Comparative effect size distributions in strength and conditioning and implications for future research

A meta-analysis

Abstract

Background Controlled experimental designs are frequently used in strength and conditioning (S&C) to determine which interventions are most effective. The purpose of this large meta-analysis was to quantify the distribution of comparative effect sizes in S&C to determine likely magnitudes and inform future research regarding sample sizes and inference methods. 

Methods Baseline and follow-up data were extracted from a large database of studies comparing at least two active S&C interventions. Pairwise comparative standardised mean difference effect sizes were calculated and categorised according to the outcome domain measured. Hierarchical Bayesian meta-analyses and meta-regressions were used to model overall comparative effect size distributions and correlations, respectively. The direction of comparative effect sizes within a study were assigned arbitrarily (e.g. A vs. B, or B vs. A), with bootstrapping performed to ensure effect size distributions were symmetric and centred on zero. The middle 25, 50, and 75% of distributions were used to define small, medium, and large thresholds, respectively.

Results A total of 3874 pairwise effect sizes were obtained from 417 studies comprising 958 active interventions. Threshold values were estimated as: small = 0.14 [95%CrI: 0.12 to 0.15]; medium:  0.29 [95%CrI: 0.28 to 0.30]; and large = 0.51 [95%CrI: 0.50 to 0.53]. No differences were identified in the threshold values across different outcome domains. Correlations ranged widely (0.06 ≤ r ≤0.36), but were larger when outcomes within the same outcome domain were considered.

Conclusions The finding that comparative effect sizes in S&C are typically below 0.30 and can be moderately correlated has important implications for future research. Sample sizes should be substantively increased to appropriately power controlled trials with pre-post intervention data. Alpha adjustment approaches used to control for multiple testing should account for correlations between outcomes and not assume independence.