Matching With Doses in an Observational Study of a Media Campaign Against Drug Abuse

Abstract

Multivariate matching with doses of treatment differs from the treatment-control matching in three ways. First, pairs must not only balance covariates, but also must differ markedly in dose. Second, any two subjects may be paired, so that the matching is nonbipartite, and different algorithms are required. Finally, a propensity score with doses must be used in place of the conventional propensity score. We illustrate multivariate matching with doses using pilot data from a media campaign against drug abuse. The media campaign is intended to change attitudes and intentions related to illegal drugs, and the evaluation compares stated intentions among ostensibly comparable teens who reported markedly different exposures to the media campaign. KeywordsCoherent signed rank test, Equal percent bias reducing, Matching with doses, Nonbipartite matching, Observational studies, Ordinal logit model, Optimal matching, Propensity score Disciplines Communication | Statistics and ProbabilityThis journal article is available at ScholarlyCommons: http://repository.upenn.edu/asc_papers/177 Matching With Doses in an Observational Study of a Media Campaign Against Drug AbuseBo Lu, Elaine Zanutto, Robert Hornik, Paul R. RosenbaumMultivariate matching with doses of treatment differs from the treatment-control matching in three ways. First, pairs must not only balance covariates, but also must differ markedly in dose. Second, any two subjects may be paired, so that the matching is nonbipartite, and different algorithms are required. Finally, a propensity score with doses must be used in place of the conventional propensity score. We illustrate multivariate matching with doses using pilot data from a media campaign against drug abuse. The media campaign is intended to change attitudes and intentions related to illegal drugs, and the evaluation compares stated intentions among ostensibly comparable teens who reported markedly different exposures to the media campaign.

Causal Inference without Balance Checking: Coarsened Exact Matching

“…This means that CEM works without modification for multicategory treatments: after the algorithm is applied, keep every stratum that contains all desired values of the treatment variable and discard the rest. This is a simple approach that can be easily used with or in place of more complicated approaches, such as based on generalizations of the propensity score (Imbens 2000;Lu et al 2001;Imai and van Dyk 2004).…”

Best Practices in Quasi–Experimental Designs: Matching Methods for Causal Inference

“…This method uses a model such as an ordinal logit model to match on a linear combination of the covariates. This is illustrated in Lu, Zanutto, Hornik, and Rosenbaum (2001), where matching is used to form pairs that balance covariates but differ markedly in dose of treatment received. This differs from the standard matching setting in that there are not clear "treatment" and "control" groups, and thus any two subjects could conceivably be paired.…”

On the use of discrete choice models for causal inference

“…The causal e ect,ˆ t , is estimated using the matching estimator described in equation (5). The structure of the matches were 1 to 1 and only considered matches if close, e.g.…”

“…Rosenbaum and Rubin [1] demonstrated that balancing on propensity scores removes bias due to all observed confounders. Only recently have propensity score methods been extended to multiple treatment settings [2][3][4][5][6]. While these extensions are promising, the choice of analytic model for the treatment assignment mechanism as well as its subsequent impact on the causal estimator remains unclear.…”