For the population aged 46 to 100 years in the study period 1996 to 2012 in the Netherlands, both the conditional effect estimate and the population averaged effect estimate indicated that being adherent to statin therapy was strongly protective against cardiovascular mortality. Including birth cohort or neighborhood SES covariates did not affect the estimate of the effect of the primary exposure on the primary outcome. The differences between the estimates from the small and the large Cox models substantially changed between time periods. Furthermore, in the simple model, the falsification end-point did not indicate bias, but in the large model there was a strong indication of bias. Confounding also appeared to differ by calendar period.
Falsification end-points and sources of bias
In this study, we avoided confounding by indication by comparing statin therapy starters amongst each other. We investigated whether healthy adherer bias affected our results by also analyzing the effect of statin adherence on falsification outcomes. In the small model, statin adherence was not protective against falsification outcomes, which means there is no indication of healthy adherer bias. However, in the large model, statin adherence did become protective against the falsification outcomes, while also becoming protective against cardiovascular mortality. A criticism of the falsification end-point approach is that the falsification end-point and the primary end-point are not necessarily affected by the same bias [
9]. However, the effect of statin adherence on both outcomes became biased in the same direction, and the relative magnitude of the bias was also the same (0.70/0.53 = 1.32 for CVD and 0.93/0.68 = 1.36 for the falsification outcomes). These indicate that both outcomes were likely affected by the same bias.
Some causal pathways from statin utilization to respiratory, endocrine and metabolic mortality are possible. We believe these pathways to be limited, and therefore to only have a minor effect on the association between statin utilization and the falsification outcomes. Nevertheless, ideally, we would have chosen deaths due to external causes (e.g. deaths as a result of accidents and physical abuse) as falsification end-points, as they are even less likely to be affected by lipid levels than the current choice. However, only ~100 events due to this cause were recorded, and hence this choice is underpowered.
Since healthy adherer bias appears to be limited, and the bias is caused by adjusting for an increased set of covariates, the source of the bias is likely either overadjustment or competing risks [
25,
26]. Overadjustment bias can be caused by conditioning on mediators or on colliders [
25]. However, none of the added variables should mediate the effect of statin adherence on cardiovascular mortality (or the falsification end-points), and none should function as colliders in this context. The bias is therefore likely caused by competing risks. Next to cardiovascular mortality (and the falsification outcomes), patients may die from a large number of other causes of death. By fitting a Cox model to data in which competing risks are present, we model the cause-specific hazard. Cause specific hazards are the hazards at time
t of a specific cause of death conditional on surviving to time
t. That is, conditional on not having died from the event under study before time
t, as well as not having died from a competing event before time
t. Therefore, the hazards of the competing cause of death affect the hazard of cardiovascular mortality. If the additional covariates in the large Cox model in this study affect the hazards of competing risks, then this also affects the hazard of cardiovascular mortality and the falsification end-points. This problem would not arise if we could model the marginal cause-specific hazard, i.e. the hazard of cardiovascular mortality where the hazards of competing causes are 0. However, the marginal cause-specific hazards are unfortunately unobservable.
We also compared bias between different calendar periods, and observed that the difference between the effect estimates of the small and the large models increased over calendar time. Overall, in the large model, the effect estimates of statin adherence on cardiovascular mortality were closer to that of clinical trials in the period prior to 2002. In the period prior to the year 2002, statins were especially indicated for patients between ages 50 to 70 years with hypercholesterolemia [
1]. Around the year 2002, important studies showed that also patients above age 70, and that diabetic patients, benefitted from statins. In the Netherlands in the year 2006, the age restrictions were formally abolished. Therefore, the patient population likely resembled the trial population more closely shortly after the introduction of statins in the population, and hence effect estimates are also more similar. Furthermore, it is possible that due to the studies and guideline changes, the patient population became more heterogeneous over time, and adjustment for potential confounders more strongly changed the effect estimate of statin therapy.
Because Cox models are non-collapsible [
22], we used the parametric G-formula (a method of direct standardization) to produce a population averaged effect estimate for the effect of statin therapy on the hazard of cardiovascular mortality. The parametric G-formula is only rarely employed [
24], but can be highly useful, as it shows the effect of a time-dependent intervention on the population level. In our study, the population-averaged estimate shows the effect on the hazard if all statin users in the population were fully adherent at all times from first dispensing onwards, compared to the situation where they were all fully non-adherent at all times. The population averaged estimate is close to the conditional effect estimate, which is likely caused by the low hazard of cardiovascular mortality (at any time point) in our sample. For this reason, we chose not to apply the parametric G-formula in subset analyses.
Birth cohort and confounding & effect modification
In this study, we conclude that non-linear birth cohort does not confound the estimates of the effect of statin adherence on the hazard of cardiovascular mortality. It may still be possible that the linear part of birth cohort confounds the outcome, however this is less problematic because age and calendar time are commonly included in analyses of drug effectiveness, and would therefore also include linear birth cohort through the dependency between the three variables [
21]. In this way, it may even be possible to model away a true birth cohort effect by using non-linear terms (including interaction effects) for age and period.
If birth cohort is (conditional on age and calendar year) related to health behavior, then this would be relevant for two reasons. First, birth cohort may affect healthy adherer bias and therefore controlling for birth cohort should result in more valid estimates of the causal effect of statin adherence on cardiovascular mortality when information on health behavior itself is unavailable. However, since we did not find evidence for confounding by birth cohort, it is less likely that birth cohort is strongly related to health behavior on the patient level. Secondly, it may mean that the effectiveness of drugs is different for different birth cohorts, because cohorts would have differences in the way they utilize drugs. Since we did not find evidence for effect modification by birth cohort, this also appears to be less likely. It could be argued that since adherence to statins and other drugs is itself an indicator of health behavior, looking at statin adherence (and adherence of other drugs) removes the effect of birth cohort on cardiovascular mortality.
Statin therapy effectiveness
Being adherent to statins appears to be protective against cardiovascular mortality, which is in line with the literature, though clinical effect may differ between subgroups (e.g. [
27-
29]). We shall here interpret the results from the small Cox model, as the larger model is known to be biased. In the small model, the population-averaged hazard ratio of statin adherence was estimated to be 0.70 (95 % CI: 0.61 to 0.81). This means that the hazard of cardiovascular mortality of a patient who is fully adherent is 30 % lower than the hazard of cardiovascular mortality in a similar patient who is completely non-adherent. This estimate is close to that of a Cochrane review of randomized clinical trials, but the confidence intervals do not overlap. In the Cochrane review, the hazard was estimated to be reduced by 17 % against fatal cardiovascular events (RR: 0.83, 95 % CI: 0.72 to 0.96). Differences may be caused by our study population being a relatively low-risk population for cardiovascular mortality (1565/49,688 * 100 % ≈ 3 % probability of CVD death).
Evaluation of data and methods
The findings of the study in regard to statin effectiveness are not directly comparable with those of earlier observational studies because the outcome definitions differed, as well as the definition of the primary exposure. To the best of our knowledge, this is the first study that used time-varying adherence to statin therapy as the primary exposure. Other studies that have related statin adherence to cardiovascular outcomes commonly calculate adherence over a fixed period, such as adherence in the first year. Using adherence in the first year is useful for predictive (and therefore clinical) purposes. However, time invariant adherence will likely be less strongly related to the outcome; a patient’s adherence in the first year should not be strongly related to his or her adherence in the 5
th year of follow up, and therefore to the hazard of mortality in the fifth or sixth year. This shows the usefulness of accounting for time-varying drug adherence [
12].
The proportion of days covered method is a method of indirect observation. It’s strength, compared to direct methods, is that direct methods are intensive and invasive, which therefore often severely limit the number of observations that can be performed [
12]. However, a limitation of indirect methods is that it cannot be observed if patients take their statin [
12]. This overestimates adherence and thereby biases statin effect estimates towards the null. However, this bias may be limited with the proportion of days covered method, given that patients who forget to take their statin are less quickly in need of new statins and therefore their adherence estimate is adjusted downwards.
More than 90 % of the patients that were censored in the study were subject to administrative censoring, which is non-informative. That is, they were still being followed when the study ended on the 31st of December 2012. The remaining number of patients were censored during the study: if this did not occur due to competing mortality, it could only occur due to patients moving out of the IADB coverage area due to the type of data sources that were used. It is unknown to what extent a move is related to impending cardiovascular mortality.
Due to data limitations, we were unable to determine the underlying health conditions of individual patients (except by using drugs dispensed as proxies for these conditions). Therefore, we could not determine which patients had familial hypercholesterolemia. However, the prevalence of this illness in the Netherlands is ~0.25 %, and therefore is unlikely to have strongly biased our estimates [
30].