Detailed description of the review methods is available elsewhere.16
Briefly, we included studies if they were randomised controlled trials of calcium supplementation (including in food) compared with placebo, with a treatment period of at least three months; participants were children (age < 18) without coexisting medical conditions or treatments affecting bone metabolism; outcome measures included areal or volumetric bone mineral density, bone mineral content, broadband ultrasound attenuation, or speed of sound measured at any site; and bone outcomes were measured after at least six months of follow-up.
We applied the search strategies (see bmj.com
for details) to the following electronic bibliographic databases: CENTRAL, (Cochrane Central Register of Controlled Trials) (issue 3, 2005), Medline (1966 to 1 April 2005), Embase (to 1 April 2005), CINAHL (1982 to 1 April 2005), AMED (1985 to 1 April 2005), MANTIS (1980 to 1 April 2005), ISI Web of Science (1945 to 1 April 2005), Food Science and Technology Abstracts (1969 to 1 April 2005), and Human Nutrition (1982 to 1 April 2005). We also hand searched conference abstract books (Osteoporosis International
91990, Journal of Bone and Mineral Research
Two reviewers (TW and KS) independently assessed articles for inclusion, extracted data, and assessed quality. Quality assessment included assessment of randomisation, allocation concealment, blinding, and description of withdrawals and dropouts and was used to give an overall rating of the risk of bias.17,18
In the absence of trials in children with fracture as an outcome, we used bone mineral density, bone mineral content, broadband ultrasound attenuation, and speed of sound as outcomes measures. These outcomes measured at any site were all considered a priori primary outcomes, but their inclusion in the analysis depended on the availability of adequate data. We had data for total body bone mineral density and for density at the femoral neck, lumbar spine, distal radius, and upper limb (defined as the distal radius or the upper limb site closest to that point). We used endpoint data rather than change data to maximise data availability.
We converted outcome measures to standardised mean differences using RevMan version 4.2.7. We assessed heterogeneity of the data with a χ2 test and conducted a meta-analysis according to a fixed effects model for the main effect outcomes. Where there was heterogeneity in subgroup analyses we used random effects models.
Subgroup analyses were performed for sex, baseline calcium intake, pubertal stage, ethnicity, physical activity, type of supplementation (milk extract compared with other calcium supplement forms), and duration of supplementation. The median mean baseline calcium intake for each study (794 mg/day) was used as the cut-off for calcium intake subgroups. We had sufficient studies to also perform this analysis using a lower cut-off (25th centile, 582 mg/day) for upper limb bone mineral density. For the physical activity analysis, where physical activity was a cointervention or subgroup in a study, we included those in the low physical activity arm in the low physical activity subgroup for the meta-analysis and those in the high physical activity arm in the high physical activity subgroup. We chose a cut point of 18 months for the study duration subgroups, so as to have exceeded the bone remodelling transient. We also performed a subgroup analysis by whether the calcium intake in the intervention group in the trial exceeded the probable threshold (1400 mg/day) below which skeletal accumulation varies with intake and above which skeletal accumulation seems constant regardless of intake.19,20
We collected data on adverse effects, where available. When necessary we contacted the authors of the primary studies to obtain additional information.
We used intention to treat data from trials wherever possible. If these were not available, we used data from the analysis of available data or, if no other data were available, data from the analysis of treatment received. For the single study in which upper limb outcomes were presented as percentage change from baseline, and no endpoint data were available,w35 we imputed endpoint data using the baseline bone mineral density and percentage change from baseline and the standard deviation (SD) of the baseline data for the endpoint SD. Where studies reported absolute change from baseline and endpoint data were not availablew1 w7 w11 w19 w33 we imputed endpoints using baseline plus change for the mean and using the SD of the baseline data for the endpoint SD. We performed a sensitivity analysis for the main effects omitting studies for which data were imputed,w1 w7 w11 w19 w33 w35 and a sensitivity analysis omitting the studyw1 that reported results only from analysis of treatment received.
To aid in the assessment of the clinical relevance of the results, we used the standardised mean difference to estimate an absolute benefit in mg/cm2
We then estimated the relative difference in the change from baseline as the absolute benefit divided by the mean of all the baseline means of the control groups, expressed as a percentage. We used funnel plots to assess publication bias.