MCMC traces for the manuscript Simulation-based inference for rapid Bayesian parameter estimation in epidemiological models: a comparison with MCMC

This dataset contains Markov chain Monte Carlo (MCMC) posterior traces used in the study Simulation-based inference for rapid Bayesian parameter estimation in epidemiological models: a comparison with MCMC. The traces were generated by calibrating a deterministic, single-age-group SECIR epidemiological model implemented with MEmilio to daily German COVID-19 intensive care unit (ICU) occupancy data. The inferred parameters comprise five disease-stage durations, the relative infectiousness of symptomatic individuals, and time-dependent contact-reduction factors defined on a 13-day changepoint grid. The dataset covers four inference settings: Three short inference windows containing 31 daily observations, beginning at offsets 0, 160, and 200 relative to 24 April 2020. One extended inference window containing 201 daily observations, beginning at offset 0. Posterior sampling used the differential-evolution Metropolis-Z algorithm implemented in PyMC. Each run used 16 chains. The short-window runs contain 100,000 retained draws per chain. The 201-day run was generated with 1,000,000 draws per chain and subsequently thinned by a factor of 100 to reduce storage requirements. The files use the NetCDF representation of an ArviZ InferenceData object. They contain posterior samples for the inferred SECIR parameters and contact-reduction factors. Where available, the posterior group also contains the derived ICU trajectory variable critical pred. These traces serve as the MCMC reference distributions for comparisons with simulation-based inference, including marginal posterior plots, first-order Wasserstein distances, Kullback-Leibler divergences, convergence diagnostics, and posterior-predictive analyses. File descriptions File Description uniform_trace_30_100000_0_False_13.nc MCMC trace for the 31-observation window beginning at offset 0. Contains 16 chains with 100,000 retained draws per chain. uniform_trace_30_100000_160_False_13.nc MCMC trace for the 31-observation window beginning at offset 160. Contains 16 chains with 100,000 retained draws per chain. uniform_trace_30_100000_200_False_13.nc MCMC trace for the 31-observation window beginning at offset 200. Contains 16 chains with 100,000 retained draws per chain. uniform_change_poins_trace_200_1000000_0_False_13_thin100.nc MCMC trace for the 201-observation window beginning at offset 0. The original run used 16 chains and 1,000,000 draws per chain; the deposited trace was thinned by a factor of 100. In these filenames, 30 and 200 denote the final simulated day. Because day zero is included, they correspond to 31 and 201 observations, respectively. The final value 13 denotes the changepoint interval in days. The corresponding github repository https://codebase.helmholtz.cloud/loki/memiliflow/-/tree/sbi-48-add-changepoint-inference