A statistical interpolation method is evaluated for routine production of ozone three-dimensional fields over western Europe. These fields are used for initializing short-term ozone forecasts issued from a chemistry-transport model. We mainly address two questions: (1) To what extent can the use of surface ozone observation data improve the description of ozone fields relative to raw simulations? (2) Does the use of ozone analysis improve short-term forecasts of the troposphere's chemical composition? The method consists of combining ozone simulations with surface ozone measurements. The resulting analyses are compared with independent observations in a statistical way over a long period of time (four consecutive summers). The improvement of the root-mean-square (RMS) error of the analyses relative to the raw simulations is ~30%. The short-term (1--2 days in advance) ozone forecasts are improved on average (by ~1 ppb of RMS error) if ozone analyses are used for initialization. The improvement is almost lost after a lead time of 36 hours. However, in cases where a model error propagates throughout the model domain, the improvement can be much larger (~10 ppb). We analyze one such case.