The protocol for this review is published elsewhere.35
Studies were eligible for inclusion if they were randomised placebo controlled trials of at least three months of vitamin D supplementation, were carried out in healthy children or adolescents aged between 1 month and 20 years, and measured primary outcomes of areal or volumetric bone mineral density, bone mineral content, or quantitative ultrasound measures (broadband ultrasound attenuation, speed of sound, stiffness index).
Although fractures in later life would be the ideal outcome measure, this would require following large numbers of children as participants for decades. No such studies have been done, so we used bone density measures as surrogate outcomes, as is commonly seen in children.36
Bone mineral density and content7 8 9
and quantitative ultrasound measures37
have all been associated with fracture risk.
Identification of studies
The electronic literature search was last updated on 9 August 2009. Without language restrictions, we searched the Cochrane Central Register of Controlled Trials, Medline (1966 to present), Embase (1980 to present), CINAHL (1982 to present), AMED (1985 to present), and ISI Web of Science (1945 to present). Web extra appendix 1 gives the search strategy used for Medline, which was adapted as appropriate for other databases.
We hand searched conference abstract issues of key journals for 2007-8: Osteoporosis International, Journal of Bone and Mineral Research, Asia Pacific Journal of Clinical Nutrition, Journal of the American Dietetic Association, Proceedings of the Nutrition Society, Journal of Nutrition. We examined the reference lists and ISI citations of all included studies. Two reviewers assessed potentially relevant articles against the inclusion criteria.
Data collection and analysis
Two reviewers independently extracted data. We extracted change in bone mineral density and bone mineral content as percentage change from baseline, as well as sex, age, pubertal stage, baseline serum vitamin D levels, ethnicity, type and dose of vitamin D given, and compliance as possible effect modifiers, along with data on adverse events. No studies reported quantitative ultrasound outcomes. Two reviewers independently assessed each trial’s risk of bias, assessing factors such as randomisation, allocation concealment, blinding, completeness of outcome assessment, and selective reporting. Where necessary we contacted authors to obtain information on primary outcome factors.
We converted bone density outcomes to standardised mean differences, calculating a standardised mean difference of the percentage change from baseline in treatment and control groups at each site for which there were sufficient data (bone mineral density for forearm, hip, and lumbar spine and total body bone mineral content). Where clinically useful, we estimated a benefit in units of percentage change since baseline from the standardised mean differences by estimating the pooled standard deviation from the means of the standard deviation of the outcomes in treatment and control groups for each study, and multiplying the standardised mean differences by this.38 39 40
We calculated statistical heterogeneity using a χ2 test on N-1 degrees of freedom, with significance conservatively set at 0.10. We also assessed inconsistency I2 using the formula [(Q-df)/Q]×100%, where Q is the χ2 statistic and df is its degrees of freedom, to describe the percentage of the variability in effect estimates due to heterogeneity. We considered a value greater than 50% as denoting substantial heterogeneity.
Meta-analysis was done according to a fixed effects model. When heterogeneity was considered substantial, we explored its causes by carrying out prespecified subgroup analyses where data were available—that is, subgroups by sex, pubertal stage, dose of vitamin D, baseline vitamin D levels, compliance, and adequacy of allocation concealment. Data were insufficient to consider ethnicity and sun exposure as possible sources of heterogeneity, as planned in the original protocol. Where there were not clear clinical reasons or study methodology reasons for substantial heterogeneity between studies, we proceeded to meta-analysis using random effects models. All analyses were carried out in Review Manager 5 (version 5.0.16).
We performed a priori subgroup analyses by sex, pubertal stage, dose of vitamin D, and baseline vitamin D levels to determine whether the effects of supplementation varied by these factors. As the numbers of studies were small, the cut-offs of 200 IU for vitamin D dose and 35 nmol/L for baseline vitamin D concentrations were chosen on the basis of their being sufficient data available at these cut-offs to allow for subgroup analyses. When possible we used intention to treat data in analyses, but if these were not available we used, in order of preference, data from available data or per protocol analyses. Assessment of publication bias was by funnel plot.
Where studies had more than one vitamin D dosage group, we combined the data from all dosage groups and compared this with placebo. The exception to this was in subgroup analyses by vitamin D dose—for those studies with a dosage group falling into each subgroup we used the data from the individual dosage group for the treatment arms and used the control group mean and standard deviation for each subgroup comparison but reduced the number in the control groups by half for each comparison.
One study was a cluster randomised controlled trial,32
with data that did not account for clustering. We used the subsequently published intracluster correlation for total body bone mineral content41
in a sensitivity analysis, correcting the sample size for the effect of clustering by the design effect of 1.946, calculated by 1+(M−1)ICC, where M is the cluster size and ICC the intracluster correlation (87 and 0.011, respectively).