Using balance statistics to determine the optimal number of controls in matching studies

Journal of Evaluation in Clinical Practice - Tập 19 Số 5 - Trang 968-975 - 2013
Ariel Linden1,2, Steven J. Samuels3
1Department of Health Policy & Management School of Public Health University of Michigan Ann Arbor Michigan USA
2Linden Consulting Group Ann Arbor Michigan USA
3Department of Epidemiology & Biostatistics School of Public Health State University of New York Albany New York USA

Tóm tắt

AbstractWhen a randomized controlled trial is not feasible, investigators typically turn to matching techniques as an alternative approach to evaluate the effectiveness of health care interventions. Matching studies are designed to minimize imbalances on measured pre‐intervention characteristics, thereby reducing bias in estimates of treatment effects. Generally, a matching ratio up to 4:1 (control to treatment) elicits the lowest bias. However, when matching techniques are used in prospective studies, investigators try to maximize the number of controls matched to each treated individual to increase the likelihood that a sufficient sample size will remain after attrition. In this paper, we describe a systematic approach to managing the trade‐off between minimizing bias and maximizing matched sample size. Our approach includes the following three steps: (1) run the desired matching algorithm, starting with 1:1 (one control to one treated individual) matching and iterating until the maximum desired number of potential controls per treated subject is reached; (2) for each iteration, test for covariate balance; and (3) generate numeric summaries and graphical plots of the balance statistics across all iterations in order to determine the optimal solution. We demonstrate the implementation of this approach with data from a medical home pilot programme and with a simulation study of populations of 100 000 in which 1000 individuals receive the intervention. We advocate undertaking this methodical approach in matching studies to ensure that the optimal matching solution is identified. Doing so will raise the overall quality of the literature and increase the likelihood of identifying effective interventions.

Từ khóa


Tài liệu tham khảo

10.2307/2529684

Abadie A.&Imbens G.(2002)Simple and bias‐corrected matching estimators for average treatment effects. Technical Working Paper T0283: NBER.

10.1093/aje/kwq224

10.1002/pds.3263

10.1111/1467-9531.271030

10.1177/0163278710376824

10.1214/09-STS313

10.1111/j.1467-6419.2007.00527.x

10.1002/sim.3697

Chambers J. M., 1983, Graphical Methods for Data Analysis

10.1177/1536867X0500500210

10.2307/2684560

10.1016/S0895-4356(00)00321-8

10.1023/A:1020363010465

Cohen J., 1988, Statistical Power Analysis for the Behavioral Sciences

10.1111/j.0006-341X.2000.00118.x

10.1002/pds.1674

10.1111/j.1365-2753.2010.01611.x

10.1016/j.jtcvs.2007.07.021

10.1002/sim.3150

10.1016/j.jclinepi.2005.07.004

10.1093/biomet/70.1.41

Lunt M.(2007)FGMATCH: stata module to perform 1:k greedy matching. Available at:http://personalpages.manchester.ac.uk/staff/mark.lunt/fgmatch.ado(last accessed 17 July 2013).

Parsons L. S.(2004)Performing a 1:N case‐control match on the propensity score.Proceedings of the Twenty‐Ninth Annual SAS® User Group International Conference Montreal Canada: SAS Institute Inc. Available at:http://www2.sas.com/proceedings/sugi29/165‐29.pdf(last accessed 17 July 2013).

Lunt M.(2007)PBALCHK: stata module to check covariate balance after matching weighting or stratifying. Available at:http://personalpages.manchester.ac.uk/staff/mark.lunt/pbalchk.ado(last accessed 17 July 2013).

10.1002/bimj.200810488

10.1111/j.1467-9531.2008.00204.x

10.1198/jasa.2011.tm09599

Linden A., 2006, What will it take for disease management to demonstrate a return on investment? New perspectives on an old theme, American Journal of Managed Care, 12, 217

10.2307/2529685

Thông tin
Thông tin xuất bản

Journal of Evaluation in Clinical Practice

Tập 19 Số 5

968-975

Thông tin tác giả