Improvements of Statistical Practices Were Strongly Needed Among Prognostic Prediction Models For Obstetric Care

Abstract

Objective: To investigate statistical practices among prognostic prediction models for obstetric care.

Study Design: We searched PubMed for studies reporting development or validation of prognostic prediction models for obstetric care published in a 10-year time span (2011-2021). Teams of investigators independently used structured, pilot-tested questionnaires to screen reports for eligibility and collect data.

Results: In total, 121 studies were eligible, including 111 reporting model development and 10 exclusively reporting external validation. In developing models, 58.6% of studies included no validation, and 35.1% conducted internal validation only. Stepwise approach (45.3%) and univariable significance test (24.5%) were frequently used for selecting predictors, and the vast majority (70.7%) simply applied p-value for the selection decision. Only 5.4% measured the correlation between predictors. Logistic regression (72.7%) was most frequently used for deriving model algorithm, and only 10.8% applied shrinkage techniques. In assessing model performance, most (81.1%) assessed model discrimination, but only 18.9% evaluated calibration.  In presenting model, 48.8% failed to present model algorithms, and 36.4% of models were recommended even without any validation. No significant differences were found in most statistical practices before and after the release of TRIPOD.

Conclusion: The statistical practices of prognostic prediction models for obstetric care were far from optimal.

Note:
Funding Information: This work was funded by the National Natural Science Foundation of China (72174132, 71974138), National Key Research and Development Program of China (2021YFC2701503), China Medical Board (CMB19-324), and Sichuan Youth Science and Technology Innovation Research Team (2020JDTD0015).

Conflict of Interests: None.

Keywords: Prognostic prediction model, Statistical analyses, Obstetric care