A Causal Approach to Nonrandom Measurement Errors

The purpose of this paper is to examine several specific kinds of nonrandom measurement errors and to note their implications for causal model construction. In doing so, my secondary purpose is to sensitize the reader to the crucial importance of making one's assumptions fully explicit and to the advantages of a causal models approach to measurement errors. It is well known that the presence of even random measurement errors can produce serious distortions in our estimates, particularly whenever one is attempting to assess the relative contributions of intercorrelated independent variables. Nevertheless, common practice is to utilize what Duncan refers to as the naive approach to the presence of measurement errors: that of acknowledging the existence of measurement errors, and even discussing possible sources of such errors, while completely ignoring them in the analysis stage of the research process. That is, measured values are inserted directly into causal models as though they adequately reflect the true values. It can easily be shown that such a practice, while leading to important simplifications, can readily lead one astray. In particular, it may blind the analyst to searching for alternative plausible explanations that allow for measurement error. There have been a number of very recent papers in the sociological literature, some of which will be briefly summarized since they may not be familiar to the reader. For the most part, these papers have dealt rather systematically with ways to handle random measurement errors, whereas nonrandom errors have been dealt with only incidentally and much less carefully.