Many studies have shown a positive association between ambient ozone levels and mortality. Typically, these findings are based on models that assume a linear relationship between log mortality and ozone level. In this study, we adapted generalized additive models in which ozone effects are presumed to occur in three different ways: as a simple linear term, as a cubic natural spline term, and as a combination of two linear terms (a threshold model). We applied these models to daily time-series data for Seoul, Korea for the years 1995-1999 and found that the threshold model always fits best among the three. A 2.6% (95% CI: 1.7-3.5) increase of estimated relative risk (RR) in the total mortality associated with a 21.5 ppb increase of daily 1-h maximum ozone lagged by 1 day was observed by linear Poisson's regression. However, a 3.4% (95% CI: 2.3-4.4) increase in the estimated RR was observed using the threshold model. Adjustments for other ambient pollutants caused little changes to these results; 2.4-2.5% in the linear models and 3.2-3.4% in the threshold models. In addition, the largest difference in the estimated RRs of the linear and threshold models was observed in the summer: 1.9% (95% CI: 0.5-3.3) by the linear model and 3.8% (95% CI: 2.0-5.7) by the threshold model. These findings indicate that the conventional time-series Poisson regression model, which dose not take threshold into consideration, could underestimate the true risk of the ozone effect on daily mortality.