Matching in cluster randomized trials using the Goldilocks Approach

Abstract

In group or cluster-randomized trials (GRTs), matching is a technique that can be used to improve covariate balance. When baseline data are available, we suggest a strategy that can be used to achieve the desired balance between treatment and control groups across numerous potential confounding variables. This strategy minimizes the overall within-pair Mahalanobis distance; and involves iteratively: 1) making pairs that minimize the distance between pairs of clusters with respect to potentially confounding variables; 2) visually assessing the potential effects of these pairs and resulting possible randomizations; and 3) reweighting variables of selecting weights to make pairs of clusters. In step 2, we plot the between-arm differences with a parallel-coordinates plot. Investigators can compare plots of different weighting schemes to determine the one that best suits their needs prior to the actual, final, randomization. We demonstrate application of the approach with the Mupirocin-Iodophor Swap Out trial. A webapp is provided.

Keywords: Baseline covariates; Matching; Randomization; Randomized trials.

Fig. 1

Possible randomizations for 3 different sets of weights for three attributes: average monthly attributable days (Pt days), Staphylococcus aureus ICU-attributable cultures per 1000 days (S aur rate), MRSA ICU-attributable cultures per 1000 days (MRSA rate). Each light gray line represents a single randomization and the black line is the mean difference between arms. The left image has no weighting and two axes exceed maximum values. The center image is matched well on the middle axis, but the first and third have some randomization draws that would exceed the desired maximum values for the mean difference between the groups. The right image reaches a happy medium.

Fig. 2

Weighting scheme used in the Mupirocin-Iodophor Swap Out Trial. The variables are: patient days (Pt days, weight = 1), Staphylococcus aureus ICU-attributable cultures per 1000 days (S aur rate, weight = 4), MRSA ICU-attributable cultures per 1000 days (MRSA rate, weight = 2), all pathogen ICU-attributable bloodstream infections per 1000 days (All Blood, weight = 4), regional mupirocin resistance estimates (Mup R, weight = 2), percent of ICU admissions with a prior history of MRSA (Hx MRSA, weight = 1), baseline usage of mupirocin (percent of mupirocin use in the first 5 days of ICU admission (Mup Adherence, weight = 1), current usage of chlorhexidine (percent adherence to daily chlorhexidine gluconate for bathing (CHG Adherence, weight = 1), median ICU length of stay (Median LOS, weight = 3), mean Elixhauser total score (Comorbidity Score, weight = 1), percent ICU patients insured by Medicaid (Medicaid, weight = 0), whether or not a facility uses polymerase chain reactions to identify MRSA in blood (PCR Blood, weight = 0), percent admissions involving a skilled nursing facility (DC SNF), percent surgical admissions (Surgery, weight = 1), whether the ICU had specialty units for oncology, bone marrow transplant, or transplant units (OncBMTTrp, weight = 2), if the ICU has bone marrow transplant or transplant units (BMTTrp, weight = 0). Note that Median LOS has the same value for all the re-randomizations. That is, for this variable, every assignment of treatment and control within the pairs results in the same mean difference in median length of stay between the control and treatment arms. This is likely due to the very small variability of this variable. The vast majority of the hospitals had the same median length of stay.