Retries#

The retries tool is a tool used to tweak the initial estimates of a given model. The initial estimates of all parameters are changed slightly which can give better models or show instabilities for the model. When comparing and ranking the new candidates, the best model is based solely on the ‘OFV’ value.

Running#

The retries tool is available both in Pharmpy and pharmr.

An example of how to initiate the tool can be seen below

from pharmpy.modeling import read_model
from pharmpy.tools read_modelfit_results, run_structsearch

start_model = read_model('path/to/model')
start_model_results = read_modelfit_results('path/to/model')
res = run_retries(model = start_model,
                  results = start_model_results,
                  number_of_candidates = 5,
                  scale = "normal")

from pharmpy$tools read_modelfit_results, run_structsearch

start_model <- read_model('path/to/model')
start_model_results <- read_modelfit_results('path/to/model')
res <- run_retries(model <- start_model,
                  results <- start_model_results,
                  number_of_candidates <- 5,
                  scale <- "normal")

Arguments#

The arguments of the retries tool are listed below.

Argument	Description
`model`	Start model to run retries on
`results`	ModelfitResults of the start model
`strictness`	Strictness criteria for model selection. Default is “minimization_successful or (rounding_errors and sigdigs>= 0.1)”
`number_of_candidates`	Number of retry-models to run. 5 is used as default
`fraction`	Determine the allowed increase/decrease for the randomly generated new initial estimates. Default is 0.1 (10%).
`use_initial_estimates`	Use initial parameter estimates when creating candidate models instead of the final parameter estimates of the input model.
`scale`	Scale to use when randomizing the initial estimates. Currently supported scales are `normal` and `UCP`.
`prefix_name`	String determining prefix of model names such that models are named {prefix_name}_retries_run2 for instance.
`seed`	A random number generator or seed to use for random sampling.

Scales#

There are different ways for the tool of computing the retry models. A shared behaviour however, is that all new initial estimates of parameters are randomly generated. The available scales can be seen in the table below.

scales	Description
`'normal'`	Will iteratively try out new, randomly generated initial estimates. If the covariance matrix for OMEGA(s) and SIGMA(s) is determined to not be positive semi-definite, new random initial estimates is generated (maximum of 20 times)
`'UCP'`	All parameters get new are transformed to the Unconstrained Parameter Scale (UCP scale) where new initial values are generated before inverting the transformation. Through this method, the values are only required to be randomized once.

Normal#

Using the normal scale, the tool is similar to the one found in PsN’s tool parallelle retries. For all parameters, the given initial estimates is used as a center point for a uniform distribution. The new random initial value is then taken from a uniform distribution with bound 10% above and below the given initial estimate (or the given upper or lower bound). The new value is allowed to be at most 10^(⁻6) units from any of the bounds. This is performed for all parameters (THETAs, OMEGAs and SIGMAs).

Once all new initial values have been computed, the covariance matrix (made from OMEGAs and SIGMAs) will be checked to see if it’s positive semidefinite or not. If it is, the new retry model will be fitted. If not, new initial values will be generated for all parameters as described above. This is performed a maxiumum of 20 times. If a positive semi definite covariance matrix has not been found by then, an error is raised.

UCP#

As opposed to the normal scale, the UCP scale will always return a model with functioning initial values for all parameters. This is ensured through the use of Unconstrained Parameters. All parameters are transformed to Unconstrained Parameters. In this value space, new initial values are computed with the same allowed variance, allowing values to be taken from a uniform distribution of values from 10% below, to 10% above the current initial estimate. With these changes applied, the inverse Unconstrained Parameters transformation is performed.

As the Unconstrained Parameter Scale is different for each parameter, the relative change might be different for all parameters and is not limited to +/- 10% of the given inital estimate.

The Retries results#

The results object contains various summary tables which can be accessed in the results object, as well as files in .csv/.json format. The name of the selected best model (based on the input selection criteria) is also included.

An example of an entire retries run can be seen below

start_model = read_model('path/to/model')
start_model_results = read_modelfit_results('path/to/model')
res = run_structsearch(model = start_model,
                        results = start_model_results,
                        number_of_candidate = 5,
                        fraction = 0.1,
                        scale = "UCP")

start_model <- read_model('path/to/model')
start_model_results <- read_modelfit_results('path/to/model')
res <- run_structsearch(model <- start_model,
                        results <- start_model_results,
                        number_of_candidate <- 5,
                        fraction <- 0.1,
                        scale <- "UCP")

The summary_tool table contains information of the model results and final ranking. It also contains information regarding how many attempts it took for the model to successfully find a positive semi definite covariance matrix:

	Number_of_retries	description	n_params	d_params	dofv	ofv	rank	parent_model
model
retries_run1	1.0	PHENOBARB SIMPLE MODEL	6	0	1.922734e-07	586.276056	1	pheno
retries_run4	1.0	PHENOBARB SIMPLE MODEL	6	0	1.383697e-07	586.276056	2	pheno
retries_run3	1.0	PHENOBARB SIMPLE MODEL	6	0	1.302390e-07	586.276056	3	pheno
retries_run5	1.0	PHENOBARB SIMPLE MODEL	6	0	1.061417e-07	586.276056	4	pheno
pheno	NaN	PHENOBARB SIMPLE MODEL	6	0	0.000000e+00	586.276056	5	pheno
retries_run2	1.0	PHENOBARB SIMPLE MODEL	6	0	-4.387580e-07	586.276057	6	pheno

To see information about the actual model runs, such as minimization status, estimation time, and parameter estimates, you can look at the summary_models table. The table is generated with pharmpy.tools.summarize_modelfit_results().

		description	minimization_successful	errors_found	warnings_found	ofv	runtime_total	estimation_runtime	PTVCL_estimate	PTVCL_SE	PTVCL_RSE	...	THETA_3_RSE	IVCL_estimate	IVCL_SE	IVCL_RSE	IVV_estimate	IVV_SE	IVV_RSE	SIGMA_1_1_estimate	SIGMA_1_1_SE	SIGMA_1_1_RSE
step	model
0	pheno	PHENOBARB SIMPLE MODEL	True	0	0	586.276056	4.0	0.32	0.004696	0.00021	0.044731	...	0.527072	0.029351	0.013415	0.457068	0.027906	0.007477	0.267918	0.013241	0.002279	0.172147
1	retries_run1	PHENOBARB SIMPLE MODEL	True	0	0	586.276056	21.0	0.87	0.004696	0.00021	0.044758	...	0.527096	0.029349	0.013418	0.457176	0.027906	0.007471	0.267734	0.013241	0.002276	0.171888
	retries_run2	PHENOBARB SIMPLE MODEL	True	0	0	586.276057	21.0	0.84	0.004696	0.00021	0.044716	...	0.527011	0.029350	0.013416	0.457114	0.027906	0.007475	0.267856	0.013241	0.002280	0.172226
	retries_run3	PHENOBARB SIMPLE MODEL	True	0	0	586.276056	21.0	0.84	0.004696	0.00021	0.044732	...	0.527151	0.029351	0.013415	0.457044	0.027907	0.007476	0.267878	0.013241	0.002280	0.172227
	retries_run4	PHENOBARB SIMPLE MODEL	True	0	0	586.276056	8.0	0.40	0.004696	0.00021	0.044728	...	0.527073	0.029351	0.013416	0.457083	0.027907	0.007468	0.267603	0.013241	0.002281	0.172295
	retries_run5	PHENOBARB SIMPLE MODEL	True	0	0	586.276056	21.0	0.89	0.004696	0.00021	0.044669	...	0.526863	0.029351	0.013413	0.457001	0.027907	0.007477	0.267944	0.013241	0.002275	0.171784

6 rows × 25 columns

Finally, you can see a summary of different errors and warnings in summary_errors. See pharmpy.tools.summarize_errors() for information on the content of this table.

			time	message
model	category	error_no