boehringer-ingelheim / bayesianmcpmod Goto Github PK

Calculation of success probabilities: A matter of taste, but some may like to see a progress bar (such as utils::txtProgressBar to get an idea of how long one still needs to wait)
Maybe worthwhile to mention that the residual SD is assumed to be the same as in the prior_list object? Or allow option to specify SD different from the one in prior list.
Again matter of taste, but maybe not use “Model Significance Frequencies” for a Bayesian approach, maybe “Success Rate per Model” or similar is more appropriate?

Potential features:

• "Nc" argument to specify the number of prior components
• is it possible to allow informative priors for non-control doses?

Unable to access the Summary of Posterior Distributions table - this would be good output to be able to programmatically access

• Output is not clear - the default return is success rate and number of sims - however using BMCP_result[1,1:5] you seem to get individual simulation results. It is not intuitive how to use this information
• For the individual simulation results, I would not call the contrast value evaluation a p-value. It's confusing and is not a p-value
• Consider adding the contrast critical value in the simulation output so that you can easily see which models are "significant"

A wrapper function, like BayesianMCPMod() that puts all of these functions together would be useful

Is there a way to assess a design on an effect that doesn't come from a parametric dose-response model (ex: custom effect with a dose-response mean vector as an input)?

Additional criteria to consider for assessDesign():

• Dose-response estimates, raw or placebo adjusted
• Quantile of the posterior distribution, ex: Pr(D_i - D_0) > threshold
• Target dose success or based on credible intervals

• I assume that this modeling is based on AIC weighting. I think the weighted probabilities for each model would be of interested to return/plot in addition to the AIC
• Individual dose summaries would be of interest (either candidate doses or interpolated doses) for the model fitting
• Control adjusted summaries would also be of interest (D_i - D_0)
• Ability to specify the quantiles of the CrI would be very useful
• Target dose estimation would also be preferred
• Posterior summaries of the form Pr(D_i > x) or placebo-adjusted summaries of the form Pr(D_i - D_0) is very useful
• Actual samples from the posterior would be of highest interest so that the user can create their own summaries and decision criteria