Re-expressing coefficients from regression models for inclusion in a meta-analysis

Area of Science:

• Biostatistics
• Epidemiology
• Health Research Methods

Background:

• Meta-analysis requires consistent data presentation across studies.
• Non-uniform variable transformations (e.g., log vs. no transformation) hinder result synthesis.
• Recent methods aim to re-express regression coefficients for comparability.

Purpose of the Study:

• To evaluate the bias of three regression coefficient re-expression methods.
• To assess the impact of independent variable skewness on re-expression bias.
• To compare re-expressed coefficients with those from untransformed models.

Main Methods:

• Simulations and 15 real-world data examples were used.
• Independent variables exhibited skewed distributions.
• Regression coefficients from log-transformed variables were re-expressed to an untransformed scale.

Main Results:

• All three re-expression methods typically yielded biased results.
• The degree of bias was predicted by the independent variable's skewness.
• Re-expressed coefficients often differed from those derived from untransformed data.

Conclusions:

• Current re-expression methods for regression coefficients are often biased.
• Variable skewness is a key factor influencing bias in meta-analysis.
• Synthesizing evidence from transformed and untransformed data remains challenging and context-dependent.