Realistic physical-biological ocean models pose challenges to statistical techniques due to their complexity, nonlinearity and high dimensionality. In this thesis, statistical data assimilation techniques for parameter and state estimation are adapted and applied to biological models. These methods rely on quantitative measures of agreement between models and observations. Eight such measures are compared and a suitable multiscale measure is selected for data assimilation. Build on this, two data assimilation approaches, a particle filter and a computationally efficient emulator approach are tested and contrasted. It is shown that both are suitable for state and parameter estimation. The emulator is also used to analyze sensitivity and uncertainty of a realistic biological model. Application of the statistical procedures yields insights into the model; e.g. time-dependent parameter estimates are obtained which are consistent with biological seasonal cycles and improves model predictions as evidenced by cross-validation experiments. Estimates of model sensitivity are high with respect to physical model inputs, e.g river runoff.