Metformin & glitazones: Do they really help PCOS patients?

Outcome measures

We studied the parameters that are needed for the diagnosis of PCOS, as well as parameters that are associated with the syndrome’s comorbidities. The measured variables included: ovulation rates, pregnancy rates, BMI, waist-to-hip ratio, lipid panel, blood pressure, fasting insulin levels, fasting blood glucose, C-peptide, glycosylated hemoglobin (Hb A_1c), LH/FSH, total testosterone, free testosterone, androstenedione, DHEAS, leptin, C-reactive protein, hirsutism (based on the Ferriman-Gallwey [F-G] score), and weight.

Methods of the review

Each included trial was evaluated in detail regarding how well it met the inclusion and exclusion criteria, the number of participants, the follow-up period, quantitative reporting of the data, and the overall methodology. The principal author rated methodological quality as good, fair, or poor on the basis of an overall assessment of these features. We did not use explicit validity checklists with summary scores because they have not been shown to predict the effect of bias on treatment differences or to provide more reliable assessments of validity.^11,12

Description of studies

All included studies met the inclusion criteria. However, it is important to note that most of the studies had low numbers of participants (only 4 studies had a sample size [n] >50, 3 studies with n >100). Some of the trials shared the same patients but analyzed different end points so the fundamental “independence” assumption required for most standard statistical analyses, including meta-analysis, was likely violated.

A few of the trials used another pharmacological agent or invasive procedure as control treatment. Some trials had designs such that in the end, the treatment and control groups both received ovulation induction agents for the patients who failed to ovulate. Even with the strict inclusion and exclusion criteria, many of the included trials still were of less than satisfactory design quality for our purposes.

Data collection/extraction

The principal author reviewed the text, tables, and figures and then collected and extracted the data from relevant publications. The data set was reviewed by the secondary authors for errors in data entry, format, outliers, or implausible values.

Statistical analysis

For each analysis, we converted all data to the same metric based on conversion formulas provided by each individual trial. We also verified the conversion factors online via data provided by standard clinical and reference laboratory values.

Some trials provided standard deviations (SD), while others provided standard error (SE) or standard error of the mean (SEM). We derived pooled variances accordingly. We converted all SDs to SE using the equation SE=SD/square root (N), where N denotes the sample size. We created data sets in the above fashion for each parameter.

Trials that did not report SD, SE, or sample variance for a given parameter were not included in the corresponding meta-analysis. We did not conduct a statistical analysis if only a small number of trials existed (ie, n <5), as it would result in less reliable conclusions.

No article provided individual level data or SD/SE for the “change” in the selected endpoint before and after treatment in the control and treatment arm. Instead, most publications presented the summary statistics separately for before and after treatment.

Since covariance (or correlation) between before and after values was not available and variability of the difference measure could not be estimated from the majority of the trials, it was not possible to perform methodologically sound meta-analysis by addressing the absolute or percent change before and after treatment and comparing this difference measure between 2 competing treatments. This is an inherent problem in many meta-analyses due to limited raw data.

However, as all the trials were randomized, we felt justified in performing the statistical analysis using the mean difference after treatment in the control and intervention arms. Our assumption: the values before treatment should be reasonably balanced between the 2 arms. Thus, we calculated the pooled estimates of treatment effect with 95% confidence intervals (CIs) for the mean differences between the control and intervention arms after treatment. We adopted the random effects model approach.¹³

Next, statistical significance was evaluated for treatment effect and heterogeneity. Publication bias was also examined by two different tests.^14,15 Sensitivity analysis was also performed to assess the impact of the identification of potential hidden studies by the trim and fill method.¹⁶

A 2-sided hypothesis with type I error of 5% was employed in all statistical testing and CI construction. Statistical analyses were carried out by STATA version 8.2.¹⁷