Patrick G. O’Malley, MD, MPH Jeffrey L. Jackson, MD, MPH James Santoro, MD, MPH Glen Tomkins, MD, MPH Erin Balden, MD, MPH Kurt Kroenke, MD Washington, DC; Bethesda, Maryland; El Paso, Texas; and Indianapolis, Indiana Submitted, revised, July 15, 1999. From the Department of Medicine (P.G.O., G.T., E.B.), Walter Reed Army Medical Center, Washington, DC; Uniformed Services University of the Health Sciences (J.L.J), Bethesda; William Beaumont Army Medical Center (J.S.), El Paso; and the Regenstrief Institute for Health Care and Indiana University School of Medicine (K.K.), Indianapolis. The opinions and assertions in this paper are the private views of the authors and should not be interpreted to reflect the views of the Department of the Army or the Department of Defense.Reprint requests should be addressed to Patrick G. O’Malley, MD, MPH, Division of General Internal Medicine, Walter Reed Army Medical Center, 6800 Georgia Avenue, Washington, DC 20307. E-mail: pgomalley@msn.com.
References
Both studies of mianserin showed no benefit for idiopathic pain.94,103 Of the 2 placebo-controlled SSRI studies, efficacy was shown for zimelidine in idiopathic musculoskeletal pain104 but not for sertraline for chronic pelvic pain.105
Tinnitus. There were 2 randomized placebo-controlled trials of a TCA for chronic tinnitus.106,107 The better quality study107 showed improvement in disability and tinnitus loudness with nortriptyline, while a study of trimipramine showed no benefit.106
Chronic Fatigue. There are 2 randomized placebo-controlled trials of antidepressants for chronic fatigue syndrome.108,109 One trial with fluoxetine showed no benefit,109 while the other trial109 using phenelzine showed improvement in multiple symptoms, illness severity, and mood.
Antidepressant Efficacy and Depression Response
Depression was assessed in 49 (52%) of the studies, of which 90% used validated tools such as the Beck Depression Inventory, Hamilton Anxiety and Depression index, Montgomery depression scale, Zung, syptom checklist 90, Minnesota Multiphasic Personality Inventory, and the Center for Epidemiologic Study depression inventory. However, an analysis of association between depression and response to treatment was performed in only 24 studies (49% of studies that assessed for depression, 25% of all studies).1,35-38,41,42,48,49,54,58,60,67,77,78,81,82,84,94,96,98,100,107,108 Of the 24 studies that assessed for an association, only 8 (33%) demonstrated a correlation between physical symptom response and depressive response;37,38,49,54,60,82,84,94 and only 3 reported any correlation statistics.38,82,94 Thus, though there appears to be little correlation of effect with depression response in the few studies where it was assessed, it is difficult to draw any conclusions on this relationship given the quality of the analyses and the small numbers of participants in the trials (Table 2).
Antidepressant Efficacy and Study Characteristics
A comparison of trials that showed a benefit with those that did not is displayed in Table 3. The following study characteristics were not associated with a greater likelihood of showing benefit: parallel design, sample size, quality rating, industry sponsorship, or country of study. However, drug class and comparison treatment (placebo or active nonantidepressant control) were associated with trial outcome. Studies with a placebo control were more likely than active nonantidepressant controls to show benefit, though this does indirectly support the benefit associated with antidepressants, since the active controls were usually medications known to be therapeutic. Studies of tricyclic antidepressants were more likely than studies of SSRIs or antiserotonin agents to have a beneficial outcome (P = .02).
Meta-Analysis
We synthesized the quantitative data from each placebo-controlled study in which data were extractable (48 studies, 49%). Data were extractable in one or both of 2 forms: a dichotomous outcome of improvement and continuous outcomes from which a standardized effect size could be calculated (standardized mean difference between placebo and active treatment). For the dichotomous outcome of improvement we used improvement in any of the following outcomes: global assessment (patient or physician), summary symptom index scores (headache index or fibromyalgia symptom score, for example), or pain severity scale scores. The pooled odds ratio (OR) for the dichotomous outcome of improvement was 3.43 (95% confidence interval [CI], 2.60-4.52; P = .04; Figure 1), while the pooled standardized mean difference was 0.87 (95% CI, 0.59-1.14; P <.001; Figure 2). The effect size was not homogeneous across all studies, and the treatment of unexplained symptoms with antidepressants was associated with a greater than 3-fold higher likelihood of improvement. For continuous outcomes, antidepressant therapy was associated with almost a full standard deviation improvement. This is considered a large effect size.112 The absolute percentage difference in improvement between the antidepressant and placebo arms was 32% (95% CI, 15%-48%), yielding a number needed to treat of 3.1 (95% CI, 2.1-6.6) before improving one patient’s symptoms.
Publication Bias
An assessment for publication bias against small studies with no or a small effect was determined by pooling all the effect sizes and standard errors from the placebo-controlled studies that had extractable data. There was evidence for significant publication bias (P <.001; Figure 3).
Sensitivity Analysis
Using meta-regression, we controlled for the following variables: drug class, withdrawal rates (>20%), quality of study, type of symptom or syndrome, year of publication (before or after 1980), and sample size. None of these variables significantly affected the summary effect size. Similarly, the effect size for each individual syndrome was not significantly different.
Using the assumptions of no effect (OR = 1.0), sample size equal to 50 (the median for all 96 trials), and a variance of 0.57 (the mean variance of the effect size among all the trials), it would take 628 trials in a random effects model to make the summary effect size statistically insignificant.
