Pooling individual participant data from randomized controlled trials: Exploring potential loss of information

Area of Science:

• Biostatistics
• Data Science
• Epidemiology

Background:

• Pooling individual participant data (IPD) is crucial for robust analyses but challenged by data heterogeneity.
• Recoding original variables is often necessary for harmonizing diverse datasets in IPD studies.
• Quantifying information loss during recoding is essential to ensure the validity of pooled analyses.

Purpose of the Study:

• To quantify the extent of information lost when recoding variables for pooled data analysis.
• To assess the impact of this information loss on the validity of subsequent analyses.
• To identify conditions under which information loss is minimized.

Main Methods:

• Information loss was quantified using R-squared values from linear regression models comparing pooled and original variables.
• The impact on analysis validity was assessed by comparing regression coefficients using original versus recoded variables.
• Simulations were conducted by recoding continuous variables and varying parameters like range and ratio of recoded values to estimate information loss.

Main Results:

• R-squared values below 0.8 were observed for 8 out of 91 recoded variables.
• A substantial impact on regression models was noted in 4 cases, particularly when continuous variables were recoded into discrete ones.
• Simulations indicated that information loss is minimized when the ratio of recoded zeroes to ones is approximately 1:1.

Conclusions:

• While large pooled datasets offer significant analytical power, data harmonization through recoding requires careful consideration.
• Recoded variables with limited explained variance from their original counterparts may compromise the validity of research findings.
• Researchers should exercise caution and validate recoded variables to ensure the integrity of pooled data analyses.