Reply to Odd Odds

David J. Carpenter, MSc, PharmD; Regan Fong, PhD; John E. Kraus, MD, PhD; John T. Davies, MSc; Christine Moore, PhD; Michael E. Thase, MD

J Clin Psychiatry 2011;72(11):1559

Article Abstract

Letter to the Editor

Because this piece does not have an abstract, we have provided for your benefit the first 3 sentences of the full text.

To the Editor: The method used to calculate odds ratios for suicidality data was described in the statistical methods section of our article. The number of suicidality-related events in each trial was low, particularly for the Definitive Suicidal Behavior endpoint, so we used exact inference and calculated odds ratios by adjusting for trial. Summary results for each trial and a prespecified analysis plan, which facilitate reproduction of our results, are publicly available and cited within our article.

See the original letter

Dr Carpenter and Colleagues Reply

To the Editor: The method used to calculate odds ratios for suicidality data was described in the statistical methods section of our article. The number of suicidality-related events in each trial was low, particularly for the Definitive Suicidal Behavior endpoint, so we used exact inference and calculated odds ratios by adjusting for trial. Summary results for each trial and a prespecified analysis plan, which facilitate reproduction of our results, are publicly available and cited within our article.

Prof Andrade points to a "substantial difference" between the crude odds ratio of 0.8 for suicidal behavior and the adjusted (reported) odds ratio of 1.2. The ability for the direction of the effect to differ in this way is widely known as Simpson’s paradox1 and highlights the importance of presenting adjusted estimates in preference to crude estimates that ignore important design factors, for example, differing randomization ratios between trials in a meta-analysis. Although the relative risk is often said to be easier to understand than the odds ratio, its choice can limit the ability to handle these issues because of the lack of suitable statistical methods or software.

Reference

1. Simpson EH. The interpretation of interaction in contingency tables. J R Stat Soc, B. 1951;13:238-241.

David J. Carpenter, MSc, PharmD

Regan Fong, PhD

[email protected]

John E. Kraus, MD, PhD

John T. Davies, MSc

Christine Moore, PhD

Michael E. Thase, MD

Editor’s note: The article discussed in these letters can be found at J Clin Psychiatry 2011;72(11):1503–1514.

Author affiliations: GlaxoSmithKline, King of Prussia, Pennsylvania (Drs Carpenter and Fong); Research Triangle Park, North Carolina (Dr Kraus); and Harlow, United Kingdom (Mr Davies); i3 Research, Cary, North Carolina (Dr Moore); and Departments of Psychiatry, University of Pennsylvania School of Medicine, Philadelphia, Philadelphia Veterans Affairs Medical Center, and University of Pittsburgh Medical Center, Pennsylvania (Dr Thase). Potential conflicts of interest: Please consult the original article [J Clin Psychiatry. 2011;72(11):0000-0000]. Funding/support: The study discussed in this letter was funded by GlaxoSmithKline.

doi:10.4088/JCP.11lr07373a

Download Free PDF

Save Save