Follow-up or change scores: does it matter which outcome is used in meta-analysis of randomized controlled trials?

Article type
Authors
Hayden J, Tomlinson G
Abstract
Background: In meta-analysis of continuous data the most commonly used techniques compare mean follow-up scores across treatment groups. This approach assumes equal baseline scores across treatment groups within studies, and does not incorporate differences in baseline scores across studies. This may be inappropriate if there are not comparable baseline scores within or between trials. Alternative methods require assumptions about variance scores or employ meta-regression analyses. Implications of the different approaches have not been explored.

Objectives: We use a recently conducted Bayesian meta-regression analysis of a Cochrane review on exercise therapy for treatment of low back pain to demonstrate and compare analytical options for outcome score used.

Methods: A systematic review was conducted within the framework of the Cochrane Back Review Group. Analyses of different choices of outcome score include graphical presentation and descriptive summary of baseline, follow-up, within- and between-group differences, and percent change scores. Four approaches to measuring treatment group outcomes are used to assess treatment effects: 1. mean outcome scores at follow-up (traditional analysis), 2. mean change from baseline 3. percent change from baseline (2 and 3 assuming correlations between baseline and follow-up of 0.25, 0.5, 0.75 and a 'best estimate'), and 4. meta-regression analysis adjusting for baseline scores. All analyses are conducted using Bayesian random-effects analyses with non-informative priors.

Results: 61 randomized, controlled trials (6390 participants) of exercise therapy were included in the review: acute (11 trials), subacute (6 trials), and chronic (43 trials) low back pain. Mean treatment effects were assessed in the more homogeneous chronic low back pain population (72 exercise treatment and 31 comparison groups). The overall exercise treatment effect using mean follow-up scores was 3.4 points (95% credible interval: 2.0,4.7) for improvement in pain outcomes and 1.4 points (0.7,2.2) for improvement in function outcomes. Results for change scores and meta-regression will be presented at the Colloquium.

Conclusions: Complete results and graphical presentation will be available and presented at the Cochrane Colloquium. Differences in the information garnered, advantages and disadvantages of the different approaches, and the interpretation of the results will be discussed.