Exposure-response modeling in occupational epidemiology is usually motivated by questions of causal inference (is there a monotonic increase of risk with increasing exposure?) or risk assessment (how much excess risk exists at any given level of exposure?). Here we focus on statistical models for exposure-response in mortality studies. Categorical analyses are useful for detecting the shape of exposure-response, but are dependent on cutpoints and are less useful for risk assessment which requires a parametric curve. Restricted cubic splines and penalized splines are useful intermediates between categorical and simpler parametric curves, may help to choose a simpler parametric curve. The shapes of spline curves will depend on the degree of "smoothing" chosen by the analyst. Exposure-response curves in occupational epidemiology often flatten out at high exposures, for a number of reasons (eg., greater misclassification at high exposure, saturation of metabolic pathways, etc). A log transformation of exposure often provides a good parametric fit to such curves, but has the disadvantage of a very high slope at low exposures, which may be the relevant exposures for environmental risk assessment. A piece-wise linear model may be a useful alternative. In general, the model with the best statistical fit may not be the "best" model for risk assessment. Another model of interest is a threshold model in which there is no risk before a certain threshold. Finally, Bayesian restrictions on exposure-response curves may prove useful in some settings in increasing precision. These points are illustrated using data from several epidemiologic studies.