Why does 4D-Var beat 3D-Var?

A set of four experiments is described which measure the expected beneficial aspects of incremental four-dimensional variational (4D-Var) compared to 3D-Var data assimilation systems: allowing for the time of each observation in the full and increment fields with which it is compared, and using time-evolved covariances. Judging each scheme by the overall accuracy of resulting numerical weather prediction forecasts compared to observations, each aspect is shown to provide benefit. On other measures of analysis quality, such as the fit of short-period forecasts and analyses to observations, the benefits of 4D-Var are less clear; it is sometimes worse. Perhaps 4D-Var is improving the analysis of growing modes, which are more important for forecasts, without improving all aspects of the analysis. Our basic 4D-Var was not provided with many observations distributed in time, and had very simple parametrizations. There is an expectation of enhanced benefits as these aspects are developed.