Data analysis

a. Exploratory data analysis

Funnel plots were used in exploratory data analysis to assess for the potential existence of small sample bias. An asymmetrical funnel plot, however, has several explanations, including true heterogeneity of effect with respect to study size, poor methodological design of small studies (Sterne 2001; Tang 2000; Thornton 2000), and publication bias. Thus, we did not place undue emphasis on this tool.

b. Statistical pooling

Where data from RCTs were thought to be sufficiently homogeneous with respect to interventions and outcomes, we calculated pooled effect sizes. For continuous variables reported on the same scale we calculated weighted mean differences. The absolute differences in outcome between each follow-up and the baseline measure for the intervention and comparison study group (ΔI and ΔC) were calculated and inserted in Review Manager Software (Review Manager 4.2). When the estimate of variance of (ΔI) and (ΔC) was not given, it was calculated from the outcome measures in each study group using the formula Vpre+ Vpost - 2r(SDpre*SDpost), where Vpre is the variance of the mean baseline outcome, Vpost is the variance of the mean follow-up outcome, r is the correlation between the baseline and follow-up values, and SDpre and SDpost are the standard deviations of the baseline and follow-up groups, respectively. Since most studies do not report r, and its true value is unknown, data are presented with r = 0.75, and a sensitivity analysis was performed as described below.

Data were pooled using the random effects model and using the DerSimonian and Laird formula for calculating between-study variance (DerSimonian 1954). Each study was weighted by the inverse of the study variance. Heterogeneity between trial results was tested using a standard chi-square test (Cochran 1954) with a significance level of alpha = 0.1 in view of the low power of such tests. When we found heterogeneity we tried to explain it by examining individual study characteristics and those of subgroups of the main body of evidence. When heterogeneity was thought to be too great to meaningfully pool the results quantitatively, the results are presented in a narrative fashion.

c. Regression analyses

We performed a meta-regression to determine whether various study-level characteristics affect weight change and GHb. The meta-regression was also weighted by the inverse of the variance of (ΔI - ΔC). Interaction terms were examined for all models. The study-level variables examined in the metaregression model included follow-up interval, the number of contacts between the care provider and participants, and the percentage attrition in the intervention group. SAS was used to perform the meta-regression (version 8.01, SAS Institute Inc., Cary, NC).

d. Subgroup analyses

We planned analyses by the following subgroups if there was a significant change in weight and the amount of data would allow meaningful analyses:

• Overweight (25.0 <BMI <30.0), obese (BMI >30.0), normal weight (BMI <25.0);

• Age: young (<40 years), middle-aged (40 to 65 years), old (>65 years);

• Treatment: on insulin, oral agents, diet only;

0 0

Post a comment