THIS PAPER IS INTENDED to stress the distinction between the conditions for lack of identification in models linear with respect to the variables but nonlinear in the parameters in the sense originally defined by Fisher [2], and the less numerous set of conditions required for first order lack of identification. The latter set of conditions involve only the first derivatives of the coefficients as functions of the parameters. It is argued that if the model suffers from first order lack of identification, it will generally be the case that the usual estimators are consistent, although not asymptotically normally distributed. In a leading special case the asymptotic distribution is discussed, and the simulation of a simple model illustrates the extent to which this asymptotic distribution approximates the actual finite sample distribution.

When estimating a single equation with an error generated by an autoregressive process of higher order than one using a sequence of likelihood ratio tests to determine the correct order, the asymptotic size of the tests will be biased because of multiple optima of the likelihood function. A new type is suggested similar to the Durbin test [2] which is not biased in this way. IN HIS ARTICLE on testing for serial correlation in the presence of lagged endogenous variables [2] Durbin proved a general theorem which gives a significance test shown to be generally asymptotically equivalent to a likelihood ratio test. This paper proposes a generalization which gives a test criterion that may be preferred to the existing test criteria insofar as it can be set up using a less arbitrary choice of the parameters to be re-estimated, and also has the advantage of being relatively simple to compute. It seems more appropriate than the general Durbin form of test for application to the dynamic specification problem discussed in the third section of this article.

Most analyses of the principal-agent problem assume that the principal chooses an incentive scheme to maximize expected utility subject to the agent's utility being at a stationary point.An important paper of Mirrlees has shown that this approach is generally invalid.We present an alternative procedure.If the agent's preferences over income lotteries are independent of action, we show that the optimal way of implementing an action by the agent can be found by solving a convex programming problem.We use this to characterize the optimal incentive scheme and to analyze the determinants of the seriousness of an incentive problem.'Support from the U.K.

Article plaidant en faveur de l'utilisation d'estimateurs d'equations simultanees pour estimer les modeles d'effets aleatoires dynamiques d'un panel de donnees

This paper considers the null hypothesis that the errors on a regression equation form a random walk. By using the standard Durbin-Watson assumptions, we derive three test statistics that are uniformly most powerful against the alternative hypothesis that the errors are being generated by the stationary first order Markoff process. Unfortunately, the tabulated lower and upper bounds are too wide apart and so we compare the powers of the three tests using simulated as well as economic data. It is then recommended that the Imhof routine should be attached to standard regression programs to calculate the exact limit of the Berenblut-Webb statistic.

J. D. Sargan, Alok Bhargava, Maximum Likelihood Estimation of Regression Models with First Order Moving Average Errors when the Root Lies on the Unit Circle, Econometrica, Vol. 51, No. 3 (May, 1983), pp. 799-820