Near-Far Matching for comparative effectiveness research

In comparative effectiveness research that attempts to estimate treatment effectiveness from observational data, the main methodological concern is confounding by indication. While Instrumental Variable (IV) methods can handle confounding due to differences in unobserved, as well as observed covariates, in many settings, the IV is only weakly predictive of the treatment assignment. Such weak instruments can lead to both inferential errors as well as bias. As an alternative to traditional IV approaches, we consider near-far matching, recently proposed in the statistical, clinical and health services research literatures for applications with weak instruments.

Use of Near-Far matching algorithms results in a matched paired study design, and has two goals: the first is so strengthen the instrument, and the second is to balance observable characteristics. To strengthen the IV, each matched pair is selected so that the individuals are distant according to the level of the IV, analogous to a well-conducted RCT where there is a substantive difference in the treatment assigned, according to randomization. Hence units that are liable to weaken the instrument are discarded, analogous to applying exclusion criteria prior to randomization. Simultaneously, matches are also selected so as to balance observed characteristics between the treatment and control groups, from the subsample of units who meet the criteria for being far apart, according to the IV.

We contrast near-far matching to the 2-stage residual inclusion approach to IV estimation for binary outcomes. We consider these approaches in a case study from the UK that attempts to identify the causal effect of reduced time to Intensive Care Unit (ICU) admission on 90-day mortality, for patients offered ICU admission. Here, the concern is that those patients admitted promptly to ICU were more acutely ill, than those whose admission was delayed, according to observed but also unobserved prognostic variables. To address the potential for unobserved confounding, we consider as an IV, the occupancy level in the ICU, as this may predict the time to ICU admission, without having an independent effect on mortality, and has previously been proposed as an instrument for the time to ICU admission. The initial results of applying the 2-stage residual inclusion approach suggest that prompt ICU admission reduced mortality. However, the occupancy level is only weakly predictive of the time to ICU (Fstatistic <10), and the estimated effect on mortality had wide 95% CI. 

Motivated by the potential for weak instrument bias, we then considered near-far matching to strengthen the IV. We matched units that were ‘far’ apart according to the occupancy level in the ICU, and within this subsample, we undertook optimal matching to balance patient-level (for example, age, gender, diagnosis, acute physiology score, season of admission), and centre-level characteristics (e.g. teaching hospital status). We conclude by discussing the implications for inference of limiting the analysis sample, and the potential for applying near-far matching in other settings for strengthening IV estimators.