Despite the hundreds of exposure-disease associations studied by perinatal epidemi ologists, few have become accepted as causal. Maternal smoking and its effects on preterm delivery and fetal growth restriction are two such causal associations, whereas direct maternal smoking and spontaneous abortion is not.' The causal role of environ mental tobacco smoke (ETS) on perinatal disease has proved particularly difficult to document despite the plausibility of the hypothesis. The current evidence for an associ ation of ETS with spontaneous abortion remains mixed. In this issue of EPIDEMIOLOGY, George et a12 report an association between ETS, as assessed by urinary cotinine level, and spontaneous abortion. Cotinine is a widely used biomarker for cigarette smoking, and this is the first study to use it to assess ETS exposure as a risk factor for spontaneous abortion. George et al report that the increased risks are similar for normal and abnormal karyotype. This may not be an implausible result given that ETS might induce the spontaneous abortion of abnormal karyotypes that survive very early pregnancy as well as the abortion of normal pregnancies. Perhaps less plausible is the finding that spontaneous abortion risk with exposure to ETS is similar, in these data, to the risk with direct smoking, an association with an order-of-magnitude greater level of cotinine exposure. In contrast, the associations of ETS on fetal growth restriction are weaker than the associations found for direct smoking,3 which seems more plausible. So what might account for the surprisingly strong risk found for ETS in the current report? The authors used a case-control study with many strong design features: it is well powered, measured several potentially important covariates, validated the spontaneous abortions, used an objective biomarker for exposure, accounted for possible bias in women seeking induced abortion, and stratified on fetal karyotype. Given that the authors have dealt well with the usual methodologic problems, what other possibilities might explain their findings? Nicotine is metabolized to cotinine by CYP (including CYP1A2) activity,4 which declines markedly in later pregnancy.5 This decline leads to slower nicotine metabolism and lower rates of cotinine for any given exposure as pregnancy progresses. Systematic bias in cases being interviewed earlier in pregnancy than controls would bias the cotinine in the observed direction. However, in this study, gestational age was closely matched,6(Table 1) So this does not seem a likely confounding factor. Residual confounding can be responsible for artifactually inflated risks if the confound ers are not modeled precisely.7 In this article, several potentially important confounders are considered: older maternal age, history of spontaneous abortion, and caffeine consumption, all of which are associated with both an increased risk of spontaneous abortion and higher cotinine levels. The article is silent as to how these factors are parameterized in the modeling; simple categorization (as shown in the tables) may have been insufficient to eliminate their effects completely. All of these would tend to bias in favor of showing risk of spontaneous abortion with higher cotinine.