The Swedish Air Passenger Tax

(1)

Bachelor Essay

The Swedish Air Passenger Tax

- The impact on passenger numbers

Author: Alexander Soto Runevall 19930617

Supervisor: Thomas Giebe Examiner: Mats Hammarstedt Term: Spring 2021

Subject: Economics Level: Bachelor

Course code: 2NA11E/2NA12E

(2)

(3)

Abstract

In this essay I have estimated what impact the Swedish air passenger tax that was introduced in Sweden on April 1^stin 2018 has had on passenger numbers in Sweden. The tax was introduced to lower demand for air travel within the public and thus, decrease the aviation industry´s greenhouse gas emissions as aviation contributes through these emissions to a large extent to climate change and is expected to have an even greater impact in the future as the number of air travels are expected to grow continuously. Besides that the emissions from aviation have this large negative impact on the climate, a large part of these emissions are not included in any pricing system or market. Therefore, aviation is a source of negative

externalities and the Swedish air passenger tax is an instrument implemented to correct for these negative externalities. Previous studies that have evaluated similar taxes and their effects have reached different conclusions. I have used the difference-in-difference method to estimate any effect from the introduced aviation tax in Sweden. Denmark is used as the control group to compare the development in Sweden with. I found that the tax has had a decreasing effect on passenger numbers in Sweden. It is discussed however if this estimated decrease in passenger numbers might be biased and thus overstated. It is also discussed if this decrease might be due to other reasons than the air passenger tax itself.

(4)

1. Introduction

In the year of 2015 the Paris Agreement was signed by a large part of the nations around the world. The goal of the agreement is to prevent the average temperature of the world from rising by another 2℃. To achieve this goal climate policies are needed in almost all sectors in all nations, aviation is no exception. Global civil aviation accounts for nearly three percent of total energy-related CO2emissions around the world (IEA, 2020), but in addition to these CO2emissions aviation also produces other greenhouse gases. These are emissions of water vapour and nitrogen oxides and the magnitude of their effect on global warming is estimated to be nearly as big as that from aviation´s CO2emissions (Lee et al., 2010). By the year of 2045, international air traffic and its emissions are expected to be about three times as high as they were before the decline of the air traffic due to the pandemic of Covid-19 (ICAO, 2019).

International air traffic is not included in the Paris Agreement and there does not seem to be any effective international policy to get emissions from aviation down to a low enough level to be in line with the goal of a maximum increase of the world´s average temperature by 2℃.

The European Union has since 2012 included aviation into its emission trading system (EU ETS) so nowadays there is a limit for CO2emissions from air traffic within the EU (European Commission, 2021). More than 80 percent of these permits were though handed out for free to the airlines and the price of carbon emissions is assumed to be too low to have any larger effect on air traffic. The International Civil Aviation Organization (ICAO), implemented in 2016 a Carbon Offsetting and Reduction Scheme for International Aviation (CORSIA) and its goal is to keep emissions from aviation at the 2020 level. Therefore, this regulation does not have the goal to reduce emissions compared with today´s level (ICAO, 2019). Also, only about 80 percent of CO2emissions from aviation are covered by this regulation. Jet fuel for international aviation is also exempt from taxes even though it is assumed that a tax on jet fuel would be the most efficient way to lower emissions from aviation. Neither the EU ETS nor CORSIA covers non-CO2emissions from aviation.

Aviation therefore causes negative externalities and a large part of these negative externalities are not covered by any pricing system that can internalize the social costs of aviation.

Furthermore, the externalities that are covered do not seem to be priced at a proper level that can internalize the social costs. Because the EU ETS does neither cover intercontinental

(6)

aviation nor non-CO2emissions, a case study of Sweden shows that only one third of

emissions from aviation are covered by this system. It is therefore concluded that the EU ETS and CORSIA do not have more than a marginal effect on emissions from aviation in Sweden (Larsson et al., 2019). Because this situation is the same across the globe some nations have taken further actions trying to reduce their emissions from aviation by implementing domestic policies. Some nations have introduced taxes on travels by air, trying to reduce demand for and emissions from aviation. Some of these nations are the UK, Germany, Norway and Sweden.

In Sweden an excise tax was introduced on air travels on April 1^stof 2018. The tax is levied on all air travels departing from Swedish airports, also domestic travels are levied with the tax, and the tax is levied for every single passenger. The amount of the tax varies depending on the final destination of the trip and there are three different levels of the tax. When the tax was introduced the amount of the first level was SEK 60 and belonging to this level are travels to destinations within Sweden and to Europe. The amount of the second level was SEK 250 and belonging to this level are travels to destinations in parts of Africa, Asia and North America.

The amount of the third level was SEK 400 and belonging to this level are travels to destinations in the rest of the world. For the year of 2021 these amounts are now SEK 63, SEK 262 and SEK 418 and the tax is being recalculated annually to follow a consumer price index (Skatteverket, 2021; Sveriges riksdag, 2017). The levels of the tax are similar to those of some other European nations that have also introduced an excise tax on air travel, but compared with the levels in the UK which has the highest tax in the world the Swedish tax is much lower.

The objective of this essay is to investigate if the introduced excise tax on aviation in Sweden has had any effect on the demand for traveling by air and the amount of travels made from Swedish airports. This will be done by analyzing data on air travels in Sweden. This kind of excise tax on aviation has been very much debated and criticized, both in Sweden and in other nations where it has been introduced. It has been criticized for being an ineffective way to decrease demand for aviation and its emissions (Brännlund, 2018). The excise tax might be ineffective if it is set at a low level or if demand for aviation is inelastic. If any of these is true for the Swedish case, then the Swedish excise tax will not have a significant impact on the demand for air travel. On the other hand, if it has had a decreasing effect on the number of air

(7)

travels made, then this kind of policy might be very useful to reduce the emissions from aviation and so other nations might consider introducing a similar excise tax. The experiences from other nations are that similar excise taxes have had some effects on reducing demand for air travels, but it has been questioned if the overall demand really has decreased or if the excise tax only has led to a substitution effect instead. My hypothesis is that the Swedish air passenger tax is set at a too low level for it to have any significant impact on the demand for air travel. The excise tax on aviation in Sweden was introduced only three years ago so it is still quite new. Therefore, little research on the Swedish case has been made so this study might bring information and conclusions about a subject that lacks desirable knowledge.

The rest of the essay has the following structure: Section 2 contains a review on previous studies made within the field of taxation on aviation. Section 3 contains the theoretical background that is used to evaluate the research question. Section 4 discusses what

methodological approach that is undertaken in this essay. Section 5 contains what data that is needed and used for the evaluation. Section 6 contains the results and a discussion of those results. Section 7 contains the conclusion.

2. Literature review

In a report from the Environmental Change Institute, Cairns and Newson (2006) evaluate the growth of air traffic in the UK and the impact it has on aspects like the economy and

environmental goals in the UK. Considering aspects for and against restraints on air traffic to reach the environmental goals in the UK the report concludes that the UK needs to implement new and stricter policies to lower demand for aviation. One such instrument is to raise the Air Passenger Duty (APD), which is a duty levied by the Government on air travel from UK airports, to make aviation less attractively priced. The report states that prices affect demand and raising this duty is the easiest and quickest way to do so. It is stated in the report that the APD is an appropriate instrument to reduce demand because nearly all airlines pass on the duty to their passengers.

The APD was doubled in 2007 and Mayor and Tol (2007) tried to estimate what effects that change would have on aviation in the UK. What they found was that emissions from aviation

(8)

would slightly increase after the charges were doubled due to the reduced relative price difference between short-distance and long-distance travels. The number of tourists traveling to nations near the UK would decrease whereas the number of tourists traveling to nations far from the UK would increase. They concluded that this duty does not have a significant effect on the overall demand for air travels and therefore the emissions from aviation. Drawbacks with the study of Mayor and Tol is that the study does not make conclusions based on observations on actual changes in air travel, but instead the model constructs simulations on different outcomes based on both observations and certain assumptions. Another drawback is that domestic air travels and business travels are excluded, and the study only considers demand for international air travels by tourists. Previous studies have shown though that business travelers are less likely than tourists to respond to price changes.

In another paper by Mayor and Tol (2010) they evaluate what effects different climate policies that have been implemented in nations in Europe have had on arrivals and emissions from aviation by using a model of international and domestic tourist numbers and flows. This model was an upgraded version of the one they used in the paper above. What they found was that these policies do not fulfill all their purposes and don´t result in any significant reduction of emissions. Flight taxes implemented in the UK and the Netherlands both have substitution effects between different destinations. These substitution effects depend on how destinations are grouped into different tax bands since both nations tax passengers depending on the final destination. For the UK the tax has the result that the amount of tourists traveling to the UK decreases when tourists travel elsewhere to avoid the tax. For British tourists there is a switch from destinations outside the EU towards destinations within the EU. For travels made within the EU there is a decrease in the number of travels to nations near the UK and an increase in the number of travels to nations further away from the UK. This has the effect that emissions from aviation in the UK increases. For the Netherlands the tax has the result that there is a decrease in arrivals into the nation as tourists to a large extent instead chose to travel

elsewhere. Therefore, the reduction in income to the Netherlands from tourism is large while the reduction in global emissions is smaller. With the tax, the prices for long-distance travels increase more than those for short-distance travels and the amount of long-distance travels decreases. Between destinations within the same tax band though, the relative prices of long- distance travels to short-distance travels decrease and result in substitution towards long-

(9)

distance travels. So, the overall effect of the tax is that long-distance travels are substituted for short-distance travels, but the very long-distance travels increases although only slightly.

Another study about the Air Passenger Duty in the UK was written by Seetaram et al. (2014) and they investigated its impact on the demand for travels to ten foreign destinations. They did this by estimating what impact the APD has on the elasticity of demand for air travels.

Income, price and tax elasticities were all estimated, and they found that income elasticities vary between 0,36 and 4,11, while price elasticities vary between -0,05 and -2,02. The income elasticities were for most destinations greater than one and these numbers indicate that air travels to foreign destinations is largely driven by income. The negative price elasticities indicate that the demand for air travel decreases when there is an increase in prices in the foreign destination. Tax elasticities estimated indicate that the APD has a decreasing effect on the demand for travels to five of the ten foreign destinations. For all destinations though, the absolute values of the tax elasticities were smaller than one which implies that the demand for air travel is inelastic with respect to changes in the APD. The study concludes that the APD only has had a marginal effect on demand for air travel.

Borbely (2019) investigates what effects the German Aviation Tax has had on passenger numbers for German airports and for airports in other nations that are located close to the German border. The German Aviation Tax is a duty that is levied on passengers departing from airports in Germany. The study finds that there is a significant decrease in passenger numbers for most German airports, but there is also a substitution effect created by the duty.

The passenger numbers for airports close to the German border have increased and for

airports within Germany there is also a substitution effect where passenger numbers for small and regional airports have decreased and passenger numbers for large hubs have remained or slightly increased. Because the duty seems to lead to a substitution of airports rather than a decrease in travels, it can be questioned if the duty really is a good instrument when trying to decrease the amount of travels and the emissions created by these. There are some drawbacks with the study, and one is that tourism related factors that could affect the demand for air travels are not accounted for in the model used. One limitation with the study is that the use of aggregated data implies that no conclusions can be made for different types of passengers and travels. That is if the duty has a different effect on holiday and business travels or short- distance and long-distance travels.

(10)

A study by Falk and Hagsten (2019) investigated what effects the flight departure taxes that were introduced in Germany and Austria in 2011 had on the number of air travel passengers.

They used a dynamic panel difference-in-difference method and found that the taxes had significant negative short-run effects in both nations. In the year of introduction, the number of passengers decreased on average with nine percent and in the following year the number decreased with five percent. They found further that regular airports and large hubs were not affected by the introduction of the tax, but instead the decrease in passenger numbers almost solely took place at airports where low-cost airlines frequently depart from. In contrast to other studies, it was not found that airports in other nations, but near the borders of Germany and Austria, had increasing passenger numbers after the taxes were introduced. Thereby the study shows no substitution effects where the demand for air travel instead increased in these nearby airports in other nations as the taxes were introduced. Why this is the case, as

compared to the experience of the Netherlands, could be that Germany only is exposed to a limited number of airports near its border in surrounding nations. For Austria there are only two airports near its border in the surrounding nations, where the number of connections from one of these is limited.

3. Theory

3.1 Externalities

If a consumer is directly affected by another consumer or producer´s consumption or production without taking part in the economic situation then this situation involves an externality to the consumer. Externalities can be of both positive and negative nature.

Externalities can also be of a kind where a producer is affected by an externality so that the production possibilities of the producer is affected by another producer´s production decision.

In both cases there are goods agents care about that are not priced and sold on any market. If all interactions between consumers and producers are taking place through different markets where all interactions are priced and priced correctly, then it is assumed that markets are functioning effectively and there should be no externalities. Unfortunately, there are many

(11)

markets where these interactions are not priced correctly and sometimes not at all and

therefore this leads to market failures. When there are no externalities the market mechanisms can in principle achieve Pareto efficient allocations, but with externalities present the market is likely to end up in an allocation that is not Pareto efficient. If this is the case, then there are actions that could be taken by institutions to try to correct the market failure. One such institution is the government (Varian, 2014).

A Pareto efficient allocation is a situation where neither agent can be made better off without the other agent being made worse off from an allocation change. In such a situation the marginal rates of substitution between the agents are the same. A common solution to achieve the Pareto efficient allocation is to give out property rights to the agents and when having these property rights the agents can trade with another until the Pareto efficient allocation is reached. Depending on how the property rights are handed out in the first place different Pareto efficient allocation are expected to be reached. All these different Pareto efficient allocations will make up the so-called contract curve and the trading with property rights will lead to an allocation somewhere on this contract curve. Where on the contract curve the allocation will end up depends both on how the property rights are handed out and on the price mechanism that is used in the trading. At the equilibrium point where supply equals demand, the competitive price will be the same as the marginal rate of substitution. If property rights are not well defined, however, then a Pareto efficient allocation will not be reached and the problem with externalities will not be solved (Varian, 2014).

One reason why externalities arise is that producers do not fully have to account for costs that arise with their production. Often producers only have to account for the production cost directly affecting the producer when trying to maximize its profit and these costs are called private costs. But in the case of externalities there are additional costs to the production and these together with the direct private costs make up the so-called social costs. Property rights and other instruments have the purpose to internalize externalities into the costs of production and the profit-maximizing decision. For a Pareto efficient allocation to be reached, the social costs and not only the private costs are to be taken into account. The price should equal the marginal social cost and if all production costs are not internalized so that private and social costs are the same then the market alone will not reach a Pareto efficient allocation (Varian, 2014).

(12)

3.2 Excise or quantity tax

Another way when trying to handle the problem with negative externalities is to implement an excise or also called a quantity tax. This is a way of trying to restrain a consumer´s budget constraint and make the consumer consume less of a good or service. The quantity tax means that the consumer has to pay the government an extra amount for every unit purchased. The quantity tax could also be designed such that it is the seller that pays the tax for every unit sold. If the consumer pays the tax then the amount the seller gets is the amount paid by the consumer minus the amount of the tax. This amount received by the seller would be the same if the tax payment was to be shifted to the seller instead of the consumer, so the equilibrium price will be the same regardless of who pays the tax. The simple equation for this is

PD- t = PSwhen the tax is being paid by the consumer and the price paid minus the tax is equal to the supply price. When the tax is levied on the seller the equation is PD= PS+ t which says that the supply price plus the tax is equal to the demand price. These two

equations are the same which implies that the equilibrium price is the same regardless of who pays the tax. If the tax is levied on the consumer, then this will shift the demand curve down with the same magnitude as the amount of the tax and if the tax is levied on the seller, then this will instead shift the supply curve up with the same magnitude as the amount of the tax.

The new equilibrium quantity will be the same regardless. The quantity will decrease as the price paid by the consumers increases and the price received by the sellers decreases (Varian, 2014).

In Figure 1 it is shown how a quantity tax will affect quantity bought or sold and the equilibrium prices. The initial situation where there is no tax is point B.

At this point the equilibrium quantity Q0

is bought at the price of P0. The price consumers pay is the same as the sellers receive. If a quantity tax is levied on every unit bought, then this will have a Figure 1. Excise tax imposed on consumers

(13)

Figure 2. Special cases of taxation

negative impact on consumer´s demand and with a decreased demand the demand curve will shift down. This can be seen in the figure where the demand curve shifts from D0to D1 and the new equilibrium situation is point A. At this point a lower quantity, Q1, is bought to a higher price for the consumers. They now pay P2instead of P0but the sellers do not receive P2but P1. The difference is the quantity tax (Varian, 2014).

In Figure 2 two special cases are shown.

The figure to the left has a perfectly elastic supply curve while the figure to the right has a perfectly inelastic supply curve. If a tax is imposed, then this will shift the supply curve up and when the supply curve is perfectly elastic the shift will be of the same size as the tax. This means that the entire tax will be passed on to the consumer who will pay p* + t while the seller will still receive p* as before. The seller in this case is willing to supply any amount to the supply price of p*. At any lower price there will not be any supply. If a tax is imposed and the supply curve is perfectly inelastic then the supplied quantity will not change. The supplied quantity will be the same regardless of the tax and the supply price, but at this supplied quantity the consumers are only willing to pay a certain price, p*. Therefore, the tax will be paid entirely by the sellers and the supply price will be p* - t.

Thus, the more elastic the supply curve is, relative to the elasticity of demand, the larger share of the tax will be passed on to the consumers (Varian, 2014).

3.3 Applying the theory on emissions from aviation

The situation with aviation and its impact on climate changes is clearly a case of negative externalities. As described before, a large part of the CO2 emissions are not covered or priced at all by any policies or taxes and thus aviation is a source of negative externalities affecting the climate in a negative and serious way. Examples of how externalities in aviation are not priced correctly within Europe is that the CORSIA regulation doesn´t cover all CO2emissions and neither CORSIA nor the EU ETS has set any price on the other types of greenhouse gas emissions that comes from aviation. Also considering that jet fuel is exempted from taxation it

(14)

is clear that aviation gives rise to large negative externalities. It should be noticed however that even if the situation about European aviation is poor due to the failure of internalizing the social costs, the situation outside Europe is even worse when it comes to internalizing the social costs as climate policies there are even weaker.

Another aspect of how the EU ETS has not been able to internalize the social costs is that the emission permits, that is the property rights, were not handed out in a good way to solve the problem with externalities. Most of the emission permits were handed out for free and because the quantity of permits was so large at the introduction of the system, this had the consequence that the social costs were not internalized into the prizing system of C02

emissions. The incentives for airlines to reduce their emissions were simply not large enough to have a major effect on emission reduction.

The Swedish air passenger tax is a way of trying to internalize the social costs with the use of a quantity tax and thus handle the problem of negative externalities not covered for by other policies. The intention is to increase the cost of air travel for the passengers through the tax and thus decrease their demand, all in line with the theory presented above. As shown above though, the impact on the demanded quantity of air travels depends on how elastic the supply curve is relative to the elasticity of demand. If the supply curve is relatively more elastic, the airlines can pass along a large share of the tax onto the passengers and this should then decrease demand. If the supply curve on the other hand is relatively inelastic the tax will not be passed on to the passengers and the demand for air travels will be unchanged. If the latter is the case then the tax will not be able to reduce emissions from aviation. Because many airfares on the market are priced very low it seems reasonable to think that airlines are not willing nor able to handle the tax burden of an introduced air passenger tax on their own and let the prices of airfares for the passengers remain at the same low level as before. This then means that the supply curve is elastic and demand for air travel should decrease as consumer prices increase. It has also been shown by previous research that airlines pass on increasing fuel costs onto the passengers by increasing the prices of airfares.

(15)

4. Method

To evaluate the Swedish air passenger tax I use a statistical panel data regression model that is commonly used when analyzing the field of economic policy changes. I use a dynamic

difference-in-difference regression model as the method to estimate if the air passenger tax has had any effect on total passenger numbers in Sweden. The model is based on the method used by Falk and Hagsten (2019). The dynamic difference-in-difference method is used to account for general trends among different groups and in my regressions, Sweden is the treated group while Denmark is the control group which Sweden will be compared with. In this case, general trends means that passenger numbers in both nations develop in the same way before the tax was introduced. Differences in passenger numbers after the introduction of the tax could then possibly be explained by the tax. Because this method compares a treated group with a non-treated group it is suitable when evaluating the kind of experiment that an introduction of an air passenger tax can be seen as being. The regression model estimates the difference over time between the treated group that has been exposed to the treatment and the non-treated group that has not been exposed to the same treatment. Therefore, in this case Sweden is the treated group because it introduced the air passenger tax during the time span of the sampling period while Denmark did not introduce such a tax and therefore act as the control group.

4.1 The difference-in-difference method

In contrast to a randomized experiment that does a simple comparison between a treated and a control group, the difference-in-difference method is non-experimental. It tries to estimate the causal effect of an implemented policy when the treatment assignment is non-random and thus, there is no obvious control group. Finding a proper control group is essential though when trying to estimate the causal effects of an introduced policy. The difference-in-

difference method subtracts the after-before difference in outcome in the control group from the after-before difference in outcome in the treated group. Therefore, data on the outcome before and after the treatment must exist for both the treated and the control group. This does not mean however, that the method is always appropriate or that it will generate unbiased estimates of the causal effect (Fredriksson; Oliveira, 2019).

(16)

The basic difference-in-difference method uses data from two groups and two time periods, and the data used can be repeated cross-sectional samples or panel data. Difference-in- difference combines a treated-control group comparison with an after-before approach. This method deducts the after-before difference in the control group from the after-before

difference in the treated group and this leads to two things. If there are other changes also happening in the control group in the same time period, then these factors are controlled for when the after-before difference in the control group is deducted from the after-before

difference in the treated group. Also, if there are characteristics that differ between the treated and the control group and that are important for the outcome, these characteristics effects are eliminated by studying changes over time as long as the treated-control group differences in these characteristics are constant over time. Treated-control group differences in time-

invariant unobservable characteristics are also eliminated as they are deducted and netted out.

The problem arising in cross-sectional studies that unobservable factors cannot be controlled for is with the difference-in-difference method possible to avoid (Fredriksson; Oliveira, 2019).

With two groups and two time periods the difference-in-difference estimate looks like the following:

DiD = (Ytreated, after- Ytreated, before) - (Ycontrol, after- Ycontrol, before) Y is the average value of the outcome variable.

The difference-in-difference method can also be extended to include several treated and control groups and several before and after time periods.

The calculation above doesn´t say anything about the significant level of the difference-in- difference estimate and therefore regression analysis is used. In the below OLS regression the difference-in-difference estimate is obtained as the β-coefficient. A are treated/control group fixed effects, B are before/after fixed effects, I is a dummy variable equal to one if an

observation comes from the treated group in the after period. Otherwise I is equal to zero. ε is the error term.

Y = A + B + β*I + ε

(17)

4.2 Parallel trends and other assumptions

There are a couple of assumptions that the difference-in-difference method relies upon. One of these assumptions is the “parallel trends” assumption and another is that there should be no spillover effects between the treated and the control group. If there are spillover effects then the treatment effect will not be identified or estimated correctly. Another assumption is that the control variables should be exogenous, meaning that they are unaffected by the treatment.

If they are not, the treatment effect will be biased. The parallel trends assumption says that the outcome variable of interest in the treated group would have followed the same time trend as the outcome variable of interest in the control group if the treated group had never been treated. The level of the outcome variable may differ between the treated and the control group due to observable and unobservable factors, but this difference must be constant over time if there is no treatment. This assumption cannot be tested for because the treated group is only observed as treated. What can be done to support this assumption is to use data from time periods before the treatment was introduced and with this data show that the treated and the control group follow a similar pattern in the time periods prior to the treatment. The conclusion that the estimated effect is due to the treatment and not due to a combination of other factors is more credible if such parallel trends are observed in the pre-treatment time periods. One single pre-treatment time period is not enough to check for parallel trends before the introduction of the treatment if a study is to be considered robust. The parallel trends assumption is often supported with the use of graphs that display the outcome variables and thus provide visual support for the assumption. Another aspect that can be visually controlled for by studying graphs of the outcome variable before treatment is that there should be no effect from the treatment before it was introduced (Fredriksson; Oliveira, 2019).

Another approach to provide support for the parallel trends assumption that is more formal is to conduct what can be referred to as placebo regressions. When doing this, the difference-in- difference method is only applied on observations from the time periods before treatment. If doing this, then no significant “treatment effect” should be estimated. So if there are enough observations available, a placebo regression can be run by excluding observations from the time period after the treatment and only include observations from the time period before the treatment (Fredriksson; Oliveira, 2019).

(18)

Additionally, if a control variable has only high values in the treated group and only low values in the control group then incomparable entities are trying to be compared. Instead, an overlap in the distribution of the control variables between the treated and control group and the time periods must exist. Also, the parallel trends assumption is scale dependent. If the outcome variable is constant over time during the time periods before the treatment in both the treated and the control group, then everything is fine. If the outcome variable is not constant over time it matters if the variable is used “as is” or if it is transformed. To solve the issue, data should be used in the form corresponding to the parameter that is to be estimated.

Data limitations, like omitted variables, might also affect the robustness of the estimates (Fredriksson; Oliveira, 2019).

When applying the difference-in-difference method one can also use control variables to achieve more robust estimates. If a treatment is not randomly assigned, then there is the concern that the treated group would have followed a different trend than the control group also without any treatment. If factors that differ between the treated and the control group and that would lead to different trends can be controlled for, then a true estimate of the effect following the treatment can be made. These factors must be exogenous. Thus, control variables should be used to control for the variables that differ between the treated and the control group and that would cause different trends in outcomes (Fredriksson; Oliveira, 2019).

In the basic OLS regression some issues with the standard errors can arise and one of these is that with many time periods the observations can exhibit serial correlation. For several typical dependent variables and especially the treatment variable in studies using the difference-in- difference method, this is true. Observation in the treated and the control group can thus be correlated over time and this means that significant levels can be largely overstated. Therefore, this fact must be corrected if significant levels are not to be overstated. Collapsing all pre- treatment time periods to one before period and all post-treatment time periods to one after period is one way to handle the issue with serial correlation. Thus, one works with two time periods of data only, after first having checked the parallel trends assumption, but this requires many treated and control groups. Even with many treated and control groups in the analysis a drawback with this approach is that the sample size is greatly reduced. Another option is to instead calculate standard errors that are robust to serial correlation and this can be done with the use of econometric software packages (Fredriksson; Oliveira, 2019).

(19)

4.3 The specifications

The two OLS specifications of the model I will use to investigate the research question look like the following:

(1) Ln(Passenger) = β0+ β1*Time + β2*Treated + β3*DiD + ε

(2) Ln(Passenger) = β0+ β1*ln(PassengerLagged) + β2*ln(Population) + β3*ln(GDP) + β4*ln(Airfare) + β5*ln(Accommodation) + β6*Time + β7*Treated + β8*DiD + ε

The first specification is used to give an estimate without using any control variables and the second specification has several control variables included in it. Below are descriptions of the outcome/dependent variable and the control/independent variables.

The outcome variable is the following:

Ln(Passenger) is the dependent variable and it is the natural logarithmic form of the total number of passengers in a nation traveling by air in a single month.

The independent variables are the following:

Ln(PassengerLagged) is the natural logarithm form of the total number of passengers in a nation traveling by air in a single month one year before the observation of the outcome variable.

Ln(Population) is the natural logarithmic form of the total size of a nation´s population and it is a variable observed on a quarterly basis.

Ln(GDP) is the natural logarithmic form of a nation´s aggregate GDP and it is a variable observed on a quarterly basis.

(20)

Ln(Airfare) is the natural logarithmic form of a consumer price index for passenger transportation by air in a nation and it is a variable observed on a monthly basis. The price index is a weighted average of airfares and the unit of measure is an index with base period in 2015 when the index equals 100.

Ln(Accommodation) is the natural logarithmic form of a consumer price index for

accommodation services in a nation such as hotels and hostels and it is a variable observed on a monthly basis. The price index is a weighted average of many different kinds of

accommodation services and the unit of measure is an index with base period in 2015 when the index equals 100. I have chosen to relate this domestic index to the same index for the European Union. With the European Union it is here referred to the 28 member nations between 2013 and 2020. The index is related to the European average by dividing the domestic index in a single month with the EU index in the same month.

Time is a dummy variable indicating if an observation belongs to the time period before or after the Swedish air passenger tax was introduced, that is if an observation belongs to the time period before or after treatment. If the observations are collected from a month before April 2018 the dummy variable is equal to zero. Otherwise, the variable is equal to one. The coefficient of Time estimates the mean value of the outcome variable for the control group in the post-treatment time period relative to the mean value for the control group in the pre- treatment time period. The coefficient therefore estimates the time trend.

Treated is a dummy variable indicating if an observation is collected from Sweden or Denmark, that is if an observation belongs to the treated group or the control group. If the observation is collected from Sweden the dummy variable is equal to one. If the dummy variable is collected from Denmark it is equal to zero. The coefficient of Treated estimates the mean value of the outcome variable for the treated group in the pre-treatment time period relative to the mean value for the control group in the pre-treatment time period. The

coefficient therefore estimates the difference between the treated and the control groups prior to the treatment.

(21)

DiD is a difference-in-difference variable signifying the treated group, that is an observation from Sweden, after the tax was introduced. This is thus the variable that will provide an estimate on if the Swedish air passenger tax has had any effect on the outcome variable.

Epsilon (ε) is the error term and β0is the intercept and the mean value of the control group in the pre-treatment period.

The specification above is similar and shares several variables with many other specifications that have been used to determine air travel and tourism demand. These other specifications have been used both to evaluate similar taxes to the one introduced in Sweden, other policies implemented to affect demand for air travel but also how tourism has been affected in other ways. The control variable Ln(PassengerLagged) is chosen because it has been shown that passenger numbers at different time periods during a year tend to turn out quite similar as what they were at the same time previous years despite events that could possibly have an effect on passenger numbers. Therefore, lagged values of this variable are sometimes used in models to predict air travel demand. Sometimes, also other lagged variables are included in the models. Ln(Population) is chosen as a control variable since passenger numbers tend to be greater in nations with bigger populations and when there is population growth passenger numbers tend to increase along with the population growth. Therefore, population is often included in models to predict tourism demand. Ln(GDP) is chosen as practically all models that predict air travel and tourism demand include this variable. It has been shown that one of the most important factors behind air travel is income and GDP is the utmost used indicator of income in air travel models. The other most important factor to determine demand for air travel is airfare prices and so, Ln(Airfare) is chosen as a control variable in the model used in this essay. When no data on airfare prices are available, price on jet fuel is often used as a proxy for this as increases in the price of jet fuel are often to a very large extent levied on to the consumers by increases in airfares. Another variable that is sometimes included in models to determine tourism demand is one that accounts for accommodation prices and therefore Ln(Accommodation) is chosen as a control variable.

(22)

5. Data

I have collected data from Sweden and Denmark to evaluate the research question. The data collected is pooled or combined data that consists of a combination of both time series and cross-section data. Because the data is collected for the same two cross-sectional units, that is the nations Sweden and Denmark, over the same time period the data is further specified as being a set of panel data. The number of time periods is 72 and these are monthly time periods from the years 2014 to 2019. The number of observations for the two nations are 72

respectively and thus the whole sample consists of 144 observations. Both nations have the exact same number of observations making this set of data a balanced panel. Because the number of time periods is greater than the number of cross-sectional subjects this data set is what is called a long panel. Most of the observations are collected on a monthly basis, but there are also observations collected on a quarterly basis.

Besides the reason that I need panel data to use the method I have chosen to evaluate the research question, there are some general advantages with panel data over pure cross-section or time series data. Panel data relate over time to the different units, whether these are individuals, households or nations as in this case, and there will be heterogeneity within the units. When performing panel data estimations heterogeneity can be taken into account by using subject-specific variables for different units. It can be said that the combination of time series and cross-section data gives more variability, less collinearity among variables and more degrees of freedom. The dynamics of change are more easily studied with the repeated cross-section observations in a panel data set because some changes cannot be observed in the observations on pure time series or cross-section data. With panel data such changes and their effects can more easily be detected and measured. Panel data also enables the study of more complicated behavioral models. If data is divided between a very large set of units, panel data is able to minimize potential bias that can result if units are aggregated into broad categories (Gujarati; Porter, 2009).

To evaluate the research question, I have used data on passenger numbers for both arrivals and departures from Swedish and Danish airports. Passenger numbers for both domestic and international travels have been used. It is aggregated data for all airports on total passenger numbers for both nations that have been used. Passenger numbers have been collected for the

(23)

years 2014 to 2019 on a monthly basis. The data is collected from Eurostat, the European Statistical Office, (2021a). It is the department of the European Commission responsible for providing statistical information to the institutions of the European Union. Because passenger numbers are aggregated and therefore not observable on airport level in this set of data the number of observations are limited to 144 in total and only 72 for each nation. Therefore the sample size is very small compared to other studies made in the same field. Because the Swedish air passenger tax was introduced on the 1^stof April in 2018 there are 51 observations before this event and 21 observations after the event for each nation. The month of December 2019 is the last observation because soon after that the Covid-19 virus struck the World and then the demand for air travel fell drastically. If observations from 2020 had been included in the data set then that would have biased the data. I will perform a placebo regression on the observation before the Swedish air passenger tax was introduced to do a further check, besides only a visual check, to see if the parallel trends assumption is fulfilled. As described in the method section of this essay, to do such a placebo regression and investigate the parallel trends assumption in general there should be enough pre-treatment time periods in the data. I will do this placebo regression on the years 2014 - 2017, thus investigating the four years in my data set prior to the treatment. A time span of three years should be enough to investigate the parallel trends assumption properly.

Total population sizes for both Sweden and Denmark have been collected from Eurostat (2021d) and these are collected on a quarterly basis for the years 2014 to 2019. From Eurostat (2021b) data on GDP for both nations has also been collected and this is quarterly data from 2014 to 2019. I have also used consumer price indices for my evaluation of the research question. These are two different consumer price indices; one is an index on air passenger transport and the other one is an index on accommodation services. They are collected separately for both nations on a monthly basis from the years 2014 to 2019 and the data is collected from Eurostat (2021c). Total passenger numbers for arrivals and departures from Swedish and Danish airports one year before a single month in the time span 2014 to 2019 have also been used. These are aggregated data for all airports in a nation for both domestic and international travels and they are collected on a monthly basis. The data for these are collected from the years 2013 to 2018 (Eurostat, 2021a).

(24)

Table 2 - Descriptive statistics Denmark

Obs Mean Std. dev. Min Max

Passenger 72 2.703.004 487.045 1.828.160 3.686.607

PassengerLagged 72 2.601.330 483.499 1.760.371 3.686.607

Population 72 5.738.458 62.223 5.631.000 5.825.000

GDP 72 72.158 4.194 65.615 79.043

Airfare 72 103,08 11,04 83,00 135,60

Accommodation 72 1,059 0,118 0,913 1,484

Table 1 and Table 2 above provide some descriptive statistics on variables included in the specifications used to evaluate the research question. Table 1 provides values for Sweden and Table 2 provides values for Denmark and the tables provide different values on the amount of observations, the means, the standard deviations, minimum values and maximum values for the variables. It can be seen that the mean for passenger numbers is greater in Sweden than in Denmark and also looking at the maximum and minimum values for passenger numbers this is the case. The mean value for lagged passenger numbers is also greater in Sweden than in Denmark, but in both nations the mean value for lagged passenger numbers is smaller than the mean value for passenger numbers. That is, the mean value for passenger numbers is smaller one year before a single month in both nations. This indicates that there is a growth in passenger numbers for each year in both nations and therefore they might share a common trend in the outcome variable. This will be investigated further later. When looking at the mean for the Airfare variable it can be seen that it is below 100 in Sweden and above 100 in Denmark. Because this is a price index variable, a value of 100 for an observation means that the observation is priced in line with the reference price in the base period. As the base period is in 2015 and 2015 is in the start of the sampling period as it is the second year of six in the

Table 1 - Descriptive statistics Sweden

Obs Mean Std. dev. Min Max

Passenger 72 3.024.242 453.689 2.142.331 3.837.942

PassengerLagged 72 2.938.533 463.536 2.028.734 3.837.942

Population 72 9.987.861 209.976 9.655.900 10.321.790

GDP 72 116.097 3.608 109.293 121.295

Airfare 72 96,65 14,05 76,46 142,39

Accommodation 72 1,008 0,044 0,907 1,094

Table 1. Descriptive statistics Sweden

Table 2. Descriptive statistics Denmark

(25)

sampling period, one might believe that a deviation from 100 in the mean value could indicate how prices have developed since 2015. If one makes such a conclusion then the prices for airfares have decreased a little in Sweden and increased a little in Denmark. There is more fluctuation in airfare prices in Sweden though, which is indicated by the greater standard deviation and the greater difference between the maximum and minimum values. Since the Accommodation variable is a domestic price index for accommodation services that is related to the same price index for the EU, it provides information about how relative prices for such services have developed in comparison to an European average. The mean value in Sweden is very close to one and gives that the price of accommodation in Sweden has become more expensive compared to an European average, but less than one percent more expensive than what it was. For Denmark the mean value is higher and slightly above one. Accommodation has become nearly six percent more expensive compared to an European average than what it was. There is also much more fluctuation in prices in Denmark, which is indicated by the greater standard deviation and the greater difference between the maximum and minimum values.

5.1 Visual inspection of pre-treatment data

Figure 3. Passenger numbers in both nations during the sampling period

(26)

In Figure 3 passenger numbers for Sweden and Denmark can be seen from the start of 2014 until the end of 2019. Time is on the horizontal axis and passenger numbers are on the vertical axis. It can be seen that passenger numbers in Sweden are greater than in Denmark in

practically all months during the time period of the study, but the difference is at some times much smaller than at other times. It is clear that both lines follow the same pattern. They both follow a seasonal pattern where passenger numbers increase during the summer months and then fall back a lot during the winter months. What can also be seen is that passenger numbers in both Sweden and Denmark follow an increasing trend as time goes by. This is at least true up till the end of the time period, that is from 2014 up till 2018. In 2018 though, the increase in passenger numbers for each year that can be seen in both nations stops in Sweden and this in contrast to what happens in Denmark. In 2019, the last year of the time period studied, there is also a decrease in passenger numbers in Sweden. Thus, this decrease happens shortly after the Swedish air passenger tax is introduced. If the difference-in-difference method is to work well then both the treated and the control group have to share the same trend. It can clearly be seen that both Sweden and Denmark do share the same trend in increasing

passenger numbers up to the point of the introduction of the tax. The fact that the increase in passenger numbers in Sweden stops around the time of the introduction of the tax is also a crucial point for the study. As was stated in the method section of this essay, for the parallel trends assumption to be fulfilled then no treatment effect should be present before the

treatment starts. From the graph it can thus be seen that no treatment effect took place before the introduction of the tax as passenger numbers continue to grow in all time periods before 2018.

(27)

As was stated in the method section of this essay, the parallel trends assumption is scale dependent and depending on if the outcome variable in both the treated and the control group is constant or not during the time period before the treatment starts, it matters in what form the outcome variable is used. If the outcome variable is not constant, then it is suggested that it should be used in the form one wishes to estimate it. As can be seen in the graph in Figure 3 and what has been stated above is that the outcome variable, that is passenger numbers, is not constant during the time period before the treatment. Therefore, I have made a second graph which is displayed in Figure 4 where passenger numbers are shown in logarithmic form. Time is still on the horizontal axis, but here the logarithmic form of passenger numbers is on the vertical axis. Even though the outcome variable is transformed, Figure 4 is still very similar to Figure 3. Passenger numbers are greater in Sweden than in Denmark at practically all times, but at times the difference is greater and at other times it is much smaller. The same pattern as was shown in Figure 3 can also be seen in Figure 4. The seasonal pattern is very clear for both nations with the big difference between the summer months and the winter months. In Figure 4 it can still be seen that passenger numbers in both nations increase up till the end of the time period investigated, that is during the time period prior to the introduction of the air passenger tax. Thus, this graph also provides visual support for parallel trends before treatment. As in Figure 3, the increase in passenger numbers in Sweden stops in 2018 and decreases slightly

Figure 4. Passenger numbers in logarithmic form in both nations during the sampling period

(28)

after that. Also, according to this graph, this development is not the same as in Denmark where no decrease in passenger numbers can be seen. Also, from this graph in Figure 4 it can be seen that the increase in passenger numbers in Sweden stops around the time the tax is introduced. This supports the parallel trends assumption as passenger numbers in both nations increase in the time period prior to the treatment and a decrease in passenger numbers in Sweden cannot be seen before the treatment starts. If an effect had been detected in the graph before the treatment started, then the parallel trends assumption would not have been fulfilled.

6. Results and discussion

6.1 Placebo regression

Table 3 below shows the result from running a placebo regression on time periods prior to the introduction of the Swedish air passenger tax. This is done to provide further support to the parallel trends assumption. The time periods included in this regression are the months in the years 2014 - 2017. Thus, this time period consists of 96 monthly observations which have been divided in half to provide two time periods with 48 observations each. Therefore, the whole years of 2014 and 2015 are included in the pre-treatment period, while the whole years of 2016 and 2017 are included in the post-treatment period. Each nation has 24 observations in each time period. It is important for the assumption of parallel trends to be fulfilled that the coefficient of the difference-in-difference variable is both statistically insignificant and insignificant in an economic interpretation. What can be seen in the table is that the

coefficient of the DiD variable, which is the difference-in-difference variable, has a value of about -0,004 which is practically zero. Thus, no treatment effect in this pre-treatment period is detected from this regression. The value of the coefficient of the DiD variable is also far from being statistically significant, as its corresponding p-value is 0,949 which is very high and thus far away from being even close to being statistically significant. These two facts support the statement made earlier that the parallel trends assumption is fulfilled, and that Sweden and Denmark share a common trend in the time period before treatment.

(29)

6.2 Main regressions

Below are two tables that show the results from running the two different OLS regression model specifications in Stata. Both regressions are run using robust standard errors, this in line with what was discussed in the method section. Table 4 shows the result from running the first specification. This specification is the really simple one and in this specification no control variables are added. When no control variables are added to the specification the regression only estimates the raw difference-in-difference of the observations from the two different groups. DiD, the difference-in-difference variable, has a negative coefficient indicating a decrease of about three percent. It is though quite far away from being

statistically significant so therefore, no conclusions about the DiD variable can be drawn from this simple model. Thus, it cannot be said that the tax has had or has not had any effect on passenger numbers. The R² value is only 0,2046 so this model does not contain enough variables to explain to any greater extent how and why passenger numbers vary. About 80 percent of the total variance of passenger numbers are not explained by this specification.

Table 3 - Placebo regression

Linear regression Number of obs = 96

F(3, 92) = 8.57

Prob > F = 0.0000

R-squared = 0.2214

Root MSE = 0.16003

| Robust

lnPassenger | Coef. Std. Err. t P>|t| [95% Conf. Interval]

Time | 0.1130737 0.0499411 2.26 0.026 0.0138863 0.212261 Treated | 0.1269706 0.047747 2.66 0.009 0.032141 0.2218003 DiD | -0.0042153 0.065331 -0.06 0.949 -0.1339683 0.1255377 _cons | 14.70153 0.0369069 398.34 0.000 14.62823 14.77483 Table 3. Placebo regression on observations in the pre-treatment period

(30)

Table 5 shows the results from running the second specification. This specification includes several control variables and should therefore be able to explain the variance of passenger numbers to a much greater extent, given that the control variables are the correct ones determining passenger numbers. To see if this is the case one can start by looking at the R² value and see how much of the variance of passenger numbers that are explained by this specification. This value of 0,9738 is very high, indicating that this specification explains about 97 percent of the total variance. When looking at the DiD variable this time it can be seen that the coefficient belonging to it still has a negative value and this value is now bigger in absolute terms. This estimate indicates that there has been about a ten percent decrease in passenger numbers. This time the coefficient is statistically significant on the 1 percent level, so according to this specification the DiD variable and therefore the tax has had a negative effect on passenger numbers. Looking at the control variables and their coefficients it can be seen that the coefficient of Ln(PassengerLagged) is positive and statistically significant on the 1 percent level. It is estimated that a one percent increase in this lagged variable of passenger numbers one year earlier increases passenger numbers by almost exactly one percent. This coefficient is also the one with the highest t-value, without any other t-value coming anywhere close. The coefficient of Ln(Population) is also statistically significant on the 1 percent level and it has a positive value. It is estimated that a one percent increase in the

population size increases passenger numbers by about 1,4 percent. The coefficient of Ln(GDP) is negative and statistically significant on the 1 percent level. It is estimated that passenger

Table 4 - Specification without control variables

F(3, 140) = 12.41

Prob > F = 0.0000

R-squared = 0.2046

Root MSE = 0.16164

Robust

lnPassenger | Coef. Std. Err. t P>|t| [95% Conf. Interval]

Time | 0.1385714 0.0437304 3.17 0.002 0.052114 0.2250289 Treated | 0.1274471 0.0328379 3.88 0.000 0.0625247 0.1923694 DiD | -0.0348141 0.0570089 -0.61 0.542 -0.1475237 0.0778955 _cons | 14.75297 0.0249841 590.50 0.000 14.70358 14.80237 Table 4. Regression without the use of control variables

(31)

numbers decrease by 0,5 percent as GDP increases by one percent. The coefficient of Ln(Airfare) is positive, but not really distinguishable from zero and it is not at all near being statistically significant. The coefficient of Ln(Accommodation) is positive and statistically significant on the 10 percent level, but it is very close to zero.

Table 5 - Specification with control variables

F(8, 135) = 535.80 Prob > F = 0.0000 R-squared = 0.9738

Root MSE = 0.02985

Robust

lnPassenger Coef. Std. Err. t P>|t| [95% Conf. Interval]

lnPassengerLagged 0.9885572 0.020156 49.05 0.000 0.948695 1.02842 lnPopulation 1.350833 0.4389969 3.08 0.003 0.4826323 2.219034 lnGDP -0.5018702 0.1464998 -3.43 0.001 -0.7916019 -0.2121386 lnAirfare 0.0043834 0.0245027 0.18 0.858 -0.0440755 0.0528422 lnAccommodation 0.0523624 0.028995 1.81 0.073 -0.0049807 0.1097055

Time 0.002704 0.0122386 0.22 0.825 -0.0215002 0.0269081

Treated -0.4834569 0.182781 -2.65 0.009 -0.8449415 -0.1219722 DiD -0.1059327 0.0212262 -4.99 0.000 -0.1479115 -0.0639539

_cons -15.22483 5.460251 -2.79 0.006 -26.02352 -4.426129

6.3 Discussion

The lagged variable is without doubt the most important one in this model when trying to explain air travel in a certain month. This is most likely because there are big seasonal variations during the whole year when it comes to air travel. People in general travel much more at certain periods and times each year and because people usually plan their trips a long time in advance, aspects such as GDP growth at the time or changing prices do not play a big role in the decision on travelling or not. As was seen above, growth in GDP is estimated to have a negative impact on passenger numbers. This is not what you would expect because previous research has shown that GDP is one of the driving factors behind demand for air

Table 5. Regression with the use of control variables

(32)

travel. Why GDP would have a negative impact on passenger numbers is hard to tell but as stated above, trips are often planned in advance because people usually like to travel at certain seasons. Summer is the most preferred season. Maybe this negative coefficient only means that when many people go on holiday at the same time during the summer, they make their trips at the same time as the economy slows down. Therefore, it is not the economy at the time of the holiday that is important for the travel decision but the economy some time before the trip is made. Another explanation to why the coefficient of the GDP variable turned out to be negative in this investigation could be that if GDP has a steady growth during the year then passenger numbers don´t follow this same path. The path of passenger numbers was shown earlier where an increase during the summer months was always followed by a decrease during the winter months. Thus, if GDP has a constant growth across the year then the decrease in passenger numbers that happens at the end of each year might make it seem like the GDP growth has a negative impact on passenger numbers. That the accommodation variable has a positive effect is also the opposite of what you would expect. If prices go up for hotels and such this should make people less willing to travel. Maybe this also has to do with the fact that people want to travel at certain times and hotel managers of course know this.

Therefore, they can charge higher prices at these times and people accept these higher prices instead of travelling at some other time. If accommodation prices are affected in this way, then the accommodation variable is not exogenous in the model. This is then a violation of one of the assumptions of the difference-in-difference model and the difference-in-difference estimate could then be biased. It should be noted however that the estimated effect of the accommodation variable on passenger numbers is very small. As stated above the DiD variable and thus the Swedish air passenger tax has had a decreasing effect on passenger numbers according to this specification. It is also clear that other variables are very important to affect travel decisions. If the government wants the air traffic to decrease further, they must probably increase the tax much more for it to have an even greater effect. Otherwise, they need to impose other policies to reduce air travel and the emissions coming from it if the climate goals are to be reached.

The Swedish Air Passenger Tax

Bachelor Essay