SMALL SAMPLE PROPERTIES OF THE SPEARMAN AUTOCORRELATION ESTIMATOR

The small sample properties of a variant of the Spearman rank correlat ion coefficient applied in the time-series context were investigated t hrough Monte Carlo methods. The rank method (r1S) has even greater bia s than the highly biased conventional parametric procedure; a traditio nal test of H-0: rho1 = 0 based on (r1S) yields unacceptable propertie s. Empirical small sample distributions associated with the rank coeff icient differ markedly from the distributions predicted by asymptotic theory. It is concluded that neither rank nor conventional parametric estimators and hypothesis tests are appropriate for very small samples in applications of time-series analysis that have been recommended in the behavioral and social science literature.