Alternatives for Dealing with Errors in the Variables: An Example Using Panel Data

Abstract

There are, as we know, sins both of omission and of commission. The thesis of this paper is that failure to seriously treat the problem of errors in the variables is one of the most common sins of omission in contemporary political science. It is not as though political scientists believe that all their variables are measured with perfect reliability. A large literature, well attended to, has shown that many (and often most) survey respondents give apparently random responses to the questions pollsters ask (Converse, 1964, 1970; Achen, 1975). Yet anomalously, a vast literature on voter behavior puts these same error-laden responses into models that assume no measurement error. The apparent reason for this anomaly is the assumption that errors in the variables create only false negatives. It is well known that random errors in the variables attenuate correlations such as Pearson's r. It is also well known that when a single error-laden variable is put on the right-hand side of a regression equation its associated coefficient is biased toward zero. In such cases, errors in the variables simply make it harder to reject the null hypothesis. Thus (the assumption seems to be) the problem can be lumped with others-small sample size or multicollinearity-which produce negative findings. Since negative findings typically are not taken as seriously as positive findings, one will not be badly misled, just impeded in rejecting null hypotheses. Alas, the situation is not nearly so benign. As Christopher Achen ( 1983) has reminded the profession, errors in a variable on the right-hand side of a