In this study of 82,012 older patients initiating use of antipsychotic medication after admission to a nursing home, those for whom conventional agents were prescribed had a 34% higher risk of death in the short term than those for whom atypical agents were prescribed, which corresponds to an additional 7.0 deaths per 100 patients treated with conventional agents. This finding of an increased risk with conventional antipsychotic agents versus atypical agents is consistent with findings from earlier studies in both predominantly community-dwelling cohorts (12
) and long-term-care cohorts (13
Nonrandomized studies using health-care utilization data are particularly scrutinized for their limited control of confounders and their potential for misclassifying diagnoses (16
). In this study context, concern has been raised that patients who are frail and at increased risk of death might be more likely to be prescribed a conventional antipsychotic agent than an atypical agent, resulting in overestimation of the association (18
). However, an earlier study that assessed residual confounding using survey data found that correction for 5 factors that are unmeasured in claims data (i.e., body mass index, smoking, Activities of Daily Living score, cognitive impairment, and physical impairment) resulted in a stronger association rather than a weaker association (35
Our study population consisted of patients eligible for Medicaid. While this might have affected the rate of usage of the cheaper conventional antipsychotic medications, this restriction should not have affected the validity of our findings, since the central issue determining internal validity is comparability between the subcohorts included. As long as socioeconomic status and its correlates do not modify the association between antipsychotic medication and short-term mortality, the findings should also be generalizable (i.e., externally valid) (36
Confounder information derived from health-care utilization or claims data was supplemented with clinical assessment data as recorded in the MDS. The MDS, which is part of the US federally mandated process for clinical assessment of all residents in Medicare- or Medicaid-certified nursing homes, provides a comprehensive assessment of each resident’s functional capabilities and helps nursing home staff identify health problems. The data are reported by the nursing homes themselves and are reviewed by nursing home inspectors but are not formally checked to ensure accuracy, yet most elements of the MDS have been shown to demonstrate good reliability (37
). As potential confounders of the association between antipsychotic medication and all-cause mortality, MDS markers of ill health, such as severity of physical and cognitive impairment, were of particular interest in this study. In addition, it has been suggested that nursing home patients may be at increased risk of death simply by being admitted to a facility with poor quality indicators, such as a higher intensity of antipsychotic medication use (39
). We accounted for potential nursing home quality indicators through use of the OSCAR database, which reflects the findings from state inspections and complaint investigations. Adjustment for these measured clinical and nursing home characteristics, however, did not meaningfully affect the effect estimates.
We used hdPS techniques in an effort to mitigate residual confounding by unobserved factors. The assumption underlying this approach is that health-care utilization or claims data can be viewed as a set of proxies that indirectly describe the health status of patients through the lenses of health-care providers and coders operating under the constraints of a specific health-care system. By measuring a large battery of proxy variables, this approach aims to increase the likelihood that in combination they serve as a good proxy for relevant unobserved confounding factors (30
In contrast to “kitchen sink” models, which indiscriminately include covariates in a PS model, the hdPS algorithm implements current knowledge regarding appropriate variable selection for PS models (e.g., variables are selected on the basis of their potential for confounding, and potential instruments are removed after review of the univariate associations between the selected variables and the exposure and outcome) (25
). In addition, by considering a multitude of variables addressing the same construct of a confounding factor, the hdPS approach also reduces the potential impact on confounding adjustment of “missing” claims data, that is, those that were not observed by the physician or not recorded by the system. Since it has been found that inclusion of interactions leaves the hdPS-adjusted effect estimates largely unchanged (30
), we did not include them in our models. By design, all empirical covariates are categorical, and therefore nonlinearity is not a concern when implementing the algorithm.
In some typical pharmacoepidemiologic studies of treatment effects, proxy adjustment via hdPS generated effect estimates closer to randomized trial findings than standard covariate adjustment of investigator-predefined variables (30
). In our study, use of hdPS changed the absolute effect from 7.79 fewer deaths per 100 persons over 180 days to 7.00 per 100 persons and the relative effect from 1.36 to 1.34. Although we have no gold standard against which to evaluate our results, the monotonic trend of a reduction in effect size with increasing levels of adjustment (see ) hints at improved confounding control with each level of adjustment. These findings suggest that there may be some confounding caused by clinical characteristics other than the measures reported in the MDS. Alternatively, the MDS variables may have been reported with sufficient nondifferential misclassification to limit the ability to control fully for the confounding.
IV estimation, which by design can control for unmeasured patient characteristics, appeared to confirm the findings of an increased mortality risk with conventional antipsychotic agents, but the instruments were too weak to permit interpretation of the estimated risk differences (40
The sensitivity analyses demonstrated that very strong mortality risk factors that are fairly imbalanced among exposure groups must be unmeasured and uncontrolled in our study to explain the observed association. The results have already been adjusted for most known, strong, independent mortality risk factors, and any unmeasured confounder of the required strength would also have to be independent of the confounders we adjusted for; that is, correlated confounders such as patient vulnerability are to some extent adjusted by factors like cognitive and functional impairment and awake time. Although it is unlikely that we missed such a strong single confounder, it is conceivable that several weaker confounders may have acted together and explain the apparent effect.
In summary, findings from nonrandomized studies assessing the safety of conventional antipsychotic agents versus atypical antipsychotic agents have been criticized for being caused by residual confounding by unmeasured factors channeling the prescribing of conventional agents to patients at higher risk of death. Use of different analytic techniques to mitigate the effects of confounding by measured factors and, to a lesser degree, unmeasured factors consistently decreased but did not eliminate the observed association in a population of nursing home patients. In this study, the addition of clinical characteristics indicative of frailty and typically not available in administrative claims data did not meaningfully affect the effect estimates.
While confirmation of these findings in other empirical examples is in order, our results suggest that hdPS adjustment based on claims data—which, in contrast to conventional confounder adjustment methods, does not rely on a limited number of investigator-specified covariates but taps into layers of information that investigators have until recently not exploited—may provide more valid estimates than conventional approaches using claims data enriched with clinical information, often believed to include important confounding variables. Given that plausible IVs have been difficult to find in epidemiology and medicine (including in this study) (31
), the hdPS approach may offer a promising and practical alternative to enhance the causal interpretation of effect estimates when residual confounding by indication is a concern.