Scientific Papers

No obvious effect on mortality from a patient choice reform expanding access to opioid disorder treatment – results from a natural experiment of policy change in Sweden | Substance Abuse Treatment, Prevention, and Policy

Description of Image

In order to facilitate comparison with the previously published paper on the same topic, we aimed to collect data for overdose deaths in Swedish adults (20–64 years of age) for the time period 2011–2021. The Cause of Death Register data is available since 1997, and we included data for all available years in order to represent the long-time trends visually.

Data sources

We collected data from two publicly available sources. Data for causes of death was collected from the public database of statistics of the Swedish National Board of Health and Welfare. This data is registered in the Swedish Cause of Death Register and it comprises data from all deaths of people registered as living in Sweden, regardless of whether they were registered as living in Sweden or not. In the register, the ultimate cause of death, as determined by a medical doctor, is registered with codes from ICD-10. In the publicly available data, the number of people who died each year with a certain ICD-10 code group (individual ICD-10 codes are aggregated in meaningful groups) are reported, divisible by sex, age, and county. We thus included data for the number of individuals who had died each year with an ultimate cause of death coded as either X42 (accidental poisoning by and exposure to narcotics and hallucinogens), X44 (accidental poisoning by and exposure to other and unspecified drugs), or Y10-Y14 (poisoning with drugs with undetermined intent), within the age rage 20–64 years. Data was divided by county (Skåne county vs. all the other 20 counties in Sweden combined).

The other data source was the Total Population Register from Statistics Sweden. This is a register on all inhabitants of Sweden, publicly available for extraction of aggregate data by various variables. For this study, we collected data on the numbers of individuals registered as living in Sweden in each county for the years 1997–2021. We then summed the data for the 20 counties that were not Skåne in order to use the population data as an offset variable in the statistical analyses described below.


The variables included from the register data were as follows: overdose deaths, population, county (Skåne vs. the other 20 counties in Sweden combined), and year (1997–2021, numbered). For purposes of statistical analysis, we included data only from 2011 and onwards.

We then created other variables based on the implementation of the intervention under study, i.e. the patient choice reform for OMT. This reform was gradually implemented in Skåne in 2014, so we excluded 2014 from the data for Skåne and designated 2011–2013 as pre-intervention and 2015–2021 as post-intervention. This dichotomous variable was labelled intervention. Because the effect of the intervention might be conceptualized as gradual rather than instantaneous, we created a new variable labelled intervention slope which was 0 for data points prior to 2014, and from 2015 to 2021 the data was 1 through 7.

The county variable can be used in order to assess the difference in means for Skåne vs. the rest of Sweden. We created interaction terms between county and the two intervention variables, to create the variables county X intervention, which has the value of 0 for all years prior to the intervention and 1 for all years post intervention for the data from Skåne county, and the value 0 for all years for the data from the rest of Sweden; and county X intervention slope, which differs from the previous variable only in that the data for the years following the intervention in the Skåne data has values from 1 to 7.

Statistical methods

We analyzed the data as an interrupted time series using a Poisson regression framework, using identical statistical methods as in the previous paper [8] but with four more years of data. Overdose deaths was used as the dependent variable with population as an offset variable, allowing us to analyze the differences in death rates. Year was included as an independent variable in order to take into account the overall temporal trend in the data. It is clear from the data prior to the intervention date that there is no difference between Skåne county and the rest of Sweden in either mean or slope, so any variables that might reflect such a difference were omitted a priori.

For the remaining four possible independent variables, i.e. intervention, intervention slope, county X intervention, and county X intervention slope, we used two analysis strategies. In the previous paper by Andersson and colleagues [8], a model selection process was utilized, indicating that the best fitting model included the intervention slope and the interaction between county and intervention slope, and this model was selected for the main results. The first analysis strategy was to use the same model as was determined as best fitting in the previous paper, and this model will henceforth be referred to as Model A. The second analysis strategy was to repeat the same model selection process and select the model that showed the best fit.

The model selection process included all the combinations of the two variable sets below:

  1. 1.

    Either intervention, intervention slope, or none of them.

  2. 2.

    Either county X intervention, county X intervention slope, or none of them.

We thus created nine models which all included year and up to two of the other four eligible independent variables. The model fit was assessed by Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), and the best fitting model was determined to be the model that included year, intervention slope, and county X intervention mean. This was the eight model of the nine models included in the model selection process, but is henceforth referred to as Model B.

The Poisson regression analyses performed in the model selection process are described in a supplementary table (table S1). The full data set is included in the supplementary material for full transparency (table S2). All analyses were performed using R 4.0.2 [9].

Description of Image

Source link