kaneplusplus / basket Goto Github PK

Do we want to remove it from the package for now?

1. Incorporate cluster analysis into full_bayes_mcmc()

2. Organize full_bayes_mcmc() output into 3 lists

   • inputs
   • basketwise output
   • clusterwise output

3. summary function
   post prob
   hpd
   median
   mean
   ESS

4. Update function to re-compute posterior prob with new p0

   • basketwise for new p0 vector and alternative input
   • clusterwise for p0 vector of same size as number of cluster and alternative input

5. Plot function

   • Basket densities
   • Prior, MAP, PEP exist as 1x3 vector
   • Segments HPDs?

6. Make full_bayes_exact() consistent with full_bayes_mcmc()
   unique to exact = PEP and posterior probability

Comment by @zabore from issue #33 .

Hi there,

I currently use the mem_exact function a lot for simulations with a relatively small number of baskets (3-5). I noticed that the function only calculates the posterior probabilities exactly, but the posterior mean, median and HPD-intervals are still estimated from MCMC samples. This slows down computation a lot, since 100.000 samples are drawn and there is no way to change this. Drawing the posterior samples takes much longer than calculating the exact posterior probabilities for a small number of baskets.

Is there a specific reason that the rest of the output isn't also calculated exactly? Since the posterior weights are already calculated in the function, it seems that it would be much faster to also calculate the posterior mean and median exactly. It would be awesome if this could be changed!

The following code shows that there is one cluster found but there are 3 Nulls (one for each basket). This is confusing.

library(basket)
data(vemu_wide)
baskets <- 1:3
vemu_wide1 <- vemu_wide[baskets,]
mcmc_res1 <- mem_mcmc(
  responses = vemu_wide1$responders,
  size = vemu_wide1$evaluable,
  name = vemu_wide1$baskets,
  p0 = 0.15)
mcmc_res1$cluster$post.prob
mcmc_res1$cluster$p0

Two things still need to be done for the SEM implementation:

1. Validate that the package code gives similar results to the reference code.
2. Add regression testing. This may include creating a reference data set from the data-raw directory.

Code is available on the Presagia One Drive.

The original code should be added to the basket/inst/ directory.

From the JSS editor:

The diverging color scale in Figure 4 is rather unbalanced and not ideal
for viewers with color vision deficiencies. For example,
grDevices::hcl.colors() offers "Green-Brown" or "Blue-Red 3". But other
packages like RColorBrewer, rcartocolors, colorspace, etc. also offer
alternatives.

logMarg.Dens
logMarg.DensSA
logMarg.DensPE

Code coverage needs to get above 90%.

cluster_PEP should take an mem object, not the individual parameters containted by an mem object.

This should be done in mem_full_bayes and mem_empirical_bayes.