Evaluating the impact of the 2010 Swedish choice reform in primary health care on avoidable hospitalization and socioeconomic inequities: an interrupted time series analysis using register data | BMC Health Services Research

Study design

This study used a multiple-group interrupted time-series (ITS) design based on individual-level data. The ITS design is a quasi-experimental evaluation design that is considered the strongest option for causal inference when randomization of subjects is not an option, and is particularly suitable for evaluation of population-level changes that occur at a specific time point, such as health reforms, if time series data is available [39, 40]. The design takes into account underlying (e.g., secular) trends of the outcome, which are not attributed to the intervention and therefore risk introducing bias the effect estimates. This is done by comparison of trends of the outcome during a period after the intervention to corresponding trends period before the intervention (rather than single observations before and after the intervention). The ITS design is commonly conducted within a single population (single-group, or uncontrolled, ITS), but can be extended to incorporate a comparison group (multiple-group, or controlled, ITS) [41, 42], which provides further control of bias from competing interventions or events occurring close to the intervention under study. The most common application is on aggregated (e.g., country-level) data, but it can also be conducted on individual-level data [42], which enables control for individual-level time-varying confounders.

The study period was 2001–2017. The choice reform was introduced nationally on Jan 1st, 2010, and the study period was divided into pre-reform period (2001–2009; 10 yearly observations), and a post-reform period (2010–2017, 8 yearly observations). Three comparison groups were constructed to capture region-level increase in private PHC provision occurring after the choice reform, following the categorization of our previous report [36]. The classification was based on public statistics from the Swedish Association of Local Authorities and Regions on the number of public and private health care centers per region and year 2009–2016.While the reform was introduced nationally in 2010, the categorization into comparison groups capitalizes on between-region heterogeneity when it comes to the establishment of new private health centers, as a central indicator of the de facto implementation of the reform. First, the proportion of private health care centers (private/total) were calculated for each region, in the year before the reform (2009 used as a baseline) and the years following the introduction of the reform (averaged across 2010–2016) [13]. Second, all regions were ranked by the absolute and relative change in proportion of private providers from before to after the reform. Lastly, the regions were classified according to tertiles, in order to achieve balanced comparison groups, and thereby minimize the risk for individual regions disproportionally affecting the group estimates. The seven regions with greatest increase in the proportion of private health care centers (> 10% absolute increase and > 60% relative increase) were categorized as high (regions of Uppsala, Södermanland, Jönköping, Kronoberg, Västra Götaland, Värmland, Dalarna); the seven counties with the smallest increase (or decrease) in the proportion (< 6% absolute increase and < 15% relative increase) was categorized as low (regions of Kalmar, Gotland, Blekinge, Halland, Örebro, Västmanland, Jämtland); and seven regions comprising the middle tertile were categorized as moderate increase of privatization (Stockholm, Östergötland, Skåne, Gävleborg, Västernorrland, Västerbotten, Norrbotten). The low group was used as the reference group in all analyses.

Study population and data

The study population included all residents in Sweden aged 18–85 years each year 2001–2017, in total comprising N = 125,438,725 observations. To facilitate the computational challenges with such large data, a random sample of 1 million individuals were drawn annually from each of the three comparison groups, resulting in an analytical sample of 51,000,000 observations uniformly distributed across three comparison groups and the 17 study years (2001–2017).

Individual-level data on the study population was retrieved for each year over the study period. The data sources were multiple registers with total population coverage, managed by Swedish public authorities. Data on hospitalizations was retrieved from the National inpatient register of The National Board of Health and Welfare, and all socioeconomic and demographic information from the Longitudinal integrated database for health insurance and labour market studies (LISA) of Statistics Sweden. All data was individually linked by the unique Swedish Personal Identity Number.

Variables

Performance outcome and socioeconomic indicators

The outcome variable comprised avoidable hospitalizations, corresponding to hospitalization due to ambulatory care sensitive conditions (ACSC), following the classification of ACSC diagnoses by The National Board of Health and Welfare [43, 44]. It was operationalized as a binary variable per year (0 = no ACSC hospitalizations; 1 = at least one ACSC hospitalization).

To estimate the bivariate phenomenon of socioeconomic inequities in ACSC hospitalizations, the ACSC hospitalization outcome was used in combination with two complementary socioeconomic indicators (procedures described in Statistical analysis). Education was based on the highest formal educational degree and classified into five levels according to Statistics Sweden’s classification SUN2000 [45] (no or basic education; primary education; secondary education; basic tertiary education (less than three years); advanced tertiary education (three years or more)). Income was based on disposable annual household income weighted by family composition and was divided into quintiles of the annual income distribution (quintile 1 = poorest to quintile 5 = richest).

Potential confounders

Several covariates were operationalized in order to further control for potential confounding. First, as regional patient choice models were implemented ahead of the 2010 national reform in eight regions (Halland, Västmanland, Stockholm, Uppsala, Kronoberg, Skåne, Östergötland, Västra Götaland [46]) and this could potentially influence the subsequent impact of the 2010 national reform, a variable indicating Early implementation was created by grouping region of residence by timing of implementation (0 = region implemented 2010, 1 = region implemented before 2010). Three early implementation regions each belonged to the high and mid implementation comparison groups, with two in the low implementation comparison group.

Second, as the ITS design relies on comparison over time and geographical regions, demographic developments and composition of the regional populations could potentially confound the results. Age was measured in years and grouped into young adulthood (18–35 years), mid-adulthood (36–64 years), and old adulthood (65–85 years); Gender as indicated by legal sex (woman or man); and Country of birth coded as Nordic countries, other high-income country (HIC), or Low- or middle-income country (LMIC). The above variables were considered as potential confounders for all analyses.

Additionally, the following confounders were identified for analyses considering population-average ACSC risk, but were not included in the analyses of inequities in ACSC to avoid the risk of overadjustment for potential mediators. Labor market position was based on the main source of income each year, with ten categories: employed; studying; care of child/close one; sickness benefits; unemployed; early retirement; social benefits; labor market program; age retirement; and no income. As a measure of Urbanicity, municipality of residence was classified into rural, mixed urban/rural, and urban [47]. Education and income, operationalized as above, were also included as covariates in the analyses of ACSC hospitalization rates only.

Statistical analysis

Descriptive statistics are reported as percentages (%). Intermediate analyses comprised estimation of ACSC hospitalizations rates as well as education- and income-related inequities in ACSC hospitalizations, by period (pre-reform and post-reform collapsed within period) and by comparison group (low, moderate, and high implementation). All analyses were performed on the individual-level sample of 51,000,000 individuals.

For the main analyses, we conducted a series of multiple-group interrupted time series analyses (ITSA) based on individual-level data [42], using generalized linear model (glm) with binomial family and log link for estimation of relative risks. In all analyses, ACSC hospitalizations was used as the dependent variable, and the low implementation group was used as the reference, with the moderate and high groups as intervention groups.

To examine the reform impact on population-average avoidable hospitalizations (Aim 1) a multiple-group ITSA with ACSC hospitalizations as the outcome was run. The glm model was as follows:

where Y_t is the outcome (annual ACSC hospitalizations), T represents time (year, 0 = 2001, 1 = 2002 … 17 = 2017), X represents period (0 = pre-reform period 2001–2009; 1 = post-reform period 2010–2017); and Z refers to a dummy variable with three categories denoting the comparison group (0 = Z₀ = reference group, omitted; 1 = Z₁ = middle group; 2 = Z₂ = high group), and their interaction effects. Here, the coefficient for TtXtZ₁ (mid vs. low) and TtXtZ₂ (high vs. low) effects, i.e., the Time*Group*Period effects, are of main interest. The corresponding estimate (β₇) tests whether the slope of the post-reform trends in ACSC hospitalizations differed between the comparison groups, taking into account the corresponding pre-reform trends, and thus represents the impact of the reform on ACSC trends. Two models were run, one crude (Model 1, shown in Eq. 1, above) as well as one model additionally adjusted for age, sex, country of birth, early implementation, labor market position, urbanicity, education, income (Model 2).

To examine the reform impact on inequities in ACSC hospitalizations (Aim 2), we performed a novel extension of the ITSA. To quantify the magnitude of inequities, we estimated the ‘Relative Index of Inequality’ (RII), which is a standardized measure of relative inequalities capturing the social gradient in an outcome [48, 49]. Note that while the present study focuses on inequities (inequalities that are avoidable and unfair), we will refer to the measure as ‘Relative Index of Inequality’, as that is its most common designation. The RII can be interpreted as the relative risk moving from the theoretically most favorable social position (0) to the most disadvantaged position (1). The relative size of the social categories is taken into account by first transforming the socioeconomic indicators into a ridit score, which uses the mid-point of the cumulative proportion of each socioeconomic category along educational level and income quintile, respectively [49]. The ridit scores were reverse coded so that a higher RII indicates a steeper social gradient in health (larger magnitude of the health inequality or inequity).

Interrupted time series analyses are conventionally used to estimate population-average impact on an outcome, as done for the first aim (Eq. 1), rather than to estimate the impact on inequities. To enable estimation of the RII within the ITS framework, we used the fact that the RII can be estimated as a relative risk in regression models. We extended the basic ITSA model (Eq. 1) to also incorporate ridit main and all interaction effects, with separate models for education and income, respectively. Specifically, we extended the ITSA model (Eq. 1) by adding the ridit as a main effect, as well as all 2-, 3- and 4-way interaction effects, as per the following model:

This extended model (Eq. 2) thus uses the same Y_t (annual ACSC hospitalizations) and contains all effects included in the model for the first aim (Eq. 1), but with the addition of ridit main and interaction effects. In this model, all effects (βs) that include the ridit can be interpreted analogously to the corresponding ITS effects without the ridit term, but instead reflecting the relative change of RII rather than of the risk of the outcome itself. As an example, and most importantly, the coefficient for the 4-way interaction term TtXtZ×ridit_t is of main interest as it represents whether the slope of the post-reform trends in RII differed between the mid and low comparison groups (β₁₄TtXtZ₁), and high and low comparison groups (β₁₄TtXtZ₂), accounting for the pre-reform trends. This estimate thus reflects the impact of the reform on the trends in inequities in avoidable hospitalizations. Two models were run for education- and income-related, respectively, including a crude model (Model 1, shown in Eq. 2), above), and a model additionally adjusted for age, sex, country of birth, and early implementation (Model 2).

As a recent study has reported that the trends of ACSC in Stockholm, Sweden, have developed in a less favorable direction for older adults [34], auxiliary analyses comprised rerunning all analyses specifically for the oldest age group (aged 65–85 years). The inferences from these auxiliary analyses were however identical to the analyses on the total sample (data available on request).