We used data from administrative health databases generated in Finland through the universal health care and drug reimbursement systems covering the 5.3 million residents. We identified prescription records since 1994 in the Prescription Register, which is managed by the Social Insurance Institution of Finland (SII) [8]. This register contains records of all reimbursed prescription drug purchases made by residents in non-institutional settings. For each purchase, the dispensing date, the Anatomical Therapeutic Chemical classification code of the WHO [9], the tablet strength, and the quantity dispensed are listed. Patients staying in a public nursing home or hospital without interruption for over 90 days are not eligible for drug reimbursement, and their purchases are not registered. We identified these patients from a separate SII register. For identifying patients entitled to higher rates of reimbursement because of certain severe, chronic conditions, we used the SII Special Reimbursement Register introduced in 1964. To be eligible for special reimbursement, a patient's condition must meet explicit predefined criteria, and a written certificate by a specialist physician is required.

We identified hospitalizations from the Finnish Care Register, managed by the National Institute for Health and Welfare. The register, covering all Finnish hospitals, includes individual-level administrative data on main and additional discharge diagnoses, as well as the admission and discharge dates. The 10th revision of the International Classification of Diseases (ICD-10) has been in use since 1 January 1996. The data from the databases were linked anonymously using encrypted personal identifiers.

Between January 1998 and June 2006 in Finland, 118 623 persons initiated atorvastatin use and 180 238 began simvastatin therapy. The mean age of the atorvastatin initiators was 60.3 (SD 8.0) years and that of simvastatin initiators was 61.4 (SD 8.0) years (Table S1). The initiators of atorvastatin tended to have slightly fewer comorbid conditions than their comparison group when the persons in all of the periods were pooled (Table S1). Over the periods, however, the prevalence of prior CVD identified in the discharge register changed remarkably. In the first half of 1998, 80.8% of the initiators with atorvastatin and 67.7% with simvastatin did not have prior CVD. In the first half of 2006, the respective proportions were 83.3% and 90.97%. As shown in Figure 1, the medians of the PSs for the atorvastatin and simvastatin initiators were the closest in period 11 (the first half of 2003) (i.e., 0.47 and 0.44, respectively). The overlap of the distributions started increasing in the first half of 2001.

After restrictions according to the strength of the initiating statin and the start of the follow-up, the atorvastatin cohort included 96 995 persons, and the simvastatin cohort had 73 868 persons. The mean age of the atorvastatin initiators was 60.4 (SD 8.0) years, and that of the simvastatin initiators was 60.9 (SD 8.0) (Table 1). As in the cohort containing all of the initiators, the persons with atorvastatin tended to have slightly fewer comorbid conditions. After the PS matching across all of the cohorts combined, 128 540 persons from both groups retained; 66.2% of the restricted atorvastatin cohort and 87% of the respective simvastatin cohort. After sequential PS matching, 54 220 persons were retained in each group (55.9% of the restricted atorvastatin cohort and 73.4% of the respective simvastatin cohort, the proportions varying across the periods) (Table S2). The covariate balance increased after the restrictions and PS matching (Table 1, Table S3).

During the 256 060 person-years of the atorvastatin (10 mg) initiators followed-up since the 270^th day after the initiation until censoring or experiencing an event, 3795 CVEs were observed, yielding a crude incidence of 15/1000 person-years. For the simvastatin (20 mg) initiators, 4261 CVEs in 253 278 person-years yielded an incidence of 17/1000 person-years. The larger person-time of the atorvastatin initiators was mainly due to the larger number of deaths among simvastatin initiators (4.01% versus 4.26%).

Atorvastatin (10 mg) was associated with an 11%–12% lower incidence of CVE in comparison with simvastatin (20 mg). The hazard ratio (HR) estimated with a conventionally adjusted Cox model was the same as the one obtained in the analysis in which the PS was used for matching across all periods and the Cox model was adjusted for strong predictors of the outcome, as well as the one from the analysis in which the PS matching was applied within sequential periods: 0.88, 0.89, and 0.88 with nearly identical confidence intervals (CI) (95% CI 0.85–0.91, 0.85–0.92, and 0.84–0.92, respectively) (Table 2). However, the HR from the analysis in which PS matching was used across all of the periods without further adjustment of the outcome model for confounders produced an HR of 0.80 (95% CI 0.77–0.83).

For each period, the HRs estimated with conventional models tended to be slightly closer to one than the HRs from the models applying PS matching, with the exception of the first two periods (Table 3).

As PS is a relative sum function, the interpretation of our findings is challenging. Since the first half of 2001 (period 7), the interquartile range of the PS distributions increasingly overlapped, although a common trend for the PSs can be found. The distribution of the PSs was visually classified into the following three phases; the phase of increasing scores from 1998 to 2001, a stable phase between 2002 and 2004, and a phase of decreasing scores since 2005 (Figure 1). The start of the plateau phase coincides with the withdrawal of cerivastatin from the market [17]. The withdrawal hardly directly affects the preferential prescribing of either statin or, consequently, the PSs. The absolute number of initiators of both statins, however, decreased during the period following the withdrawal (period 8 in Table S2). This decrease possibly reflects mistrust towards the safety of statin drugs. The end of the plateau phase coincided with a substantial decrease in the prices of the generic simvastatin products [18]. Generic substitution was launched in Finland in April 2003. We assume that the price gradient between the branded atorvastatin and generic simvastatin (and other generic statins) was great enough to affect prescribing practices as late as 2005. During the last three periods (Table S2), initiations clearly shifted toward simvastatin products. A policy change restricting reimbursement for atorvastatin was implemented in October 2006; however, the fact that this change was forthcoming was not made known until the preceding June.

The effect estimate did not reach unity in any of our comparative effectiveness analyses. The strengths of atorvastatin (10 mg) and simvastatin (20 mg) may not compare equally. For feasible analyses we had, however, too few of those initiating simvastatin (40 mg) use in the first years of the study. In a systematic review [11], the percentage of reduction in serum low-density lipoprotein by simvastatin (20 mg) was 32% and that for atorvastatin (10 mg) was 37%, yielding a potency ratio of 0.86, near the HRs in our exploratory comparative effectiveness analyses.

We compared, however, the effectiveness of atorvastatin (10 mg) versus simvastatin (40 mg) in the whole cohort of initiators; that is, we estimated a PS across all periods combined and used the PS for matching within the whole cohort. In this analysis, HRs approached the unity (HR 0.94; 95% CI 0.84–1.06) (Table S4).

As the analyses were all intention-to-treat, neither switching between statins, changes in the doses, nor discontinuation of, or adherence to statin therapy was accounted for. Therefore, a misclassification of person-time was possible.

Head-to-head comparisons of the effects of atorvastatin versus simvastatin on cardiovascular outcomes are scarce. In the only large enough comparative randomized trial, 75% of the participants had pre-randomization statin therapy, which diluted the possible difference [19]. In several observational studies conducted among privately insured or employer insured populations in the US at the beginning of the 2000's, the initiators of atorvastatin (10/20 mg) use tended to have a lower risk of cardiovascular outcomes than those initiating simvastatin (20/40 mg] use during a few years of follow-up [20]–[23]. In the conventionally adjusted predictive models, the HRs ranged from 0.87 [21] to 0.98 [22]. In a study in which PS matching was applied to the whole cohort, the relative risk was 0.91 (p = 0.02) [23].

For valid effect estimates, correct specification of the PS model is essential. If the associations between the covariates and the exposure substantially differ across subgroups, applying cohort-wide PS in subgroup analyses may lead to biased estimates. In our exploration, the effect estimates from the cohort-wide PS analysis improved (i.e., approached unity) when we added strong predictors of the outcome (i.e., age, sex, prior CVD, drugs used for the modification of CVD risks, number of hospital days, and number of different drugs redeemed) to the outcome model (Table 2). Actually, the estimate became identical to the one from the sequential cohort model. Correct matching for PS, on the other hand, can be estimated by covariate balance. In our study, the balance of the key patient characteristics predicting CVEs, calculated as the standardized mean difference, improved after cohort-wide PS matching when compared with that of the unmatched study population. The balance was, however, further improved after matching within periods (Table S3). It must be noted that, due to the limitations of the register data, we did not have access to information on many important risk factors for CVEs. We did not have data on lipid profiles, family history of CVD, lifestyle-related risk factors nor on the use of acetylsalicylic acid of the study population.

We reached the same results with the conventional regression model applying standard adjustments as we did with the period-stratified (sequential cohort) model and the cohort-wide PS matched model adjusted for strong predictors of the outcome. Our period-stratified model could probably not discriminate between the characteristics of the atorvastatin and simvastatin initiators and other factors associated with calendar time. Although period was a weak confounder, we did, however, find prescribing dynamics over time as the odds ratios of some of the covariates (number of hospital days, prior CVD, diabetes) varied across the PS models. Furthermore, omitting period from the conventional Cox model did not affect the HR of the outcome. However, a traditional PS matched analysis produced an effect estimate that was the furthest from unity.

In a recent analysis of the effect estimates of cancer chemotherapies that applied a calendar time-specific PS approach [3], a clear confounding effect of the calendar time was found. The definition of the periods was based on drug policy events. Policy events may substantially affect the cost of relatively expensive cancer drugs and thus modify the selection of a therapy. In our study, the length of the accrual periods was based on calendar time, and only one substantial policy change took place during the study period. Although we could identify channeling of the atorvastatin and simvastatin after the policy event, calendar time did not act as a strong confounder in our comparative effectiveness analyses.

This study expands upon the previous work of Rassen et al. [24], who compared the effect estimates for subgroups from the outcome models by applying a PS estimated for a full cohort and also within the subgroups. They observed practically the same effect estimates for the subgroups in the subgroup-specific PS models as in the full cohort PS models. The model for the full cohort PS included interaction terms between the subgroups and strong confounders. When the subgroups were large enough, the validity of the full cohort PS approach was not threatened, and it was more efficient than the stratified analyses. The calendar-time-specific PS matched approach in an analysis of the effectiveness of second-generation antipsychotics on cardiovascular outcomes yielded the same effect estimates as the conventional PS matched approach [4]. Furthermore, preliminary results on the effectiveness of inhaled, long-acting beta-agonists on asthma exacerbations [25] showed that the effect estimates were the same whether or not they were derived from the model using a PS estimated for the entire study period or a PS estimated for specific years and used for matching in the models.