Drug Treatment of Primary Insomnia: A Meta-Analysis of Polysomnographic Randomized Controlled Trials

Abstract

Context

Although insomnia is a frequent health complaint that is often treated with drugs, little is known about differences in treatment efficacy of various drug classes on objective versus subjective outcome measures.

Objective

Our aim was to compare treatment efficacy of classical benzodiazepines, benzodiazepine receptor agonists (zopiclone, zolpidem and zaleplon), antidepressants (including low-dose doxepin), neuropeptides, progesterone receptor antagonists, hormones, melatonin receptor agonists, antihistamines, antiepileptics, and narcotics addressing primary insomnia.

Data Sources

We conducted a comprehensive literature search (up to 5 April 2013) using PubMed, Cochrane Clinical Trials, PQDT OPEN, OpenGREY, ISI Web of Knowledge, PsycINFO, PSYNDEX, and the WHO International Clinical Trials Registry Platform.

Eligibility Criteria

Only polysomnographic, parallel-group, randomized controlled drug trials were included; eligibility was determined by two independent authors.

Data Synthesis

We used a random effects model, based on 31 studies reporting 80 treatment conditions, covering 3,820 participants.

Results

Effect size estimates for the total sample of pooled drug classes suggest that there is a small-to-moderate, significant, and robust effect for objective outcomes (sleep onset latency g = −0.36, total sleep time g = 0.27) and subjective outcomes (sleep onset latency g = −0.24, total sleep time g = 0.21). Results indicate higher effect sizes for benzodiazepine receptor agonists and classical benzodiazepines compared with antidepressants (including low-dose doxepin) and for classical benzodiazepines compared with benzodiazepine receptor agonists. Benzodiazepine receptor agonists demonstrated higher effect sizes for objective outcomes.

Limitations

Data on drug safety were not analyzed.

Conclusions

Future studies should use objective and subjective assessment. Focusing on efficacy, clinicians should favor benzodiazepine receptor agonists and classical benzodiazepines over antidepressants (including low-dose doxepin) for primary insomnia treatment, but the additional consideration of different side effect profiles can lead to alternative treatment decisions.

Appendix

1.1 Detailed Information on Quantitative Data Synthesis and Moderator Analyses

Since comparative effectiveness research (CER) trials result in a higher clinical efficacy of the drug compared with conventional placebo-controlled trials [74], we decided a priori to restrict the searches to placebo-controlled trials.

The intergroup effect sizes were computed using the following formula: $d = \frac{{\bar{X}_{1} - \bar{X}_{2} }}{{\sqrt {\frac{{\left( {n_{1} - 1} \right) S_{1}^{2} + \left( {n_{2} - 1} \right) S_{2}^{2} }}{{n_{1} + n_{2} - 2}}} }}$, where $\bar{X}_{1}$ and $\bar{X}_{2}$ are the sample means, $S_{1}$ and $S_{2}$ are the SDs, and $n_{1}$ and $n_{2}$ are the sample sizes in the intervention condition and the control condition, respectively.

For studies reporting mean change, SD difference, and N in each group, the intergroup effect size was calculated using the following formula: $d = \frac{{\bar{X}_{1} - \bar{X}_{2} }}{{\sqrt {\frac{{\left( {n_{1} - 1} \right) S_{1}^{2} + \left( {n_{2} - 1} \right) S_{2}^{2} }}{{n_{1} + n_{2} - 2}}} }}$, where $\bar{X}_{1}$ and $\bar{X}_{2}$ are the sample mean changes, $n_{1}$ and $n_{2}$ are the sample sizes in the intervention condition and the control condition, respectively, and $S_{1}$ and $S_{2}$ are the SDs determined by the following formula: $S_{x} = \frac{{{\text{SD change}}_{x} }}{{\sqrt {2 (1 - r)} }}$, where SD change _x is the given SD change and r is the pre-post correlation. To calculate controlled effect sizes, the correlation between pre- and post-treatment measures is called for; however, it could not be determined from the study reports. As recommended by Rosenthal [33], we used a conservative estimate of r = 0.70 instead.

Hedges’ g can be computed by multiplying d by a correction factor $J = 1 - \frac{3}{4df - 1}$, where df is the degrees of freedom to estimate the intra-group SD.

Q is determined by the following formula: $Q = \sum\nolimits_{i = 1}^{k} {W_{i} Y_{i}^{2} } - \frac{{\left( {\sum\nolimits_{i = 1}^{k} {W_{i} Y_{i} } } \right)}}{{\sum\nolimits_{i = 1}^{k} {W_{i} } }}$, with W _i being the weight of the study, Y _i the effect size of the study, and k the number of studies included. To determine the expected value of Q, we used the degrees of freedom ($df = k - 1$, with k being the number of studies included). A significant Q test (p value less than alpha set at 0.05) indicates heterogeneity in effect sizes.

I ² is determined by using the following formula: $I^{2} = \left( {\frac{Q - df}{Q}} \right) \times 100\;\%$. I ² is expressed as a ratio, with a range of 0–100 %, and describes what proportion of the observed variance reflects real differences in effect sizes. Higgins et al. [30] suggest that values of 25, 50, and 75 % can be considered as low, moderate, and high, respectively.

We computed the fail-safe N using the following formula: $X = \frac{{K(K\bar{Z}^{2} - 2.706)}}{2.706}$, where K is the number of studies in the meta-analysis and $\bar{Z}$ is the mean Z obtained from the K studies. The effect size can be characterized as robust if the number of studies (X) required to reduce the overall effect size to a non-significant level exceeds 5K + 10 [33].

We used the Trim and Fill method, which examines whether negative or positive trials are over- or under-represented, depending on the sample size. This information can then be used to re-calculate the effect size estimates, if the funnel plot is asymmetric. The divergence of the original effect size and the re-calculated effect size shows the degree of robustness of the results.

Instead of conducting a power analysis, we report the observed effect size with its CI, which is more informative than the statement that power was low [31]. We also did not report Ms and SDs for measurement artifacts because construct-level relationships were not the focus of this analysis.