For various reasons, data on smoking are frequently missing, or only partially available, in retrospective epidemiologic studies of occupational risk factors. In such situations, indirect methods may be used to evaluate the magnitude and direction of the potentially confounding effects of smoking. Such an evaluation can be made quantitatively or qualitatively. Here we describe both approaches. A specific problem relates to case-referent studies, where sampling variation in referent selection may limit the possibility of controlling for confounding by smoking, even when smoking data are available. We present data showing that estimates of risk from occupational exposures which are not controlled for smoking may be as accurate as estimates derived after controlling for smoking, when the number of referents is relatively small. The problem of interaction is also discussed. In the absence of smoking data, the investigator has no indication of how smoking and occupation jointly affect disease risk (eg, additively or multiplicatively). The multiplicative model is usually assumed. However, if exposure and smoking act independently (additively), rate ratios are diminished. In such situations, in the presence of negative confounding by smoking, rate ratios may actually even be less than one--also when exposure and disease are strongly related.