Ethics statement

In the Netherlands, medical research is governed by the Medical Research Involving Human Subjects Act (‘Wet Medisch-wetenschappelijk Onderzoek met mensen'), which is based on the principles of the declaration of Helsinki. This law only applies if study participants are subjected to any action, treatment or behaviour. Written confirmation from the Medical Ethics Review Committee of the Academic Medical Center that the Medical Research Involving Human Subjects Act was obtained stating that this law does not apply to this study and therefore no official approval was required.

Data

Internet search query data collected from Google Trends have previously been used to predict influenza epidemics [19], unemployment rates [20] and seasonal depression [21]. Google Trends search query data has recently been used to analyse behavioural patterns like consumer interest in non-cigarette tobacco products [22], the use of electronic cigarettes [23,24] and the relation between tax avoidance and a cigarette tax increase [25,26]. This previously conducted research [19–26] shows Google Trends to be a useful tool to analyse behavioural patterns which are difficult to measure otherwise.

Google Trends data is publicly available and can be downloaded directly from http://www.google.com/trends. Google Trends data are scaled to the average search traffic (the number of searches conducted at a certain point of time) for a selected point in time. The outcome variable, relative search volume (RSV) measures the total number of searches conducted for the selected query compared with the total number of Google searches executed at that point in time. The time period with the highest relative amount of search queries of the chosen term is assigned a value of 100. Other time periods obtain a score relative to 100 [27]. Because of the indexing used, Google Trends has the possibility to compare search queries with other search queries, between geographic regions and within broader periods of time [27,28].

We used Google Adwords (google.com/adwords), to collect the average monthly search volume for various smoking cessation related search queries [26]. Because the Dutch equivalent to ‘quit smoking’ (‘stoppen met roken’) was by far the most used search query (93% for the Netherlands and 90% for Belgium), compared to alternatives like ‘how do I quit smoking’, ‘help quit smoking’ and ‘smoking withdrawal symptoms’, this term was used as outcome in the analysis.

Analytical design

This study was conducted using a quasi-experimental design to examine trends in search queries around interventions. Interrupted time series analysis allows the comparison of outcome measures before and after the implementation of an intervention [29,30]. Therefore, this method can be used to analyse the effects of the introduction of new tobacco control policies by using RSV data. Time series analyses have been previously used to measure the impact of tobacco control policies on smoking prevalence in Australia [31]. RSV data were made available by Google Trends from 2004 onwards. Dutch RSV data for the query ‘quit smoking’ (‘stoppen met roken’) were retrieved for the 2004 to 2013 period on a weekly scale. Based on the availability of Google Trends data, three Dutch smoking cessation interventions could be included: the smoking ban of 2008, the reimbursement of SCS in 2011 and the reintroduction of the reimbursement of 2013 (Table 1). In January 2004, a national smoke-free legislation at the workplace was implemented. Since no pre-intervention data were available, the effect of this legislation could not be analysed. Therefore, January 2004 was excluded from the analysis. In total, data from 517 weekly time points were retrieved, with the highest RSV observed in the middle of 2008 (Fig 1). Sensitivity analysis showed that our results were robust with exclusion of the first 6 months of 2004 (instead of January only).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Google Trends RSV data of ‘quit smoking’ for the Netherlands.

Dutch RSV (A), Dutch seasonally adjusted RSV (B) and as ARIMA (1,0,1) model (C).

https://doi.org/10.1371/journal.pone.0148489.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Tobacco control policies implemented in the Netherlands and Belgium in the period 2004–2013.

https://doi.org/10.1371/journal.pone.0148489.t001

We added the Dutch speaking part of Belgium as the control group, since Belgium is the country that is most comparable to the Netherlands in terms of geography, language, history and culture. Moreover, Belgium and the Netherlands have a similar smoking prevalence [18]. We used the same Dutch term on ‘quit smoking’ to obtain Belgian RSV data. The Belgian RSV ‘quit smoking’ data comprised a large number of missing data points until mid-2006. Therefore the period of January 2004 to August 2006 was excluded from the Belgian analysis. In total, data from 387 weekly time points were retrieved, with the highest RSV observed in the beginning of 2007 (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Google Trends RSV data of ‘quit smoking’ for Belgium.

Belgian RSV (A), Belgian RSV seasonally adjusted (B) and as ARIMA (1,0,2) Model (C).

https://doi.org/10.1371/journal.pone.0148489.g002

Statistical analysis

To correct for seasonality, a seasonal decomposition in SPSS (version 21) was performed and the seasonally adjusted series were used for the analyses [32]. This enabled us to adjust for the peak in search queries around, among others, the first day of January when people make New Year’s resolutions to quit smoking. The seasonal adjusted series were log transformed in order to stabilize variance over time [33].

An interrupted time series analysis was performed on the adjusted data [29,30]. Autoregressive integrated moving average (ARIMA) modelling was used to account for dependency between data points in time series. ARIMA modelling is able to account for non-independence between the data points in the time series, which would otherwise cause standard errors to be estimated incorrectly. When using ARIMA modelling, the length between measured time periods should be constant. Furthermore, a substantial amount of data points is necessary. Time series analysis using ARIMA modelling is accomplished in three phases: identification, estimation and diagnosis. In the identification phase ACF (autocorrelation function) and PACF (partial autocorrelation function) plots were examined to see which patterns are present in the data. The ACF plot is a bar chart of the coefficients of correlation between a time series and lags of itself. The PACF plot is a bar chart of the correlation coefficients between the series and lags of itself that are not explained by correlation at all lower order lags. Based on the visual inspection of the ACF and PACF, initial autoregressive (AR) and moving average (MA) terms were determined as a tentative model. In the estimation phase the tentative model was fitted to the seasonally adjusted RSV series to determine the fit of the model. Extra AR and MA terms were added to make sure no terms were left out of the model. In the diagnosis phase the best fitting model was determined based on residual ACF and PACF and the final model was used for further analyses [33,34].

The models were used to estimate the association with various tobacco control policies implemented in the Netherlands and Belgium (Table 1). Based on findings in literature, effects were estimated to be relatively short lived, with an increase in search queries before and shortly after the implementation of the legislation and a decrease in search queries a few weeks or months [15–17] afterwards. To assess the timing and duration of the effect of the tobacco control policies, twelve intervention effect periods were distinguished, ranging in duration from sixteen weeks up to one week before the intervention and one week up to 52 weeks afterwards. This method was chosen to minimize the amount of effect periods, in order to avoid the risks of multiple testing. Moreover, a similar approach is used in an article by Szatkowski et al. [15]. For each of the tobacco control policies, the respective periods were modeled as binary intervention dummies coded ‘1’ for the duration of the intervention period and coded ‘0’ for the period before and after that period of intervention. To clarify, for the effect period of one week after the intervention, this specific intervention variable was only coded as ‘1’ for the first week after the intervention. All other weeks were coded as ‘0’. Similarly, for the effect period of 5 to 8 weeks after the intervention, a variable was added in which week 5, 6, 7 and 8 were coded as ‘1’ and all other weeks were coded as ‘0’.

The dummies for the different periods were added simultaneously to the final model to estimate the intervention effects according to each period [35]. Output was measured in percentage points relative to the expected level without the influence of the tobacco control policies. This process was repeated for each of the Dutch tobacco control policies.

The Belgian ARIMA analysis was used as a control. The same regression model was applied to the Belgian model to study whether the changes observed around the Dutch interventions would be restricted to the Netherlands or not. Dummy variables related to Belgian policies (Table 1) were also added to the Belgian model to control for the effect of these Belgian interventions. Effect sizes and confidence intervals were estimated and a p-value smaller than 0.05 was considered to imply statistical significance. All analyses were performed using SPSS 21.

Evaluation of methodology

ARIMA interrupted time series modelling is appropriate to evaluate several types of community level health interventions [36]. ARIMA modelling is especially suitable when using detailed time series data such as Google Trends data [37]. Our study shows that Google data, measured in RSV, could be a useful tool to measure the effect of tobacco control policies, based on its ability to represent large populations. This is in agreement with Cavazos-Rehg et al. [23] who state that the results of their study validate the potential of Google trends as a valuable monitoring tool for tobacco use. However, RSV is a proxy and therefore does not give information on actual smoking cessation behaviour.

Time series analyses that distinguish multiple exposure periods are susceptible to the risk of multiple testing. However, we think that the results of this study are not influenced by multiple testing, since statistically significant results were found in far more than five per cent of the exposure periods. Another concern with time series is the statistical power. In time series, statistical power is primarily affected by the amount of post and pre intervention data points [38]. In our study, a total of 517 weekly data points of the Netherlands and 387 data points of Belgium were available, which seem sufficient for our purposes. According to Thomas [39], a method to estimate power retrospectively is to look at the width of confidence intervals While large effects could be demonstrated, we may not have been able to demonstrate smaller effects with statistical significance. For example, in the case of the introduction of the smoking ban in 2008, we could demonstrate changes occurring within 1 to 4 weeks after the intervention, but smaller changes in the same direction at the longer term could not be demonstrated with statistical significance, possibly due to inadequate power.

One of the main concerns of using internet search query data is the representativeness of such data. The prevalence of internet use is higher among individuals who are younger, more educated and have a higher income [40]. Moreover, compared to older people, those in their twenties may be more likely to think about quitting in response to smoking bans in bars and restaurants [11]. Therefore, our results may not be representative for the entire Dutch population. However, it has recently been demonstrated that individuals of sixty years and older are equally likely to use internet search engines to search for health related information compared to adolescents [26]. Furthermore, in the Netherlands and in Belgium, the percentage of the population having internet access is very high, 95% and 80% respectively [41].

Out of the seven Dutch and Belgian tobacco control policies used in our analyses, four were implemented at the first of January. Since they coincided with the New Year’s resolutions peak, it was difficult to separate the impact of the policies from this peak. We corrected for this by performing a seasonality correction, which eliminated the average peak found at the first week of January. However, this was not a completely accurate process, since Google Trends provides weekly data from Monday till Sunday and each subsequent year started on a different weekday. Because of the potential of residual confounding, results found in the first week of January should be interpreted with caution.

After the implementation of the reimbursement of SCS in 2011, an increase in RSV was found not only in the Netherlands, but also in Belgium. We could speculate that the increase in RSV could be caused by Dutch speaking media influencing not only the Dutch, but also Belgian residents. However, after the implementation of the reimbursement of SCS in 2013 no effects on Belgian RSV were found. A possible explanation might be that the introduction of the Dutch reimbursement policy of 2011 was accompanied by a media campaign [42], as opposed to the reimbursement policies of 2013. This could have influenced the exposure of Belgian residents towards this policy. Nonetheless, the Belgian results show that despite our time series study design, mid-term effects such as observed in relationship to the Dutch reimbursement policies of 2011 and 2013, may be subject to residual confounding, since more information is necessary to correct for media exposure.

Substantive interpretations

The findings of this study are consistent with other types of studies on the effect of tobacco control legislation on quit attempts and cessation rates. Earlier research on the Dutch smoke-free legislation in the hospitality industry found a temporary increase in quit attempts and a temporary decrease in smoking prevalence [10,11]. Bar visitors were more likely to quit smoking after the implementation of the smoking ban compared to the period before the implementation [10]. In New Zealand, a large increase in quit line telephone call registrations was found in the six-month period after the implementation of a smoke free legislation [12,13]. In England, a significant increase in quit attempts was found for two months after the start of smoke-free legislation [14]. This suggests that the effect of the smoking ban on smoking cessation related behaviour is short to mid-term, which is comparable with the short-term effect that was found in this study. The short duration and large magnitude of increase in research queries is suggestive of a “shock effect”. This would imply that the effect weakens quickly when the population adjusts to the new situation [43,44].

A Dutch study found that the implementation of SCS led to a decrease in smoking prevalence in the six months afterwards, thus suggesting long-term effects on smoking prevalence [14]. Also, a randomized controlled trial on free offer of nicotine patches through telephonic smoking quit lines found an increase in quit attempts and a decrease in smoking prevalence when the period of offering free nicotine patches increased [45]. A study in Massachusetts found that reimbursement of SCS was followed by a significant decline in smoking prevalence for at least two years with a yearly relative decline of 15,2% [46].

The available evidence thus suggests that the reimbursement of SCS had an effect on quit attempts and smoking prevalence for at least one year. This is consistent with our results of the reimbursement of SCS in 2011, but not with 2013. The shortened duration of the effects of the SCS policy in 2013 comparison with the SCS policy in 2011 may have been due to the fact that the same type of policies was repeated within a short time period. Moreover, a media campaign accompanied the introduction of the policy in 2011 but not in 2013 [46].

In contrast with the Dutch tobacco control policies, the Belgian tobacco control policies were not clearly associated with increases in Belgian RSV. This might be due to the complex history of tobacco control policies in Belgium, where new smoke-free policies were introduced with multiple exceptions and a limited enforcement [47].

Conclusion and implications

Our results provide some empirical support to believe that the introduction of a smoking ban in restaurant and bars is associated with a short-term increase in online searching for information about smoking cessation. The reimbursement of SCS costs seems to have a mid to longer-term effect on searching behaviour. This evidence is consistent with earlier results of previous studies that report a positive impact of such policies in trends on smoking cessation rates and specifies with more accuracy the magnitude and duration of these effects.

Information on the time frame during which tobacco control policies may affect smoking cessation behaviour and its precursors, can be used to improve the implementation of these policies. Periods when smokers are more likely to contemplate smoking cessation may provide windows of opportunity for interventions targeted at smokers who contemplate smoking.