Super learning to hedge against incorrect inference from arbitrary parametric assumptions in marginal structural modeling

详细信息    查看全文

• 作者：Romain Neugebauer ; Bruce Fireman ; Jason A. Roy ; Marsha A. Raebel ; Gregory A. Nichols ; Patrick J. O'Connor
• 关键词：Super learning ; Marginal structural model ; Inverse probability weighting ; Comparative effectiveness research ; Time-dependent confounding ; Selection bias
• 刊名：Journal of Clinical Epidemiology
• 出版年：2013
• 出版时间：August, 2013
• 年：2013
• 卷：66
• 期：8_supp_S
• 页码：S99-S109
• 全文大小：628 K

文摘

<h4 class=""h4"">Objectiveh4>Clinical trials are unlikely to ever be launched for many comparative effectiveness research (CER) questions. Inferences from hypothetical randomized trials may however be emulated with marginal structural modeling (MSM) using observational data, but success in adjusting for time-dependent confounding and selection bias typically relies on parametric modeling assumptions. If these assumptions are violated, inferences from MSM may be inaccurate. In this article, we motivate the application of a data-adaptive estimation approach called super learning (SL) to avoid reliance on arbitrary parametric assumptions in CER.<h4 class=""h4"">Study Design and Settingh4>

Using the electronic health records data from adults with new-onset type 2 diabetes, we implemented MSM with inverse probability weighting (IPW) estimation to evaluate the effect of three oral antidiabetic therapies on the worsening of glomerular filtration rate.<h4 class=""h4"">Resultsh4>

Inferences from IPW estimation were noticeably sensitive to the parametric assumptions about the associations between both the exposure and censoring processes and the main suspected source of confounding, that is, time-dependent measurements of hemoglobin A1c. SL was successfully implemented to harness flexible confounding and selection bias adjustment from existing machine learning algorithms.<h4 class=""h4"">Conclusionh4>

Erroneous IPW inference about clinical effectiveness because of arbitrary and incorrect modeling decisions may be avoided with SL.