Autocorrelation in Empirical Studies of Wage Determination

The need for careful specification in econometric analyses is well-known. If some component of a particular specification is incorrect, then conventional inferential procedures may be invalid. In particular, the presence of autocorrelation in linear economic models may lead to the use of inappropriate formulae for estimates of standard errors of estimated coefficients and for F-statistics which provide the bases for tests of significance. Concern over this problem has generally been represented in recent years by two checks of sensitivity, namely, the calculation of Durbin-Watson d-statistics and the use of either autoregressive transformations of the type introduced by Cochrane and Orcutt (1949) or similar approaches such as the Hildreth-Lu (1960) scan procedure, which can be associated with maximum likelihood methods. At the same time, econometric studies of wage determination provide an excellent illustration of certain specific deficiencies which may be introduced by conventional checks of sensitivity. Clarification of the quantitative significance of these deficiencies depends critically upon adequate investigation of institutional features of the labour market, and we cannot claim to have achieved this final goal. However, incompleteness of knowledge does not preclude a demonstration that current practices in empirical studies of wage-determination leave much to be desired. Two well-known studies of Phillips curves, those of Perry and Bodkin et al., provide the bases for discussion.