现在的位置: 首页临床统计, 临床话题>正文
[MEDSCAPE述评]:倾向性评分:用处何在?
2015年10月09日 临床统计, 临床话题 暂无评论

COMMENTARY

Propensity Scores--Here, There, and Everywhere: But Are They Useful?

Aaron B. Holley, MD

October 01, 2015

Propensity-Score-Graphic

What Are Propensity Scores?

I have been meaning to write about propensity scores for some time now. As a pulmonary and critical care physician in the United States, I dutifully read my Chest journal whenever possible. The editor sends emails to members of the ChestCollege that highlight important articles as they are published online.

This past week, two of the highlighted studies[1,2] used propensity scores to analyze their data.[3] The next email I received was from the pulmonary medicine fellowship director at my hospital. He had attached the article we will be discussing in our journal club this week. That's right, another study using a propensity score. I figured it was time to talk more about this technique.

Propensity scores are used to reduce confounding in observational studies.[4] When measuring the effect of an intervention using observational data, researchers need to "match" the patients with a population of control individuals who did not receive the intervention. Although it's possible to match for specific covariates (eg, age, race, and sex), two problems will inevitably arise: (1) If the controls are sampled from a comparable group of people who did not receive the intervention, there is a selection bias (unmeasured factor or factors that might cause clinicians to withhold the intervention and that can systemically influence group differences); and (2) depending on the size of the population being sampled, matching each patient to another using specific covariates may reduce the sample size. Only randomization can truly eliminate selection bias, but propensity matching offers a popular solution to both problems.

How Are Propensity Scores Used?

Generating a propensity score starts with identifying covariates that are associated with receiving the intervention being studied. These covariates are entered into a regression model, and a score is derived. The value of the score corresponds to the probability that a patient would receive the intervention. Then the score is used to adjust or match patients with controls before analyzing differences in the outcome under study.

We can take one of the studies published online in Chest[1] as an example. Using a large database, the investigators attempted to determine which drug—beta blockers (BBs), calcium channel blockers (CCBs), amiodarone, or digoxin—provides the best outcomes for treating atrial fibrillation (AF) in the presence of sepsis. Among other factors, treatment choice varied by year, geographic location, physician specialty, and hospital. It's not hard to imagine any one of these factors creating bias. What if internal medicine physicians use CCBs and intensivists use BBs to treat AF, but intensivists achieve better sepsis outcomes? Without some sort of matching or adjustment, one might erroneously conclude that BBs lead to better outcomes, when in fact it's the intensivist's care that is driving the difference. The same can be said about myriad other factors that drive prescribing practices and affect outcomes. Propensity scores allow researchers to eliminate all of these recognized biases. Unfortunately, only true randomization eliminates all selection bias, recognized or unrecognized. The volume of covariates used to derive the propensity score in this study is impressive, and the authors claim that their score accounts for 90% of the treatment variation seen in their data set.

Are propensity scores simply another statistical "weapon of mass destruction?" At quick glance and in my opinion, the answer is "no." They aren't nearly as "dirty" as meta-analyses can be. There is nuance, and certain areas deserve more study,[5,6] but propensity scores are powerful tools to help investigators properly interpret observational data. One need only to look at the three studies cited below and see how many covariates influence the selection of an intervention, thus creating bias, to see the value of propensity scores. I expect that we will see this statistical method often moving forward.

References

1. Walkey AJ, Evans SR, Winter MR, Benjamin EJ. Practice patterns and outcomes of treatments for atrial fibrillation during sepsis: a propensity-matched cohort study. Chest. 2015 Aug 13. [Epub ahead of print]

2. Hsu DJ, Feng M, Kothari R, Zhou H, Chen KP, Celi LA. The association between indwelling arterial catheters and mortality in hemodynamically stable patients with respiratory failure: a propensity score analysis. Chest. 2015 Aug 13. [Epub ahead of print]

3. Veluswamy R, Ezer N, Mhango G, et al. Limited resection versus lobectomy for older patients with early-stage lung cancer: Impact of histology. J Clin Oncol. 2015 Aug 3. [Epub ahead of print]

4. Rosenbaum P, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika. 1983;70:41-55.

5. Griswold ME, Localio AR, Mulrow C. Propensity score adjustment with multilevel data: setting your sites on decreasing selection bias. Ann Intern Med. 2010;152:393-395. Abstract

6. Austin PC. A critical appraisal of propensity-score matching in the medical literature between 1996 and 2003. Stat Med. 2008;27:2037-2049. Abstract

给我留言

您必须 [ 登录 ] 才能发表留言!

×
腾讯微博