How Reliable The Measurement of Predictors Should Be: Monte Carlo Study in the Regression Framework
Current study investigate the effect of measurement error on the estimation of predictors that are measured either with or without error, type 1 error rates, and power to detect non-null parameters. Author also looked for minimum value of measurement reliability needed for the analysis to provide desired results. Such value used to be based only on subjective judgements without any empirical study to support them. Simulation was conducted by manipulating the reliability of one predictor, the sample sizes, and the correlation between predictors. The model used in current study included only two predictors. The results showed that the higher the reliability of predictor measurement, the lower the bias of estimates of the two predictors and the type 1 error rates. Increasing reliability was also followed by increased power. Author also demonstrated that the minimum reliability to achieve desired results should be .8 to .9, not .7 as suggested by others.