While ethical and practical limitations prevent researchers from literally randomizing subjects to quitting or continuing to smoke, successful smoking cessation trials effectively randomize subjects to having a higher probability of quitting (if they are in the intervention arm) or lower probability of quitting (if they are in the control arm). An IV analysis applied to such trials can yield an unbiased estimate of the effect of quitting on weight gain.
IV analysis applied to randomized trials is essentially intention to treat (ITT) analysis taken one step further. In ITT analysis, outcomes are compared based on the initial treatment assignments, despite the fact that some subjects may not have adhered to their assignments (Fisher et al. 1990
). IV analysis scales up the ITT estimate of the treatment effect by dividing it by the difference across groups in probabilities of receiving the treatment, recognizing that the ITT estimate is likely to be an underestimate of the true treatment effect. Newhouse and McClellan (1998)
provide an overview of the application of IVs to health outcomes research. The use of the technique in this field has grown in recent years (McClellan, McNeil, and Newhouse 1994
; Schoenbaum et al. 2002
; Brooks et al. 2003
; Hadley et al. 2003
The validity of the IV approach rests on two assumptions (Angrist, Imbens, and Rubin 1996
). First, the IV must be correlated with the treatment being evaluated. Second, the IV must be correlated with the outcome of interest only through the IV's relationship with the treatment. That is, the IV must not have a direct effect on the outcome of interest.
In the context of our analysis of smoking cessation trials, the IV is the random assignment to the intervention or control, the “treatment” is whether a person successfully quits smoking, and the outcome of interest is weight gain. The IV estimate of the effect of quitting on weight gain is thus the ITT estimate (difference in average weight gain across the intervention and control groups) divided by the difference in probabilities of quitting across the two groups.
In our primary analysis we defined quitters as “sustained quitters.” These individuals did not smoke at any follow-up points in their respective study periods. In sensitivity analysis, we expanded the definition to “cross-sectional quitters.” These were people who had quit as of the last follow-up, but not necessarily as of intermediate follow-ups.
We sought to reanalyze data from previous studies using this IV approach. We searched the medical literature for smoking cessation randomized control trials that would be suitable. Such studies had to fulfill several criteria. First, they had to satisfy the two critical assumptions in IV analysis: (1) the smoking cessation intervention had to be effective (i.e., the IV—assignment to the intervention—had to be correlated with the “treatment,” quitting); (2) the intervention could not have a significant direct effect on body weight (i.e., the IV had to be correlated with weight gain only through its effect on the “treatment,” quitting).
Regarding the first criterion, we considered an intervention “effective” if it produced quit rates at least 15 percent higher than in the control group. In general, a high correlation between the IV and the “treatment” is necessary because the standard errors of the estimated treatment effect are inversely related to the strength of this correlation (Angrist, Imbens, and Rubin 1996
). As a result of the second criterion, we excluded studies in which the central component of the intervention included pharmaceutical cessation aids, because we assumed that these were likely to have direct effects on body weight. Even though some nicotine replacement strategies are not clearly associated with weight changes (Fiore et al. 2000
), we chose to be conservative by assuming that nicotine in any form may act as an appetite suppressant. We also excluded studies explicitly designed to modify multiple lifestyle risk factors instead of just smoking.
In addition we confined attention to studies for which the data necessary to construct IV estimates were available. In particular, we needed the following information to construct an IV estimate: (1) average weight gain for the intervention group; (2) average weight gain for the control group; (3) quit rate for the intervention group; (4) quit rate for the control group. As noted earlier, previous studies of weight gain associated with smoking cessation have focused on the comparison between quitters and continuing smokers, ignoring the randomized assignments. Thus, in most published studies the first two pieces of information listed above were not available. Finally, we required that the studies include the data necessary to construct the conventional estimate of the effect of cessation on weight gain, for comparison purposes, and that the follow-up period be at least 6 months. This last condition reflected our interest in more lasting changes in body weight.
We searched Medline for all smoking cessation studies published since 1975 that included information on weight gain. We performed additional searches based on lists of references in these articles. In cases in which published studies did not display all the data necessary for our study but appeared potentially suitable, we contacted the authors to inquire about unpublished data.