Anxiety disorders are common disorders, which pose a major burden to society and the individual. An anxiety disorder may be treated with medication, in particular with antidepressants such as the selective serotonin reuptake inhibitors (SSRIs). However, much of what is known about antidepressants is derived from research in depression rather than anxiety. In recent years, researchers have found that antidepressants are more effective for severely depressed patients than they are for patients with milder symptoms. It is possible that a similar relationship between symptom severity and antidepressant efficacy exists for anxiety disorders, but there is currently little evidence available to answer this question. As antidepressants are frequently prescribed to patients with mild or moderate anxiety, a clear understanding of their effectiveness across the severity range is vital to inform treatment decisions. Therefore, the purpose of this meta-analysis is to examine whether initial symptom severity affects antidepressant efficacy for anxiety disorders.

Study design:

The proposed project is an individual patient data meta-analysis. The investigators will collate data from 29 randomized controlled trials of second-generation antidepressants (specifically: paroxetine, paroxetine controlled release (CR), duloxetine, and fluoxetine) for the short-term treatment of an anxiety disorder, including a total of approximately 8,800 participants. The anxiety disorders that are included in the proposed project are generalized anxiety disorder (GAD), social anxiety disorder (SAD), obsessive-compulsive disorder (OCD), post-traumatic stress disorder (PTSD), and panic disorder (PD). Generalized linear mixed models will be used to investigate whether initial symptom severity is related to antidepressant efficacy.

Statistical analysis plan:

A separate longitudinal analysis will be conducted for each disorder. For GAD, SAD, OCD, and PTSD, linear mixed models will be used; for PD, a generalized linear mixed model (multilevel negative binomial regression) will be used, as the dependent variable for this disorder (number of panic attacks) is a discrete count variable. Maximum likelihood estimation will be used as the estimation method for the linear mixed models, while Laplace approximation will be used as the estimation method for the multilevel negative binomial regression for PD. In all models, measurement occasion represents level 1, participants represent level 2, and trial represents level 3. The effect measure of interest is the change in symptoms from baseline, except for PD, for which the effect measure of interest is the total number of panic attacks per two weeks.

The initial model will be built by including all the fixed effects of interest, regardless of significance. These include initial severity, treatment group and covariates (see below). Linear and quadratic terms for time (in days since baseline) will be included. For each participant, the actual visit dates will be used (if available) rather than the intended weekly visit date. The following two- and three-way interactions will also be included: severity x group, severity x linear time, group x linear time, severity x group x linear time, severity x quadratic time, group x quadratic time, severity x group x quadratic time.

Using this first model, the variance-covariance structure of the nested data will be modeled by including random effects. Random effects for study, subject and (linear and quadratic) time, as well as various covariance structures (unstructured, autocorrelated errors, Toeplitz, etc.) will be considered. Restricted maximum likelihood (REML) will be used for estimation, and the best-fitting model will be selected based upon the Akaike Information Criterion (AICc).

In this best-fitting model, backward selection with maximum likelihood (ML) will be used to select the significant fixed effects. Non-significant interaction terms will be removed from the model (unless the three-way interaction of group x severity x (linear or quadratic) time is significant, in which case all two-way interactions and main effects that use these variables will be retained). The best-fitting model will again be selected based upon the Akaike Information Criterion (AICc).

Covariates: Models with and without the following covariates will be tested: age, gender, and duration of illness (if available). Only the main effect of these covariates will be included; no interactions with other variables will be included.

Missing data: Baseline variables (such as initial severity) are likely to be essentially complete, but some change scores are likely to be missing due to dropout or missed measurement occasions. The assumption is made that these data are missing at random (MAR), that is: missingness of the dependent variable may depend upon observed variables (such as previous symptom scores or covariates), but it does not depend upon the value of the unobserved (missing) variable. When MAR holds, the mixed model yields unbiased estimates of coefficients and standard errors even when some data is missing, and no other methods for handling missing data are required.

Inclusion Criteria:

• Inclusion in the intent-to-treat population of a randomized controlled trial of a second-generation antidepressant for the short-term treatment of an anxiety disorder
• Diagnosed with an anxiety disorder (GAD, SAD, OCD, PTSD or PD)

Exclusion Criteria:

• Assignment to an active comparator that is not a second-generation antidepressant (e.g. a benzodiazepine)

The study population consists of adults aged 18 and older (with the exception of OCD trials, which may include adolescents aged 14 and older) who have been diagnosed with generalized anxiety disorder (GAD), social anxiety disorder (SAD), obsessive-compulsive disorder (OCD), post-traumatic stress disorder (PTSD), or panic disorder (PD) and who were randomized to placebo or to treatment with a second-generation antidepressant.