Four methodological limitations need to be noted in evaluating the substantive meaning our results. First, our predictors are limited by the fact that no information was collected about smokeless tobacco use or about quantity-frequency of smoking (e.g., number of cigarettes, cigars, or pipes smoked per day; number of years smoked). The prospective associations found here might have been stronger if the predictors had included these refinements. Second, we did not consider the possibility that some people who make suicide plans and attempts might deny ever having suicide ideation, as the skip logic used in our surveys only assessed plans and attempts among respondents who reported a history of suicide ideation. Third, we did not control for all DSM-IV disorders. Non-affective psychosis (NAP), for example, was not included in the core NCS-R assessment. Yet we know that both smoking44
are comparatively common among people with NAP. Exclusion of NAP, then, presumably led to an overestimation of the net effects of smoking. Fourth, although control data on history of pre-existing mental disorders and suicidality were gathered prospectively in our two-wave panel survey, the measures in each survey were based on retrospective reports. For example, the aspect of baseline tobacco use most strongly predictive of later SROs was early-onset
nicotine dependence, a measure that required respondents to provide retrospective reports in the baseline survey about their age when they first experienced symptoms of dependence. Similarly, the outcome measures required respondents to make retrospective reports about the occurrence of SROs in the decade since the baseline survey. Systematic recall errors in these reports could have introduced bias into our estimates of predictive associations. Some indication that retrospective recall of smoking history is likely to be unbiased comes from our investigation of the fact that the detailed baseline assessment of smoking history was obtained only from the respondents in the tobacco supplement. Our finding that the associations of baseline smoking with subsequent SROs were equivalent in magnitude whether smoking was assessment prospectively or retrospectively argues against the existence of retrospective recall bias at least for this part of the assessment of smoking history.
In the context of these limitations, we were able to reproduce the same basic data pattern that has been documented in previous prospective studies of smoking and SROs: a significant time-lagged dose-response association between smoking-related variables and subsequent SROs. The associations involving suicide ideation disappeared when we introduced controls for baseline risk factors. The conditional association of the highest level of smoking involvement (lifetime nicotine dependence with early ages of onset of use, daily use, and dependence) with suicide plans among ideators, in comparison, remained statistically significant even in a model that included a comprehensive set of controls.
Several previous prospective epidemiological studies found, unlike us, that significant gross prospective associations between smoking and SROs disappeared entirely after controlling for a series of risk factors.15, 17, 18
The fact that we found at least one of these associations to remain, between the highest level of smoking involvement and subsequent suicide plans, might be due to the fact that the smoking variables in these earlier studies all measured use rather than nicotine dependence, whereas the significant net association in our study involved early-onset nicotine dependence. Another possible explanation for the difference between our result and the results of these earlier studies is that our finding of a significant net association is a chance finding due to the fact that we examined associations between many different measures of smoking and four different measures of SROs. Replication of our finding that early-onset nicotine dependence predicts suicide plans among ideators is needed before we can reject the hypothesis of chance association. It is also noteworthy that one other previous prospective epidemiological study found more consistent evidence than we did of statistically significant effects of smoking on SROs after controlling for risk factors.11
This difference might be due to the fact that the controls used in that earlier study were much less complete than those used here.
By using more refined measures of both smoking and SROs than previous studies, we were able to expand our understanding of these associations by discovering that early-onset nicotine dependence is the aspect of smoking most strongly predictive of subsequent SROs and that suicide ideation and, among ideators, suicide plans, are the only SROs predicted. Importantly, neither suicide gestures among ideators nor suicide attempts among ideators were significantly predicted by any of the smoking-related variables we considered. Significant unconditional time-lagged associations of smoking-related variables with subsequent suicide gestures and suicide attempts are, in fact, present in our data due to the people who make suicide gestures and attempts being a subset of the people who have suicide ideation, but our decomposition of the SROs shows that the significant associations are actually only with ideation and plans. Previous research has shown that other predictors of ideation differ from the conditional predictors of attempts among ideators.46
The predictive effects of smoking now have to be conceptualized in terms of this emerging evidence of differential effects.
Although, as noted above, some previous prospective studies15, 17, 18
showed, like us, that the association between smoking and subsequent suicide ideation can largely be explained by baseline risk factors, those previous studies made no attempt to distinguish the explanatory effects of common causes from the mediating effects of controls that might have been both consequences of smoking and determinants of subsequent suicide ideation. Our decomposition of these different types of effects is consequently especially useful in showing that the subset of these variables most reasonably conceptualized as common causes account for more of the explanatory effect than the variables that might, at least in part, be mediators. This finding argues against the otherwise plausible possibility that the significant gross effects documented here are due to causal associations of smoking-related predictors that are mediated by intervening mental disorders.
We found, in comparison, that our control variables did not explain the statistically significant association between early-onset nicotine dependence and subsequent suicide plans among ideators. While this failure might be seen as indirectly arguing that nicotine dependence might have a causal effect on suicide plans, the finding that remitted nicotine dependence was as strong a predictor as active nicotine dependence is inconsistent with this interpretation. A more plausible interpretation in light of this specification is that the determinants of nicotine dependence, which are presumably indicated by respondents having either active dependence or a history of remitted dependence, rather than dependence itself are the true causal factors. Our failure to explain the association between baseline history of early-onset dependence and subsequent suicide plans, under this interpretation, might be seen as due to the fact that we did not measure the actual common causes of the two variables. It remains for future research to determine what those common causes might be, but it seems likely based on the current results that smoking is not itself of causal importance in this regard, at least in predicting SROs. It is important to reiterate the caution in the introduction, though, that the same results might not hold in predicting suicide deaths, as only a small fraction of the people who have SROs go on to complete suicides and the predictive associations of smoking with suicide deaths might go through different causal pathways than those involving SROs.