Regression artifacts in time-series and longitudinal data

Abstract

The problem of “regression artifacts” in causal inference, otherwise known as the problem of error and reliable irrelevant variance in “independent” variables used for matching or regression adjustment, is illustrated first in the time-series data where a treatment is triggered by an extreme measure. The “offset effect” in psychotherapy, Underwood's scalloped learning curve, and potential pseudo-effects in AIDS therapies are used as illustrations. The magnitude of such artifacts is computable if the autocorrelation pattern for various lags is known, and thus could be distinguished from genuine effects. For longitudinal studies in which a population of respondents is repeatedly measured, the problem of anchoring the matching or regression adjustments on a single wave of measurement (usually the first) is illustrated as affected by the proximally autocorrelated nature of such measures. Data from a famous study of the effects of job training are reinterpreted in light of this consideration.