Comparison of Bayesian approaches for developing prediction models in rare disease: application to the identification of patients with Maturity-Onset Diabetes of the Young

Cardoso, Pedro; McDonald, Timothy J.; Patel, Kashyap A.; Pearson, Ewan R.; Hattersley, Andrew T.; Shields, Beverley M.; McKinley, Trevelyan J.

doi:10.1186/s12874-024-02239-w

Table 4 Summary of key characteristics of the several modelling approaches

From: Comparison of Bayesian approaches for developing prediction models in rare disease: application to the identification of patients with Maturity-Onset Diabetes of the Young

Modelling Approaches	Advantages	Disadvantages
Original	- Model can be built in one training dataset (e.g. case-control) but probabilities can be updated based on information of prior probability from other studies.	- Individuals cannot have a recalibrated probability lower than the estimated prevalence in the general population.
		- Probabilities are aggregated into groups.
		- The recalibrated probabilities can be sensitive to the choice of grouping used.
Albert Offset	- Model can be built in one training dataset (e.g. case-control) and probabilities can be updated based on information of prior probability from other studies.	- Assumes the likelihood ratio for any given set of covariates is the same in the training data and calibration/target population.
Re-estimation	- Model can be built on one training dataset (population-representative) so no recalibration necessary.	- Requires a large sample size from the general population if model development required.
Re-estimation		- A lot of uncertainty in a rare disease setting due to very low number of positive cases.
Recalibration	- Scales odds ratios allowing for easy recalibration so that the model can be used in different settings (e.g.: general population or lab referral usage).	- Requires two datasets: a training dataset (e.g. case-control) and recalibration dataset (e.g. general population).
Re-estimation mixture	- The model can be specified to include additional information on biomarker testing.	- Requires a large sample size from the general population.
Re-estimation mixture	- Model can be built on one training dataset (population-representative) so no recalibration necessary.	- A lot of uncertainty in a rare disease setting due to very low number of positive cases.
Recalibration mixture	- The model can be specified to include additional information on biomarker testing.	- Requires two datasets: a training dataset (e.g. case-control) and recalibration dataset (e.g.: general population).
Recalibration mixture	- Scales odds ratios allowing for easy recalibration so that the model can be used in different settings (e.g.: general population or lab referral usage).

Back to article page

ISSN: 1471-2288

Contact us

General enquiries: journalsubmissions@springernature.com

BMC Medical Research Methodology

Contact us