A Catch-22—the Test–Retest Method of Reliability Estimation

Abstract

This article addresses the problems with the traditional reinterview approach to estimating the reliability of survey measures. Using data from three reinterview (or panel) studies conducted by the General Social Survey, we investigate the differences between the two-wave correlational approach embodied by the traditional reinterview strategy, compared to estimates of reliability that take the stability of traits into account based on a three-wave model. Our results indicate that the problems identified with the two-wave correlational approach reflect a kind of “Catch-22” in the sense that the only solution to the problem is denied by the approach itself. Specifically, we show that the correctly specified two-wave model, which includes the potential for true change in the latent variable, is underidentified, and thus, unless one is willing to make some potentially risky assumptions, reliability parameters are not estimable. This article compares the two-wave correlational approach to an alternative model for estimating reliability, Heise’s estimates based on the three-wave simplex model. Using three waves of data from the GSS panels, which were separated by 2-year intervals between waves, this article examines the conditions under which the wave-1, wave-2 correlations which do not take stability into account approximate the reliability estimate obtained from three-wave simplex models that do take stability into account. The results lead to the conclusion that the differences between estimates depend on the stability and/or fixed nature of the underlying processes involved. Few if any differences are identified when traits are fixed or highly stable, but for traits involving changes in the underlying traits the differences can be quite large, and thus, we argue for the superiority of reinterview designs that involve more than 2 waves in the estimation of reliability parameters.