The (in)stability of Bayesian model selection criteria in disease mapping

Vranckx, Maren; Neyens, Thomas; Faes, Christel

doi:10.1016/j.spasta.2021.100502

Cited by 8 publications

(10 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Although here DIC and WAIC calculated with R-INLA and NIMBLE show very small differences and clearly point towards the model with a Type IV interaction, Vranks et al 34 warn about disparities among model selection criteria values provided by different software packages. Differences may be due to software-specific posterior samples or softwarespecific calculations of the model selection statistics.…”

Section: Spanish Breast Cancer Mortality Datamentioning

confidence: 57%

“…Table 2 displays the values of the DIC and the WAIC together with the mean deviance (D(θ)) and the effective number of parameters (p D ) for the spatio-temporal models with the ICAR and BYM spatial priors, the two forms of including the RW1 in NIMBLE, Nimble 1 and Nimble 2, the four types of interaction, and the set of hyperpriors H1 (results for hyperpriors H2 and H3 are shown in Tables A.1 and A.2 respectively in the Supplementary Materials). In general, DIC and WAIC values that differ by less than five point towards the same model (see for example Vranks et al 34 ). According to these criteria, the ICAR and BYM spatial priors lead to very similar results, and the Type IV interaction is preferred over the others.…”

Section: Spanish Breast Cancer Mortality Datamentioning

confidence: 93%

See 1 more Smart Citation

Space-time interactions in Bayesian disease mapping with recent tools: Making things easier for practitioners

Urdangarin

Goicoa

Ugarte

2022

Stat Methods Med Res

View full text Add to dashboard Cite

Spatio-temporal disease mapping studies the distribution of mortality or incidence risks in space and its evolution in time, and it usually relies on fitting hierarchical Poisson mixed models. These models are complex for practitioners as they generally require adding constraints to correctly identify and interpret the different model terms. However, including constraints may not be straightforward in some recent software packages. This paper focuses on NIMBLE, a library of algorithms that contains among others a configurable system for Markov chain Monte Carlo (MCMC) algorithms. In particular, we show how to fit different spatio-temporal disease mapping models with NIMBLE making emphasis on how to include sum-to-zero constraints to solve identifiability issues when including spatio-temporal interactions. Breast cancer mortality data in Spain during the period 1990-2010 is used for illustration purposes. A simulation study is also conducted to compare NIMBLE with R-INLA in terms of parameter estimates and relative risk estimation. The results are very similar but differences are observed in terms of computing time.

show abstract

Section: Spanish Breast Cancer Mortality Datamentioning

confidence: 57%

Section: Spanish Breast Cancer Mortality Datamentioning

confidence: 93%

Space-time interactions in Bayesian disease mapping with recent tools: Making things easier for practitioners

Urdangarin

Goicoa

Ugarte

2022

Stat Methods Med Res

View full text Add to dashboard Cite

show abstract

“…The variables age and gender have significant effects, while the interaction effect between age and gender is not significant. Leaving out the interaction term gives similar model results which shows robustness in terms of model fit, and a very small difference in WAIC estimates is observed, leading to no clear support for the model with or without the interaction term (Vranckx et al., 2021 ). Therefore, we here present the model with the interaction term between age and gender due to the primary purpose of using the age and gender variable as confounding variables in this epidemiological process.…”

Section: Resultsmentioning

confidence: 91%

“…Moreover, we investigated which random‐effects structure performs better when prediction is the objective using the LMPL (log marginal predictive likelihood; Geisser & Eddy, 1979 ). Only very small differences between the LMPL estimates of the different options were noticed, which were too small to prefer a single structure (Vranckx et al., 2021 ). Hence, based on the WAIC, this section gives the results of the model where the processes generating the confirmed cases and symptoms are linked via a bivariate spatial correlation, in addition to a univariate normal uncorrelated heterogeneity term for the symptoms' processes and a gamma uncorrelated heterogeneity term for the confirmed cases' process.…”

Section: Resultsmentioning

confidence: 99%

A spatial model to jointly analyze self‐reported survey data of COVID‐19 symptoms and official COVID‐19 incidence data

et al. 2022

Self Cite

View full text Add to dashboard Cite

This work presents a joint spatial modeling framework to improve estimation of the spatial distribution of the latent COVID‐19 incidence in Belgium, based on test‐confirmed COVID‐19 cases and crowd‐sourced symptoms data as reported in a large‐scale online survey. Correction is envisioned for stochastic dependence between the survey's response rate and spatial COVID‐19 incidence, commonly known as preferential sampling, but not found significant. Results show that an online survey can provide valuable auxiliary data to optimize spatial COVID‐19 incidence estimation based on confirmed cases in situations with limited testing capacity. Furthermore, it is shown that an online survey on COVID‐19 symptoms with a sufficiently large sample size per spatial entity is capable of pinpointing the same locations that appear as test‐confirmed clusters, approximately 1 week earlier. We conclude that a large‐scale online study provides an inexpensive and flexible method to collect timely information of an epidemic during its early phase, which can be used by policy makers in an early phase of an epidemic and in conjunction with other monitoring systems.

show abstract

“…With the same PI coverage, the model with the smaller MIS has less uncertainty [12]. In addition, goodness-of-fit of models are compared using WAIC, with smaller values (difference larger than 5) corresponding to a better fitting model [26,27]. Figure 3 presents the 5-day ahead predictions for the number of new hospitalisations at different phases of the epidemic.…”

Section: Epidemiology and Infectionmentioning

confidence: 99%

Multivariate phenomenological models for real-time short-term forecasts of hospital capacity for COVID-19 in Belgium from March to June 2020

et al. 2021

Self Cite

View full text Add to dashboard Cite

Phenomenological models are popular for describing the epidemic curve. We present how they can be used at different phases in the epidemic, by modeling the daily number of new hospitalisations (or cases). As real-time prediction of the hospital capacity is important, a joint model of the new hospitalisations, number of patients in hospital and in intensive care unit is proposed. This model allows estimating the length of stay in hospital and ICU, even if no (or limited) individual level information on length of stay is available. Estimation is done in a Bayesian framework. In this framework, real-time alarms, defined as the probability of exceeding hospital capacity, can be easily derived. The methods are illustrated using data from the covid-19 pandemic in March-Jun 2020 in Belgium, but are widely applicable.

show abstract

The (in)stability of Bayesian model selection criteria in disease mapping

Cited by 8 publications

References 19 publications

Space-time interactions in Bayesian disease mapping with recent tools: Making things easier for practitioners

Space-time interactions in Bayesian disease mapping with recent tools: Making things easier for practitioners

A spatial model to jointly analyze self‐reported survey data of COVID‐19 symptoms and official COVID‐19 incidence data

Multivariate phenomenological models for real-time short-term forecasts of hospital capacity for COVID-19 in Belgium from March to June 2020

Contact Info

Product

Resources

About