Perturbation Analysis: a Novel Method to Detect and Correct the Bias of Unmeasured Factors

Abstract

Randomized controlled study is the gold standard research method in biomedicine. By contrast, the validity of a (non-randomized) observational study is often threatened by unknown/unmeasured factors which have confounding and/or effect-modifying potentials. In this talk, I will propose a ‘perturbation test’ to detect the bias of unmeasured factors and a ‘perturbation adjustment’ to correct it. The perturbation analysis is an ingenious method; it circumvents the problem of measuring the unknowns by collecting the ‘perturbations’ of the unmeasured factors instead. Any variable can be counted as a perturbation variable even if it is only very weakly associated with the unmeasured factors in question. Computer simulation shows that as the number of the perturbation variables increases, the operating characteristic of the perturbation test increases (progressively up to ~1.0), and the bias after the perturbation adjustment decreases (progressively down to nearly zero). Perturbation analysis can be a valuable tool to deal with the pervasive problem of residual bias in observation studies.