Response to Correspondence by Aursnes et al., BMC Psychiatry 2006
John Kraus, GlaxoSmithKline
7 March 2007
The correspondence by Aursnes and colleagues (BMC Psychiatry 2006, 6:55 (28 November 2006); doi:10.1186/1471-244X-6-55), which refers to and uses data from GSK’s 2006 analysis of adult suicidality and paroxetine (www.gsk.com), contains some comments regarding that analysis that require clarification. The data used for the Aursnes analysis is a subset of the total dataset used in GSK’s 2006 analysis, containing only the clinical trials with adult patients for the Major Depressive Disorder (MDD) indication (3455 paroxetine patients, 1978 placebo patients, 19 trials). Within this MDD population, GSK identified evidence of a possible increased risk for suicidal behavior in adult patients with MDD (11/3455 (0.32%) for paroxetine, vs. 1/1978 (0.05%) for placebo; odds ratio = 6.7 (95% CI 1.1, 149.4); p=0.058). Eight of the 11 paroxetine-treated patients were aged 30 years or less. When looking at the entire clinical trial data used in GSK's 2006 Analysis (8958 paroxetine patients, 5953 placebo patients, 57 trials), there was no evidence of treatment difference in suicidal behavior or ideation, or in suicidal behavior alone (www.gsk.com). Across all indications, although not statistically significant, this analysis showed a higher frequency of suicidal behavior in young adults (prospectively defined as age 18-24) treated with paroxetine compared with placebo (17/776 [2.19%] versus 5/542 [0.92%]). In the older age groups (25-64 years and ≥65 years), no such increase was observed. However, it is difficult to conclude a causal relationship between paroxetine and suicidality due to the small incidence and absolute number of events, the retrospective nature of the 2006 analysis, and potential for confounding by the fact that the events of interest are a symptom of the psychiatric illnesses themselves.
The GSK 2006 analysis examined suicidality within individual indications as well as for a pooled “all depressive disorders” group. Therefore, MDD, intermittent brief depression (IBD), dysthymia, and bipolar depression were examined individually and as a pooled group (full results can be found at www.gsk.com). Presumably, the group Aursnes et al. refer to as “patients with a history of suicide attempts” is the IBD group of GSK's 2006 Analysis. Additionally, the cases comprising GSK's 2006 analysis were individually reviewed by external experts who were blinded to treatment and who classified cases into predefined suicidality categories (e.g., suicide attempt, suicidal ideation, etc). As a result of this classification process, three paroxetine cases previously identified as suicidal behavior were not classified as suicidal behavior by the expert raters, whereas an additional two events in the paroxetine group and one event in the placebo group were newly identified by the expert raters (see Briefing Document at www.gsk.com).
Aursnes and colleagues claim that GSK is unwilling to accept Bayesian principles in the analysis of clinical trials and refer to FDA guidance for the use of Bayesian statistics. Bayesian applications are indeed becoming more common within the pharmaceutical industry, including within GSK, but it should be noted the FDA guidance refers to medical device clinical trials rather than to pharmacological clinical trials (http://www.fda.gov/cdrh/osb/guidance/1601.html). Indeed, in the recently completed FDA meta-analysis of adult suicidality and antidepressant treatment, which assessed data from all manufacturers of antidepressants (http://www.fda.gov/ohrms/dockets/ac/06/briefing/2006-4272b1-01-FDA.pdf), FDA’s statistical group used the exact method, as GSK had done in the 2006 Analysis; FDA did not use Bayesian statistical methods. FDA’s analysis not only corroborated GSK's finding of no increased risk of suicidality in patients 25 and over taking antidepressants compared to placebo, but, in fact, found a significantly protective effect for antidepressants in these patients that increases with increasing age.
Despite the utility of Bayesian principles in many circumstances, we feel their use adds little to the understanding of the data from GSK’s 2006 analysis. It is noteworthy that the difference between the results presented by Aursnes and those presented by GSK is not so much an issue stemming from the use of frequentist vs. Bayesian methods, but from the use of one vs. two-sided probabilities. GSK, in common with FDA and current convention, quoted two-sided probabilities.
Respectfully Submitted,
John E. Kraus, MD, PhD
Director, Psychiatry, Neurosciences Medicines Development Center, GlaxoSmithKline
Jack Modell, MD
Vice President, Psychiatry, Neurosciences Medicines Development Center, GlaxoSmithKline
John T. Davies, MSc
Director, Biomedical Data Sciences, GlaxoSmithKline
James Roger, PhD
Director, Research Statistics Unit, GlaxoSmithKline
William A. Ball, MD
Senior Director, Psychiatry, Neurosciences Medicines Development Center, GlaxoSmithKline
Alun W. Bedding, BSc
Director, Biostatistics and Programming Development Partners, GlaxoSmithKline
Competing interests
GlaxoSmithKline is the manufacturer of Seroxat, Paxil, and Paxil CR (paroxetine HCl)
Aursnes and Natvig answer th comment by Kraus
Ivar Aursnes, University of Oslo
22 August 2007
We acknowledge most of the arguments made by Kraus and others in their comments to our correspondence. It should, however, be noted that according to GSK`s own Briefing Document (www.gsk.com) the group we referred to as "patients with a history of suicide attempts" was included in the earlier so-called Article 31 analysis, the last part of which was submitted to EU regulatory authorities in 2004. If some of these patients still enter into the intermittent brief depression (IBD) group of GSK`s 2006 analysis, even the latter analysis suffers from what we have denoted "data-drowning".
Furthermore, we are a little hesitant to include confounding as a mechanism of bias when referring to randomized trials, and concerning the FDA meta-analysis, we find it appropriate to quote from table 16 in that publication:
“Suicidal Behavior Risk for Active Drug relative to Placebo
Drug Class/Drug Odds Ratio 95% Confidence Interval p value
SSRI 1.23 0.82 - 1.85 0.31
Paroxetine 2.76 1.16 – 6.60 0.02”
- which we believe settles the issue: the re-analysis of data performed by FDA concludes for paroxetine with a two-sided p-value of 0.02. Note that the one-sided version of this p-value equals 0.01 which corresponds well with our conclusion of at least 98 % probability that paroxetine increases suicide attempts.
The back history of the drug under discussion might also be of public interest. Relevant for the debate are quotes from unpublished literature. The documentation following the new drug application for registration in 1989 contained a table 24 on page 203 that listed 6 patients under the heading “PLACEBO” with the columns patient identity, sex, age, dosage, duration and description. We gradually learned from studying the material that there was in fact only one suicide attempt among patients randomized to placebo (number 0201009 in the table), as now confirmed by GSK. The other five had occurred among patients during the run in stages of studies of various kinds, including studies with active control groups and studies without control groups.
We are aware that such practice of presentation has been brought to light during a US lawsuit, and that it was claimed by GSK that it was a standard procedure. All the same, this does not justify the presentation in table 22 of the same report:
“Table 22
Attempted Suicide
Worldwide Clinical Trials Program
PAROXETINE PLACEBO ACTIVE CONTROLS
n = 2963 n = 554 n = 1151
N0. (%)
40 (1.3) 6 (1.1) 12 (1.0)”
The denominator in the placebo column is definitely wrong: instead of 554 the number should, according to our calculations, be several thousands, the exact figure being difficult to extract from the report. It follows that the true data concerning the risk of suicide activities connected with paroxetine obviously should have implied concerns as to its safety. If the suspicion had been acted upon at that time, the now established risk of suicidal activity among children and young adults could have been revealed at a much earlier stage.
Respectfully
Ivar Aursnes
Professor of clinical pharmacology
Bent Natvig
Professor of statistics
Competing interests
The authors declare that they have no competing interests
Response to Correspondence by Aursnes et al., BMC Psychiatry 2006
7 March 2007
The correspondence by Aursnes and colleagues (BMC Psychiatry 2006, 6:55 (28 November 2006); doi:10.1186/1471-244X-6-55), which refers to and uses data from GSK’s 2006 analysis of adult suicidality and paroxetine (www.gsk.com), contains some comments regarding that analysis that require clarification. The data used for the Aursnes analysis is a subset of the total dataset used in GSK’s 2006 analysis, containing only the clinical trials with adult patients for the Major Depressive Disorder (MDD) indication (3455 paroxetine patients, 1978 placebo patients, 19 trials). Within this MDD population, GSK identified evidence of a possible increased risk for suicidal behavior in adult patients with MDD (11/3455 (0.32%) for paroxetine, vs. 1/1978 (0.05%) for placebo; odds ratio = 6.7 (95% CI 1.1, 149.4); p=0.058). Eight of the 11 paroxetine-treated patients were aged 30 years or less. When looking at the entire clinical trial data used in GSK's 2006 Analysis (8958 paroxetine patients, 5953 placebo patients, 57 trials), there was no evidence of treatment difference in suicidal behavior or ideation, or in suicidal behavior alone (www.gsk.com). Across all indications, although not statistically significant, this analysis showed a higher frequency of suicidal behavior in young adults (prospectively defined as age 18-24) treated with paroxetine compared with placebo (17/776 [2.19%] versus 5/542 [0.92%]). In the older age groups (25-64 years and ≥65 years), no such increase was observed. However, it is difficult to conclude a causal relationship between paroxetine and suicidality due to the small incidence and absolute number of events, the retrospective nature of the 2006 analysis, and potential for confounding by the fact that the events of interest are a symptom of the psychiatric illnesses themselves.
The GSK 2006 analysis examined suicidality within individual indications as well as for a pooled “all depressive disorders” group. Therefore, MDD, intermittent brief depression (IBD), dysthymia, and bipolar depression were examined individually and as a pooled group (full results can be found at www.gsk.com). Presumably, the group Aursnes et al. refer to as “patients with a history of suicide attempts” is the IBD group of GSK's 2006 Analysis. Additionally, the cases comprising GSK's 2006 analysis were individually reviewed by external experts who were blinded to treatment and who classified cases into predefined suicidality categories (e.g., suicide attempt, suicidal ideation, etc). As a result of this classification process, three paroxetine cases previously identified as suicidal behavior were not classified as suicidal behavior by the expert raters, whereas an additional two events in the paroxetine group and one event in the placebo group were newly identified by the expert raters (see Briefing Document at www.gsk.com).
Aursnes and colleagues claim that GSK is unwilling to accept Bayesian principles in the analysis of clinical trials and refer to FDA guidance for the use of Bayesian statistics. Bayesian applications are indeed becoming more common within the pharmaceutical industry, including within GSK, but it should be noted the FDA guidance refers to medical device clinical trials rather than to pharmacological clinical trials (http://www.fda.gov/cdrh/osb/guidance/1601.html). Indeed, in the recently completed FDA meta-analysis of adult suicidality and antidepressant treatment, which assessed data from all manufacturers of antidepressants (http://www.fda.gov/ohrms/dockets/ac/06/briefing/2006-4272b1-01-FDA.pdf), FDA’s statistical group used the exact method, as GSK had done in the 2006 Analysis; FDA did not use Bayesian statistical methods. FDA’s analysis not only corroborated GSK's finding of no increased risk of suicidality in patients 25 and over taking antidepressants compared to placebo, but, in fact, found a significantly protective effect for antidepressants in these patients that increases with increasing age.
Despite the utility of Bayesian principles in many circumstances, we feel their use adds little to the understanding of the data from GSK’s 2006 analysis. It is noteworthy that the difference between the results presented by Aursnes and those presented by GSK is not so much an issue stemming from the use of frequentist vs. Bayesian methods, but from the use of one vs. two-sided probabilities. GSK, in common with FDA and current convention, quoted two-sided probabilities.
Respectfully Submitted,
John E. Kraus, MD, PhD
Director, Psychiatry, Neurosciences Medicines Development Center, GlaxoSmithKline
Jack Modell, MD
Vice President, Psychiatry, Neurosciences Medicines Development Center, GlaxoSmithKline
John T. Davies, MSc
Director, Biomedical Data Sciences, GlaxoSmithKline
James Roger, PhD
Director, Research Statistics Unit, GlaxoSmithKline
William A. Ball, MD
Senior Director, Psychiatry, Neurosciences Medicines Development Center, GlaxoSmithKline
Alun W. Bedding, BSc
Director, Biostatistics and Programming Development Partners, GlaxoSmithKline
Competing interests
GlaxoSmithKline is the manufacturer of Seroxat, Paxil, and Paxil CR (paroxetine HCl)
Aursnes and Natvig answer th comment by Kraus
22 August 2007
We acknowledge most of the arguments made by Kraus and others in their comments to our correspondence. It should, however, be noted that according to GSK`s own Briefing Document (www.gsk.com) the group we referred to as "patients with a history of suicide attempts" was included in the earlier so-called Article 31 analysis, the last part of which was submitted to EU regulatory authorities in 2004. If some of these patients still enter into the intermittent brief depression (IBD) group of GSK`s 2006 analysis, even the latter analysis suffers from what we have denoted "data-drowning".
Furthermore, we are a little hesitant to include confounding as a mechanism of bias when referring to randomized trials, and concerning the FDA meta-analysis, we find it appropriate to quote from table 16 in that publication:
“Suicidal Behavior Risk for Active Drug relative to Placebo
Drug Class/Drug Odds Ratio 95% Confidence Interval p value
SSRI 1.23 0.82 - 1.85 0.31
Paroxetine 2.76 1.16 – 6.60 0.02”
- which we believe settles the issue: the re-analysis of data performed by FDA concludes for paroxetine with a two-sided p-value of 0.02. Note that the one-sided version of this p-value equals 0.01 which corresponds well with our conclusion of at least 98 % probability that paroxetine increases suicide attempts.
The back history of the drug under discussion might also be of public interest. Relevant for the debate are quotes from unpublished literature. The documentation following the new drug application for registration in 1989 contained a table 24 on page 203 that listed 6 patients under the heading “PLACEBO” with the columns patient identity, sex, age, dosage, duration and description. We gradually learned from studying the material that there was in fact only one suicide attempt among patients randomized to placebo (number 0201009 in the table), as now confirmed by GSK. The other five had occurred among patients during the run in stages of studies of various kinds, including studies with active control groups and studies without control groups.
We are aware that such practice of presentation has been brought to light during a US lawsuit, and that it was claimed by GSK that it was a standard procedure. All the same, this does not justify the presentation in table 22 of the same report:
“Table 22
Attempted Suicide
Worldwide Clinical Trials Program
PAROXETINE PLACEBO ACTIVE CONTROLS
n = 2963 n = 554 n = 1151
N0. (%)
40 (1.3) 6 (1.1) 12 (1.0)”
The denominator in the placebo column is definitely wrong: instead of 554 the number should, according to our calculations, be several thousands, the exact figure being difficult to extract from the report. It follows that the true data concerning the risk of suicide activities connected with paroxetine obviously should have implied concerns as to its safety. If the suspicion had been acted upon at that time, the now established risk of suicidal activity among children and young adults could have been revealed at a much earlier stage.
Respectfully
Ivar Aursnes
Professor of clinical pharmacology
Bent Natvig
Professor of statistics
Competing interests
The authors declare that they have no competing interests