Reported methods for handling missing change standard deviations in meta-analyses of exercise therapy interventions in patients with heart failure: A systematic review

Abstract

BackgroundWell-constructed systematic reviews and meta-analyses are key tools in evidenced-based healthcare. However, a common problem with performing a meta-analysis is missing data, such as standard deviations (SD). An increasing number of methods are utilised to calculate or impute missing SDs, allowing these studies to be included in analyses. The aim of this review was to investigate the methods reported and utilised for handling missing change SDs in meta-analyses, using the topic of exercise therapy in heart failure patients as a model.MethodsA systematic search of PubMed, EMBASE and Cochrane Library from 1 January 2014 to 31st March 2018 was conducted for meta-analyses of exercise based trials in heart failure. Studies were eligible to be included if they performed a meta-analysis of change in exercise capacity in heart failure patients after a training intervention.ResultsTwenty two publications performed a meta-analysis on the effect of exercise therapy on exercise capacity in heart failure patients. Eleven (50%) publications did not directly report the approach for dealing with missing change SDs. Approaches reported and utilised to deal with missing change SDs included imputation, actual and approximate algebraic recalculation using study level summary statistics and exclusion of studies.ConclusionChange SDs are often not reported in trial papers and while in the first instance meta-analysts should attempt to obtain missing data from trial authors, this information is frequently not forthcoming. Meta-analysts are then forced to make a decision on how these trials and missing data are to be handled. Whilst not one approach is favoured for dealing with this matter, authors need to clearly report the approach to be utilised for missing change SDs. Where change SDs are imputed meta-analyst are encouraged to explore several options and have a sound rationale as to the choice, and where data is imputed, sensitivity analysis should be conducted.

Physical relaxation for occupational stress in healthcare workers: A systematic review and network meta-analysis of randomized controlled trials

“…For the primary outcome, we obtained mean changes of stress scores and standard deviation from baseline for both arms, with conversion from pre-and post-intervention stress scores if not otherwise available using a correlation coefficient of zero. 38,39 When multiple scales are present in a study, preference was given to measures more specific for stress and those that are more commonly used in other identified studies. In decreasing order of preference, these are the Perceived Stress Scale (PSS), the Maslach Burnout Inventory for emotional exhaustion (MBI-EE), the Nursing Stress Scale (NSS), and others if these are not available.…”

Effectiveness and tolerability of rectal ointment and suppositories containing sucralfate for hemorrhoidal symptoms: a prospective, observational study

“…Regarding sample size, the aim was to achieve a precision estimate of at least 0.6 points in terms of the halfwidth of the 95% confidence interval (CI). As the effect size was unknown and no previous studies with the products were available, the mean of the baseline standard deviations was used [ 26 ] from a study of a different ointment (containing sucralfate) with similar indications [ 20 ]. Thus, under the assumption that the standard deviation of the change in mSSS is 1.5 points, and the precision estimate in terms of the 95% CI is ± 0.6 points, a minimum of 24 patients had to be included in the analysis for both products.…”

A Note on Estimating Unreported Sample Statistics for Meta-Analysis

“…Sixty-eight percent of the researchers surveyed by Weir, et al [3] had encountered studies with unreported sample statistics when conducting systematic reviews. Batson and Burton [4] discovered that 41 percent of meta-analyses of diabetes studies reported specific efforts, such as imputation, to address the problem of missing variances, and Pearson and Smart [5] obtained similar findings among meta-analyses of studies that examined the effect of exercise in heart failure patients. Sangnawakij, et al [6] found 38 studies of surgical procedures that reported neither variances nor standard deviations, and Wiebe, et al [7] examined more than 100 systematic reviews that were plagued by missing variance data.…”

“…Such work is often seriously impeded, however, when the primary studies fail to report basic statistical information from their samples, such as means or standard deviations. This problem of unreported sample statistics in published clinical trials and other basic research is well-documented ( [1][2][3][4][5][6][7]). For example, 87 percent of the randomized, controlled trials of antidepressants reviewed by Streiner and Joffe [1] failed to fully report the mean, standard deviation, and/or sample size; similarly, more than 80 percent of the clinical trials reviewed by Chan, et al [2] reported insufficient statistical information for inclusion in meta-analysis.…”