To help assess the effectiveness of the model-based analysis and prediction procedures at the National Meteorological Center (NMC), we compare the seasonal and nonseasonal components of sea level from 44 tide gauges in the tropical Pacific with those of the dynamic heights output by two 11-year model reanalyses (1982--1992) at the same locations, which differ mainly in their wind forcing. Both reanalyses assimilate ocean thermal data and incorporate most of the procedures used by NMC in producing operational ocean aalyses and experimental coupled model climate forecasts. The reanalyses reproduce the broad patterns of annual amplitude and phase and of seasonal and nonseaonal variance, except for severe underestimates along the eastern boundary, especially north of the equator. The annual cycles and interannual departures of zonal flow indices estimated from selected island pairs near the dateline show good correspondence for the North Equatorial Countercurrent (NECC) and somewhat flawed and noisy comparisons for the North Equatorial Current (NEC) and South Equatorial Current (SEC). The reanalyses also reproduce the large-scale time and space patterns of nonseasonal variability in the first three empirical orthogonal functions (EOFs), which together explain about 65% of the anomalous variability and characterize the El Ni¿o-Southern Oscillation cycle.

The first two EOF modes describe the westward migration of three ENSO episodes, and the third mode appears to capture differences between episodes. However, the reanalysis based on the anomalous winds generated by the NMC medium-range forecast model shows significant discrepancies in the large-scale spatial and temporal variability. These discrepancies disappear in the reanalysis based on departures of the Florida State University analyzed wind fields. Hence the wind forcing critically affects the reanalysis in spite of the assimilation of ocean thermal data. Future improvements in the atmospheric model to produce a more realistic evolution of the wind field can therefore lead to significantly better model integrations in the analysis and initializtion mode (with data assimilation) as well as in the coupled model forecast mode.