Carbon Taxes and CO₂ Emissions: A Replication of Andersson (American Economic Journal: Economic Policy, 2019)

3.1 Data

The data and code used by Andersson (2019) are posted at: https://www.aeaweb.org/articles?id=10.1257/pol.20170144. The materials were well documented and clearly explained. Data were provided for 15 OECD countries (excluding Norway), with Sweden being the treated unit and the other 14 carefully chosen OECD countries serving as controls. Andersson set the time of treatment as 1990, the year when the VAT was introduced.

As part of my verification analysis, I attempted to retrieve all data from the original data sources in Andersson (2019). My goal was to build up a dataset for 16 OECD countries (including Norway) for the period 1960–2005. This would enable me to confirm that the data used by Andersson could be recreated from the original sources, so that a researcher working from the same data sources would get the same results. I also needed to be certain that the same variables were available for Norway as part of my extension analysis. In doing so, I discovered that data on the level of CO₂ emissions from transport are no longer available in the World Bank WDI database (2020). As a result, I sought alternative data sources.

Andersson subsequently made available to me the original data he downloaded from the World Bank WDI Database (2015). These were virtually identical to the data publicly packaged with Andersson (2019) (Table 1), which also includes CO₂ emissions for Norway’s transport sector. My extension analysis relies on these latter data.

In order to base my analysis as much as possible on publicly available data, I also indirectly calculated CO₂ emissions from transport (million metric tons) as the product of “CO₂ emissions from fuel combustion” (from IEA) and “CO₂ emissions from transport (% of total fuel combustion)” (from the 2020 World Bank WDI Database). The compiled CO₂ emissions data from transport starts from 1971, while the WDI data from Andersson dates to 1960. The values of the constructed CO₂ emissions variable are very similar, with a difference of only 0.6% on average over the period 1971–2005. My initial analysis uses these publicly sourced, CO₂ emissions data.

Later, I use Andersson’s emissions data from 1960 to 2005. The longer pre-treatment period enables me to better assess whether emissions from the synthetic control follow that of the treated unit. It also allows me to be more consistent with Andersson (2019) in my extension to Norway. For all the other variables, I use the publicly available data that I directly retrieved from the original sources.

Aside from past emissions data, Andersson’s SCM analysis relies on the following set of predictor variables: real GDP per capita, number of motor vehicles, gasoline consumption, and urban population. The data for the predictors cover the years 1980–1989. Table 1 provides summary statistics of Andersson’s dataset and my compiled dataset for the 15 OECD countries (excluding Norway). As shown in the table, my compiled data and Andersson’s data are identical for the first three predictors, and only slightly different for Urban Population.

3.2 Synthetic Control Method

The SCM was originally proposed in Abadie and Gardeazabal (2003) and Abadie et al. (2010) to estimate the effect of large-scale or aggregate interventions. It uses a data-driven approach to construct a control unit (synthetic control), which is a weighted average of the untreated units in the donor pool. The intuition is that the combination of unaffected units provides a more appropriate comparison than any single unaffected unit alone. The following is the basic framework of SCM.

Suppose there are J + 1 units, j = 1 , 2 , … , J + 1 , with the first unit ( j = 1 ) being the treated unit and the others untreated units consisting of the “donor pool”. The time periods are t = 1 , 2 , . . . , T , and the first T 0 periods are the pre-intervention period. For each unit j and time period t , we define Y jt I as the potential outcome when the unit is under intervention (“I”), and Y jt N as the potential outcome without intervention (“N”). Then the treatment effect for the treated unit in period t ( t > T 0 ) is given by the difference between its observed outcome under intervention and its unobserved potential outcome without intervention. Then our goal is to estimate Y 1 t N for t > T 0 . For this purpose, SCM constructs a synthetic control group as a weighted average of the untreated units in the donor pool, as shown in the following equation:

A set of weights W = ( w 2 , … , w J + 1 ) ′ are assigned to the control units with 0 ≤ w j ≤ 1 and w 2 + … + w J + 1 = 1 . Abadie and Gardeazabal (2003) and Abadie et al. (2010) propose to choose the vector W * to minimize the following distance:

where X 1 is a k × 1 vector of predictors of the outcome variable for the treated unit, including the set of covariates Z 1 and the pre-intervention values of the outcome variable for the treated units. In our case, the predictors consist of some key characteristics affecting the level of posttreatment CO₂ emissions. By minimizing the objective function defined in equation (2), we aim to construct a synthetic control that most resembles the treated unit in those characteristics that help predict the outcome variable prior to the treatment.

Notice that equation (2) includes a vector of coefficients V = ( v 1 , … , v k ) . This is the predictors’ weights representing their relative importance in predicting the values of the outcome variable. V is chosen by minimizing the distance between the observed outcome for the treated group Y 1 t and the counterfactual outcome from the synthetic control group in the pre-treatment period.

When this distance, namely the mean squared prediction error (MSPE) is small, the outcomes for the treated and synthetic control will follow a similar trend. Then, the synthetic control estimate of the treatment effect τ ˆ 1 t is as follows:

The validity of the SCM relies on the “convex hull assumption,” which states that the values of predictors for the treated unit X 1 should not be far away from the control units (Abadie et al., 2010; Bonander et al., 2021). If that is not the case, the fit may be poor and the use of SCM is not recommended. Moreover, bias can arise if the donor pools consist of units with very different characteristics or the relationship between the potential outcome and Z 1 is highly nonlinear (Abadie et al., 2010). Lastly, the risk of over-fitting may also increase with the size of the donor pool, especially when T 0 is small (Abadie, 2021). While there is no generally accepted guidelines on how many pre-treatment periods are sufficient ( T 0 ), the examples used by Abadie and colleagues have T 0 equal to 15 (Abadie et al., 2011), 19 (Abadie et al., 2010), and 31 (Abadie, 2021). Andersson (2019), and this replication, have T 0 = 30.

When the above assumptions are fulfilled, SCM provides an alternative estimation method to DID when the assumption of PT for treatment and control units is suspect. As Andersson (2019) notes,

“The parallel trends assumption is difficult to verify, which is a drawback for the DiD method. It is sometimes possible pretreatment by analyzing the trends of the outcome variable, but obviously impossible after treatment. When the treated unit and the control group do not follow a common trend, the DiD estimator will be biased. Therefore, finding a method that relaxes the parallel trends assumption is preferable for comparative case studies” (page 10).

By reweighting the controls to match the pre-treatment trends in the treated unit, SCM increases the likelihood that the parallel trends assumption holds (Bonander et al., 2021). The other merits of SCM include providing data-driven and formal criteria for selecting control units, showing us the contribution of each control unit in constructing the synthetic control, and the relative importance of the predictors in predicting the pre-treatment outcomes. Although classic SCM does not produce standard errors and confidence intervals, it is common for researchers to conduct in-time, in-space, and leave-one-out placebo tests (Abadie et al., 2010, 2011).

3.3 Regression Analysis

While SCM constitutes the main focus of Andersson’s analysis, he also employs a regression analysis of gasoline consumption. The regression analysis has two uses. First, it provides a robustness test of the SCM analysis. In the absence of endogeneity, the two methods should produce similar estimated effects. The regression analysis also allows one to isolate the separate effects of the VAT and the carbon tax. This is important given that it is the latter which is of primary interest.

Previous research recognizes the possibility that tax-induced price changes may generate distinct demand responses compared with equivalent, market-determined price movements (Rivers & Schaufele, 2015). This phenomenon is called “tax salience.” In the Swedish case, gasoline consumption could respond differently to a rise in gasoline prices induced by the introduction of VAT and a rise due to the carbon tax. Andersson’s analysis allowed for this possibility.

Specifically, he estimated a log-linear gasoline demand model. The retail price of gasoline was decomposed into the carbon tax-exclusive price p t v = ( p t + τ t , energy ) VAT t and the carbon tax τ t , CO 2 v = ( τ t , CO 2 ) VAT t , where τ t , energy and τ t , CO 2 are the values of energy tax and carbon taxes, respectively, and VAT is a multiplier and is added to each price component. The log-linear model is specified as follows:

where x t is the gasoline consumption per capita; D t , CO 2 is a treatment dummy that takes the value of 1 for years from 1991 and onward and zero otherwise; X t includes a vector of control variables (real GDP per capita, urban population (%), and the unemployment rate) and a time trend; and ϵ t is the error term.

To address the possible endogeneity of gasoline prices, Andersson used crude oil prices and the energy tax rate as instruments for the carbon tax-exclusive gasoline price and performed two-stage least squares estimation. Based on the regression results, Andersson conducted a simulation, in which he approximated the amount of (counterfactual) CO₂ emissions in three cases: Sweden without carbon taxes and VAT, Sweden with VAT but no carbon taxes, and Sweden with carbon taxes and VAT. In the simulation, I multiply gasoline consumptions (kg) by the emission factor of gasoline to get simulated CO₂ emissions.^[2] The gasoline consumptions under Case 1 and Case 2 are predicted from the gasoline consumption-price regression. Then, the difference in the simulated emissions between Case 2 and Case 3 identifies the effect of carbon taxes on CO₂ emissions.

4.1 Verification with Andersson’s Data and Code

As noted above, Andersson (2019) provided his data and code as supplementary materials to his published journal article. This allowed me to confirm that his results were “push-button” replicable. Because I obtained identical results to him, I do not report a side-by-side comparison of Andersson’s original results and my replication.

4.2 SCM with Alternative Synthetic Swedens

Figure 1a is a replication of Figure 4 in Andersson (2019). It sets 1990, the year when the VAT was extended to gasoline and diesel, as the treatment year. Figure 1a demonstrates that synthetic Sweden successfully reproduced the trajectory of CO₂ emissions for Sweden before the treatment. As shown in the graph, per-capita CO₂ emissions from transport in Sweden experienced an immediate drop after 1990 compared to what would have been expected without a carbon tax. The emission gap between Sweden and synthetic Sweden increased gradually in the early 1990s and then remained constant. The resulting average reduction in CO₂ emissions from transport in Sweden is −0.29 metric tons per capita, which accounts for 10.9% of the emissions in the absence of VAT and carbon taxes on average.

Figure 1

SCM analysis (Swedish Transport). (a) Replication of Andersson’s Figure 4. (b) Three Synthetic Swedens.

My replication of Figure 1a adds two additional synthetic control groups. In addition to Andersson’s control group, I created a second control group that drops Denmark from the donor pool. Denmark introduced a carbon tax on energy products in 1992, though it exempted the transport industry. Andersson, noting that the carbon tax rate was relatively low, decided to keep it in the donor pool. However, Denmark has the largest weight in the construction of its synthetic Sweden. Therefore, I check how conclusions may change if Denmark is excluded from the control group. My third synthetic Sweden uses the data I compiled and reported in Table 1.

Figure 1b recasts Figure 1a in terms of differences in CO₂ emissions from transport between actual Sweden and the three synthetic counterparts. As shown in the figure, when Denmark is excluded from the donor pool, the gap becomes wider. This is consistent with the bias one would expect from including a carbon tax adopter in the controls, which would mute the estimated effect of the carbon tax.

Table 2 reports the values for the predictors corresponding to the three synthetic control analyses above.^[3] The first four columns reproduce the results in Table 1 of Andersson (2019). They show that compared to the population-weighted average of the 14 OECD countries, Andersson’s synthetic Sweden more closely resembles Sweden with respect to the means of the predictors and per-capita CO₂ emissions during the pre-treatment years.

Columns (5)–(7) of Table 2 show that when we exclude Denmark from the donor pool, the predictor values of the synthetic counterpart are still close to that of Sweden, while the MSPE increases slightly from 0.0012 to 0.0026. It indicates that we are able to reduce potential bias at a small cost. For this reason, going forward, my analysis excludes Denmark from the donor pool.

Returning to Figure 1b, we see that when I use the data that I compiled rather than Andersson’s data, the emissions gap is larger in the early 1990s but decreases in the late 1990s. The corresponding average treatment effect is −0.39 metric tons, larger than what Andersson reports (−0.29 metric tons).

Columns (8)–(10) of Table 2 present data on the predictors using my compiled data. While the overall values are similar to Andersson (2019), the corresponding synthetic Sweden has a poorer fit during the pre-treatment years (MSPE of 0.0037 versus 0.0012 and 0.0026). Given that the goodness of fit is worse for the compiled data, the subsequent analysis will use Andersson’s data, excluding Denmark.

4.3 Regression Analysis of Gasoline Consumption Using an Alternative Estimator

When SCM is used in Andersson (2019), by its nature, two effects could not be distinguished: the introduction of a carbon tax in Sweden and the adoption of the VAT in gasoline and diesel, both of which coincided. What further complicates this case is other tax changes that affected gasoline during the post-treatment period. As the tax rate of the carbon tax rose in 2000, it was accompanied by a simultaneous decrease in the tax rate of the energy tax.

In order to separate out the price effects of the carbon tax from the VAT and general price increases, Andersson estimates a demand equation for gasoline consumption that includes both components. To estimate this equation, he uses time series data on Brent Crude oil prices and gasoline consumption from 1970 to 2011. To address endogeneity, he uses two instruments: the energy tax rate and the price of crude oil. However, the IV estimates are similar to the OLS estimates and a Hausman test finds no statistical difference, so he concludes that endogeneity is not a problem.

To account for autocorrelation, Andersson (2019) used the Newey-West procedure to obtain “serial correlation-robust” standard errors for the OLS estimates. Alternatively, I adopt the Prais-Winsten estimation procedure, which is a form of generalized least squares (GLS). Its key assumption is that the errors follow a first-order autoregressive process. In practice, the Prais-Winsten method first estimates the correlation between the errors at t and t − 1 and then applies a linear transformation to decorrelate the error term (Bottomley et al., 2023).

Table 3 presents my verification of the regression analysis in Andersson’s article. The results in columns (1), (2), and (3) are exactly the same as in the study by Andersson (2019). Additionally, Column (4) shows the Prais-Winsten regression results. Across Table 3, the two price components (“Carbon tax-exclusive gasoline price” and “Carbon tax”) have negative and statistically significant coefficients.

According to the Prais-Winsten estimates in Column (4), if the carbon tax and the non-tax price components each increase by one Swedish Krona, gasoline consumption would decrease by 10.7 and 4.6%, respectively. These effects are smaller in size compared to Andersson’s estimates, which are 18.6 and 6.0%, respectively. I note that Prais-Winsten assumes that the error terms are AR(1). However, there is mixed evidence that the error terms follow a higher order autoregressive process. In that case, Prais-Winsten should still improve efficiency over OLS/Newey-West, which makes no adjustment for serial correlation in its coefficient estimates.

Based on the estimates in Column (1), Table 3, Andersson (2019) approximated CO₂ emissions from three scenarios.

As shown in Figure 2a, my calculation produces results quite similar to Andersson (2019). It indicates that the effect of the VAT remained relatively constant after its extension to gasoline in 1990. On the other hand, the effect of the carbon tax experienced a significant increase after 2000 as the real tax rate increased from 0.92 SEK/liter in 2000 to 2.11 SEK/liter in 2005. As Andersson (2019) noted, the sharp increase reflects the separate effect of carbon tax while keeping the (real) rate of energy tax constant. Figure 2b replicates Figure 14 in the study by Andersson (2019) by comparing the effects of carbon tax and VAT from the SCM and the simulation based on OLS regression. Note that the estimated VAT + Carbon tax emission effects from the SCM and regression analyses are quite similar, especially in the first half of the post-treatment period when the real tax rate of the energy tax remained constant.

Figure 2

Disentangling the effects of VAT and Carbon Tax (Swedish Transport). (a) Regression Analysis. (b) SCM and Regression Analysis.

Figure 2b also adds my estimates of emission reduction based on the Prais-Winsten price estimates. The associated reduction is only half the size of the estimated emission reduction from the SCM and the OLS estimates over most of the treatment period. This is a direct consequence of the smaller estimated price and tax effects in Column (4) of Table 3. However, it also shows a sharp increase in impact after 2,000 due to increasing carbon taxes.

Andersson (2019) used the price estimates to assign the proportions of total emission reduction estimated from the SCM (−0.29 metric tons) to VAT and to carbon taxes. He concluded that carbon taxes alone reduced per-capita CO₂ emissions from transport in Sweden by 6.3% (0.17 metric tons) on average in the post-treatment period. I repeat this calculation based on my synthetic control analysis (without Denmark) and the Prais-Winsten estimates, and conclude that the Swedish carbon taxes alone reduced CO₂ emissions from transport by 7.7% (0.21 metric tons).

Why is a larger estimated emission effect of the carbon tax obtained, compared to Andersson (2019), despite estimating a smaller coefficient of the carbon tax from using the Prais-Winsten method (cf. Table 3)? The reason is that according to the simulated counterfactual emissions, the Swedish carbon taxes account for 59% of the mixed emission reduction from VAT + carbon tax on average during 1991–2005. This share is roughly the same as that from Andersson’s approach (60% on average). As my SCM estimate (without Denmark) is larger, the resulting carbon tax-induced emission reduction is larger. Following Andersson, I apply this larger share to the SCM estimate of the total effect, and that produces a larger carbon tax estimate. While Andersson (2019) fortuitously found that his SCM and regression estimates were similar, he placed greater confidence in the SCM estimates and used the results for the regression analysis solely to determine the split between VAT and carbon taxes. This issue will reappear later when I extend Andersson’s analysis to Norway’s transport sector.

In conclusion, while I modify Andersson’s analysis in several substantive ways, my final estimate of the effect of carbon taxes on CO₂ emissions ends up being very close to the 6.3% reduction Andersson estimated.

5.1 Norway’s Carbon Tax

Norway levied CO₂ taxes on petroleum, mineral fuel, and natural gas in 1991. In 1991, the tax rate was 39.6 USD per ton of CO₂ for natural gas offshore on the continental shelf, 35 USD for oil offshore on the continental shelf and 15–17 USD for heating oil. Petrol was also subjected to a heavy tax of 259 NOK per ton of CO₂ (namely, 40 USD/ton) (Andersen et al., 2001). After the implementation, CO₂ tax on petrol had increased steadily to a rate of 405 NOK per ton of CO₂ (46 USD/ton) in 2,000 and 336 NOK per ton of CO₂ (52 USD/ton) in 2005. It is worth noting that the initial rate of Norwegian CO₂ tax on petrol was even higher than that of Sweden’s, while it grew more slowly than Sweden’s. This leads to an interesting comparison to Sweden’s carbon tax.

Similarly to Sweden, there are extensive exemptions and differentiation of carbon tax rates in Norway (Bruvoll & Larsen, 2004; Lin & Li, 2011). Since Norway is one of the world’s major oil and natural gas producers and exporters, 29% of total CO₂ emissions from oil and gas extraction in 2001 (Statens, 2003). Carbon taxes on oil and gas extraction are set at a comparatively high level, 49 USD for natural gas and 43 USD for oil in 1999 (Bruvoll & Larsen, 2004). Other high-polluting industries, such as the metal-producing industry, are partly or totally exempted for fear of losing competitiveness. There are also exemptions for fishing, air, and ocean transport. As a result, only 60% of the total CO₂ emissions in Norway are subjected to CO₂ tax. On average, tax revenue from CO₂ emissions accounts for 16.9% of total environmental tax revenues.

5.2 SCM Analysis

The predictor variables. The first step in the SCM analysis is choosing a set of predicting variables to construct a synthetic Norway. I apply the same predictors used for constructing synthetic Sweden: real GDP per capita, Motor vehicles (per 1,000 people), Gasoline consumption per capita, Urban population (%), per-capita CO₂ emissions from transport in 1970, 1980, and 1989. Table 4 shows the values of the key predictors in the pre-treatment period. The majority of the weights (75%) are given to the past values of CO₂ emissions.

In contrast, the number of motor vehicles and urban population receive weights of 0.9 and 0%, respectively. It seems as if these two predictors do not play much of a role in predicting per-capita CO₂ emissions from transport in Norway. Nevertheless, I decide to keep them on the predictor list to maintain comparability with Andersson (2019).

Overall, Table 4 suggests that synthetic Norway is a better comparison group than the simple average of the untreated, OECD countries. According to the weights given to the 13 OECD countries, synthetic Norway is best reproduced by a combination of Belgium (0.771), Switzerland (0.113), and the United States (0.116).

Checking the parallel trends assumption in the pre-treatment period. While we are unable to test the assumption of PT in the post-treatment period, it is possible to check for it in the pre-treatment period. Figure 3a compares per-capita CO₂ emissions from Norway’s transport sector with the average per-capita CO₂ emissions from transport for the 13 OECD countries. It is clear from the figure that the PT assumption does not hold during the pre-treatment period because Norway’s per-capita CO₂ emissions clearly grew faster than the OECD average. This provides support for using SCM over DID or a two-way fixed effects model.

Figure 3

SCM analysis (Norwegian Transport). (a) Norway and the OECD average. (b) Norway and Synthetic Norway.

Results from the SCM analysis. Figure 3b shows the results from SCM applied to Norway’s transport sector. CO₂ emissions in Norway and synthetic Norway follow a common trend prior to 1990 and diverge after 1990. In 1990, 1 year before the implementation of Norwegian CO₂ tax, per-capita CO₂ emissions from transport in Norway dropped below that of synthetic Norway. This is consistent with there being an anticipation effect prior to the actual implementation of the carbon tax.

Contrary to Sweden, the SCM analysis finds that emissions in the Norwegian transport sector exceeded that of the synthetic control counterfactual during the years 1996–1999. This constitutes evidence against the effectiveness of carbon taxes to reduce CO₂ emissions. Nevertheless, I still estimate an accumulated effect of 0.96 metric tons of emission reduction per capita over the full, 1991–2005, post-treatment period. The corresponding annual per-capita emission reduction is −0.064 metric tons and is on average 2.4% lower than the scenario without the carbon tax. My conclusion is that the Norwegian carbon tax had an overall small, negative effect on per-capita CO₂ emissions from transport in Norway.

Placebo robustness tests. To check the credibility of the SCM results for Norway, I conduct placebo tests. For the in-time placebo tests, the treatment is assigned to the years 1970 and 1980. Figure 4 shows that although the fits are not perfect, possibly due to the large variations in the outcome, there is no sign of placebo effects after 1970 and 1980.

Figure 4

In-time placebo tests (Norwegian Transport). (a) Year of Intervention is 1970. (b) Year of Intervention is 1980.

The in-space placebo test iteratively assigns the treatment to untreated countries in the donor pool and compares the sizes of the resulting “treatment” effects. Figure 5a shows that when we focus on seven placebo units (including Norway) with MSPEs less than 0.01, the effect in Norway (−0.064 metric tons per capita) is the second largest, smaller than Switzerland with a placebo effect of −0.24 metric tons per capita. The “treatment” effects for the other placebo units are either close to zero or positive.

Figure 5

In-space and leave-one-out placebo tests (Norwegian Transport). (a) In-Space Placebo Test. (b) Leave-One-Out Placebo Test.

Figure 5b shows the results of the leave-one-out placebo test, in which the untreated units with positive weights are dropped from the donor pool one at a time to reconstruct the synthetic controls. This practice gives us a range of the estimated effects of carbon tax on per-capita CO₂ emissions from transport in Norway, which is −0.093 to −0.037 metric tons per capita. These placebo tests prove the robustness of my baseline results.

Possible confounders. Andersson noted that economic growth could be a confounder that affected his estimates of the impact of the Swedish carbon tax. While SCM is not well-suited to handle confounders, I nevertheless follow Andersson’s approach to investigate whether Norwegian economic performance might bias the carbon tax effect estimated by SCM analysis.

For constructing a synthetic Norway’s real GDP per capita, the same weights used for the synthetic Norway’s CO₂ emissions per capita are applied. As shown in Figure 6, the relative trends in per-capita CO₂ emissions from transport and per-capita real GDP in Norway were closely co-moving from the late-1980s to the early-1990s (grey-shaded area in the figure). This suggests that the observed downturn in CO₂ emissions in 1990 might not have been an “anticipatory effect” associated with the introduction of the carbon tax, but rather was a reflection of a relative downturn in Norwegian economic activity. If so, by spuriously attributing this to the carbon tax, SCM overstates its impact.

Figure 6

Co-movements between GDP and CO₂ emissions (Norwegian Transport).

On the other hand, starting in the mid-1990s and continuing through to the end of the post-treatment period, Norway’s actual GDP per capita did substantially better than synthetic Norway. To the extent that the increased, relative economic activity contributed to greater CO₂ emissions, SCM analysis understates the impact of the carbon tax. While it is possible that the economy is a confounder, the mixed effects described above do not provide clear evidence that SCM either under- or over-estimates the emission effects of the carbon tax for the Norwegian transport sector.

In conclusion, while I find a small, overall negative effect on emissions due to the carbon tax in the Norwegian transport industry, the higher emissions of the Norwegian transport sector compared to its synthetic control counterfactual from 1996 to 1999 is noteworthy.

5.3 Regression Analysis

As noted above, Andersson (2019) used regression analysis to disentangle the effect of the carbon tax from other taxes. Fortuitously, the overall, estimated effect of VAT + carbon tax from his regression analysis was approximately equal to that of the SCM analysis.

There is no need to do the same for Norway because the introduction of carbon taxes in 1991 was not accompanied by major changes in other taxes that might affect the SCM analysis. Nevertheless, in order to maintain comparability with Andersson (2019), I use regression analysis to calculate an alternative estimate of the emissions effect of Norway’s carbon tax.

Andersson (2019) used annual time series data for Sweden from 1970 to 2011. In contrast, the data available to me only span the period from January 1995 to December 2017, though most of the variables are available monthly. Statistics Norway provides monthly data on gasoline prices, motor gasoline consumption, and quarterly data, which are used for the control variables in this regression analysis, resulting in a dataset of 276 observations. The data on the carbon tax rate are from the IEA (2009).

Table 5 shows the estimation results. The OLS estimates with Newey-West standard errors are shown in Column (1), indicating that a one Norwegian Krone (NOK) increase in the real carbon tax-exclusive gasoline price and carbon tax would result in decreases in gasoline consumption of 2.3 and 27.3%, respectively. Column (2) shows the Prais-Winsten estimates. The coefficients for the real carbon tax-exclusive price and carbon tax are −0.019 and −0.289, both statistically significant at the 1% level. This means that one NOK increase in the carbon tax is estimated to reduce gasoline consumption by 28.9%. In contrast, if the gasoline price rises by one NOK yet this change is not caused by the carbon tax, gasoline consumption would only decrease by 1.9%.

Column (3) presents results from instrumental variable (IV) estimation, using crude oil prices as an instrument for the carbon tax-exclusive price. As with Sweden, the results are quite similar to the OLS estimates in Column (1). A test for weak instruments shows that the crude oil price is not a weak instrument.

Assuming a roughly one-to-one exchange rate between the Norwegian Krone and the Swedish Krona, these estimated price effects suggest that gasoline consumption is more responsive to carbon taxes in Norway. Also, the persistent finding that carbon taxes have a larger impact than other price components of gasoline is consistent with the estimates from Sweden and “tax salience” theory.

Using the Prais-Winsten estimates in Column (2), I simulated CO₂ emissions from Norway’s transport sector in two scenarios: with carbon tax and without carbon tax. The effect of a Norwegian carbon tax in reducing per-capita CO₂ emissions is obtained by taking the difference between these two scenarios, which is −0.55 metric tons per capita on average during 1991–2005. This effect size accounts for 27% of per capita CO₂ emissions from transport that would occur without a carbon tax. In light of the SCM estimates this seems implausibly large (Figure 7 for a comparison of the two methods).

Figure 7

Comparing SCM and regression (Norwegian Transport).

As I obtained two contrasting estimates from two different methods, a natural question is: To which should one attach greater weight? There are reasons to prefer the SCM estimate. First, despite efforts to address endogeneity bias through IV estimation, there remain concerns about misspecification and endogeneity with regression estimates of a gasoline demand equation. This was why Andersson (2019) preferred his SCM estimates to the regression estimates. But there is a second reason to prefer SCM.

The regression analysis is based solely on gasoline consumption. It does not include diesel consumption. CO₂ emissions from road transportation in Norway increased 19% from 1990 to 2001. This rise is primarily attributed to an increase of 73% in emissions from diesel, whereas emissions from gasoline vehicles decreased by 7% during this period (Statens, 2003).^[4] This substitution from gasoline to diesel is not captured in the regression analysis, which focuses solely on the carbon tax effect on gasoline consumption. In contrast, SCM has the advantage of capturing the impact on emissions for the whole transport sector, not just the portion due to gasoline consumption and its attendant CO₂ emissions.

In conclusion, I estimate that the carbon tax had different effects on the transport sector in Sweden and Norway: a relatively large effect for Sweden, and a relatively small effect for Norway (as shown in Table 7). The reasons for this difference are not clear. The goal of this analysis was to extend Andersson’s methods to see if I could obtain similar results for Norway.