Impact of Amsterdam’s Low Emission Zone on NOx

(1)

Bachelor Thesis University of Amsterdam Economics and Business Economics

Impact of Amsterdam’s Low Emission Zone on NO

x

concentrations

June 30, 2022 M. J. Nauta

12188980

Supervisor: K.H.L. Sommer

(2)

2

Statement of originality

This document is written by Student Mark Nauta who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Mark Nauta, June 2022

(3)

3

Abstract

The following research examines whether the 2018 rule changes in Amsterdam’s low emission zone have led to a reduction in the concentration of nitrogen oxides in the regulated parts of the city. To answer the research question, a difference in difference model has been used. The results show that the 2018 rule changes have led to a 6.3% reduction in emissions of nitrogen oxides. This effect is statistically significant with robust standard errors and statistically insignificant with clustered standard errors. However, the effect of Amsterdam’s low emission is likely underestimated in this research, due to spillover effects. Therefore, it is expected that the actual effect is larger and statistically significant with clustered standard errors. This research recommends doing a similar analysis after controlling for the exact spillover effects.

In this way, unbiased results about the effectiveness of a low emission zone are obtained.

(4)

4

1. Introduction

In 2019, the Highest Administrative Court of the Netherlands ruled that the government’s current nitrogen rules were too lenient. The government was condemned to introduce stricter legislation on nitrogen emissions. Since then, the Netherlands has struggled with a nitrogen crisis (Reijneker, 2021). As a result of this crisis, many infrastructure projects and building projects have either been postponed or cancelled. The solution to this crisis has not yet been found, due to the many issues the government faces implementing the right restrictions to decrease nitrogen emissions (Boerep, 2019).

Several cities in the Netherlands have introduced a so-called ‘low emission zone’

(hereafter: LEZ). The most polluting vehicles are not allowed to enter this restricted area. LEZ’s aim to reduce air pollution; One of the key factors of air pollution is nitrogen oxide (hereafter:

NOx) (Luchtmeetnet, n.d.). Therefore, an effective LEZ would be a helpful measure for the Dutch government to reduce nitrogen emissions in the Netherlands. For this reason, it is relevant to study the effect of such a LEZ on NOx emission.

In 2008, a LEZ in Amsterdam was introduced to improve air quality. The first restrictions consisted of banning the most polluting trucks in the centre of Amsterdam.

However, the reductions in NOx and particulate matter were not significant. Therefore, Amsterdam made the existing rules stricter in 2018 in order to achieve its goal of being emission-free in 2025. The new rules consist of banning the most polluting diesel cabs, diesel trucks and diesel delivery vans (Gemeente Amsterdam, 2018). The open-source data regarding NOx emissions in Amsterdam is available from 2014 onwards. Therefore, the research only focuses on the 2018 rule changes in Amsterdam’s LEZ. This research tries to answer the following research question: Did the 2018 rule changes in Amsterdam’s low emission zone lead to a reduction in nitrogen oxides in the regulated parts of the city?

This question is important for two main reasons. First of all, a reduction in NOx

concentrations as a result of a LEZ would imply that it’s a helpful tool for (local) governments to improve air quality and reduce NOx emissions. This is especially relevant as the Netherlands faces problems with the concentration of NOx emissions. Secondly, research has shown that NOx is harmful to nature and human beings. Exposure to high NOx concentrations has short- and long-term effects, such as headaches and breathing problems (Staff, 2015). Waersted et al.

(2022) found that diesel vehicles produce the highest NOx emissions in Oslo. Moreover, Dallmann et al. (2019) also noted that diesel vehicles contribute to the highest NOx emissions in their study in Paris. However, they recommend that further research is needed to obtain a

(6)

6 clearer picture. The findings of Wearsted et al (2022), and Dallman et al (2019) look promising for this research, the LEZ in Amsterdam bans diesel vehicles and these vehicles have the highest emissions. It’s therefore expected that the LEZ leads to a significant drop in NOx emissions.

To answer the research question, a difference in difference model (hereafter: DiD model) is used. The DiD model compares measurement stations within the LEZ (treatment group) to stations outside (control group) over time. This research uses national institute for public health and the environment of the Netherlands (RIVM) data on NOx emissions in Amsterdam from 2015 until 2019. Dutch national weather service (KNMI) weather data is used to control for changes in temperature, wind speed and humidity during the years of interest. The data of both RIVM and KNMI has been merged and is used to run the DiD model. This regression model indicates the effect of the 2018 rule changes in Amsterdam’s LEZ on NOx

emissions.

Results have shown that the 2018 rule changes in Amsterdam’s LEZ decreased NOx

emissions on average with 6.3% in the regulated parts of the city compared to the non-regulated parts. This decrease in NOx emissions is statistically significant with robust standard errors and statistically insignificant when using clustered standard errors. However, it is expected that the actual decrease in NOx emission due to the 2018 rule changes will be larger as spillover effects cause for an underestimation of the effectiveness.

All relevant theories, models and scientific researches are described in section 2. This section also contains a policy segment in which the LEZ of Amsterdam is explained in depth.

The data and variables which have been used in this research are described in section 3. The validations of the model assumptions are also described in section 3. The regressions, output and results of the model are described in section 4. Finally, the discussion and conclusion of the conducted study are presented in section 5 and 6, respectively.

(7)

7

2. Theoretical Framework and Literature

2.1 Low Emission Zone Research

To improve air quality, many European cities have implemented a LEZ. In such LEZ’s, the most polluting vehicles are not allowed to enter the restricted area (RIVM, 2021).

The effectiveness of LEZ’s has already been studied multiple times in scientific research.

Griffiths et al. (2016) investigated the impact of London’s LEZ on air quality and children’s respiratory health. They carried out a sequential yearly cross-sectional study with around 2000 children. Although Griffiths et al. (2016) found that London’s LEZ led to higher air quality, they found no evidence of an increase in children’s respiratory health. Santos et al. (2019) used a DiD model to research the effect of Lisbon’s LEZ on air quality. The air quality increased significantly in Lisbon due to the LEZ. Researchers found significant reductions in annual average concentrations of PM10 (particulate matter 10 micro-meters or less) and NO2. However, insignificant reductions were observed for NOx and PM2.5 (particulate matter 2.5 micro-meters or less). Morfeld et al. (2014) examined the effect of 17 German LEZ’s on NO, NO2 and NOx

emissions. They used multiple linear and log linear fixed effects models to compare pairs of emissions before the LEZ to pairs after the implementation of the LEZ. Each pair consisted of data within the LEZ and reference data outside the LEZ. Although Morfeld et al. (2014) concluded that the decrease in NO, NO2 and NOx emissions due to the LEZ’s in Germany was statistically significant, the decrease in emissions was small. Tarriño-Ortiz et al. (2022) took a different approach in their research and investigated the effect of Madrid’s LEZ by surveying to explore changes in modal shift. The findings of Tarriño-Ortiz et al. (2022) show that the LEZ in Madrid caused an increase in demand for environmentally friendly alternatives and a decrease in the demand for polluting vehicles. However, the increase in air quality was not significant in Madrid due to the LEZ.

To conclude, research has shown mixed evidence about the effectiveness of LEZ’s in several European cities. Both Griffiths et al. (2016) and Santos et al. (2019) recommend implementing stricter restriction standards, because the LEZ was not completely effective in their conducted research. Ezeah et al. (2015) addressed this mixed evidence of the effectiveness of the LEZ’s in their research. They looked at several LEZ’s in European cities and concluded that the observed differences in the effectiveness of LEZ’s are a result of different implementations. They argue that it is extremely hard to establish clear criteria of the factors which make a LEZ effective. Therefore, further scientific research is needed to obtain better criteria. However, the researchers mention one key factor which is a likely reason why some

(8)

8 LEZ’s are effective: regions with many LEZ’s close to each other are more effective, because firms cannot reassign their old polluting vehicles to non-LEZ areas (Ezeah et al., 2015).

This research focuses on the LEZ in Amsterdam and uses a DiD model. Therefore, the research of Santos et al. (2019) and Morfeld et al. (2014) are most similar to this research, because they use a DiD model and a modified DiD model, respectively. In addition to the published scientific research about LEZ’s, this research controls for three weather variables which influence NOx emissions (see section 2.3). This research distinguishes itself from other scientific papers for this reason.

2.2 Low Emission Zone Amsterdam

Amsterdam has the worst air quality of all large cities in the Netherlands according to a yearly study from the European Environment Agency (2021). Therefore, a LEZ in Amsterdam was introduced in 2008 to reduce air pollution and improve air quality. The LEZ first only covered the centre of Amsterdam. As of 2020, the LEZ in Amsterdam includes the complete area within the ring road A10, this is roughly 50 % of Amsterdam’s total land area (Gemeente Amsterdam, 2018). The first restrictions in 2008 consisted of banning the most polluting trucks in the centre of Amsterdam. However, the reduction in NOx concentrations was not significant, because the non-banned trucks were more polluting than expected (Goudappel Coffeng and Buck Consultants International, 2010). Therefore, Amsterdam introduced stricter rules in 2018 to achieve its goal of being emission-free in 2025. In addition, the new rules banned the most polluting diesel cabs and diesel delivery vans. Moreover, more kinds of trucks were banned in Amsterdam’s LEZ from 2018 onwards. (Gemeente Amsterdam, 2018).

This research focuses solely on the 2018 rule changes in Amsterdam’s LEZ, because the open-source RIVM data started in 2014. Moreover, the amount of measurement stations has increased over time (RIVM, 2021).

The RIVM (national institute for public health and the environment of the Netherlands) measures many components of air quality (including NOx) all over the Netherlands.

Luchtmeetnet is a website created by the RIVM to check the air quality of all measurement stations in the Netherlands at any given time. The region of Amsterdam has seventeen measurement stations, which is the most for any region in the Netherlands (Luchtmeetnet, n.d.).

To put this into perspective, the LEZ’s of Utrecht and Arnhem only have two and three measurement stations, respectively. Therefore, the region of Amsterdam is the most suitable to examine the effect of a LEZ in the Netherlands.

(9)

9 As mentioned above, the region of Amsterdam has seventeen measurement stations: six stations are located within the LEZ and eleven stations are located outside of it. Therefore, the six stations within the zone are the treatment group of this research and the other eleven stations are the control group. Appendix A1 contains a table with information about all the measurement stations used for this research. A map with the marked measurement stations of the RIVM is shown in Appendix A2.

2.3 Nitrogen Oxides: NOx

This thesis focuses on the effect of a LEZ on NOx emissions. NOx is the generic term for nitrogen oxides and it is the combination of nitrogen dioxide (NO2) and nitrogen monoxide (NO). NOx is one of the key factors of air pollution and research has shown that NOx is harmful to nature and humanity. Exposure to high NOx concentrations has direct short- and long-term effects, such as headaches, eye irritation and respiratory diseases. Indirectly, NOx emissions affect humanity by damaging the ecosystems (Staff, 2015). Jonson et al. (2017) estimated that almost 10.000 premature deaths in the adult population in the European Union plus Switzerland and Norway can be attributed to NOx emissions in 2013.

Fuel consumption of vehicles is the greatest contributor to NOx emissions in big cities (RIVM, 2021). Waersted et al. (2022) found that diesel vehicles are the most polluting vehicles with the highest emissions. Dallmann et al. (2019) drew a similar conclusion with their research in Paris. Furthermore, this study showed that Euro 6 diesel cars emit 4.8 times as much NOx

emissions compared to similar Euro 6 petrol cars. All of these findings are in line with the restrictions of Amsterdam’s LEZ, as these rules banned the most polluting diesel vehicles in this zone (milieuzone, 2022).

However, NOx emissions are not solely affected by vehicle emissions. Although Wearsted et al (2022) found diesel cars to be the most important source of NOx emissions, they also found that there was a temperature dependency. According to their research in Norway, NOx emissions increase on cold winter days. Secondly, The Leibniz Institute for Tropospheric Research (2020) found that wind speed affects NOx emissions as well. They concluded that wind speeds are negatively correlated with NOx emissions from a cross-sectional study in two German cities. Lastly, Nasriddinov et al. (2020) found that humidity is a major variable in the concentrations of NOx emission. Moreover, they concluded that high humidity might even affect the accuracy of the sensors that measure NOx emissions. Therefore, humidity should also be taken into account when researching NOx emissions.

(10)

10 To conclude, omitting temperature, humidity and wind variables would lead to omitted variable bias. By including those weather variables in the regression as control variables, this bias can be solved. Even though the primary goal of this thesis is to examine the effect of the LEZ, it is also possible to verify the weather dependence of NOx emissions.

2.4 Difference in Difference Model

As previously described, this thesis focuses on the effect of the LEZ in Amsterdam.

Therefore, a model is needed to estimate the causal effects of the LEZ on NOx concentrations.

Randomized controlled trials (hereafter: RCTs) are the best methods to estimate causal effects, because this method provides the lowest biases in outcomes. However, one of the key assumptions to running a valid randomized controlled trial is that treatment should be assigned randomly (Angrist & Pischke, 2014). The averages between the control group and treatment group should be the same before the treatment took place. This assumption is violated, because measurement stations within the LEZ had higher NOx concentrations before the treatment took place (see section 3, Figure 1 and section 3, Table 1). Implementing a RCT in this research is not possible, because we cannot randomly assign some measurement stations to the LEZ.

Therefore, a RCT method would cause selection bias and wrong estimations of the treatment effect, so a model is needed that takes non-randomly assigned treatment into account.

A DiD model is the solution to non-randomly assigned treatment. The DiD model takes into account that the treatment and control group were different at the start of the treatment.

Designing a DiD requires three key variables: a dummy for the treatment group, a dummy for the post-treatment period and an interaction dummy (Angrist & Pischke, 2014). Firstly, the dummy for the treatment group indicates the pre-treatment difference between the measurement stations within the LEZ and the measurement stations outside of this zone. This dummy allows thus for initial differences between the control group and the treatment group. Secondly, the post-treatment period dummy indicates changes over time that both affected the treatment- and control group. A negative value for the post-treatment period dummy would indicate that NOx

emissions in the complete region of Amsterdam decreased during the treatment. Thirdly, the interaction dummy indicates the actual effect of the treatment. The interaction term indicates the average treatment effect of the treated (ATT) and is the key variable to check the effect of the LEZ. A negative coefficient for the interaction term would indicate that the new rules regarding the LEZ in Amsterdam introduced in 2018 have caused a decrease in NOx emissions in the regulated areas.

(11)

11 Although the DiD model seems to solve the addressed problem of non-randomly assigned treatment, it relies on stronger assumptions than RCTs. In addition to the four assumptions of the Best Linear Unbiased Estimators (BLUE), the DiD model also relies heavily on the parallel trend assumption. The parallel trend assumption states that, although the treatment and control group have different concentrations of the outcome before the treatment, their trends should behave the same over time (Columbia Public Health, n.d.). In other words, the measurement stations outside the LEZ and the stations inside the LEZ should move parallel in absence of the LEZ. There are no statistical tests to prove the parallel trend assumption. As the treatment always takes place, it is impossible to know the results without this treatment.

However, a visual inspection to check if the trends of the treatment and control group move parallel overtime before the treatment took place is the best option to verify the parallel trend assumption. This visual inspection has been done in section 3.3.1, the output is shown in figure 1. The likelihood of the parallel trend assumption is also tested with so-called placebo tests, these results are shown in section 3.3.2.

Another important assumption of the DiD model is that there should be no spillovers.

This occurs when the treatment also affects the control group (Columbia Public Health, n.d.).

In other words, spillovers occur when the rule changes in Amsterdam’s LEZ also affect the measurement stations outside the LEZ. As mentioned by Tarriño-Ortiz et al. (2022), LEZ’s increase demand for eco-friendly vehicles in the complete region of a LEZ. Moreover, it is likely that the stations close to the LEZ are also affected, as these are on the way to the LEZ. In addition, polluting vehicles are likely to avoid these routes, because they are not allowed to enter the LEZ. A DiD model would underestimate the effect of a LEZ, because the NOx

emission in the non-regulated stations also dropped due to the LEZ and this decreases the total difference in NOx emissions between the regulated and non-regulated stations. Therefore, the coefficient of interaction dummy would be too small. However, this spillover effect in Amsterdam’s LEZ might be smaller, because the main corridor of Amsterdam (A10) is excluded from the LEZ. Vehicles can still drive around Amsterdam and do not have to avoid the region completely. Firms and car users are therefore not obligated to replace their polluting vehicles when they only have to drive around Amsterdam.

The effect of spillovers is tested in section 4.3. The main regression of this research is shown in that section, with several exclusions of the measurement stations in the control group nearest to the LEZ. Therefore, the effect of spillovers in this research can be tested, because the measurement stations closest to the LEZ have the highest probability of being affected by the 2018 rule changes in Amsterdam’s LEZ. Spillovers are likely to play a role when the

(12)

12 regressions, which include only the stations far away from the LEZ, are the most statistically significant and show the biggest decrease in NOx concentrations.

3. Methods

3.1 Data

Six datasets have been used for this research. The first five datasets are open-source datasets that have been retrieved from Luchtmeetnet (n.d.), which is a special website from the RIVM with datasets of all components of air quality. The datasets from Luchtmeetnet contain hourly NOx data of all measurement stations in the Netherlands for a given year. The five excel datasets that are used cover the years from 2015 up to 2019. The years 2020 and 2021 have been excluded from this research for two main reasons. Firstly, the Netherlands struggled with the COVID crisis in 2020 and 2021, which caused a decrease in traffic jams by up to more than 50%, because most people had to work at home. This, of course, had consequences for car usage and NOx emissions in the Netherlands (Ministerie van Infrastructuur en Waterstaat, 2020).

Secondly, the area of Amsterdam’s LEZ was expanded in 2020. The area between the north of the river ‘t IJ and south of the ring road A10 were added to the LEZ (Gemeente Amsterdam, 2020). This expansion would have moved measurement station number two (see appendix A1.) from the control into the treatment group. Therefore, the years 2020 and 2021 have been excluded from this research and it focuses solely on the years 2015 until 2019.

As mentioned in section 2.3, several weather variables influence the amount of NOx

emissions and these variables should be taken into account to solve the omitted variable bias.

This bias occurs when a regression model doesn’t include a relevant variable, which will then lead to bias estimators. To obtain all relevant weather variables, the last (sixth) dataset is retrieved from the Dutch national weather service, KNMI. The KNMI has registered many weather variables across the Netherlands for over 70 years (KNMI, 2022). The region of Amsterdam is closest to the KNMI measurement station of Schiphol, therefore this station has been used for the conducted research. The KNMI dataset is an open-source text file with all relevant daily weather information of Schiphol for the last 70 years. This weather dataset was used to control for daily changes in temperature, wind speed and humidity.

To get meaningful data, the datasets were processed. All datasets were merged to obtain one dataset. This final dataset contains daily NOx emissions of all measurement stations in the

(13)

13 region of Amsterdam combined with relevant weather data. The analysis has been done with daily data, as only daily values of the weather variables were available.

3.2 Descriptive Statistics

The variables which were used for the regression of the DiD model are stated below in Table 1. This table contains the number of observations, mean, standard deviation, minimum value and maximum value of all relevant variables.

Table 1: Descriptive Statistics

Variable Obs. Mean Std. Dev. Min Max

NOx (µg/m³) 30842 49.348 44.536 0 473.696

NOx out LEZ pre-2018 (µg/m³) 11982 42.118 40.826 1.62 473.70 NOx in LEZ pre-2018 (µg/m³) 6552 71.586 53.521 6.87 398.94 NOx out LEZ post-2018 (µg/m³) 7937 36.809 33.521 0 334.69 NOx in LEZ post-2018 (µg/m³) 4371 58.602 42.822 6.94 435.71

Temperature (ºC) 31042 11.168 6.029 -6.6 29.5

Humidity (%) 31042 79.777 9.614 36 99

Windspeed (m/s) 31042 4.878 2.148 1 14.6

Distance (km) 31042 3.459 3.55 0 10.8

LEZ 31042 0.353 0.478 0 1

After 31042 0.4 0.49 0 1

LEZ*After 31042 0.141 0.348 0 1

First of all, the ‘NOx’ variable is the dependent variable for the regressions in section 4.

This variable is collected from the five RIVM excel files and is measured in microgram (µg) per cubic meter (m³). The NOx variable misses 200 observations, due to maintenance of the measurement stations. After a visual inspection of the NOx data, the missing values appeared to be random. Therefore, the missing values likely do not lead to non-randomization problems.

The NOx data has been split into four different variables to obtain the descriptive statistics for both the control group (outside LEZ) and treatment group (in LEZ), before and after the 2018 rule changes. These four variables are shown below the general NOx variable.

(14)

14 The means of ‘NOx out LEZ pre-2018’ and ‘NOx in LEZ pre-2018’ show that measurement stations in the LEZ had on average 69.7% higher NOx emissions before the start of the 2018 rule changes

The variables ‘Temperature’, ‘Humidity’ and ‘Windspeed’ were collected from the KNMI weather file and are measured in Celsius, percentages and meters per second, respectively. ‘Distance’ shows the average distance of all measurement stations to the LEZ and it is measured in kilometers (km). The last three variables are all dummies, which can only be equal to zero or one. The dummy ‘LEZ’ is equal to one for all NOx data measured in the LEZ.

‘After’ equals one for all NOx data measured after the start of 2018. The dummy ‘LEZ*AFTER’

is the variable of interest for this research and is only equal to one for all NOx data measured in the LEZ after the start of 2018. The means of the three dummies show the fraction of the data equal to one for a specific dummy; Circa a third of the seventeen stations were treated, for 40%

of the time. Lastly, the regressions in section 4 also contain the season dummies: ‘winter’,

‘spring’, ‘summer’, and ‘fall’, because then NOx data contains seasonality (see figure 1). These final four dummies are not shown in the descriptive table, because they do not have relevant descriptive statistics.

3.3 Validating Parallel Trend Assumption 3.3.1 Visual Inspection

As mentioned in section 2.4, the DiD model relies heavily on the parallel trend assumption. The most standard way to check this assumption is to plot the data over time. The figure below shows the monthly averages of NOx data of the treatment and control group. The vertical line indicates the start of the stricter rules in Amsterdam’s LEZ. The blue dotted line shows the expected values of the treatment group if the LEZ had not been implemented and the parallel trend assumption holds. The first point of the blue dotted line was created by taking the difference at the start of 2018 between the measurement stations in the LEZ and the stations outside the LEZ. The next points were created by plotting the points parallel to the points of the measurement stations outside the LEZ.

(15)

15 Figure 1: Validating the Parallel Trend Assumption

The treatment and control group move similar over time before 2018. Due to this similarity before 2018, the treatment and control group would likely have moved parallel overtime after 2018. Therefore, the parallel trend assumption likely holds in this DiD model.

Seasonality in the NOx data is also noticeable in figure one, because the NOx data rise in winter periods and drop in summer periods.

The NOx data of the treatment group decreased compared to the expected parallel trend data of the treatment group (blue dotted line). A drop in the treatment group compared to the control group would indicate a decrease in NOx due to the stricter rules of Amsterdam’s LEZ.

Although this drop looks promising in figure one, the significance of this drop must be tested.

3.3.2 Placebo Tests

Running so-called placebo tests is another way to check whether it is likely that the parallel trend assumption holds. All post-treatment data is excluded in the placebo tests, so the tests only contain NOx data between 1 January 2015 and 1 January 2018. The placebo tests pick several treatment periods and pretend that the treatment took place at that time. The placebo time periods for this research are: 1 July 2016, 1 January 2017 and 1 July 2017. An insignificant value for the interaction dummy in a placebo time period indicates the likeliness of the parallel trend assumption. However, finding multiple significant values for the interaction term

(16)

16 indicates that something is wrong about the parallel trend assumption in the research (Huntington-Klein, 2013).

The main DiD regression model (see section 3.5) is used for each placebo time period and the interaction dummy ‘LEZ*After’ is shown in the table below. This table shows the value, clustered SE and p-value for the interaction term for each placebo test.

Table 2: Placebo Tests Parallel Trend Assumption

LEZ*After Coefficient Clustered SE p-value

1 July 2016 -.0551715 .0173181 0.006

1 January 2017 -.0268645 .0227542 0.255

1 July 2017 .0238097 .0358092 0.516

The placebo tests on the first of January 2017 and on the first of July 2017 are both statistically insignificant: p-values of 0.255 and 0.516, respectively. This result indicates that it is likely the parallel trend assumption holds for these time periods. Moreover, these time periods are the closest to the actual start of the treatment (1 January 2018), which increases the likelihood that the measurement stations of the treatment and control group would have moved parallel in absence of the treatment. However, the placebo test on the first of July 2016 is statistically significant, which indicates that the measurement stations in this time periods did not move parallel over time.

To conclude, a visualization of the NOx data and the placebo tests closest to the actual treatment period indicate that the parallel trend assumption likely holds. However, it is recommended for further research to investigate why a placebo test on the first of July 2016 is statically significant.

3.4 Validating Other Model Assumptions 3.4.1 Distribution NOx Data

Figure 2 shows a histogram before the 2018 rule changes with the NOx distribution of both the treatment (blue) and control group (orange). The x-axis indicates the NOx (µg/m3) values and the y-axis indicates the frequency of every bin. The distribution of both the treatment and control group are right-skewed (positive skewness). Therefore, both distributions are not symmetrical normal distributions. However, the distributions seem to fit neatly for log-normal

(17)

17 distributions (Angrist & Pischke, 2014). The scientific Python library SciPy was used to calculate the best fit log-normal distribution values of mu (mean) and sigma (standard deviation) for both distributions. The probability density functions (PDF) with the best-fitted values of μ and σ have been plotted as dashed lines in Figure 2. The formula, used for plotting the PDF functions, and the exact values for μ and σ are stated in appendix A3. The dashed lines fit well, so both distributions are likely to follow a lognormal distribution.

Figure 2: Histogram of NOx Distribution, before 2018

Figure 3 shows a histogram before the 2018 rule changes with a log transformation of the NOx distribution of both the treatment (blue) and control group (orange). The x-axis indicates the log NOx (µg/m3) values and the y-axis indicates the frequency of every bin. The scientific Python library SciPy was used to calculate the best fit normal distribution values of mu and sigma. The formula used for plotting the PDF function and the exact values for μ and σ are stated in appendix A4. The dashed lines fit well, so both distributions are likely to follow a normal distribution after the log-transformation of the data. Therefore, the regression equation used in this research uses the log of the dependent value, NOx.

(18)

18 Figure 3: Histogram of NOx Distribution with Log-Transformation, before 2018

3.4.2 Serial Correlation

Time series data often have serial correlation in the residuals and this needs to be solved.

Serial correlation (also called: autocorrelation) occurs when an error term in period one is correlated with an error term in period two (Angrist & Pischke, 2014). The data contains less information due to serial correlation, because some of its information correlates with the previous data point(s). This leads to too large t-values and too small SE’s (standard errors) of the estimates. Serial correlation occurs often in time series, because there are often variables that influence the independent variables, but move slowly over time.

Wooldridge’s test in Stata has been done to check for serial correlation in residuals of the NOx data (Drukker, 2003). This test confirmed the occurrence of serial correlation in the residuals. Therefore, the DiD regression model with regular robust SE’s, which only controls for heteroscedasticity in the residuals, would probably lead to too small SE’s of the estimators without intervention.

Clustering the SE’s on the individual measurement stations can solve the serial correlation problem. This allows for arbitrary correlation within the individual measurement stations (Angrist & Pischke, 2014). However, Nichols & Schaffer (2007) concluded from their paper that the number of clusters is essential for the accuracy of the clustered SE’s. The number

(19)

19 of clusters should be circa 50 or more to get accurate SE’s. However, the cure can be worse than the disease for a cluster size of much less than 50, making the clustered SE’s inaccurate.

The cluster size of this research is limited to seventeen (the number of measurement stations). Therefore, even clustered SE’s are not a perfect solution for the serial correlation problem. For this reason, both the standard robust SE’s and the clustered SE’s are shown in section 4.

3.5 Regression Equation

To analyse the effect of the 2018 rule changes in Amsterdam’s LEZ, a DiD regression model is used. The important regression equation used for this research is stated below.

Regression Equation (1):

(1) log (NOxit) = β0 + β1Aftert + β2LEZ * Afterit + β3Temperaturet + β4Humidityt + β5Windspeedt + β6Wintert + β7Springt + β8Summert + ϒi Station ii + ϵit

As described before, NOx is the dependent variable in the regression equation and the log of the NOx data is used for the regression analysis. The post-treatment dummy ‘After’

indicates changes over time that both affected the treatment- and control group. The variable of interest is the interaction dummy ‘LEZ*After’ (coefficient β2). This dummy shows whether the 2018 rule changes in Amsterdam’s LEZ are effective or not. The variables ‘Temperature’,

‘Humidity’ and ‘Windspeed’ are included in the regression equation to control for weather changes during the research. The dummies ‘Summer’, ‘Spring’ and ‘Winter’ are included in the regression equation to capture the seasonality of the NOx data. The sixteen ϒ_i (i goes from 1 until 16) parameters are the fixed effects of the individual measurement stations. These individual fixed effects control for any time invariant differences between the measurement stations. The ‘LEZ’ dummy (indicating the initial difference between the treatment and control group) does not add any information to the regression equation. The equation already has individual fixed effects, so the ‘LEZ’ dummy is excluded from the main regression equation and column 2.

The dummies for ‘Fall’ and the seventeenth measurement stations have been omitted in the regression equation to avoid the Dummy Variable Trap (Angrist & Pischke, 2014). This trap occurs when one variable can be predicted from the others which will result in co-linearity.

(20)

20

4. Results

4.1 Results with Clustered SE’s

The output of regression equation (1) with clustered SE’s is shown in Table 3. This table also contains a second and third regression equation to show the robustness of the estimators. The third equation also has the ‘LEZ’ dummy, because this equation does not have individual fixed effects. Therefore, the ‘LEZ’ dummy is needed in the third equation to capture the initial difference between measurement stations outside the LEZ and the stations inside the LEZ.

The first equation/column is the preferred specification for this research. The individual fixed effect of every measurement station is summarized as one variable in the table, because the individual value of every measurement station is irrelevant. All other variables with clustered SE’s are shown in Table 3.

First of all, the variables ‘Temperature’ and ‘Windspeed’ are both negative and significant. A decrease in temperature and/or wind speed leads on average to a decrease in NOx

emissions in Amsterdam. The way temperature and windspeed affect NOx emissions are in line with the findings of Wearsted et al. (2022) and The Leibniz Institute for Tropospheric Research (2020), respectively. Moreover, the ‘Humidity’ variable is positive significant. An increase in humidity leads on average to an increase in NOx. This is in accordance with the findings of Nasriddinov et al. (2020).

As stated before, the variable of interest is the interaction dummy, ‘LEZ * After’.

Although the variable ‘LEZ After’ is negative, the variable is statistically insignificant with clustered SE’s (p-value = 0.148). Therefore, the effectiveness of the 2018 rules changes in Amsterdam’s LEZ is not statistically significant when using clustered SE’s.

(21)

21 Table 3 Regression Output Clustered SE’s

(1) (2) (3)

Log (NOx) Log (NOx) Log (NOx)

LEZ 0.612^*

(0.233)

After -0.0922^*** -0.0746^*** -0.0759^***

(0.0140) (0.0140) (0.0136)

LEZ*After -0.0648 -0.0642 -0.0630

(0.0427) (0.0428) (0.0428)

Temperature -0.00887^** -0.0277^*** -0.0278^***

(0.00278) (0.00250) (0.00249)

Humidity 0.0115^*** 0.0159^*** 0.0159^***

(0.00149) (0.00157) (0.00159)

Windspeed -0.151^*** -0.141^*** -0.141^***

(0.0110) (0.0111) (0.0112)

Winter 0.164^***

(0.0197)

Spring -0.148^***

(0.0126)

Summer -0.342^***

(0.0138) Measurement

Stations FE

Yes Yes No

N 30841 30841 30841

R² 0.415 0.380 0.359

Clustered Standard Errors in parentheses

* p < 0.05, ^** p < 0.01, ^*** p < 0.001

4.2 Results with Robust SE’s

The output of regression equation (1) with robust SE’s is shown in Table 4. The robust standard errors control for heteroscedasticity in the residuals This table also contains a second equation to show the robustness of the estimators. However, the first column is the most essential for this research. The equation estimators of equation one in Table 4 correspond exactly with the estimators of equation one in Table 3. However, the SE’s of the estimators are different due to the robust SE’s.

The intersection dummy ‘LEZ * After’ is now statistically significant. This is the main difference from the conclusions with clustered SE’s. The 2018 rule changes in Amsterdam’s

(22)

22 LEZ are effective with robust SE’s. The calculation of the percentage change in NOx emissions due to the rule changes in Amsterdam’s LEZ is stated below.

Percentage change in NOx: 100 ∗ (e^β2− 1) = 100 ∗ (e^−0.0648− 1) ≈ − 6.3%

The 2018 rule changes in Amsterdam’s LEZ lead to an average decrease in NOx of 6.3%

in the regulated areas compared to the non-regulated areas.

Table 4 Regression Output Robust SE’s

(1) (2) (3)

Log (NOx) Log (NOx) Log (NOx)

LEZ 0.612^***

(0.00990)

After -0.0922^*** -0.0746^*** -0.0759^***

(0.00786) (0.00808) (0.00976)

LEZ*After -0.0648^*** -0.0642^*** -0.0630^***

(0.0113) (0.0117) (0.0155)

Temperature -0.00887^*** -0.0277^*** -0.0278^***

(0.000782) (0.000547) (0.000686)

Humidity 0.0115^*** 0.0159^*** 0.0159^***

(0.000375) (0.000358) (0.000446)

Windspeed -0.151^*** -0.141^*** -0.141^***

(0.00136) (0.00131) (0.00170)

Winter 0.164^***

(0.00960)

Spring -0.148^***

(0.00885)

Summer -0.342^***

(0.00935) Measurement

Stations FE

Yes Yes No

N 30841 30841 30841

R² 0.629 0.606 0.359

Robust Standard Errors in parentheses

* p < 0.05, ^** p < 0.01, ^*** p < 0.001

(23)

23 4.3 Robustness: Spillover effects

As described in section 2.4, spillover effects play a role in this research. The effect of spillovers is tested in this section. As stated before, a DiD model would underestimate the effect of a LEZ with spillovers. As explained, the NOx emissions in the non-regulated stations also dropped due to the LEZ. This results in a decrease of the total difference in NOx emissions between the regulated and non-regulated stations. Therefore, the coefficient of the interaction dummy would be too small, leading to a downwards bias of the interaction dummy. This bias indicates that the actual value of the coefficient for the interaction dummy, β₂

,

is largerthan the estimated value.

In order to analyse the effect of spillovers, the main regression (see section 3.5) of this research is shown with different exclusions of measurement stations in the control group nearest to the LEZ. Column one in Table 5 shows the same regression output as in Table 3 column one, which is shown as a reference. Columns 2, 3, 4 and 5 all have restrictions in which measurement stations in the control group are included in the regression. For example, column 2 only uses those measurement stations in the control group which are at least 2 km away from the LEZ.

As a result of the distance restrictions, the amount of measurement stations in the control group drops with steps of two. Columns 1, 2, 3, 4 and 5 include eleven, nine, seven, five and three measurement stations in the control group, respectively. All columns use clustered SE’s for the regression and the columns use all the complete treatment group of six measurement stations.

The results in Table 5 indicate the likeliness of spillover effects in this research. The coefficient for the interaction dummy, ‘LEZ*After’ is negatively larger for every additional distance restriction. The p-values of the interaction dummy for columns 1, 2, 3, 4 and 5 are:

0.148, 0.103, 0.093, 0.028 and 0.012, respectively. The interaction dummy thus, is statistically significant in columns 4 and 5, because these columns have the largest distance restrictions for the control group.

These results show the likeliness of spillover effects in this research. Therefore, the estimated decrease of 6.3% in NOx due to the 2018 rule changes in Amsterdam’s LEZ is likely larger than presented in this research. For example, if only measurement stations in the control group which are at least 5 km away from the LEZ were included, the estimated decrease in NOx

due to the 2018 rule changes would be circa 10.4%.

(24)

24 Table 5: Spillover Effects

(1) (2) (3) (4) (5)

Log (NOx) Log (NOx)

>2km

Log (NOx)

>3.5km

Log (NOx)

>5km

Log (NOx)

>8km After -0.0922^*** -0.0813^*** -0.0765^** -0.0492^** -0.0289^**

(0.0140) (0.0146) (0.0186) (0.0117) (0.00620)

LEZ*After -0.0648 -0.0754 -0.0821 -0.110^* -0.133^*

(0.0427) (0.0432) (0.0450) (0.0430) (0.0414) Temperature -0.00887^** -0.00831^* -0.00939^* -0.00853^* -0.0113^*

(0.00278) (0.00299) (0.00314) (0.00366) (0.00369) Humidity 0.0115^*** 0.0117^*** 0.0108^*** 0.0105^*** 0.00888^***

(0.00149) (0.00146) (0.00138) (0.00162) (0.00146) Windspeed -0.151^*** -0.150^*** -0.152^*** -0.149^*** -0.152^***

(0.0110) (0.0115) (0.0114) (0.0131) (0.0157)

Winter 0.164^*** 0.172^*** 0.162^*** 0.165^*** 0.146^***

(0.0197) (0.0165) (0.0174) (0.0192) (0.0173) Spring -0.148^*** -0.142^*** -0.143^*** -0.133^*** -0.129^***

(0.0126) (0.0134) (0.0148) (0.0157) (0.0184) Summer -0.342^*** -0.336^*** -0.337^*** -0.331^*** -0.313^***

(0.0138) (0.0149) (0.0165) (0.0192) (0.0184) Measurement

Stations FE

Yes Yes Yes Yes Yes

N 30841 27237 23634 19993 16386

R² 0.415 0.411 0.426 0.410 0.456

Clustered Standard Errors in parentheses

* p < 0.05, ^** p < 0.01, ^*** p < 0.001

(25)

25

5. Discussion

The 2018 rule changes in Amsterdam’s LEZ are statistically significant with robust SE’s and statistically insignificant with clustered SE’s. There are two main reasons why the effectiveness of the 2018 rule changes in Amsterdam’s LEZ is not significant with clustered SE’s.

Firstly, the number of measurement stations used in this research is a likely reason why the 2018 rule changes are not significant with clustered SE’s. In order to get accurate SE’s circa fifty or more measurement stations should be used, while this research only includes seventeen.

Increasing the number of stations would improve the accuracy of clustered SE’s. However, RIVM measurement stations are costly and it will take time to increase the number of measurement stations in Amsterdam.

Secondly, spillover effects are likely to influence the effectiveness of the 2018 rule changes in Amsterdam’s LEZ. Tarriño-Ortiz et al. (2022) have a possible explanation as to why spillovers play a role in this research. They found that a LEZ increases the demand for environmental alternatives and decreases the demand for polluting vehicles in the region of the LEZ. Therefore, measurement stations close to the LEZ are probably influenced by the LEZ.

A DiD model would underestimate the effect of a LEZ with spillovers, because the NOx

emission in the non-regulated stations also dropped due to the LEZ. The effect of spillovers has been tested and the effectiveness of the LEZ increased when omitting the measurement stations in the control group closest to the LEZ. Therefore, it is expected that the drop of 6.3% is larger when controlling for the spillover effects.

The findings of this research are in line with a similar study by Morfeld et al. (2014).

They found a 4% reduction in NOx emissions associated with German LEZ’s. Morfeld et al.

argued that this reduction was statistically significant, but rather small. Almost the same conclusions were drawn in the research of Santos et al. (2019). They found reductions in NOx

emissions due to the Lisbon LEZ, but this reduction was small and not statistically significant.

To conclude, there are two recommendations for further research. Firstly, it is advised to do comparable research with controlling for the exact spillover effects caused by the LEZ.

This likely leads to more unbiased results and it is expected that the decrease in NOx emissions will be larger. Secondly, it is up to the policy maker to determine whether they find a 6.3%

reduction in NOx emissions significant enough. In case they are not pleased with this decrease in NOx emissions, they may want to implement further restrictions. For example, in Griffiths et al. (2016), the researchers recommended implementing a no emission zone after their conducted study regarding London’s LEZ.

(26)

26

6. Conclusion

This research has tried to answer the following research question: ‘Did the 2018 rule changes in Amsterdam’s low emission zone lead to a reduction in nitrogen oxides in the regulated parts of the city?’ Five datasets of the RIVM and one dataset of the KNMI have been used in this research.

In order to answer the research question, an econometrical DiD regression model has been used. The DiD model compares measurement stations within the LEZ (treatment group) to stations outside the LEZ (control group) over time. This regression model indicates the effect of the 2018 rule changes in Amsterdam’s LEZ on NOx emissions. The main model assumptions of the DiD model have been validated in this research. The results of the regression are shown with robust SE’s and clustered SE’s. The data used for the DiD model has serial correlation issues. Clustered SE’s on an individual level solve this serial correlation. Therefore, clustered SE’s were preferred for this research.

Results have shown that the 2018 rule changes in Amsterdam’s LEZ decreased NOx

emissions on average with 6.3% in the regulated parts of the city compared to the non-regulated parts. This decrease in NOx emissions is statistically significant with robust SE’s and statistically insignificant when using clustered SE’s. However, this research likely underestimates the effect of the 2018 rule changes in Amsterdam’s LEZ, due to spillover effects. Therefore, it is expected that the drop of 6.3% is larger and also statistically more significant with clustered SE’s when controlling for the spillover effects. This research recommends doing a similar study with controlling for the exact spillover effects to obtain an unbiased result on the effectiveness of a low emission zone.

(27)

27

Bibliography

Angrist, J. D., & Pischke, J. S. (2014). Mastering ’Metrics. Amsterdam University Press

Boerep, L. (2019, September 25). Waarom staan er door stikstof 18.000 bouwprojecten op losse schroeven? 4 vragen en antwoorden. Business Insider Nederland. Retrieved April 24, 2022, from https://www.businessinsider.nl/stikstof-uitspraak-pas-uitleg/

Columbia Public Health. (n.d.). Difference-in-Difference Estimation. Retrieved May 15, 2022, from https://www.publichealth.columbia.edu/research/population-health-methods/difference- difference-estimation

Dallmann, T., Bernard, Y., Tietge, U., & Muncrief, R. (2019). Remote sensing of motor vehicle emissions in Paris. TRUE.

https://theicct.org/wcontent/uploads/2021/06/TRUE_ParisRS_study_20190909.pdf

Drukker, D. M. (2003). Testing for Serial Correlation in Linear Panel-data Models. The Stata Journal:

Promoting Communications on Statistics and Stata, 3(2), 168–177.

https://doi.org/10.1177/1536867x0300300206

European Environment Agency. (2021). European city air quality viewer. Retrieved June 10, 2022, from https://www.eea.europa.eu/themes/air/urban-air-quality/european-city-air-quality-viewer Ezeah, C., Finney, K., & Nnajide, C. (2015). A Critical Review Of The Effectiveness Of Low

Emission Zones (LEZ) As A Strategy For The Management Of Air Quality In Major

European Cities. Journal of Multidisciplinary Engineering Science and Technology (JMEST).

https://www.researchgate.net/publication/280575118_A_Critical_Review_of_the_Effectivene ss_of_Low_Emission_Zones_LEZ_As_A_Strategy_for_the_Management_of_Air_Quality_in _Major_European_Cities

Gemeente Amsterdam. (2018). Gemeentelijk ontheffingenbeleid milieuzone bestelauto’s, autobussen en taxi’s Amsterdam 2018. Overheid.nl. Retrieved March 28, 2022, from

https://zoek.officielebekendmakingen.nl/gmb-2018-66151.html

Gemeente Amsterdam. (2020, January 6). Staatscourant 2020, 1173 | Overheid.nl > Officiële bekendmakingen. Overheid.nl. Retrieved May 31, 2022, from

https://zoek.officielebekendmakingen.nl/stcrt-2020-1173.html

Goudappel Coffeng and Buck Consultants International. (2010, October). Landelijke effectstudie milieuzones vrachtverkeer 2010 Effecten op de luchtkwaliteit. F. Nieman.

https://www.milieuzones.nl/sites/default/files/Effectstudie_2010.pdf

Griffiths, C., Mudway, I., Wood, H., Marlin, N., Dundas, I., Walton, R., Grigg, J., Sheikh, A., Lee, T.,

& Kelly, F. (2016). P180 Impact of the london low emission zone on children’s respiratory health: a sequential yearly cross sectional study 2008–2014. Thorax, 71(Suppl 3), A182.1- A182. https://doi.org/10.1136/thoraxjnl-2016-209333.323

(28)

28 Huntington-Klein, N. (2013, June 13). Chapter 18 - Difference-in-Differences | The Effect.

Theeffectbook. Retrieved June 20, 2022, from https://theeffectbook.net/ch- DifferenceinDifference.html?panelset=stata-code

Jonson, J. E., Borken-Kleefeld, J., Simpson, D., Nyíri, A., Posch, M., & Heyes, C. (2017). Impact of excess NOx emissions from diesel cars on air quality, public health and eutrophication in Europe. Environmental Research Letters, 12(9), 094017. https://doi.org/10.1088/1748- 9326/aa8850

KNMI. (2022). KNMI - About KNMI. Retrieved April 25, 2022, from https://www.knmi.nl/over-het- knmi/about

Leibniz Institute for Tropospheric Research (TROPOS). (2020). Traffic density, wind and air stratification influence concentrations of air pollutant NO2. Sciencedaily.

https://www.sciencedaily.com/releases/2020/06/200626114750.htm Luchtmeetnet. (n.d.). Airquality index (LKI). Retrieved April 24, 2022, from

https://www.luchtmeetnet.nl/informatie/luchtkwaliteit/luchtkwaliteitsindex-(lki)

Ministerie van Infrastructuur en Waterstaat. (2020, August 13). Meer thuiswerken = minder file - Werken aan Duurzame Mobiliteit. rijksoverheid. Retrieved June 1, 2022, from

https://magazines.rijksoverheid.nl/ienw/werkenaanduurzamemobiliteit/2020/03/meer- thuiswerken-is-minder-file

Morfeld, P., Groneberg, D. A., & Spallek, M. F. (2014). Effectiveness of Low Emission Zones: Large Scale Analysis of Changes in Environmental NO2, NO and NOx Concentrations in 17 German Cities. PLoS ONE, 9(8), e102999. https://doi.org/10.1371/journal.pone.0102999 Nasriddinov, A., Rumyantseva, M., Konstantinova, E., Marikutsa, A., Tokarev, S., Yaltseva, P.,

Fedorova, O., & Gaskov, A. (2020). Effect of Humidity on Light-Activated NO and NO2 Gas Sensing by Hybrid Materials. Nanomaterials, 10(5), 915.

https://doi.org/10.3390/nano10050915

Nichols, A., & Schaffer, M. (2007, September 10). Clustered Errors in Stata. Stata. Retrieved June 1, 2022, from https://www.stata.com/meeting/13uk/nichols_crse.pdf

Reijneker, M. (2021). Stikstof. Wageningen University & Research. Retrieved April 18, 2022, from https://www.wur.nl/nl/Dossiers/dossier/Stikstof-

1.htm#:~:text=Lees%20door%20over%20het%20huidige%20stikstofprobleem&text=Die%20 condities%20zijn%20in%20Nederland,steeds%20te%20veel%20stikstof%20terecht.

RIVM. (2021). Stikstof - Stikstofoxiden (NOₓ). Retrieved April 20, 2022, from https://www.rivm.nl/stikstof/stikstofoxiden-nox

Santos, F. M., Gómez-Losada, L., & Pires, J. C. (2019). Impact of the implementation of Lisbon low emission zone on air quality. Journal of Hazardous Materials, 365, 632–641.

https://doi.org/10.1016/j.jhazmat.2018.11.061

(29)

29 Staff, S. X. (2015, September 23). NOx gases in diesel car fumes: Why are they so dangerous?

Phys.Org. Retrieved April 25, 2022, from https://phys.org/news/2015-09-nox-gases-diesel-car fumes.html#:%7E:text=NOx%20has%20direct%20and%20indirect,land%E2%80%94harming

%20animals%20and%20plants.

Tarriño-Ortiz, J., Gómez, J., Soria-Lara, J. A., & Vassallo, J. M. (2022). Analyzing the impact of Low Emission Zones on modal shift. Sustainable Cities and Society, 77, 103562.

https://doi.org/10.1016/j.scs.2021.103562

Uitleg milieuzones | Milieuzones in Nederland. (2022). milieuzones. Retrieved April 30, 2022, from https://www.milieuzones.nl/english

Waersted, E. G., Sundvor, I., Denby, B. R., & Mu, Q. (2022). Quantification of temperature dependence of NOx emissions from road traffic in Norway using air quality modelling and monitoring data. ScienceDirect. https://doi.org/10.1016/j.aeaoa.2022.100160

(30)

30