 Research
 Open access
 Published:
Comparison of Bayesian approaches for developing prediction models in rare disease: application to the identification of patients with MaturityOnset Diabetes of the Young
BMC Medical Research Methodology volume 24, Article number: 128 (2024)
Abstract
Background
Clinical prediction models can help identify highrisk patients and facilitate timely interventions. However, developing such models for rare diseases presents challenges due to the scarcity of affected patients for developing and calibrating models. Methods that pool information from multiple sources can help with these challenges.
Methods
We compared three approaches for developing clinical prediction models for population screening based on an example of discriminating a rare form of diabetes (MaturityOnset Diabetes of the Young  MODY) in insulintreated patients from the more common Type 1 diabetes (T1D). Two datasets were used: a casecontrol dataset (278 T1D, 177 MODY) and a populationrepresentative dataset (1418 patients, 96 MODY tested with biomarker testing, 7 MODY positive). To build a populationlevel prediction model, we compared three methods for recalibrating models developed in casecontrol data. These were prevalence adjustment (“offset”), shrinkage recalibration in the populationlevel dataset (“recalibration”), and a refitting of the model to the populationlevel dataset (“reestimation”). We then developed a Bayesian hierarchical mixture model combining shrinkage recalibration with additional informative biomarker information only available in the populationrepresentative dataset. We developed a method for dealing with missing biomarker and outcome information using prior information from the literature and other data sources to ensure the clinical validity of predictions for certain biomarker combinations.
Results
The offset, reestimation, and recalibration methods showed good calibration in the populationrepresentative dataset. The offset and recalibration methods displayed the lowest predictive uncertainty due to borrowing information from the fitted casecontrol model. We demonstrate the potential of a mixture model for incorporating informative biomarkers, which significantly enhanced the model’s predictive accuracy, reduced uncertainty, and showed higher stability in all ranges of predictive outcome probabilities.
Conclusion
We have compared several approaches that could be used to develop prediction models for rare diseases. Our findings highlight the recalibration mixture model as the optimal strategy if a populationlevel dataset is available. This approach offers the flexibility to incorporate additional predictors and informed prior probabilities, contributing to enhanced prediction accuracy for rare diseases. It also allows predictions without these additional tests, providing additional information on whether a patient should undergo further biomarker testing before genetic testing.
Background
Clinical prediction models can be useful in rare diseases to aid earlier diagnosis and more appropriate management. However, developing these models can be challenging as suitable data sources for model development may be difficult to acquire. The prevalence of a rare disease in a populationofinterest can be informed by population cohorts [1], but low numbers of cases in these datasets limit the ability to identify risk factors and produce robust predictive models for disease risk in the general population [2]. Casecontrol studies [3] enrich the study population with more disease cases than a random sample from the population, facilitating more robust estimates of associations between patient features and disease risk using measures such as odds ratios. Furthermore, the rise of rare disease registries [4] makes recruiting larger case numbers for these studies easier. However, from a clinical perspective, disease risk probabilities are more natural metrics than odds ratios for diagnosis or screening purposes, but estimated risk probabilities from casecontrol data will be overestimated as they are not based on random samples from the general population [5, 6]. A key challenge is, therefore, how to produce wellcalibrated estimates of individual disease risk probabilities for rare diseases in the general population, utilising information from different data sources.
Various methods have been developed that borrow information from one population and recalibrate their outputs to be valid in another population [7,8,9,10,11,12,13]. These approaches include simple methods such as adjustments of the likelihood ratio based on the sensitivity and specificity of the test at various thresholds [14] or offset updating to adjust the overall model probabilities according to a more appropriate population prevalence [10, 11]. However, these approaches are limited and would not account for differences in patient characteristics that may occur in different datasets, which could be a particular problem in casecontrol studies when enriching for a particular disease or when only collecting specific controls, which would ignore more “greyarea” patients that may be seen in a population setting. More complex techniques are available, such as shrinkage methods to adjust the intercept and model coefficients [7, 12], or previous studies could be used to inform the prior belief of model parameters in Bayesian modelling [15]. Although more sophisticated, these approaches would need data from multiple sources that may not be available for rare diseases. In addition, datasets may not always contain the same information for rare diseases, and specific testing or features may only be available to a subset of patients. More flexible approaches are needed that would allow modelling in these situations.
We use a specific motivating example of developing a prediction model for a rare form of diabetes called MaturityOnset Diabetes of the Young (MODY) that can be used to inform referral decisions for genetic screening for the condition. In this study, we 1) evaluate a range of approaches for appropriately recalibrating model probabilities in prediction models for rare diseases utilising different data sources (including casecontrol data, prevalence estimates, and population datasets) and 2) develop a Bayesian hierarchical mixture modelling approach which can combine a clinical features risk model with additional informative biomarker test information, utilising prior information from other data sources to account for missing data and ensure that the recalibrated probabilities are clinically plausible given specific test results. This latter approach also allows for predictions for new individuals who do not have biomarker test results (since these are not currently routinely collected for MODY), which greatly facilitates using such a prediction model in clinical practice. We also show how the model can help inform on the utility of additional biomarker testing before making a final screening decision for MODY.
Motivating example
Our motivating disease system in this manuscript is a rare youngonset genetic form of diabetes called MaturityOnset Diabetes of the Young (MODY) [16], which is estimated to account for 1–2% of all diabetes cases [17, 18]. MODY is challenging to identify and is estimated to be misdiagnosed in up to 77% of cases [19]. Diagnostic genetic testing is expensive; however, it is crucial to properly diagnose as these patients do not require treatment with insulin injections [20], unlike the most common youngonset form of diabetes, type 1 diabetes (T1D).
Statistical models that use patient characteristics to predict the probability of having MODY can aid decisions regarding which patients to refer for diagnostic MODY testing. One such set of models is routinely used in clinical practice via an online calculator [14] (found at: https://www.diabetesgenes.org/exeterdiabetesapp/ModyCalculator) and has been shown to improve positive test rates of new MODY cases [19]. These prediction models for MODY were developed using casecontrol data and recalibrated to population prevalences using conversion tables derived from the sensitivities and specificities at different probability thresholds [14]. There are several consequences of this approach for prevalence adjustment: i) the recalibrated probabilities end up being grouped; ii) individuals cannot have a recalibrated probability that is lower than the estimated prevalence in the general population; and iii) the recalibrated probabilities can be sensitive to the choice of grouping used. Addressing these limitations would be important, but the most appropriate approach for adjusting for the prevalence is unclear.
In addition, since the original model development, biomarker screening tests (Cpeptide and islet autoantibodies [21, 22]) have become routinely available clinically, and the results of these tests could significantly alter the probability of MODY. Cpeptide is a measure of endogenous insulin secretion, and islet autoantibodies are markers of the autoimmune process in T1D. MODY is characterised by noninsulin dependence, so these patients produce significant amounts of their own endogenous insulin (have positive Cpeptide), and they do not have the autoimmune process associated with T1D (negative islet autoantibodies), whereas being Cpeptide negative (i.e. insulin deficient) or having positive islet autoantibodies is characteristic of T1D. Finding approaches to build these test results into the recalibration would have considerable advantages.
Methods
Setting
The diagnosis of MODY requires expensive genetic testing. Currently, patients are referred for diagnostic genetic testing on an adhoc basis when the clinician considers a MODY diagnosis. In line with guidelines (ISPAD [23] and NHS genomic testing criteria [24]), criteria for referring can include:

Clinical presentation and patient features (including age at diagnosis, BMI, treatment, measures of glucose control (HbA_{1c}) and family history of diabetes);

Results of biomarker testing (Cpeptide and islet autoantibodies);

The use of prediction models in the form of the MODY calculator (can be found at: https://www.diabetesgenes.org/exeterdiabetesapp/ModyCalculator).
Study population
For model development and recalibration, we used data from two sources comprising patients with confirmed MODY and insulintreated patients with T1D, the predominant alternative diagnosis in youngonset patients:
Casecontrol dataset (Fig. 1a)
This dataset was used to develop the original MODY prediction model [14]. All participants were diagnosed with diabetes between the ages of 1 and 35. T1D was defined as occurring in patients treated with insulin within 6 months of diagnosis [14]. The dataset includes 278 patients with T1D and 177 probands with a genetic diagnosis of MODY obtained from referrals to the Molecular Genetics Laboratory at the Royal Devon and Exeter NHS Foundation Trust, UK. The dataset comprises the following variables: sex, ageatdiagnosis, ageatrecruitment, BMI, parents affected with diabetes and HbA_{1c} (%). No biomarker data are available.
Populationrepresentative dataset (UNITED – Fig. 1b)
The UNITED study [25] was a populationrepresentative cohort that recruited 62% of all patients with diabetes diagnosed between the ages of 1 and 30 in two regions of the UK (Exeter and Tayside) (n=1418). Due to the expense of genetic testing, a screening strategy with Cpeptide and islet autoantibody testing was used to narrow down the cohort eligible for MODY testing (Fig. 1b).
For this model, consistent with the original model [14], we analysed all patients insulintreated within 6 months of diagnosis, corresponding to 1171 patients, of which 96 were tested for MODY (given that they were Cpeptide positive and antibody negative) and 7 MODY cases were diagnosed (Fig. 1b). The dataset is comprised of the following variables: sex, ageatdiagnosis, ageatrecruitment, BMI, parents affected with diabetes and HbA_{1c} (%), with additional Cpeptide and islet autoantibodies test results.
Approaches for recalibration
The analysis in this paper was split into three different scenarios to enable populationappropriate probabilities to be calculated with and without the additional biomarker information:

Scenario a) Clinical features model ignoring biomarker information. For this analysis, we used all patients in the populationrepresentative dataset (UNITED). This scenario assumes all those not MODY tested are \({\text{MODY}}^{}\) in the population cohort, i.e. 7 MODY positive patients and 1,164 MODY negative (of which 1,075 were not tested for MODY but are assumed to be MODY negative for the analysis since the biomarker results are inconsistent with MODY).

Scenario b) Clinical features model in only those prescreened to be at increased probability of MODY based on the biomarkers. This included 96 patients, of which 7 are MODY positive and 89 MODY negative. This scenario only analyses patients in the population cohort who had genetic testing (i.e. tested Cpeptide positive and autoantibody negative), so it provides more appropriate model probabilities in patients with these test results indicating a higher risk of MODY, but simply rules out MODY (does not provide a probability) in those who are Cpeptide negative or antibody positive.

Scenario c) Model fully incorporating both clinical features and biomarker information. We analysed all patients in the population cohort and included biomarker information. For this analysis, we included 96 patients who had testing for MODY (7 MODY positive and 89 MODY negative) and 1,075 patients who did not have testing for MODY. The biomarker information of those not MODY tested was used to more appropriately adjust the model probabilities (151 Cpeptide positive and autoantibody positive, 924 Cpeptide negative) (Fig. 1b).
We explored six approaches for producing predictions using different degrees of data availability, which fall into three groups:

1.
Approaches that only utilise casecontrol data and adjust to a known population prevalence: Original and Offset.

2.
Approaches that utilise a casecontrol dataset and additional calibration dataset (e.g. the UNITED population dataset in this study): Reestimation and Recalibration.

3.
Approaches that utilise additional data on informative diagnostic tests and provide biologically plausible constraints: mixture model approaches (for both Reestimation and Recalibration). This mixture model splits individuals into two groups according to their diagnostic test information (a Cpeptide negative or antibody positive group: \({C}^{}\cup {A}^{+}\); and a Cpeptide positive and antibody negative group: \({C}^{+}\cap {A}^{}\)). We use an informative prior distribution to constrain the probability of having MODY in the \({C}^{}\cup {A}^{+}\) group and use one of the other recalibration methods in the \({C}^{+}\cap {A}^{}\) group.
The Supplementary Materials Notation section contains a glossary of mathematical symbols used throughout the article. We fit these models using the package NIMBLE [26, 27] (version 1.0.1) in the software R [28] (version 4.3.2).
1. Training dataset only approaches
Training data model
Let \({M}_{j}^{C}\) be a binary variable denoting whether an individual \(j\) in the casecontrol data set has MODY or not, such that
We then model:
where the log odds are given by:
with \({X}_{jv}^{C} (v=1,\dots ,p)\) a set of \(p\) covariates for individual \(j\). We put independent vague \({\text{Normal}}\left(\mu =0, {\text{sd}}=10\right)\) priors on the regression parameters.
The posterior for this model then takes the form:
where \(\theta =\left({\beta }_{v}^{C};v=0,\dots ,p\right)\), with \(\pi \left(\cdot \right)\) denoting the relevant probability (density) mass functions derived above for the model and joint prior distribution.
Original approach
This method was implemented by Shields et al. (2012) [14] during the development of the original MODY prediction model. The approach fitted a model to a casecontrol dataset using the patients’ characteristics and used the relationship:
or
where \({M}^{+}\) is the event that the patient has MODY, and \({R}^{+}\) is whether a hypothetical “test” is positive (and similarly for \({M}^{}\) and \({R}^{}\)). In this case, \({R}^{+}\) is derived by applying a threshold, \({p}^{*}\), to the predicted probabilities \({p}_{j}^{C}\) obtained from a training model (see eq. (2)) for a given individual \(j\), such that an individual is classed as positive if \({p}_{j}^{C}>{p}^{*}\) and negative otherwise.
Therefore, for a given choice of \({p}^{*}\), estimates of the sensitivity, \(P\left({R}^{+}{M}^{+}\right)\), and specificity, \(P\left({R}^{}{M}^{}\right)\), of these classifications at a range of thresholds were calculated using the casecontrol data. \(P\left({M}^{+}\right)\) is then chosen as an estimate of the prevalence of MODY in the general population, which in the original model was given by 0.7% [14], which assumed no knowledge of biomarker test results. In this paper, we adjusted slightly differently depending on scenario a) or scenario b). In scenario a), we estimated the pretest probability to be 0.6% (informed by the prevalence of MODY in the UNITED dataset). For scenario b), we estimated the pretest probability to be 7.3% (informed by the prevalence of MODY in those who were Cpeptide positive and antibody negative in UNITED).
For a new individual in the general population, with covariates \({X}_{i}^{*}\) say, then one can derive an estimate for \({p}_{i}^{C}\) (based on eq. (2)) as
before using Table 1 to map their predicted \({p}_{i}^{C}\) from the casecontrol model to a recalibrated probability of having MODY in the general population.
Albert Offset approach
This approach was proposed by Albert (1982) [10] and similarly to the method above, leverages the relationship:
where \(X\) is a set of explanatory variables. In words:
For the training data (C – casecontrol dataset), we use the same model for \({M}_{j}^{C}\) and \({p}_{j}^{C}\) as before (see eqs. (1) and (2)), and then the idea is that if we know the disease odds in the training data, then we can rewrite eq. (2) as:
hence the original \({\beta }_{0}^{C}={\eta }_{0}^{C}+\mathrm{log }(\text{pretest odds training})\). Therefore, under the assumption that the likelihood ratio for any given set of covariates is the same in the training and calibration datasets, then for a new individual \(i\) in the general population, with covariates \({X}_{i}^{*}\), we can recalibrate as:
This approach gives individuallevel recalibration probabilities that do not rely on thresholding. The Albert Offset approach requires a training dataset for fitting the original model and an estimate of the disease odds in both the training data and the populationofinterest. For this cohort, as before, you could adapt the offset based on the prevalence of MODY of 0.6% based on scenario a) or 7.3% based on scenario b). We also explore an example where the likelihood ratio assumption is not maintained between datasets for illustrative purposes. We put independent vague \({\text{Normal}}\left(\mu =0, {\text{sd}}=10\right)\) priors on the regression parameters.
2. Populationrepresentative dataset approaches
Reestimation approach
This approach fits a new model directly to the populationrepresentative dataset (UNITED), ignoring the casecontrol dataset entirely. Given sufficient cases and controls in a given dataset, this model fitted using, e.g. maximum likelihood, will give asymptotically unbiased estimates for the odds ratios and probabilities. However, for rare diseases, one would have to have very large sample sizes to get sufficient numbers of cases to develop an entirely new model. As a comparison, we use the model structure developed in the casecontrol dataset and then refit the model to the populationrepresentative dataset (UNITED). Here, we denote the MODY status for individual \(i\) in the UNITED dataset as \({M}_{i}^{U}\), and model this as
where
with \({X}_{iv}^{U}\) \(\left(v=1,\dots ,p\right)\) a set of \(p\) covariates for individual \(i\). We place independent vague \({\text{Normal}}\left(\mu =0, {\text{sd}}=10\right)\) priors on the regression parameters.
The posterior for this model then takes the form:
where \(\theta ={(\beta }_{v}^{U}; v=1,\dots ,p)\), with \(\pi (\cdot)\) denoting the relevant probability (density) mass functions derived above for the model and joint prior distribution.
Recalibration approach
In the context of the models developed here, the Recalibration approach [7] uses a model fitted to the casecontrol dataset to generate predictions of the linear predictor for each individual in the populationrepresentative data set (UNITED). In the training data, for individual \(j\), \({M}_{j}^{C}\) and \({p}_{j}^{C}\) are modelled as before (see eqs. (1) and (2)), and then for each individual \(i\) in the calibration dataset (UNITED), with predictors \({X}_{i}^{U}\), the linear predictors \({z}_{i}=\widehat{{\upbeta }_{0}^{\text{C}}}+\widehat{{\upbeta }_{1}^{\text{C}}}{X}_{i1}^{U}+\dots +\widehat{{\upbeta }_{p}^{\text{C}}}{X}_{ip}^{U}\) are calculated, where \(\widehat{{\beta }_{v}^{C}}\) is a point estimate of the \(v\) ^{th} regression parameter from the casecontrol model. These \({z}_{i}\) terms are then used as covariates in a second (shrinkage) model:
This approach [7] can have the effect of scaling the odds ratios and intercept terms where necessary, and a sideeffect is that if no recalibration is required, then \({\gamma }_{0}=0\) and \({\gamma }_{1}=1\). Again, these approaches could be built using scenarios a) and b), dependent on the assumptions we are willing to make with UNITED. The method used by Steyerberg et al. (2004) [7] uses the point predictions for \({z}_{i}\) based on the maximum likelihood estimates from the casecontrol data, which ignores the uncertainty in the estimations of \({z}_{i}\). Instead, we develop a joint Bayesian hierarchical model where we simultaneously fit both models and propagate the uncertainties directly from the casecontrol model to the recalibration model [29]. We put independent vague \({\text{Normal}}\left(\mu =0, {\text{sd}}=10\right)\) prior distributions on the regression parameters, with a \({\text{Normal}}\left(\mu =0, {\text{sd}}=10\right)\) prior for \({\gamma }_{0}\) and a \({\text{Normal}}\left(\mu =1, {\text{sd}}=10\right)\) prior for \({\gamma }_{1}\).
The posterior for this joint model then takes the form:
where \(\theta =\left({\theta }^{U},{\theta }^{C}\right)\) corresponds to the full vector of parameters, with \({\theta }^{U}=\left({\gamma }_{0},{\gamma }_{1}\right)\) and \({\theta }^{C}=\left({\beta }_{v}^{C};v=1,\dots ,p\right)\), with \(\pi \left(\cdot \right)\) denoting the relevant probability (density) mass functions derived above for the different component models and joint prior distribution.
3. Mixture model approach
One area of development in this manuscript is how to incorporate biomarker test information into the model when the biomarker tests can place very strong constraints on the postrecalibration probabilities depending on their specific values. For example, a simple way to include a binary test result would be to add another covariate into the linear predictor in one of the previous methods. In the analysis, the biomarker data only exists in the calibration data (UNITED) but not the training data (casecontrol), so this approach would only use information from the calibration data to estimate the parameters relating to the biomarkers. Since there are few cases in the calibration data, this would necessarily result in large standard errors for the estimated effects and could lead to biologically implausible estimates. For example, an individual who is Cpeptide negative or antibody positive can be considered to have a very low probability of having MODY, justified through prior data and biological plausibility (Cpeptide negativity means that an individual is producing negligible amounts of their own insulin, which defines T1D). In clinical practice, an individual with these biomarker results would be treated as having T1D, which is equivalent to assuming that the probability of having MODY given these results is exactly zero. However, this approach does not allow for the rare (but possible) event that an individual has a positive genetic MODY test but is antibody positive or Cpeptide negative (which would ideally also allow for imperfect sensitivities and specificities of the biomarker tests).
Using the mixture model approach in scenario c), it is possible to incorporate a nonzero prior probability of having MODY in these cases, where we use independent data sets to inform the prior distribution for this probability. We note that the mixture model allows for different prior constraints to be used for different subsets of the data: here the prior probability of having MODY is very low for Cpeptide negative or antibody positive individuals [21, 22], but is not similarly constrained for Cpeptide positive and antibody negative individuals. Similar ideas could be used for other diseases where the prior information may not be as strong.
For the UNITED data, we let
with
and
We then set:
Letting \({X}_{i}^{U}\) be a vector of additional covariates for individual \(i\), we can model \({M}_{i}^{U}\) as
where
We model \({p}_{{M}^{+}{C}^{}\cup {A}^{+}}\) using a \({\text{Beta}}(\alpha = 2.2,\beta =7361.3)\) prior probability distribution (see Supplementary Materials Prior elicitation section for a justification of this choice). We then model \({p}_{{M}^{+}{C}^{+}\cap { A}^{},{X}_{i}}\) differently, depending on whether we use the Reestimation or Recalibration approaches (see below).
Reestimation approach
For the Reestimation approach we model
and to finalise, we put independent vague \({\text{Normal}}\,\left(\mu =0, {\text{sd}}=10\right)\) priors on the regression parameters.
Recalibration approach
For the Recalibration approach we also utilise the casecontrol data. If we let \({M}_{j}^{C}\) be the MODY status for individual \(j\) in the casecontrol dataset, with vector of covariates \({X}_{j}^{C}\), then \({M}_{j}^{C}\) and \({p}_{j}^{C}\) are modelled as before (see eqs. (1) and (2)). Then, for individual \(i\) in the UNITED dataset (with \({C}_{i}^{+}\cap {A}_{i}^{}\)), we model
where
Incorporating biomarker test results
To allow for predictions in the absence of biomarker test results (which are not routinely collected in clinical practice), we model
with \({X}^{*U}\) comprised of the variables BMI, ageofdiagnosis, ageofrecruitment and parents affected with diabetes (here we use restricted cubic splines with 3 knots to model the continuous variables). In this case the predicted probability of MODY for an individual with unknown test results will be a weighted average of the \({p}_{{M}^{+}{C}^{}\cup {A}^{+}}\) and \({p}_{{M}^{+}{C}^{+}\cap { A}^{},{X}_{i}}\), weighted by the probability of being \({C}^{}\cup {A}^{+}\) based on suitable individuallevel characteristics. We place independent vague \({\text{Normal}}\left(\mu =0, {\text{sd}}=10\right)\) prior distributions on the regression parameters, with a \({\text{Normal}}\left(\mu =0, {\text{sd}}=10\right)\) prior on \({\gamma }_{0}\) and a \({\text{Normal}}\left(\mu =1, {\text{sd}}=10\right)\) prior on \({\gamma }_{1}\).
For the Reestimation mixture, the posterior then takes the form:
where \(\theta =\left({\theta }^{U},{\theta }^{T},{p}_{M^+{C}^{}\cup {A}^{+}}\right)\) corresponds to the full vector of parameters, with \({\theta }^{U}=\left({\beta }_{v}^{U};v=1,\dots ,p\right)\) and \({\theta }^{T}=({\beta }_{v}^{*};v=1,\dots ,r)\) with \(\pi \left(\cdot \right)\) denoting the relevant probability (density) mass functions derived above for the different component models and joint prior distribution.
For the Recalibration mixture, the posterior then takes the form:
where \(\theta =({\theta }^{U},{\theta }^{T},{p}_{M^+{C}^{}\cup {A}^{+}},{\theta }^{C})\) corresponds to the full vector of parameters, with \({\theta }^{U}=\left({\gamma }_{0},{\gamma }_{1}\right)\), \({\theta }^{C}=\left({\theta }_{v}^{C};v=1,\dots ,p\right)\) and \({\theta }^{T}=\left({\beta }_{v}^{*};v=1,\dots ,r\right)\), with \(\pi \left(\cdot \right)\) denoting the relevant probability (density) mass functions derived above for the different component models and joint prior distributions.
Assessment of model performance, calibration and stability analysis
In scenario a), we validate fitted probabilities for all patients in UNITED (setting those with missing MODY testing to \({\text{MODY}}^{}\)). In scenarios b) and c), we only validate fitted probabilities on \({C}^{}\cup {A}^{+}\) patients as these were the only patients who had prescreening based on biomarkers and had genetic testing of MODY genes. The area under the receiver operating characteristic (AUROC) curve was used as a measure of overall discrimination performance. Calibration curves were plotted to visualise how well the predicted probabilities were calibrated against the observed data. For the calibration curves, predicted probabilities were grouped by quintiles and plotted against the observed probability of positive individuals within each quintile. To assess convergence, we monitored the available parameters for evidence of convergence and GelmanRubin \(\widehat{R}\) values [30]. Further validation procedures are explained in the Supplementary Materials Stability analysis section.
Results
Comparing datasets
In the casecontrol dataset, 177 out of 455 patients had MODY, leading to an enriched proportion with MODY of 40%. In contrast, in the recalibration population (UNITED) cohort, 7 out of 1171 patients (0.6%) had MODY, much more consistent with the prevalence of MODY in the population. The characteristics of patients in the two datasets were broadly similar (sFig. 1).
Models and their recalibration from the 6 different approaches
All models in this study converged quickly, so we ran four chains of 500,000 iterations, with the first 300,000 discarded for burnin in each case (sFig. 24).
The first, recalibration approach, the Original approach achieved an \(\widehat{R}=1.0\) for all parameters. As expected, the choice of prevalence used for recalibration affected the conversion probabilities. Table 1 shows the different postrecalibration probabilities of having MODY using the different prevalences for both scenarios a) and b), with postrecalibration probabilities more appropriately higher in scenario b) to allow for the biomarker results in those who were Cpeptide positive and antibody negative.
Table 2 describes the model parameter estimates for the Albert Offset and Reestimation approaches in scenarios a) and b). The Albert Offset approach achieved an \(\widehat{R}=1.0\), and the Reestimation approach achieved an \(\widehat{R}=1.0\) for all parameters. Coefficients were quite different in the various approaches.
The Recalibration approach achieved an \(\widehat{R}=1.0\) for all parameters. In scenario a), estimates were \({\gamma }_{0}=4.39\) (95%CI 5.41; 3.56) and \({\gamma }_{1}=0.96\) (95%CI 0.49; 1.54) and in scenario b), the estimates were \({\gamma }_{0}=2.26\) (95%CI 3.33; 1.37) and \({\gamma }_{1}=0.86\) (95%CI 0.31; 1.57).
For scenario c), fully incorporating the biomarker information into the model, probabilities could be obtained using the Recalibration and Reestimation mixture approaches. The Recalibration mixture approach achieved an \(\widehat{R}<1.01\) for all parameters, with an estimated \({\gamma }_{0}=2.26\) (95%CI 3.33; 1.37) and \({\gamma }_{1}=0.86\) (95%CI 0.32; 1.58). The Reestimation mixture approach achieved an \(\widehat{R}<1.01\). The model that estimates \(T\) achieved an AUROC of 0.76 (95%CI 0.75; 0.77) in both mixture approaches, with model parameters described in sTable 1.
Discrimination and calibration of the models developed using the 6 different approaches
All approaches led to good model discrimination, with the Reestimation approaches having the highest AUROC (Table 3).
In scenario a), for the approaches that used only the casecontrol dataset and adjusted for a known prevalence, the Original approach overestimated the observed probability of MODY in the UNITED population and had large uncertainty at higher percentages. In contrast, the Albert Offset approach slightly underestimated the observed probability of MODY in the UNITED population. Looking at the approaches that used the populationrepresentative dataset (UNITED), both the Reestimation and Recalibration approaches slightly underestimated the observed probability of MODY with slightly more uncertainty in the predictions from the Reestimation approach (Fig. 2).
In scenario b), for the approaches that used the casecontrol dataset alone, with adjustment for known prevalence, the Original approach overestimated the observed probability of MODY in the UNITED population. In contrast, the Albert Offset approach was able to calibrate well. Looking at the approaches that used an additional calibration dataset (UNITED), both the Reestimation and Recalibration approaches calibrated well, but the Reestimation approach demonstrated more uncertainty in the probability predictions (Fig. 3). In scenario c), the Reestimation and Recalibration mixture approaches demonstrated similar performance to the equivalent models that did not use a mixture model approach, with similar levels of uncertainty in probability predictions (Fig. 3). In this case, the Albert Offset method worked well, but it relies on the assumption that the likelihood ratio is the same in the two populations. For illustrative purposes, we also provide an example setting where the likelihood ratio is different between the training and calibration datasets (violating the assumption). In this latter example, the Albert Offset approach fails to calibrate well. In contrast, the Recalibration approach can scale the odds ratios and calibrates well (sFig. 5), so this method would be preferred if a recalibration dataset is available.
Stability plots for the mixture models in scenario c)
The bootstrap stability test was made for both mixture approaches. Both mixture approaches were ran 50,000 iterations with the first 30,000 discarded for burnin, with an average \(\widehat{R}=1.02\) (95% 1.0; 2.1) for the Reestimation mixture approach and an average \(\widehat{R}=1.01\) (95% 1.0; 1.6) for the Recalibration mixture approach (the higher \(\widehat{R}\) values occurred in bootstrapped datasets with less than 3 positive MODY cases, but this only occurred in 8/1000 datasets and made no difference to the plots in Fig. 4, and so we left these runs in). Both recalibration approaches showed some variability in the estimated probabilities, with the Reestimation mixture approach demonstrating higher uncertainty across all estimated probability levels (Fig. 4). However, we can see that because the Recalibration mixture approach borrows weight from the casecontrol data, the estimates were more stable than the Reestimation mixture method. We also noted that by using the hierarchical modelling approach, the Recalibration mixture model uncertainty included the uncertainty in the casecontrol predictions. Thus, these uncertainty estimates are larger than a model where this additional predictive uncertainty is ignored.
Final recalibrated probabilities
The approach chosen for our final models was the Recalibration mixture approach, which incorporated the most information with the lowest uncertainty in probability predictions. The mixture model ensures that those with biomarkers consistent with T1D (the \({C}^{}\cup {A}^{+}\) individuals) are predicted to have a very low probability of MODY, consistent with independent prior information. Fig. 5 shows the predicted probability of MODY in the remaining \({C}^{+}\cap {A}^{}\) individuals. Considering only 0.6% of the cohort had MODY, the model produced a wide range of probabilities. Most nonMODY cases were predicted to have a low probability of MODY, with 97.2% (1,132/1,164) of individuals having an upper 95% CI probability of MODY under 10%. In contrast, 7 out of the 7 MODY cases had an upper 95% CI probability >10%. This would mean that if using a >10% threshold to initiate MODY testing for the population, 39 patients would be tested, giving a positive predictive value of 17.9% (Fig. 5), equivalent to the Original approach.
Discussion
This paper explored recalibration methods for adapting a statistical model from casecontrol data to the general population for rare disease prediction.
We have shown that the calibration of disease risk probabilities can be improved via various methods, and in particular, our results show the added benefits of utilising a secondary (recalibration) dataset that corresponds to a random sample from the general population despite there being few cases in the latter. In addition, the recalibration data contains additional information on biomarker tests, which are highly informative about disease risk, but only for certain subsets of test results; because of this, some biomarker information is only available for subsets of individuals. Our Recalibration mixture model allows the inclusion of (incomplete) biomarker information and informative prior information (derived from previous studies) about disease risk for specific subsets of test results to ensure clinically valid risk probabilities in those cases.
The Recalibration mixture model has several other advantages. It allows for predictions to be made in clinical practice even if the biomarker information is not available. We can also propagate parameter uncertainties from the casecontrol model to the recalibrated population predictions by utilising the Bayesian framework. This gives a more robust estimate of the underlying predictive uncertainty than classical models that ignore this uncertainty. Furthermore, the predictions for individuals without biomarker test information also propagate the uncertainties from the missing information. Finally, since this model is used to help inform which individuals should be screened for MODY using expensive genetic testing, for those individuals who have missing biomarker information, we show how the mixture model can also be used to inform clinicians about the added utility of performing a biomarker test before making a final decision of whether to send individuals for genetic testing. Although highlighted with a specific application, these ideas could be adapted to other rare diseases.
We compared several approaches for recalibrating probabilities when developing prediction models for rare diseases. We showed that the Original method tends to overestimate the probabilities in the general population, but that the Albert Offset [10], Reestimation and Recalibration [7] approaches achieve good calibration of MODY probability predictions in both the model of the overall population and also in the model examining only the subset who were \({C}^{+}\cap {A}^{}\) (those genetically tested for MODY). The Albert Offset [10] and Recalibration [7] approaches achieved the smallest uncertainty around the observed probability of MODY. The Recalibration mixture model showed stability in our study and was the only approach that appropriately constrained the probability of MODY in \({C}^{}\cup {A}^{+}\) individuals to be consistent with the strong prior information available in this setting. When developing prediction models for rare diseases in practice, different approaches will be plausible in different scenarios based on the available data sources. Table 4 provides an overview of the advantages and disadvantages of all modelling approaches explored in the manuscript.
When only a training dataset (casecontrol dataset in our setting) is available, and the aim is to adjust probabilities based on population prevalence, then the Albert Offset approach was the preferred method as it estimated the probability of MODY well in both our scenarios, with reasonable uncertainty in the predictions. In contrast, the Original approach [14] relies on thresholding probabilities and overestimates the probability of MODY in both scenarios. The Albert Offset approach has also been compared to other recalibration methods in other studies. Chan et al. (2008) had similar findings and deemed the Albert Offset the best approach [11]. In contrast, Grill et al. (2016) described this approach as the worstperforming one in their study [12]. These differences may relate to the Albert Offset approach's strong assumption that the covariate distribution is the same in the training dataset as in the population for which the probabilities are adjusted [10, 11], and as we showed in sFig. 5, the Albert Offset approach can perform poorly for datasets where the covariate distribution is different. When recalibrating models for different settings, the likelihood ratio assumption could become harder to justify depending on the specific setting, and particular caution would be required in populations where the clinical characteristics of the patients differ substantially from the casecontrol dataset used for original model development. Establishing whether the similarities between covariate distributions are sufficient for preassessing the performance of the Albert Offset method is of interest for future research.
When only a populationrepresentative dataset is available, the Reestimation approach would be necessary. The Reestimation approach calibrates well in the general population dataset for both scenarios but demonstrates high uncertainty surrounding the model predictions. In contrast, Grill et al. (2016) describe the Reestimation approach as having equivalent performance to the Albert Offset, with both performing worse than all other approaches [12]. The high uncertainty in our analysis can be attributed to the fact that there are only 7 positive MODY cases in the general population dataset (prevalence of 0.6%) and that the distribution of predicted probabilities is skewed towards zero. This highlights the problem with fitting models for rare diseases to population data [31], where the low prevalence means very large sample sizes would be required to reduce the uncertainty around the predictions, and it would be important to assess the adequacy of the sample size prior to model fitting [32]. When both training and populationrepresentative datasets are available (as in our study), we showed that the Recalibration approach demonstrates good calibration in the populationrepresentative dataset for both scenarios. This approach combines the information captured from the training data with information from the calibration dataset [7, 33], producing relatively low uncertainty in the model predictions compared to other approaches explored in this paper.
We also explored a scenario where additional biomarker testing was available but performed only on a limited subset of patients. Screening using biomarkers is common in clinical practice and often used in rare diseases where universal testing is not costeffective or could be invasive (e.g. screening for chromosomal defects in pregnancy [34]). We developed a Bayesian hierarchical mixture model to follow the referral process involved in MODY testing and, therefore, utilise the additional biomarker tests for further refinement in the prediction of MODY probabilities. As a Bayesian model, we can incorporate additional information from other studies into the prior distributions for certain parameters, something previously explored by Boonstra et al. [15] in a different setting where additional information is only present for a subset of individuals. This approach has a further advantage in that predictions can still be made for patients with missing additional biomarkers, which are modelled using patient characteristics. This is important for our setting in which instead of ignoring the biomarker results altogether, the model has used this information to improve predictions so that even when biomarker information is missing, the MODY probabilities are a weighted sum across the latent biomarker test results, where the weights are informed by a model relating potential biomarker test outcomes conditional on a set of clinical features. We also combined the mixture model with the Reestimation and the Recalibration approaches for just \({C}^{+}\cap {A}^{}\) individuals. Both approaches showed uncertainty levels in the probability predictions consistent with the previously observed uncertainty estimates. Furthermore, both approaches were tested for stability using bootstrapped versions of the populationrepresentative dataset [35], demonstrating that the Recalibration mixture approach was more stable with the predicted probabilities of MODY than the Reestimation mixture approach.
Other approaches for recalibration have been considered in previous work. Chan et al. (2008) [11] compared three methods to update pretest probability with information on a new test: the Albert (Albert Offset in our study) [10], Spiegelhalter and KnillJones (SKJ) [36] and Knottnerus [37] approaches. The SKJ represents an alternative to the Albert Offset approach, with similar performance in their paper. The Knottnerus approach was more suited to cases with sequential biomarker testing, which was not appropriate for our work since we did not have data on some combinations of tests, instead we grouped biomarkers into a composite measure \(T\). The Knottnerus approach could be compared to the mixture model approaches (allowing for nonindependence between both biomarkers), and examining these approaches when considering sequential testing of more than one biomarker could be considered in future work. Grill et al. (2016) compared several methods for incorporating new information into existing risk prediction models: logisticnew (equivalent to Reestimation), LRjoint, LRoffset (Albert Offset), and LRshrink (equivalent to SKJ from reference [11, 36]). In contrast to our study, their original models were built in population data, and the new datasets with additional features were either cohort or casecontrol data [12]. In the context of rare diseases, casecontrol data is likely to provide the best dataset for initial model development since this gives the most power for estimating model parameters, and the populationrepresentative model can then borrow information from the casecontrol model. In cases where the additional data are only available in the casecontrol setting, and original models were built on population data, the joint model approach (Recalibration) could be adapted to this scenario.
The model we recommend for the available MODY data is the Recalibration mixture approach. A major strength of this procedure is that it allows predictions for patients with missing biomarker testing, and this weighted prediction of MODY probability can be used to inform whether a patient should be referred for further testing [38]. This model allows for the incorporation of strong prior information regarding the probability of having MODY for \({C}^{}\cup {A}^{+}\) individuals, propagates uncertainties regarding the missing data in the UNITED study, and borrows weight from the casecontrol model through the recalibration procedure [7], thus improving the stability of predictions [35]. This model provides sensible predictions for the probability of MODY for patients with/without additional testing for Cpeptide and antibodies. Patients with missing MODY testing (i.e. \({C}^{}\cup {A}^{+}\)) could have been set as a negative result test for all approaches due to the strong clinical knowledge of these tests being consistent with a T1D diagnosis. However, this may not be the case for other settings, where patients with missing outcomes could be believed to have a higher probability of the outcome, and therefore, assuming that the outcome is negative may be less justifiable.
There are some limitations to the Recalibration Mixture approach. We currently use biomarker tests as binary (positive/negative) results; in practice, biomarkers may be on a continuous scale. As such, the model could be adapted to include the biomarker results as additional covariates, which could be numerically integrated out if predicting to an individual that was missing this information in practice [38]. We are also limited by the small sample sizes in rare diseases [39], and even with our final model utilising two datasets, model predictions still have some uncertainty. However, we still saw good separation between MODY and T1D, and even accounting for the uncertainty, probability thresholds could be defined that rule out clear nonMODY cases and can be used to determine positive test rates at different probabilities in practice. These thresholds would balance the amount of testing to be carried out against the potential for missing genuine MODY cases, depending on how conservative the choice of threshold is. The model has yet to be validated in a holdout dataset, but the stability plots using bootstrapped datasets provide some insight into the stability of model predictions [35]. Although the 95% credible interval of bootstrapped probabilities is relatively wide at higher values, the 50% credible interval is narrow for all probabilities around the equal line.
This paper provides a comparison of several recalibration approaches. The development of our recalibration approach uses established methodologies, and we have shown how it could apply to identifying patients with a high probability of MODY to allow more targeted diagnostic testing, but these ideas could be applied to other diseases. In practice, other settings could benefit from a similar Bayesian hierarchical model structure where informative biomarkers or additional testing information are available but only in a subset of patients due to its invasive nature or high cost of testing. With this structure, the model can be used to consider whether additional testing should be carried out when the individual already has a low probability (not on the cusp of referral), something explored previously in treatment selection for Type 2 diabetes [38]. Furthermore, this modelling structure could be particularly useful in other rare diseases with low sample sizes since it borrows weight from multiple datasets through recalibration, improving predictions.
Conclusion
We have compared several approaches to developing prediction models for rare diseases. We found the Recalibration mixture model approach to be the best approach, combining casecontrol and populationrepresentative data sources. This approach allows the incorporation of additional data on biomarkers and appropriate prior probabilities.
Availability of data and materials
Data are available through application to the Genetic Beta Cell Research Bank https://www.diabetesgenes.org/currentresearch/geneticbetacellresearchbank/ and the Peninsula Research Bank https://exetercrfnihr.org/about/exeter10000prb/. Example R code for fitting the approaches used in this study is available on GitHub within the repository “ExeterDiabetes/PedroMODY_recal_approaches”.
References
Johnson SR, Ellis JJ, Leo PJ, Anderson LK, Ganti U, Harris JE, Curran JA, McInerneyLeo AM, Paramalingam N, Song X, Conwell LS, Harris M, Jones TW, Brown MA, Davis EA, Duncan EL. Comprehensive genetic screening: the prevalence of maturityonset diabetes of the young gene variants in a populationbased childhood diabetes cohort. Pediatr Diabetes. 2018;20(1):57–64.
Mitani AA, Haneuse S. Small data challenges of studying rare diseases. Diabetes Endocrinol. 2020;3(3):e201965.
Schulz KF, Grimes DA. Casecontrol studies: research in reverse. Epidemiology. 2002;359(9304):431–4.
Kölker S, Gleich F, Mütze U, Opladen T. Rare disease registries are key to evidencebasec personalized medicine: highlighting the european experience. Front Endocrinol. 2022;13:832063.
Greenland S. Modelbased estimation of relative risks and other epidemiologic measures in studies of common outcomes and in casecontrol studies. Am J Epidemiol. 2004;160(4):301–5.
Rothman KJ, Greenland S. Modern Epidemiology. Philadelphia: LippincottRaven; 1998.
Steyerberg EW, Borsboom GJ, van Houwelingen HC, Eijkemans MJ, Habbema JDF. Validation and updating of predictive logistic regression models: a study on sample size and shrinkage. Stat Med. 2004;23(16):2567–86.
Steyerberg E. Clinical Prediction Models: a practical approach to development, validation and updating. Springer International P. 2009.
Schuetz P, Koller MT, ChristCrain M, Steyerberg EW, Stolz D, Müller CA, Bucher HC, Bingisser RM, Tamm M, Müller B. Predicting mortality with pneumonia severity scores: importance of model recalibration to local settings. Epidemiol Infect. 2008;136(12):1628–37.
Albert A. On the use and computation of likelihood ratios in clinical chemistry. Clin Chem. 1982;28(5):1113–9.
Chan SF, Deeks JJ, Macaskill P, Irwig L. Three methods to construct predictive models using logistic regression and likelihood ratios to facilitate adjustment for pretest probability give similar results. J Clin Epidemiol. 2008;61(1):52–63.
Grill S, Ankerst DP, Gail MH, Chatterjee N, Pfeiffer RM. Comparison of approaches for incorporating new information into existing risk prediction models. Stat Med. 2016;36(7):1134–56.
Cheng W, Taylor JM, Gu T, Tomlins SA, Mukherjee B. “Informing a risk prediction model for binary outcomes with external coefficient information”, Journal of the Royal Statistical Society. Series C Appl Stat. 2019;68(1):121–39.
Shields BM, McDonald TJ, Campbell MJ, Hyde C, Hattersley AT. The development and validation of a clinical prediction model to determine the probability of MODY in patients with youngonset diabetes. Diabetologia. 2012;55:1265–72.
Boonstra PS, Barbaro RP. Incorporating historical models with adaptive Bayesian updates. Biostatistics. 2020;21(2):e47–64.
Colclough K, Patel K. How do I diagnose maturity onset diabetes of the young in my patients? Clin Endocrinol. 2022;97(4):436–47.
Gardner D, Tai ES. Clinical features and treatment of maturity onset diabetes of the young (MODY). Diabetes Metab Syndr Obes. 2012;2012(5):101–8.
Naylor R, Johnson A, Gaudio D, Adam M, Feldman J, Mirzaa G, Pagon R, Wallace S, Bean L, Gripp K and Amemiya A. Maturityonset diabetes of the young overview, University of Washington; Seattle, 19932023.
Pang L, Colclough KC, Shepherd MH, McLean J, Pearson ER, Ellard S, Hattersley AT, Shields BM. Improvements in awareness and testing have led to a threefold increase over 10 years in the identification of monogenic diabetes in the U.K. Diabetes Care. 2022;45(3):642–9.
Shepherd M, Shields B, Hudson M, Pearson E, Hyde C, Ellard S, Hattersley A, Patel K. A UK nationwide prospective study of treatment change in MODY: genetic subtype and clinical characteristics predict optimal glycaemic control after discontinuing insulin and metformin. Diabetologia. 2018;61(12):2520–7.
Thanabalasingham G, Pal A, Selwood MP, Dudley C, Fisher K, Bingley PJ, Ellard S, Farmer AJ, McCarthy MI, Owen KR. Systematic assessment of etiology in adults with a clinical diagnosis of youngonset type 2 diabetes is a successful strategy for identifying maturityonset diabetes of the young. Diabetes Care. 2012;35(6):1206–12.
Besser RE, Shepherd MH, McDonald TJ, Shields BM, Knight BA, Ellard S, Hattersley AT. Urinary Cpeptide creatinine ration is a practical outpatient tool for identifying hepatocyte nuclear factor 1α/hepatocyte nuclear factor 4α maturityonset diabetes of the young from longduration type 1 diabetes. Diabetes Care. 2011;34(2):286–91.
Greeley SA, Polak M, Njølstad PR, Barbetti F, Williams R, Castano L, Raile K, Chi DV, Habeb A, Hattersley AT, Codner E. ISPAD clinical practice consensus guidelines 2022: the diagnosis and management of monogenic diabetes in children and adolescents. Pediatr Diabetes. 2022;23(8):1188–211.
National Health Service. National Genomic Test Directory: testing criteria for rare and inherited disease.,” [Online]. Available: https://www.england.nhs.uk/wpcontent/uploads/2018/08/rareandinheriteddiseaseeligibilitycriteriav2.pdf. Accessed 6 Aug 2023.
Shields B, Shepherd M, Hudson M, McDonald T, Colclough K, Peters J, Knight B, Hyde C, Ellard S, Pearson E, Hattersley A and UNITED study team. Populationbased assessment of a biomarkerbased screening pathway to aid diagnosis of monogenic diabetes in youngonset patients. Diabetes Care. 2017; 40(8): 10171025, 2017.
de Valpine P, Turek D, Paciorek C, AndersonBergman C, Temple Lang D, Bodik R. Programming with models: writing statistical algorithms for general model structures with NIMBLE. J Comput Graph Stat. 2017;26(2):403–13.
de Valpine P, Paciorek C, Turek D, Michaud N, AndersonBergman C, Obermeyer F and et al., “NIMBLE: MCMC, particle filtering, and programmable hierarchical modeling,” 2022. [Online]. Available: https://cran.rproject.org/package=nimble.
R Core Team, “R: a language and environment for statistical computing,” 2021. [Online]. Available: https://www.Rproject.org/.
Gelman A, Carlin J, Stern H, Rubin D. Bayesian data analysis. New York: Chapman and Hall/CRC; 1995.
Gelman A, Rubin D. Inference from iterative simulation using multiple sequences. Stat Sci. 1992;7(4):457–72.
Griggs R, Batshaw M, Dunkle M, GopalSrivastava R, Kaye E, Krischer J, Nguyen T, Paulus K, Merkel P. Clinical research for rare disease: opportunities, challenges, and solutions. Mol Genet Metab. 2009;96(1):20–6.
Mitani A, Haneuse S. Small data challenges of studying rare diseases. JAMA Network Open. 2020;3(3):e201965.
Moons K, Kengne A, Grobbee D, Royston P, Vergouwe Y, Altman D, Woodward M. Risk prediction models: II. external validation, model updating, and impact assessment. Heart. 2012;98:691–8.
Wright D, Kagan K, Molina F, Gazzoni A, Nicolaides K. A mixture model of nuchal translucency thickness in screening for chromosomal defects. Ultrasound Obstet Gynecol. 2008;31(4):376–83.
Riley RD and Collins GS. Stability of clinical prediction models developed using statistical or machine learning methods. Biometric J. 2023;65(8):2200302.
Spiegelhalter D, KnillJones R. Statistical and knowledgebased approaches to clinical decisionsupport systems, with an application in gastroenterology. J R Statl Soc Series A. 1984;147(1):35–77.
Knottnerus J. Application of logistic regression to the analysis of diagnostic data: exact modeling of a probability tree of multiple binary varibles. Med Decis Mak. 1992;12(2):93–108.
Cardoso P, Dennis JM, Bowden J, Shields BM and McKinley TJ. Dirichlet process mixture models to impute missing predictor data in counterfactual prediction models: an application to predict optimal type 2 diabetes therapy. BMC Med Inform Decis Mak. 2024; 24(12). https://doi.org/10.1186/s12911023024003.
Riley RD, Snell KI, Burke DL, Harrel FE Jr, Moons KG, Collins GS. Minimum samples size for developing a multivariate prediction model: part II  binary and timetoevent outcomes. Stat Med. 2019;38:1276–96.
Acknowledgements
For the purpose of open access, the author has applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission.
Funding
PC and TJM (McKinley) are funded by Research England’s Expanding Excellence in England (E3) fund. KAP is a Wellcome Trust fellow (219606/Z/19/Z). This work was further supported by Diabetes UK (reference 21/0006328), the National Institute for Health and Care Research (NIHR) Exeter Biomedical Research Centre (BRC) and the National Institute for Health and Care Research Exeter Clinical Research Facility. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR, the Wellcome Trust or the Department of Health and Social Care.
Author information
Authors and Affiliations
Contributions
PC, BMS and TJM (McKinley) conceived and designed the study. PC, BMS and TJM (McKinley) analysed the data and developed the code. KAP and TJM (McDonald) provided the cohort data partly used to estimate the prior distribution of a positive MODY test result, into those with Cpeptide negative or autoantibody positive test results. ATH and ERP led the UNITED population study. All authors contributed to the writing of the article, provided support for the analysis and interpretation of results, critically revised the article, and approved the final article.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
For the UNITED study, ethics approval was granted by the NRES Committee South West–Central Bristol (REC no. 10/H0106/03). For the casecontrol data, this study was approved by the Genetic Beta Cell Research Bank, Exeter, UK (ethical approval was provided by the North Wales Research Ethics Committee, UK (IRAS project ID 231760)) and the Peninsula Research Bank (Devon & Torbay Research Ethics Committee, ref: 2002/7/118).
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Cardoso, P., McDonald, T.J., Patel, K.A. et al. Comparison of Bayesian approaches for developing prediction models in rare disease: application to the identification of patients with MaturityOnset Diabetes of the Young. BMC Med Res Methodol 24, 128 (2024). https://doi.org/10.1186/s1287402402239w
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s1287402402239w