Methodological considerations
Our diabetes outcome definition, consisting of a diagnosis of diabetes in a hospital admission or a prescription for a diabetes medication, probably did not capture some patients diagnosed in an ambulatory setting but not treated with pharmacotherapy. Ascertaining diabetes from physician claims data typically requires multiple visits over a number of months in order to remove false positives. We excluded such potential cases from our outcome definition to avoid potential bias in the rate ratio due either to exposure misclassification (had we defined such ambulatory cases on the second or third visit for diabetes) or immortal time bias (after multiple visits going back to the first visit to define the index date). Some of the excluded cases could have been selected as controls, but this should have been infrequent and therefore of little consequence given the large pool of controls.
Although our outcome definition omitted some cases, we did not expect that our definition would be different between patients prescribed higher potency or lower potency statins. Our definition probably captured patients with more serious disease because they were either admitted to hospital or required medication. Furthermore, patients included in our study were all hospitalised for an occlusive vascular event or procedure, and those patients would have likely undergone serum blood glucose testing while in hospital. While provinces and hospitals vary in how well secondary diagnoses are captured, the fact that all patients in the study were hospitalised before starting their statin should have substantially eliminated prevalent diabetes cases at cohort entry.
Rigorous definitions for identifying diabetes patients in administraive claims for physician office visits typically require multiple such visits over time, in order to rule out cases where diabetes was suspected but where the disease was not confirmed in subsequent testing. Using multiple office visits to identify ambulatory cases in an as-treated analysis would have required the introduction of immortal person time into the design,30 hence those potential cases did not meet the study outcome definition. However, diagnoses for diabetes recorded in physician office visits were still used to exclude patients from entering the secondary prevention cohorts, to minimise the risk that the patients who entered the study were not free of diabetes at baseline. More importantly, there was no diabetes diagnosis on the hospital discharge abstract for all patients who entered the study.
In administrative claims data, as in clinical practice, it is usually impossible to determine the exact timing of the onset of diabetes. As with many other diseases in observational research, the date of a first encounter with the healthcare system that coincides with the first recorded occurrence of the disease is used as a proxy for the timing of disease onset. Further, our endpoint was defined as the date of a first antidiabetic prescription in many cases, which would have required a prior diagnosis of diabetes. Given the possibility that some cases in our study might have had undiagnosed diabetes when they started taking a statin, we conducted a sensitivity analysis in which diabetes cases captured in the first 90 days of treatment were not counted. This alternative approach effectively eliminated most cases in the ≤120 day category of current cumulative statin exposure. Still, diabetes risk remained significatly elevated in patients prescribed higher potency statins in the 121-365 day exposure category. This sensitivity analysis provided some reassurance that our rate ratios were capturing a true relative association.
Even though our results are much more compatible with the maximum likelihood estimates reported in previous meta-analyses than with a null hypothesis, confounding by indication remains a possible threat to the validity of our results. To minimise this bias, our reference group consisted of patients who also received a statin. We adjusted for a broad spectrum of possible confounders using high dimensional propensity scores, which included both pre-specified and empirically identified confounders. There was probably a trend over time towards use of higher potency statins, and we therefore included calendar year of cohort entry as a covariate in our propensity score models.