Two recent cohort studies using large linked health care utilization databases have consistently found that conventional APM use increases the risk of death in elderly patients compared with atypical APMs.9,10
One of these studies has been questioned because of the lack of information in claims data on important potential confounders such as cognitive, physical, and functional impairment.11
Our analysis demonstrates that confounding bias by patient factors that are not measured in claims data will not lead to an overestimation of the increased risk of death by conventional APMs but rather suggests a moderate underestimation. This result is mainly based on the observation that users of atypical APMs have more cognitive and functional impairment than those treated with conventional APMs.
A tendency that frail elderly patients with some behavioral disturbances are more likely to use APMs and are more likely to use atypical APMs has been observed before.4
In the present study we used three patient factors, cognitive, physical, and functional impairment, as proxies for frailty in elderly patients. Applying the resulting bias estimates to a recent cohort study based on Medicare claims data on the use of conventional versus atypical APMs5
resulted in an increase in the relative risk measure from 1.37 to 1.44 (). This result demonstrates the relatively small change of the association after adjusting for 5 important additional patient factors not observed in Medicare claims data. More importantly, it underlines that the observed increased risk of death in patients initiating atypical APMs is unlikely to be explained by confounding.
Reported Associations between Conventional APM Users vs. Atypical APM Users and Death in Elderly Patients.
Our approach to assess direction and magnitude of unobserved confounding in claims data makes several simplifying assumptions. Exposure, confounder, and outcome were all coded as dichotomous variables. While this may not be of concern for the outcome of interest (mortality) and the drug use categories, for some confounders it may be an oversimplification, e.g., ADL score. Choosing alternative cut-points in confounder variables like ADL score may change the strength of an association; this is not likely if the underlying dose-response relationship is monotonic.26
Because of the “U”-shaped relationship between body mass index and mortality that would violate this assumption of monotonicity of a dose-response relationship, we computed the weighted bias average without the binary obesity indicator variable, which substantially increased the weighted bias estimate for the comparison between any APM and non-users. For our estimation of bias we assumed that the unobserved “true” drug-death association was 1. If the unknown true association is different from 1, our estimation of bias may be slightly inconsistent. However, the closer an association is to the null, the less our bias estimate will diverge from the true bias. Finally, we did not consider the joint distribution of unmeasured confounders. Instead, we computed a weighted sum of each potential confounder observed in the survey data as an approximation of the net bias. Although the extremes of this range assuming independence of individual biases are unlikely, their use will lead to a conservative interpretation of the data.25
Our bias estimates regarding cognitive and functional impairment when comparing any APM versus non-users were somewhat sensitive to the choice of the independent effect of cognitive and functional impairment on mortality as derived from the medical literature. This may not be surprising since in older patients APMs are frequently prescribed to patients with dementia and behavioral disturbances who have a higher risk of death so that any small change in the independent effect of the confounders must have influenced the estimate of bias meaningfully. The confounding bias arising in the comparison between conventional versus atypical APMs was much less sensitive to variation in the literature estimate.
Valid bias assessment depends on the survey being performed in a representative sample of the main claims data study. Given that the MCBS was designed to be representative for Medicare beneficiaries as well as the high response rate and data completeness, it is a valid and readily available source for bias assessment. Generalizability is slightly compromised by the fact that the MCBS time periods were not entirely overlapping with the studies the results were applied to, and the fact that the oldest-old were oversampled in MCBS. This analysis is limited to 5 unobserved confounders and does not address other sources of bias that may affect each of the observational studies in different ways in addition to covariates that should be routinely adjusted in claims data studies, including prior medication use, co-medications, and health services use.27
Each of the 5 unobserved confounders may have been reported with some degree of random misclassification, which would limit the ability to capture fully the confounding factor in the survey and would lead to an underestimation of bias. The survey used for this analysis was of limited size, occasionally resulting in wide confidence intervals. For the assessment of bias, the width of the confidence intervals as a measure of estimation precision has no implications as long as associations are estimated validly. Another limitation of MCBS is its cross-sectional nature that does not fully rule out the possibility that drug exposure resulted in changes of the measured characteristics and not vice versa.
Claims data studies tend to underestimate the association of conventional APMs with death compared with atypical APMs because of residual confounding by frailty measures. Studies comparing APM use with non-users may substantially overestimate harmful effects of APMs.