search

stuwrighthealthecon / MANC-RISK-SCREEN

This repository contains the files required to run a discrete event simulation based economic evaluation of a number of different screening strategies to detect breast cancer in women in the UK
3 stars 1 forks source link

Investigate variance reduction #3

Closed stuwrighthealthecon closed 1 year ago

stuwrighthealthecon commented 2 years ago

Look into variance reduction techniques to reduce model-related variance in results and reduce required number of runs for stable results

stuwrighthealthecon commented 2 years ago

Suggestion from validation review

Whilst R offers the ability to fix streams of random numbers with set.seed this model relies on large number of randomly drawn patients to be modelled to reduce Monte Carlo error to acceptably small values which negates the requirement to fix streams of random numbers. Despite this there is no reason that random streams could not be fixed to more closely align each model run if desired.

The user guide to the model suggests running many million simulations to minimise Monte Carlo error which should negate any requirement to draw the sample of patients for each screening option. Drawing the same random stream of patients would seem logical, given that there are only 15613 risk profiles to choose from, it may make more sense to the reviewer to sample each patient a given number of times. When combined with the R function set.seed() identical random draws could be applied to each treatment arm. See Appendix 1 for more details.

stuwrighthealthecon commented 1 year ago

Significant changes made in speedboost branch to be version 1.1. Now pre-generate a sample which can be run through different strategies rather than re-generating each time. Also changed discounting and removed women who would not differ between strategies (i.e. cancer before screening age)