Bivariate regression is used to estimate energy expenditure from doubly labeled water data. Two straight lines are fitted to the logarithms of the enrichments of oxygen-18 and deuterium simultaneously as a bivariate regression, so that the correlations between the oxygen and deuterium regression coefficients can be estimated. Maximum likelihood methods are used to extend bivariate regression to unbalanced situations caused by missing observations and to include replicate laboratory determination from the same urine samples, even if one of the replicates is missing. Use of maximum likelihood allows the determination of a confidence interval for the energy expenditure based on the log likelihood surface rather than use of the propagation of variance methods for nonlinear transformations. The model is extended to include the subject's deviations from the two lines as a bivariate continuous-time first-order autoregression to allow for serial correlation in the observations. The analysis of data from two subjects, one without apparent serial correlation and one with serial correlation, is presented.