Methods to analyze continuous outcomes by incorporating baseline data in individual participant data meta-analysis of non-randomized studies

Tags: Oral
Syrogiannouli L1, Wildisen L1, Efthimiou O2, Del Giovane C1
1Institute of Primary Health Care (BIHAM), University of Bern, 2Institute of Social and Preventive Medicine (ISPM), University of Bern

Background: contrary to randomized studies, non-randomized studies, where a continuous outcome of interest is assessed at baseline and follow-up, are likely to show baseline imbalance between the treatment or exposure and the control group. This may confound the study estimate and consequently the meta-analysis (MA) estimate. Follow-up data as the outcome, adjusting for baseline (ANCOVA) or the change from baseline are commonly used to analyze such data. While recommendations on these methods have been provided for MA of randomized studies, less is known for MA of non-randomized studies. Studies may have used different methods to analyze continuous outcomes and therefore aggregate estimates derived differently may influence the pool result. Analysis of individual participant data (IPD) allows review authors to standardize the statistical method across studies.


1) to identify the methods used in published IPD-MA of non-randomized studies for continuous outcomes;

2) to compare several methods that take account of baseline data to analyze continuous outcomes in IPD-MA of non-randomized studies.

Methods: we searched Embase, MEDLINE Ovid, and Cochrane from inception up to April 2019 to identify studies that applied IPD-MA to synthesize continuous outcomes from non-randomized studies and incorporated baseline outcome data in the analysis.

We will describe the statistical methods used in each IPD-MA. For the second objective, we will consider the following:

1) follow-up values as outcome, ignoring baseline values;

2) ANCOVA, assuming linear confounding effect;

3) change from baseline as outcome; and

4) propensity score, including the conditional probability of being assigned to a treatment/exposure group given the baseline data as covariate in analysis.

We will also consider a variation of ANCOVA that includes the baseline values as a spline term, to allow for potential non-linear confounding effect. We will apply each method to IPD of cohort studies to investigate if patients with thyroid dysfunction (exposure) are at risk of depressive symptoms (outcome), measured on a validated depression scale similar across studies. We will describe the differences in results across methods.

Results: for the second objective, we collected IPD from six cohort studies, including 24,524 participants, with data on depressive symptoms at baseline and first available follow-up, and thyroid status at baseline. We will present the results of the two objectives at the Colloquium.

Conclusions: different methods that are used to deal with baseline data in the analysis of continuous outcomes in non-randomized studies may lead to different and potentially inappropriate MA estimates. We plan to provide guidance on choice of methods.

Patient or healthcare consumer involvement: we did not involve patients or healthcare consumers