Bias in comparative effectiveness studies due to regional variation in medical practice intensity: a legitimate concern, or much ado about nothing?
Bias in comparative effectiveness studies due to regional variation in medical practice intensity: a legitimate concern, or much ado about nothing?
复制标题
DOI：
10.1161/circoutcomes.112.966093
复制
发表时间：
2012-09-01
期刊：
影响因子：
--
通讯作者：
Schneeweiss S
中科院分区：
文献类型：
--
作者：
Huybrechts KF;Seeger JD;Rothman KJ;Glynn RJ;Avorn J;Schneeweiss S
e62 Circ Cardiovasc Qual Outcomes September 2012 met and whether accounting for region using quintiles of the proposed EoLEI would alter effect estimates (implying the presence of bias). These studies derive from 2 different data sources, each incorporating regions with different medical care intensity:(1) a population of patients≥ 65 years of age enrolled in both Medicare and the Pennsylvania Pharmaceutical Assistance Contract for the Elderly (PACE) programs between 1995 and 2002 (Pennsylvania, example 1); and (2) a population of patients≥ 65 years of age enrolled in Medicare with prescription drug coverage through either a stand-alone Part D plan or a retiree drug plan in 2005 to 2007 that was administered by a pharmacy benefits management company (all US states, examples 2 and 3). All 3 analyses used an intentto-treat approach with initial exposure status as determined from pharmacy claims carried forward until censoring at the end of follow-up or the occurrence of the outcome of interest, whichever came first. All patients were new users defined as not having a prescription for the medication of interest in the prior 180 days. Baseline covariates were assessed using healthcare use data within 6 months before treatment initiation. Time to death was ascertained using Medicare claims, which are routinely cross checked with Social Security data. Myocardial infarction was defined as a hospitalization≥ 3 days (unless the subject died during the hospitalization) with a principal or secondary diagnosis of ICD-9-CM 410. x1. We implemented a high-dimensional propensity score (hdPS) algorithm to adjust for confounding using claims-based clinical covariate information. 12, 13 Healthcare use or claims data indirectly describe the health status of patients through the lenses of healthcare providers operating under the constraints of a specific healthcare system. Measuring a large battery of variables should increase the likelihood that in combination they will serve as a good proxy for relevant unobserved confounding factors. The hdPS algorithm identifies thousands of diagnoses, procedures, and pharmacy claim codes, eliminates covariates of very low prevalence, and minimal potential for causing confounding using well-established methods, 14, 15 and then uses propensity score techniques to adjust for a large number of target covariates. These newly identified covariates can result in better confounding adjustment and opportunity for valid causal inference than conventional approaches based solely on investigator-defined covariates. 13, 16, 17 Despite these advantages, it should be noted that any systematic difference between regions in intensity of healthcare use will be transferred to the hdPS as it is estimated using diagnoses and procedures documented during and medication use resulting from actual encounters with the healthcare system. Table 1 provides a basic description of the study populations. Both datasets include hospital referral regions with different medical care intensity. All 306 hospital referral regions are represented in the nationwide data (examples 2 and 3), versus only 23 regions in the regional dataset (example 1). In both of these datasets, variation in medical care intensity across regions similar to that reported by Song is seen, but the differences are less pronounced for the regional dataset. The distributions of treatments studied, however, did not differ meaningfully by region. To assess the effect of accounting for region, we mapped each patient’s zip code of residence to
