Suicidal Thoughts and Behavior With Antidepressant Treatment:  Reanalysis of the Randomized Placebo-Controlled Studies of Fluoxetine and Venlafaxine

Dr Gibbons and colleagues’ analysis (1) of selected placebo-controlled randomized studies of antidepressants (n=9,185) is framed as a counterpoint to and advance over the much larger meta-analyses conducted by the FDA (n=99,839) that found a two-fold increased risk of suicidal thoughts and behaviors among children and young adults randomized to antidepressants, and a decreased risk in older adults. Critics of the FDA’s findings, and of the black box warning that resulted (2,3) have argued that the results were subject to ascertainment bias that favored finding suicidality among subjects randomized to antidepressants. Because the outcome of interest in the current study encompasses routinely collected assessments of suicidal thoughts, the main-effects analyses test this central critique (Table 3, first row). What these data show, in direction and magnitude (but contrary to the impression readers will get from the authors’ stated conclusions), is what one would expect from a study this size if the ascertainment bias critique was unsubstantiated: results mirroring the FDA findings, suggesting a harmful effect on youth. That the harmful effect estimate reported by Gibbons et al. does not quite reach statistical significance (p=0.17) is not surprising given their sample size is approximately one-tenth that used by the FDA.In addition to drawing misleading conclusions from their main-effects analysis, noted above, the authors’ extensions of these findings via mediation analyses are subject to bias. In their discussion, the authors state: “Randomization brings about balance in both observable and unobservable potential confounders,” supposedly strengthening their findings. While randomization may on average balance confounders for treatment, no such balance occurs for the mediator. Thus, mediator-outcome confounding is likely to have been introduced (4,5). Although the need to account for confounders of the mediator and the outcome has been long acknowledged in the field of psychology (6), the authors do not appear to take any analytic control for such confounding, nor do they mention this potential bias in their discussion. It is worth noting, nevertheless, that if the confounding introduced by the mediation analysis is minimal, the findings as presented suggest that there is a statistically significant harmful direct effect of treatment on suicide risk in youth that is not discussed in the text (Table 3, last row, p=0.05).Putting aside the mediation analyses, which are suspect for the methodological reasons discussed above, we are left with results from Gibbons et al that align with the FDA findings: in adults and geriatric patients, there is a protective effect of antidepressants on suicidality, while in youth the effect appears harmful.

1. Gibbons RD, Brown CH, Hur K, Davis JM, Mann JJ. Suicidal Thoughts and Behavior With Antidepressant Treatment: Reanalysis of the Randomized Placebo-Controlled Studies of Fluoxetine and Venlafaxine. Arch Gen Psychiatry. Feb 9 2012.

2. Friedman RA, Leon AC. Expanding the black box - depression, antidepressants, and the risk of suicide. N Engl J Med. Jun 7 2007;356(23):2343-2346.

3. US Food and Drug Administration. Memorandum: relationship between antidepressant drugs and suicidality. 2006; http://www.fda.gov/ohrms/dockets/ac/06/briefing/2006-4272b1-01-FDA.pdf. Accessed May 30, 2012.

4. Cole SR, Hernan MA. Fallibility in estimating direct effects. Int J Epidemiol. Feb 2002;31(1):163-165.

5. Vanderweele TJ, Vansteelandt S. Odds ratios for mediation analysis for a dichotomous outcome. Am J Epidemiol. Dec 15 2010;172(12):1339-1348.

6. Judd CM, Kenny DA. Process analysis: Estimating mediation in treatment evaluations. Evaluation Review. 1981;5:602-619.

Gibbons and colleagues (1) reanalyzed suicidality short-term data from 4 pediatric trials of fluoxetine (3 Eli Lilly-funded and the Treatment of Adolescents With Depression Study, TADS (2,3)) and found no increased risk of suicidality. The investigation, however, is fraught with problems. First, they did not solicit suicidal ideation adverse events reports (AERs), despite the fact that suicidal ideation is considered a suicide event by the Columbia Classification Algorithm of Suicide Assessment (C-CASA) (4), FDA analyses (5,6) and the TADS. Consequently, the study’s AER data necessarily underestimates suicidality by only capturing suicidal behaviors (attempts or suicides).

Second, the primary analysis used one item of a clinician-rated scale. This is unusual for this type of analysis. FDA meta-analyses examining suicidality in pediatric SSRI trials have relied on coded AER narratives, with suicide item score data secondarily analyzing worsening or emergent suicidality (5), or electronic search and manual review for serious adverse event descriptions (6). The study coded AERs (7) on the CDRS-R, but, lacking ideation AERs, no other supplemental data was added to the scale scores.

Third, the authors did not describe procedures to minimize bias, especially since AER data was obtained from Ely Lilly, funder of 3 of the 4 analyzed studies. As a comparison, an evaluation of safety in 23 pediatric antidepressant trials used blinded pharmaceutical personnel to identify suicidal events; Columbia University suicidology experts then blindly analyzed that data (5). Without these types of safeguards, bias cannot be eliminated.

Fourth, coding AERs as 7 on the CDRS-R item 13 hides suicidal behaviors instead of evaluating them as separate variables. In TADS, suicide events are analyzed separately, making them accessible for contrast analysis. For example, an analysis of the 12-week TADS suicidal behavioral events reported 6 suicide attempts for SSRI groups compared with 1 for non-SSRI groups (2). Though under-powered for statistical analysis, such data becomes meaningful within multiple contexts of analysis (e.g., by week 36 in TADS, there were 17 suicide attempts for fluoxetine; 1, non-fluoxetine, counting those switched to fluoxetine from original assignment (7).

Finally, the authors failed to mention that the short time frame (12 weeks) as a limitation. In TADS, time to first suicidal event ranged from 0.4 – 31.1 weeks (7). Beyond 12 weeks, only fluoxetine groups experienced a suicidal event. Counting those switched to fluoxetine, there were 36 suicidal events for fluoxetine takers and 8 for non-fluoxetine takers (7).

In sum, this study is inconsistent with standard pediatric suicidality assessment methodology, resulting most notably in under-representation of suicide-related events. Collapsing the first two categories in item 13 may have further undercounted these by downgrading suicidal ideation when angry to a non-event, though this is unclear (a table of revised response categories is not included). As derived, the study’s conclusion lacks credibility, especially given its contradiction of previous studies (5,6,8). The disclosure that Gibbons has served as an expert witness for Wyeth and Pfizer Pharmaceuticals in cases related to antidepressants and suicide further draws into question this study’s credibility. Unfortunately, the media blitz in its wake stands to mislead the public and prescribers.

References

1. Gibbons RD, Brown CH, Hur K, Davis JM, Mann J. Suicidal thoughts and behavior with antidepressant treatment: reanalysis of the randomized placebo-controlled studies of fluoxetine and venlafaxine. Arch Gen Psychiatry. 2012;69(6):580-587.

2. March J, Silva S, Petrycki S, Curry J, Wells K, Fairbank J, Burns B, Domino M, McNulty S, Vitiello B, Severe J; Treatment for Adolescents With Depression Study (TADS) Team. Fluoxetine, cognitive-behavioral therapy, and their combination for adolescents with depression: Treatment for Adolescents With Depression Study (TADS) randomized controlled trial. JAMA. 2004; 292(7):807-820.

3. March JS, Silva S, Petrycki S, Curry J, Wells K, Fairbank J, Burns B, Domino M, McNulty S, Vitiello B, Severe J. The Treatment for Adolescents With Depression Study (TADS): Long-term effectiveness and safety outcomes. Arch Gen Psychiatry. 2007;64:1132–43.

4. Posner K, Oquendo MA, Gould M, Stanley B, Davies M. Columbia Classification Algorithm of Suicide Assessment (C-CASA): classification of suicidal events in the FDA’s pediatric suicidal risk analysis of antidepressants. Am J Psychiatry. 2007;164(7):1035-1043.

5. Hammad TA, Laughren T, Racoosin J. Suicidality in pediatric patients treated with antidepressant drugs. Arch Gen Psychiatry. 2006;63:332-339.

6. Mosholder AD, Willy, M. Suicidal adverse events in pediatric randomized, controlled clinical trials of antidepressant drugs are associated with active drug treatment: a meta-analysis. J Child Adolesc Psychopharmacol. 2006 Feb-Apr;16 (1-2):25-32.

7. Vitiello B, Silva SG, Rohde P, Kratochvil CJ, Kennard BD, Reinecke MA, Mayes TL, Posner K, May DE, March JS. Suicidal events in the Treatment for Adolescents With Depression Study (TADS). J Clin Psychiatry. 2009;70(5):741-747.

8. Olfson M, Marcus SC, Shaffer D. Antidepressant drug therapy and suicide in severely depressed children and adults. Archives of General Psychiatry. 2006 Aug. 63:865-72.

Sparks and colleagues criticize our paper for relying upon prospective clinician ratings of suicidal ideation and behavior instead of corresponding spontaneous reports (adverse event reports – AERs). As indicated in our paper, we did include AERs of suicidal behavior (e.g., suicide attempts) based on the Columbia Classification Algorithm of Suicide Assessment (C-CASA), and coded them as the highest level of suicidal behavior (category 7 – “Has made suicide attempt within the last month or is actively suicidal”) . The reason for the use of the augmentation of the clinician ratings with the corresponding AERS data for suicidal behavior is that we were concerned that a suicide attempt might lead to study drop-out and therefore the suicide attempt might not be recorded by the clinician. We did not share the same concerns regarding spontaneous reports of suicidal ideation and therefore did not include these data. Furthermore, there is serious concern that such spontaneous reports are subject to ascertainment bias in which patients randomized to drug have more side effects in general and therefore more contact with study staff, and as a consequence greater opportunity to report suicidal ideation. Our paper was quite clear with respect to these analytic decisions and we strongly disagree with the remarks of Sparks and colleagues in this regard. However, unlike all previous analyses of the randomized controlled trial (RCT) data, we preserved the ordinal nature of the progression between suicidal thoughts and behavior giving increased weight to suicidal behavior vs. suicidal thoughts. We see this as an added strength of our analysis which failed to show increased risk of suicidal thoughts and behavior with antidepressant treatment.

Sparks and colleagues suggest that we did not describe procedures to minimize bias in that we did not use Columbia University suicidology experts to blindly review the data. In fact, these were exactly the data used in our analysis and provided to us by Eli Lilly.

Sparks and colleagues suggest that coding AERs of suicidal behavior as a 7 on the CDRS-R “hides suicidal behavior.” We fail to see why coding AERs of suicidal behavior is inconsistent with the rating of “Has made suicide attempt within the last month or is actively suicidal.” We can imagine no more accurate and parsimonious categorization of the AERs data on suicidal behavior, and indeed, it places these events at the highest category of suicidal risk in our mixed-effects ordinal regression analysis.

Sparks and colleagues misrepresents the TADS study data by indicating that there were 17 suicide attempts for fluoxetine and only one for non-fluoxetine. Vitello (1) indicate that SSRIs were given to 2 CBT patients and 9 placebo patients (who ultimately had a suicide event), due to non-response to the randomly assigned treatment. They do not indicate that the SSRI was fluoxetine, and these patients were given an SSRI because of clinical non-response, making the conclusion that this is a “signal” invalid. Indeed, the TADS authors conclude that “Most depressed adolescents presenting with suicidal events during treatment are still significantly depressed, with no or minimal signs of improvement and without evidence of drug associated behavioral activation.” This conclusion is consistent with our findings.

In terms of our collapsing the first two categories in item 13 “Feels life is not worth living” and “Wishes he or she were dead or any thoughts of possible death to self,” we followed the Columbia Classification System (2) which Sparks and colleagues criticized us for ignoring. Finally, Sparks and colleagues suggest that our analysis was biased based on restriction to 12 weeks. In fact, our analyses included all available data through the end of the observation period and indicated that the rate of suicidal ideation and behavior were in fact almost three times greater following week 8 for youth randomized to placebo (13.6%) than fluoxetine (4.8%); see Table 2 of our paper.

1. Vitiello B, Silva S, Rohde P, Kratochvil C, Kennard B, Reinecke M, Mayes T, Posner K, May DE, March JS. Suicidal events in the treatment for adolescents with depression study (TADS). J Clin Psychiatry 2009; 70 (5): 741-747.

2. Posner K, Oquendo MA, Gould M, Stanley B, Davies M. Columbia Classification Algorithm of Suicide Assessment (C-CASA): classification of suicidal events in the FDA’s pediatric suicidal risk analysis of antidepressants. Am J Psychiatry. 2007;164(7):1035-1043.

Swanson and colleagues are correct in noting that the mediational effect is not protected completely from bias due to randomization, and our statement that randomization protects against unobserved confounders should include a qualifier that this protection is limited to confounders of the treatment by mediator relationship. As for potential confounders of the depressive symptom/suicide risk relationship, this study, as well as any randomized trial, cannot completely rule these out. However, we are confident that the strong mediational findings for adults and geriatric patients would not be seriously affected by an unobserved confounder since the correlation between depressive symptoms and suicide risk is nearly 0.5. As for the findings for youth, Swanson and colleagues have stated that the data suggest “a harmful effect on youth.” The basis for their statement is that “the harmful effect estimate reported by Gibbons et.al, does not quite reach statistical significance (p=0.17),” which they suggest is not surprising because the sample size in our study (708) was one-tenth the size of that used by the FDA. This is a curious statement given there is no consistent pattern of suicide risk by time and treatment other than a marked reduction for both conditions (see Table 2 from our original paper which reports the observed proportions of patients with suicidal thoughts and behavior over time, stratified by age and drug group). For youth, the suicidal thoughts and behavior rates are similar at baseline and week 1, lower on treatment during weeks 2-5, higher on treatment during weeks 6-8, and lower on treatment thereafter. It is hard to imagine how these data justify the statement by Swanson and colleagues that “putting aside the mediation analysis, …, in youth the effect appears harmful.” Rather, these data seem quite consistent with the null hypothesis of no effect, given that there are four post-baseline weeks in which the suicidal event rate for treatment was lower than placebo and four post-baseline weeks in which the suicidal event rate for treatment was greater than placebo. Finally, it is important to note that while FDA did find a harmful effect for suicidal thoughts and behavior in youth that were spontaneously reported to study staff, they did not find a harmful effect for the prospective clinical ratings of suicidal thoughts and behavior. Our study combined the prospective ratings with spontaneous reports of suicidal behavior in a more unified analysis and like FDA’s analysis of the prospective data; we also did not find a significant effect of treatment on suicidal thoughts and behavior in youth.

Swanson and colleagues are correct in noting that the mediational effect is not protected completely from bias due to randomization, and our statement that randomization protects against unobserved confounders should include a qualifier that this protection is limited to confounders of the treatment by mediator relationship. As for potential confounders of the depressive symptom/suicide risk relationship, this study, as well as any randomized trial, cannot completely rule these out. However, we are confident that the strong mediational findings for adults and geriatric patients would not be seriously affected by an unobserved confounder since the correlation between depressive symptoms and suicide risk is nearly 0.5. As for the findings for youth, Swanson and colleagues have stated that the data suggest “a harmful effect on youth.” The basis for their statement is that “the harmful effect estimate reported by Gibbons et.al, does not quite reach statistical significance (p=0.17),” which they suggest is not surprising because the sample size in our study (708) was one-tenth the size of that used by the FDA. This is a curious statement given there is no consistent pattern of suicide risk by time and treatment other than a marked reduction for both conditions (see Figure https://webshare.uchicago.edu/users/rdg/AGP%20Articles%20Table/Table2.pdf, derived from Table 2 which reports the observed proportions of patients with suicidal thoughts and behavior over time, stratified by age and drug group). For youth, the suicidal thoughts and behavior rates are similar at baseline and week 1, lower on treatment during weeks 2-5, higher on treatment during weeks 6-8, and lower on treatment thereafter. It is hard to imagine how these data justify the statement by Swanson and colleagues that “putting aside the mediation analysis, …, in youth the effect appears harmful.” Rather, these data seem quite consistent with the null hypothesis of no effect, given that there are four post-baseline weeks in which the suicidal event rate for treatment was lower than placebo and four post-baseline weeks in which the suicidal event rate for treatment was greater than placebo. Finally, it is important to note that while FDA did find a harmful effect for suicidal thoughts and behavior in youth that were spontaneously reported to study staff, they did not find a harmful effect for the prospective clinical ratings of suicidal thoughts and behavior. Our study combined the prospective ratings with spontaneous reports of suicidal behavior in a more unified analysis and like FDA’s analysis of the prospective data; we also did not find a significant effect of treatment on suicidal thoughts and behavior in youth.