average monthly price data at the district level for 18 Lithuanian districts, including interior and border observations.I compare prices of interior districts with those districts closest to the Polish and Belarusian borders, as these two countries had lower excise tax rates than Lithuania during the COVID-19 pandemic.Specifically, the excise tax in Belarus at the time was €0.0648/0.5 litre (as long as the alcohol content is below 7%).In Poland the excise tax was €0.1092 and in Lithuania it was €0.1778 for a 0.5 litre of 5% strength beer.
For the Belarusian border municipalities I use Varėna and Ignalina, while for the Polish border municipalities I use Alytus and Marijampolė.The location of these districts is illustrated in Figure A4 in Appendix A in the supplemental data online.One problem with these data is that the Polish border municipalities do not lie directly on the Polish/Lithuanian border, but are connected with the two major highways that go to Poland, and are theoretically close enough for residents to engage in cross-border shopping.In addition, these data collected from the National Statistics Bureau of Lithuania consist of average prices from all beer types, while the present study is focused on beers sold from the Carlsberg group (which has a 40% market share in the Lithuania).
The results of this robustness check indicate that there are no statistically significant changes to prices along the low-tax borders.The data also show that the standard deviation of price changes increase during the lockdowns by only a few euro cents, which suggests that no large price differentials emerged during this time period.For these results, see Figures A5 and A6 and Tables A3 and  A4 in Appendix A in the supplemental data online.

Study 2: Excise tax changes and heterogeneous border effects
An ideal way to study the effect of excise tax changes on cross-border beer sales would be to take a country without any excise tax and then introduce an exogenous tax change.However, all the countries in the study begin in a treated stateeveryone is already taxed.Thus, I am only able to measure the effects of an increase in the intensity of treatment as opposed to a more desirable situation of going from no treatment to full treatment.
Earlier related research has relied on TWFE models to overcome the problem of multiple tax rate changes within the same country/state/municipality.Most notably, two prominent studies that investigate the effect of corporate taxation on employment have employed a distributed-lag model (Fuest et al., 2018;Serrato et al., 2016).As an alternative study design, an event study may also be used.However, given certain assumptions, it has been shown that distributed-lag models are equivalent to event study designs (Schmidheiny & Siegloch, 2019).Regardless, TWFE models with heterogeneous effects and staggered adoption have already been shown to be biased (de Chaisemartin & D'Haultfoeuille, 2020).Thus, instead of including every tax change in the sample, I propose to break down the data into three individual tax event DiD studies as it avoids many of the OLS-related weighting issues associated with TWFE models already mentioned (Goodman-Bacon, 2021).In addition, the effects of large tax changes are potentially more relevant to policymakers in democratic societies due to the salience induced by media reports.
As Latvia is sandwiched between Lithuania and Estonia and has few large tax changes of its own, I will use this country as the basis for measuring tax changes in its neighbours.I measure how tax changes in Lithuania and Estonia affect Latvian border regions relative to Latvia's interior regions.This proposed method is similar to that used by Ye and Kerr (2016) in a study on a law change on cross-border alcohol sales.When there is an increase in the excise tax in a border region, it is difficult to disentangle how much of the decrease was due to cross-border substitution and how much to a pure decrease in consumption.However, this problem can be avoided by simply looking at sales on the other side of the border.
In order to reduce the impact of other tax changes on the study results, the data sample is limited to 12 months before and after each tax change in question.I further limit the data window for the tax increase in Estonia (July 2017) because the post-period overlaps with tax changes in Finland (January 2018) and Latvia (March 2018).In an unpublished study by TAK Research Pasi Nurkka in Finland, 8.0% of all alcohol brought back to Finland from abroad was originally purchased in Latvia, implying that concurrent tax changes in Finland would contaminate the results.
A general formula for a DiD study is presented below: In the DiD set-up for an individual tax change, y i,t is the logarithm of the number of beer unit sales each month in each region.I include tax i,t as a dummy variable with values of 0 before the implementation of the tax, and 1 afterwards.I also include the dummy treatment i,t that codes for the border region.In this case, b 3 measures the causal DiD effect, and 1 i,t is the unobserved error term.
In addition to testing how excise taxes affect sales across the border, I also test the hypothesis that border districts experiencing a domestic tax increase will reduce consumption at a greater rate than interior districts of the same country.I predict that many residents of the border districts will choose to substitute their domestic beer purchases for equivalents on the other side of the border.Thus, for the Lithuanian excise tax increase, I will designate the northern Lithuanian border as the treatment group and the interior districts will act as the control group.A corresponding set-up will be implemented with the Estonian tax changes, where the southern Estonian border will be designated the treatment group and the interior Estonian regions will be the controls.It should be noted that this design relies on the common shock of the domestic excise tax to be equivalent between the border and interior regions so that it can be differenced out in the DiD process.
As a robustness check, I test two placebo tax changes for the Latvian northern and southern borders.This entails running the same regressions at times when there were no cross-border tax changes.The placebo for the first Lithuanian tax increase from March 2017 will instead be modelled as happening January 2019, and the Estonian tax decrease in July 2019 will instead be modelled as taking place in October 2016.These dates (and the length of the pre-and post-periods) were specifically chosen to minimise the overlap with any other excise tax changes.Examining the effects of beer excise taxation on cross-border sales in border regions of the Baltic States 629

REGIONAL STUDIES
As this particular study only uses data from one country at the district level, the normal practice of using clustered standard errors does not apply.Instead, the standard errors will be computed using a wild bootstrap combined with the six-point weight distribution method (Webb 2013).This type of inference is recommended under conditions of heteroscedasticity and small sample sizes.I also check the inference using the method of Donald and Lang (2007), which I do not report as the results are roughly the same as using the wild bootstrap.

Results
Table 5, columns 1, 3 and 5, represent the DiD estimated cross-border effects, that is, the degree to which Latvian border sales were affected relative to interior district sales by neighbour tax changes.The effects are all the expected sign and differ relative to the size of the respective tax changes.Columns 2, 4 and 6 suggest that for both the Lithuanian and Estonian tax changes, the domestic border regions did not experience any significant change in beer sales relative to their respective interior districts.These findings show that cross-border trading effects are strong on the lower tax side of the border and that changes in beer purchases on the higher tax side of the border are not significantly different to interior districts.As an explanation: lower tax border districts receive shoppers from the entire neighbouring country, while on the high tax border, the drop in sales is only coming from local residents.Nevertheless, the results here show that the border effects of tax changes are asymmetric.Finally, the placebo tests from columns 7 and 8 both produce statistically insignificant coefficients, bolstering confidence in the previously mentioned results.
In order to compare the effects of the three main tax changes, I take the coefficients from Table 5 (columns 1, 3 and 5) and divide them by the size of the tax changes.This provides the approximate percentage change in cross-border beer sales per €0.01 change and can be summarised as follows: . Lithuania, March 2017: 2.23%.
All three tax changes are within the same range in terms of effect per euro cent change.One surprising aspect of this result is that effect of the tax decrease in Estonia 2019 (the only tax decrease in the sample) is similar in size to the tax increases.It has been argued that retailers tend to pass on tax increases to consumers but pocket tax decreases (Ardalan & Kessing, 2021), but this does not appear to be the case for the Estonia 2019 tax reduction.
In order to examine the parallel trends assumption, I include a graphical diagnostic of the observed means and linear trends for the tax changes identified in the study.Figures A7-A9 in Appendix A in the supplemental data online show that while the pre-trends do not perfectly match, they follow each other closely in the provided examples.There is also a clear change in slope between the treatment and control groups after the implementation of the tax changes.Finally, Table 6 includes an F-test for the parallel trends assumption for the three main regressions.This test rejects the parallel trends hypothesis for the Estonian tax increase in 2017, which calls into question the causality implications.The difference in pre-trends is possibly due to the fact that SuperAlko, a major player in the cross-border alcohol trade, opened two large stores on the northern Latvian border in October and December 2016, which had a large effect on beer sales in the border districts relative to the interior.As the construction of the new cross-border alcohol shops increased the growth of beer sales in the pre-treatment period, which means this particular DiD estimate may be biased.

Robustness checks using price data
In study 2, I rely on the assumption that excise tax increases are passed onto consumers as price changes, which then affect cross-border consumption.Thus, it is important to check that prices are in fact changing in line with excise tax changes.In order to test this, I obtained country-level average retail price data for the Carlsberg group in monthly intervals from Nielsen for all three countries.I employ the DiD method discussed above to test the impact of excise tax changes in Lithuania and Estonia, except that for each tax change I use the other two countries as the control group.
The main limitation of these data are that Nielsen only has prices aggregated at the national levels.However, these prices represent averages for the beers in the Carlsberg group and are significantly superior to having average prices for all beer types.The lack of price data at the municipal level does not allow it to be tested whether price changes are symmetric for border and interior regions.However, the findings from the previous robustness check in section 3.1.3for Lithuanian price data suggests that border prices did not significantly differ from interior prices.Thus, I will assume that this finding applies to the other Baltic States as well.In addition, as there are only three groups in the sample, the standard errors produced by the DiD estimator are biased and will therefore not be reported here.The history of the price data for each country is presented in Figure 6.The red vertical lines indicate the three excise tax changes under study in addition to the date of the border closures.Visually, the impact of the excise tax changes is apparent at each tax change, but no impact is immediately discernible from the border closures.
The DiD estimates presented in Table A5 in Appendix A in the supplemental data online show that the excise tax changes resulted in a €0.0625 increase in Lithuania (2017), a €0.166 increase in Estonia (2017) and a €0.128 decrease in Estonia (2019).
These price changes suggest a price pass through rate summarised below: Examining the effects of beer excise taxation on cross-border sales in border regions of the Baltic States

REGIONAL STUDIES
Comparing these results with those of study 2, the high rate of price pass-through is consistent with the strong changes in beer purchases observed in the study.This pass-through finding is also in line with studies that show that premium type beers (higher initial price) have higher rates of pass through than cheap brands (Ally et al., 2014).

CONCLUSIONS
By exploiting the exogenous COVID-19 induced border closures and a series of large excise tax changes, this study evaluates the impact of excise taxation on beer sales in the different border regions of the Baltics using data from the Carlsberg group and representing a large portion of the total beer market.Starting with an eventstudy design to examine the pre-existing state of the cross-border beer trade, I find that the average full treatment effect of border closures in April 2020 increased beer sales by 14% in the higher tax border areas and reduced sales by 52% in the lower tax border districts.These figures are not directly applicable to individual borders as their treatment strengths in the study depend on the level of the border tax differential.However, the findings can be interpreted to mean that the effect of excise tax on cross-border shopping is on average 3.7 times stronger on the low tax side of the border than on the high tax side.This suggests that cross-border shoppers are willing to travel great distances (not only from the neighbouring district) to buy cheaper beer across the border.
Next, I use a simplified DiD approach to investigate the causal impact of large tax increases in Lithuania (March 2017) and Estonia (July 2017), as well as a large tax decrease in Estonia (July 2019) on cross-border sales in neighbouring Latvia.In all three cases, I establish significant changes in beer sales in the Latvian border regions as compared with the interior districts.Interestingly, In Lithuania and Estonia there were no significant effects on the domestic side of the border as compared with domestic interior districts from these tax changes.For all three tax changes, a €0.01 change in excise tax corresponds to a 2% (approximately) change in beer sales just across the border.
As part of the robustness checks along the Russian and Belarusian borders, where cross-border shopping is hampered by customs limits and visa restrictions, I found no evidence of sales changes during the border closures.It is noteworthy that regions adjacent to non-Schengen countries did not experience significant changes in beer sales, as excise tax differentials were quite large along the border.This result points to the effectiveness of the import and customs controls, which only allow 16 litres of beer to be imported from non-EU states, in preventing cross-border shopping.
A further check on the absence of significant price changes between border and interior districts during the  COVID-19 border closures suggests that price trends were parallel before and during the closures.This finding suggests that no large-scale price discounts occurred in border areas, which would bias the results of the study if they did exist.In addition, price data also supports the existence of a high degree of pass-through of excise tax changes to Carlsberg group beer prices.Thus, the price data are consistent with the evidence of large changes in beer sales reported on the other side of a border following domestic excise tax changes.The present study contributes to the literature on the regional impact of excise taxation in several ways.First, the policy setting of the Baltic States engaging in large and relatively frequent tax changes is atypical.No other study that I am aware of uses multiple large excise tax change events in a small geographical area.In addition, the COVID-19 border closures represent a perhaps once in a lifetime opportunity to study an exogenous shock to cross-border beer sales.Finally, the use of the state-ofthe-art weighted DiD estimator proposed by de Chaisemartin and D'Haultfoeuille (2022) has made this study possible due to its ability to produce estimates that are robust to heterogeneity in several dimensions, which are present in the data.
Cross-border shopping can impact border regions as a source of investment and jobs with the construction of specialised alcohol shops, but this investment is highly susceptible to changes in tax rates.For example, the results show that when Lithuania increased its excise tax rate in March 2017, a cross-border trade developed on the Latvian side of the border.However, Latvia slowly raised its rates over the years, and by the time of the COVID-19 border closures in 2020, the tax rates were almost equal and evidence of a cross-border beer trade ceased to exist in the data.
The results of this paper imply that higher tax border regions suffer from a loss of excise and VAT revenue, in addition to the likely negative effects of local residents consuming more alcohol purchased across the border.Furthermore, the border regions in the Baltics are predominantly rural, and thus already suffer from lower levels of economic development and emigration (Smętkowski, 2013) with higher levels of alcohol abuse (Grigoriev et al., 2020), compounding any negative effects associated with the cross-border alcohol trade.This problem likely fuels continued regional inequality, which is a challenge that needs to be addressed by both policymakers and academics.Future research in this area can focus on wine, spirits and the non-premium beer brands that were not included in the present study to develop a clearer understanding of the overall effect of excise taxation and its effect on cross-border shopping.
Ultimately, governments not only need to take into account domestic considerations when planning excise tax legislation, as the present study has illustrated that large differences in cross-border excise tax rates fuel inefficient cross-border shopping by consumers.One potential policy suggestion involves the devolution of alcohol taxation powers to the local border district authorities.This would allow districts on the higher tax side of the border to reduce the tax differential enough that cross-border shopping becomes uneconomic when considering time and transportation costs, while allowing internal districts the ability to maintain higher tax rates.EU excise tax harmonisation is another policy that would limit the inefficiencies associated with cross-border beer sales.An elementary form of beer excise tax harmonisation exists within the EU today: €0.0187/100 litres and 1% alcohol.However, Baltic State tax levels already far exceed this threshold.It appears unlikely that this prisoner's dilemma-type game will result in harmonised tax rates without the imposition of a higher minimum tax threshold at the EU level, meaning that the cross-border beer trade is here to stay for the foreseeable future.

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the author.

NOTES
1.The Schengen Zone consists of 26 European countries that have abolished internal border checks and passport controls with a common visa policy.2. Counties, municipalities and districts are the English translations used by the respective governments.3.These figures come directly from the Carlsberg group.4. In the United States, the analogue would be 40 oz.containers of malt liquor.5.This was confirmed in an interview with Gediminas Kulikauskas from Lithuanian Customs (https:// lrmuitine.lt/).Written informed consent was obtained from the interviewee regarding publication of material related to his interview and to be identified.6.Assuming a 5% alcohol content.

Figure 2 .
Figure 2. Baltic States at the district level showing the interior (light shading) versus border districts (dark shading) used in the study.Source: Mapmaker Klimantas, www.mapklimantas.com

Figure 3 .
Figure 3. Excise tax levels in the Baltics over time.Note: Excise tax is calculated in euros for a 0.5-litre beer with 5% alcohol content.

Figure 4 .
Figure 4. Effect of border closures on higher tax borders.Note: The month of full border shutdown corresponds to t ¼ 2 (April 2020).All changes are relative to January 2020 (t ¼ −1).To the right of t ¼ 1 are the difference-in-difference (DiD) estimates of the effect of the border closure on the logarithm of beer sales.To the left are the DiD placebo estimates that should not significantly differ from zero if the parallel trends condition is satisfied.Standard errors are estimated using 100 bootstrap replications.The 95% confidence intervals based on a normal approximation are shown in red.

Figure 5 .
Figure 5.Effect of border closures on lower tax borders.Note: The month of full border shutdown corresponds to t ¼ 2 (April 2020).All changes are relative to January 2020 (t ¼ −1).To the right of t ¼ 1 are the difference-in-difference (DiD) estimates of the effect of the border closure on the logarithm of beer sales.To the left are the DiD placebo estimates that should not significantly differ from zero if the parallel trends condition is satisfied.Standard errors are estimated using 100 bootstrap replications.The 95% confidence intervals based on a normal approximation are shown in red.

Figure 6 .
Figure6.Beer price history.Note: 2017m3 is when Lithuania raised its excise tax rates; 2017m7 and 2019m7 are when Estonia changed its excise tax rates; and 2020m3 marks the closing of the borders due to the COVID-19 pandemic.This retail price data are for beers from the Carlsberg group in the Baltics, measured in euros price per litre.
Note: Price data are in euros and only reported at the national level, while all other data are reported at the district level.

Table 2 .
Cross border tax differentials in March 2020.
Note: Values are the tax amount differential for a 0.5 litre beer of 5% alcohol in euros.

Table 3 .
Effect of border closures on higher tax borders.
Note: LB CI, lower bound 95% confidence interval; UB CI, upper bound 95% confidence interval; and N, number of districts.'Switchers' refers to the number of border districts included in the treatment group.

Table 4 .
Effect of border closures on lower tax borders.Note: LB CI, lower bound 95% confidence interval; UB CI, upper bound 95% confidence interval; and N, number of districts.'Switchers' refers to the number of border districts included in the treatment group.

Table 5 .
Difference-in-difference (DiD) results of excise tax changes in Lithuania and Estonia.The far left column describes the treatment versus control groups in the country of study.Column titles refer to the individual tax change locations and times.