Description of the two interventions
In the second quarter of 2012, Strama Västra Götaland initiated two interventions directed toward PHCCs in the Västra Götaland region. In intervention A, a Strama peer visited each healthcare center and informed physicians and nurses about appropriate antibiotic prescribing behavior and how the center prescribed compared with other PHCCs in the region. Intervention B was initiated by Strama but conducted by the healthcare centers themselves, where clinicians shared their prescribing rates for different indications and antibiotics. They also discussed how well they had succeeded in following the guidelines for antibiotic prescription and possible ways to improve. Each PHCC was asked to send a report on this self-evaluation to Strama. After doing so, the center received a reward of a fixed sum plus an adjusted amount based on the number of listed patients at the healthcare center.
The initial selection of PHCCs that would get a visit by Strama was not random; instead, centers that expressed an interest and those with high prescription levels were more likely to be targeted first. Selection later was also not randomized. At the end of 2012, about 35% of the centers had been visited. At the beginning of 2013, two things happened: more PHCCs were targeted by intervention A, and intervention B was started. As of the second quarter of 2014, 97 and 92% of the PHCCs had been targeted by interventions A and B, respectively. The interventions where never done simultaneously. However, they could have happened during the same observed quarter.
Description of the sample
We have data on prescription of all types of antibiotics and those related to treatment of respiratory tract infection at 206 PHCCs between the first quarter of 2011 and the second quarter of 2014. From the second quarter of 2011 and onwards, we also have information about characteristics of the PHCCs The main analysis is based on the balanced panel-that is, we included only healthcare centers for which we have observations for the whole-time period from the third quarter of 2011 (because of the need to include a lagged dependent variable in the analysis) to the fourth quarter of 2013.
Characteristics of PHCC
We investigated whether there were statistically significant differences between characteristics between private and public PHCC and between the last quarters of 2011 and 2013, using a t-test for continuous variables and Wilcoxon rank-sum test for discrete variables (shares).
Econometric framework
Using an econometric analysis, we investigated the effects of the two interventions on prescriptions while controlling for the relationship between characteristics of the healthcare centers and prescription levels, in general, and the difference between private and public PHCCs.
We began by investigating the effects of the two interventions on total antibiotics prescriptions per patient visits. In addition, we investigated the level of respiratory tract infection related antibiotic prescriptions per patient visits of an average PHCC in Region Västra Götaland. We focused on respiratory tract infection related antibiotic prescriptions per patient visits because there is evidence that the rate of inappropriate prescriptions is higher in respiratory tract infections than in other type of infections [25]. We used respiratory tract infection related antibiotics as a proxy for respiratory tract infections, we are aware that these antibiotics to a lesser extent are used in other infections as well.
Because the assignment of PHCCs in both interventions was to some extent based on their past levels of prescriptions, and given the panel structure of the data, we evaluated the effects of the interventions by means of linear dynamic panel-data models. These models include p lags of the dependent variable as covariates and contain unobserved panel-level effects to take account of the feedback from past prescriptions. The specification to be estimated is presented as follows:
$$ {\displaystyle \begin{array}{c}{y}_{it}={\rho y}_{i,t-1}+{\alpha}_1{T}_{1i}+{\alpha}_2{T}_{2i}+\beta {x}_{it}+{\mu}_t+{v}_i+{\varepsilon}_{it}\\ {}\left(i=1,\dots, N;t=1,\dots, T\right),\end{array}} $$
(1)
where yit denotes PHCC, i’s total antibiotics prescriptions per patient visits/respiratory infection related antibiotic prescriptions per 100 visitsFootnote 1 in quarter t, yi,t–1 is the level of prescriptions of healthcare center i in quarter t–1; T1i and T2i are intervention status indicators that are equal to 1 if the healthcare center was targeted by intervention A or B, respectively, and 0 otherwise; and xit is a vector of characteristics at both the center and patient levels, which are intended to capture the observed time-varying selection criteria used by Strama to determine the treatment status of clinics. Similarly, μt denotes quarterly dummy variables accounting for seasonal effects of antibiotic prescriptions; vi are panel-level unobserved effects that are correlated with the lagged dependent variable and take into account unobserved characteristics of healthcare centers, such as the motivation of directors to participate in the interventions; and εit is the error term. The direct effects of the interventions are estimated by the parameters α1 and α2. Standard errors include the Windmeijer’s finite-sample correction, which are suitable in presence of both heteroskedasticity and autocorrelation [26].
This equation is estimated by means of the extended system generalized method of moments estimator (GMM) derived by Arellano and Bover [27]) and Blundell and Bond [28]), an estimator that is suitable with data with few time periods and many panels. It uses extra moment conditions and lagged differences of yit as instruments for equations in levels, as well as lagged levels of yit as instruments for equations in first differences. The use of the system GMM estimator has several advantages with respect to the standard first differences GMM, including substantial efficiency gains, which makes the level restrictions informative even in the presence of weak instruments [29].Footnote 2
We estimated eq. (1) for all, public, and private centers. Moreover, we evaluated the duration of the effects by separating the total effects of the interventions into short-run effects (i.e., separate effects for the intervention quarter and the first quarter after the intervention) and long-run effects (the remaining quarters).
We then investigated the effects of the two interventions on PHCCs with similar characteristics. We were interested mainly in public and private centers, but also those that did not already have low levels of antibiotic prescription before the interventions started. We therefore also analyzed the prescribing behavior of centers when removing the centers with low levels of prescription, in our case the bottom quartile with the lowest levels. This gave us information on the effects of the interventions when the type of the PHCC matters.
Definition of respiratory tract infection related antibiotics according to anatomical therapeutic chemical classification (ATC)
Phenoxymethylpenicillin (ATC J01CE02), doxycycline (ATC J01CE02), amoxicillin (ATC J01CA04), cephalosporines (J01DB-DE), macrolides (ATC J01FA), amoxicillin with beta-lactamase inhibitor (ATC J01CR02).