Meta-analyses of randomized clinical trials (RCTs) are crucial for informing clinical guidelines and policies by synthesizing evidence. RCTs aim to minimize bias by creating similar groups at baseline, implying that any post-intervention differences are attributable to the interventions. However, baseline imbalances in prognostic factors (age, disease severity) can arise by chance or due to non-random reasons. Insecure allocation concealment (investigators influencing participant allocation) can introduce bias. Deviation from intention-to-treat analysis (e.g., per-protocol analysis) and selective reporting of data also distort results. Meta-analysis of baseline values can reveal bias, potentially compromising the validity of results. Previous research has demonstrated such imbalances; for example, a meta-analysis of calcium supplements for weight loss showed a lower baseline body mass in treatment groups, and a meta-analysis of oseltamivir for influenza revealed bias in treatment allocation. The authors observed baseline imbalance in their previous meta-analysis comparing isometric exercise and controls for hypertension, where the exercise group had significantly higher baseline blood pressure. This imbalance was largely driven by a single large study. Therefore, this meta-epidemiological study investigated baseline imbalance in comparisons of various exercise types and antihypertensive medicines, also exploring associations with sample size and selection bias.

The introduction cites several prior studies demonstrating baseline imbalances in meta-analyses. A meta-analysis of calcium supplements showed a significant difference in baseline body mass between treatment and control groups, influencing the overall effect size. Another study on oseltamivir for influenza found allocation bias evident in baseline data. The authors also refer to their previous work, which revealed baseline blood pressure imbalances in a meta-analysis of isometric exercise for hypertension, highlighting the potential for such biases to affect the validity of results and underscoring the need for the current study.

The study, preregistered on the Open Science Framework, analyzed data from 391 RCTs (197 exercise trials and 194 antihypertensive medicine trials) from a previously published network meta-analysis. Two authors independently extracted baseline data (systolic blood pressure (SBP), diastolic blood pressure (DBP), and age) from each RCT, prioritizing data from all randomized participants. Data transformations were applied where necessary (e.g., converting median and range to mean and standard deviation). Intervention groups within RCTs using the same exercise mode or medicine type were combined. Risk of bias in allocation concealment was assessed using the Cochrane Risk of Bias tool. Duplicate baseline data from four studies were removed. Fixed-effect meta-analyses assessed baseline imbalance and inconsistency. Meta-regression analyses examined associations between baseline imbalance and sample size, allocation concealment risk, and whether data were reported for all randomized participants. Four studies provided duplicate baseline data, and these duplicates were removed.

The analysis of 190 exercise RCTs revealed no baseline SBP imbalances across comparisons. However, substantial inconsistency was observed in comparisons of resistance exercise vs. control ($I^2$ = 33.0%), endurance exercise vs. control ($I^2$ = 14.4%), and resistance exercise vs. combined exercise ($I^2$ = 22.0%). No significant moderator effects were found for SBP. For DBP, no baseline imbalances were found, but substantial inconsistency existed in several comparisons (endurance vs. control, resistance vs. control, resistance vs. combined). Sample size significantly moderated the DBP differences in endurance vs. control and resistance vs. control comparisons. Reporting data for all participants also significantly moderated DBP differences in endurance vs. control and resistance vs. combined comparisons. In the age analysis, a baseline imbalance was found in the resistance vs. control comparison, with the resistance group being 0.3 years younger (95% CI 0.6 to 0.1). Inconsistency was also present in the endurance vs. control comparison ($I^2$ = 14.1%). The analysis of antihypertensive medicine trials showed less data availability; however, no baseline imbalances and few inconsistencies were identified.

The findings indicate the presence of baseline imbalances and inconsistency in some exercise comparisons, particularly concerning DBP and age in resistance training studies. The inconsistency in the exercise comparisons suggests that the effect of exercise on blood pressure may vary considerably across studies. These imbalances and inconsistencies highlight the importance of considering baseline characteristics when interpreting meta-analyses of exercise interventions for blood pressure. The lack of similar imbalances in the medication analysis could be due to stricter methodological approaches in pharmaceutical trials. This study underscores the need for meta-analyses to include analyses of key baseline prognostic variables to ensure balance and enhance the reliability of treatment effect estimates. While this study has focused on blood pressure, these findings have implications for other meta-analyses across diverse fields.

This meta-epidemiological study revealed baseline imbalances and inconsistencies in some exercise comparisons within a network meta-analysis of interventions for blood pressure management. These findings highlight the importance of considering and reporting baseline prognostic variables in future research and meta-analyses to ensure balance across trials and improve the validity of effect estimates. Future research should focus on exploring the reasons for these imbalances and inconsistencies and developing strategies to mitigate them in future RCTs.

The study's reliance on previously published data limits its ability to address specific methodological choices made in the original RCTs. The data extraction process relied on available published data, and some data were extracted from figures, which might introduce error. The large number of studies and the age of some included articles prevented the authors from contacting original authors for clarification or missing data. The heterogeneity observed across studies could be due to factors not considered in the current analysis.