Sepia shared/hierarchical theta models

Collections of multiple models with shared thetas or hierarchically-modeled thetas are created using these classes.

class sepia.SepiaSharedThetaModels.SepiaSharedThetaModels(model_list=None, shared_theta_inds=None)

Container for multiple models with sharing of selected thetas between models.

Variables

  • model_list (list) – list of instantiated sepia.SepiaModel objects

  • shared_theta_inds (numpy.ndarray) – indices showing which thetas are shared across models, shape (n_hier_theta, n_models)

  • n_hier (int) – number of shared theta groups

  • n_models (int) – number of models

  • to_update (numpy.ndarray) – 0/1 matrix indicating which variables need to be shared

  • to_sample (numpy.ndarray) – 0/1 matrix indicating which variable should be sampled (then copied to others)

Instantiate shared theta model object.

Parameters

  • model_list (list) – list of instantiated sepia.SepiaModel objects

  • shared_theta_inds (numpy.ndarray) – indices showing which thetas are shared, shape (n_shared_theta, n_models)

Raises

TypeError – if number of models doesn’t match shared_theta_inds shape or if variable types aren’t the same (categorical vs continuous)

Note

In shared_theta_inds, each row corresponds to one group of shared thetas, and each column gives the index of the theta within a particular model, with -1 used to indicate no theta from a particular model is part of the shared group. Example: shared_theta_inds = np.array([(1, 1, 1), (2, -1, 4)) for 3 models, theta index 1 shared across all models, theta indices 2/4shared across models 1 and 3 but no corresponding theta in model 2.

do_mcmc(nsamp, do_propMH=True, prog=True)

Do MCMC for shared theta model.

Parameters

  • nsamp (int) – number of MCMC samples

  • do_propMH (bool) – use propMH sampling for params with stepType propMH?

  • prog (bool) – show progress bar for sampling?

get_samples(nburn=0, sampleset=None, numsamples=None, flat=True, includelogpost=True)

Extract MCMC samples into dictionary format for each model in self.model_list. By default, all samples are returned, or samples can be subset using in various ways using the optional input arguments.

Parameters

  • nburn (int) – number of samples to discard at beginning of chain

  • sampleset (list) – list of indices of samples to include

  • numsamples (int) – number of samples to include, evenly spaced from first to last

  • flat (bool) – flatten the resulting arrays (for parameters stored as matrices)?

  • includelogpost (bool) – include logPost values?

Returns

dict of dict – one dict per model, each dict is array of samples for each parameter, keyed by parameter name

Raises

TypeError if no samples exist or nburn inconsistent with number of draws

Note

Adds key theta_native with theta rescaled to original range.

class sepia.SepiaHierarchicalThetaModels(model_list, hier_theta_inds)

Container for multiple models with hierarchical Normal model on selected thetas.

Variables

  • model_list (list) – list of instantiated sepia.SepiaModel objects

  • hier_theta_inds (numpy.ndarray) – indices showing which thetas are hierarchically linked, shape (n_hier_theta, n_models)

  • n_hier (int) – number of hierarchical groups

  • n_models (int) – number of models

  • to_update (numpy.ndarray) – 0/1 matrix indicating which variables need to be hierarchically updated

  • hier_mu (list) – list of sepia.SepiaParam objects for hierarchical mu parameters

  • hier_lambda (list) – list of sepia.SepiaParam objects for hierarchical lambda parameters

  • hier_delta (list) – list of sepia.SepiaParam objects for hierarchical delta (lockstep update) parameters

Instantiate hierarchical model container.

Parameters

  • model_list (list) – list of instantiated sepia.SepiaModel objects

  • hier_theta_inds (numpy.ndarray) – indices showing which thetas are hierarchically linked, shape (n_hier_theta, n_models)

Raises

TypeError – if number of models doesn’t match hier_theta_inds or if user attempts to model categorical variable hierarchically.

Note

In hier_theta_inds, each row corresponds to one group of hierarchically modeled thetas, and each column gives the index of the theta within a particular model, with -1 used to indicate no theta from a particular model is part of the hierarchical group. Example: hier_theta_inds = np.array([[1, 1, 1], [2, -1, 4]]) for 3 models, theta index 1 hierarchical across all models, theta indices 2/4 hierarchical across models 1 and 3 but no corresponding theta in model 2.

do_mcmc(nsamp, do_propMH=True, prog=True, do_lockstep=True)

Does MCMC for hierarchical model.

Parameters

  • nsamp (int) – number of MCMC samples

  • do_propMH (bool) – use propMH sampling for params with stepType propMH?

  • prog (bool) – show progress bar for sampling?

  • do_lockstep (bool) – do lockstep updates?

get_samples(nburn=0, sampleset=None, numsamples=None, flat=True)

Extract MCMC samples into dictionary format for each model in self.model_list. By default, all samples are returned, or samples can be subset using in various ways using the optional input arguments.

Parameters

  • nburn (int) – number of samples to discard at beginning of chain

  • sampleset (list) – list of indices of samples to include

  • numsamples (int) – number of samples to include, evenly spaced from first to last

  • flat (bool) – flatten the resulting arrays (for parameters stored as matrices)?

Returns

dict of dict – one dict per model, each dict is array of samples for each parameter, keyed by parameter name

Raises

TypeError if no samples exist or nburn inconsistent with number of draws

Note

Adds key theta_native with theta rescaled to original range.