Skip to main content

Table 2 Methodological challenges when developing and validating a risk prediction model using IPD from multiple studies as identified from those 15 articles in our review (written below in a framework similar to recommendations by Abo-Zaid et al.[37] for prognostic factors)

From: Developing and validating risk prediction models in an individual participant data meta-analysis

Methodological issue Challenge
Identifying relevant studies • Unavailability of IPD in some studies
Issues within studies • How to assess quality of studies available
• Inability of IPD to overcome deficiencies of original studies, such as missing participant data or of being low methodological quality.
Heterogeneity across studies • Dealing with different definitions of disease or outcome
• Dealing with different (or out-dated) treatment strategies, especially when a mixture of older and newer studies are combined
Statistical issues for meta-analysis • Dealing with a mixture of IPD from retrospective and prospective studies
• Missing data, including: missing predictor values and missing outcome data for some participants within a study, and completely unavailable predictors in some studies
• Difficulty in using a continuous scale for continuous factors in meta-analysis when some IPD studies give values on a continuous scale and others do not
• Dealing with IPD from trials where both control and treatment groups are available
Assessment of potential biases • How to assess the impact of excluded studies who did not provide IPD
Model development • Accounting for clustering of patients within different IPD studies
• Allowing for heterogeneity in baseline risk (intercept term) across studies
• Allowing for heterogeneity in predictor effects across studies
Model validation • Lack of external validation if all studies used for model development
• Sample size required to implement the internal-external approach (i.e. sample size of studies to be excluded, and also the total number of IPD studies needed)