Packet randomized experiments for eliminating classes of confounders

Abstract

Background: Although randomization is considered essential for causal inference, it is often not possible to randomize in nutrition and obesity research. To address this, we develop a framework for an experimental design-packet randomized experiments (PREs), which improves causal inferences when randomization on a single treatment variable is not possible. This situation arises when subjects are randomly assigned to a condition (such as a new roommate) which varies in one characteristic of interest (such as weight), but also varies across many others. There has been no general discussion of this experimental design, including its strengths, limitations, and statistical properties. As such, researchers are left to develop and apply PREs on an ad hoc basis, limiting its potential to improve causal inferences among nutrition and obesity researchers.

Methods: We introduce PREs as an intermediary design between randomized controlled trials and observational studies. We review previous research that used the PRE design and describe its application in obesity-related research, including random roommate assignments, heterochronic parabiosis, and the quasi-random assignment of subjects to geographic areas. We then provide a statistical framework to control for potential packet-level confounders not accounted for by randomization.

Results: Packet randomized experiments have successfully been used to improve causal estimates of the effect of roommates, altitude, and breastfeeding on weight outcomes. When certain assumptions are met, PREs can asymptotically control for packet-level characteristics. This has the potential to statistically estimate the effect of a single treatment even when randomization to a single treatment did not occur.

Conclusions: Applying PREs to obesity-related research will improve decisions about clinical, public health, and policy actions insofar as it offers researchers new insight into cause and effect relationships among variables.

Keywords: Causal inference; experimental design; methods; multidisciplinary; statistics.