dwinter / mdm Goto Github PK

objects returned by fit_mdm() and fit_dm() should have methods including (at least):

• print
• coef
• logLik
• AIC
• BIC
• vcov

and possibly summary() ?

At present the mdmParam building function would return Inf for all counts when \phi = 0. To avoid this we test that phi must be in (0,1]. We should probably do something special for the 0 case so we can have contibunuty between D-M and Multinomial models.

An example:

hets <- data.frame(
  ref = c(4260, 3161, 3166, 3152, 3167, 3744, 3536, 3908, 3420),
  alt = c(1800, 1239, 1331, 1327, 1380, 1892, 1673, 2033, 1761)
)
fit_mdm(hets,2)

Error in svd(X) : infinite or missing values in 'x'
In addition: Warning message:
In fit_mdm(X, 2) : Cycle 1: Log-Likelihood decreased by Inf!

traceback gives this... which looks an intermediate matrix is getting bad values

4: stop("infinite or missing values in 'x'")
3: svd(X)
2: ginv(Io)
1: fit_mdm(X, 2)

A little more sleuthing suggests this is occruing in mdmAugmentData but that's as far as I got.

It would be useful to have a utility function to calculate the probability that given observation belongs to a particular model-component.

This calculation is already performed as part of the EM algorithm , but should be exposed as a function for users.

mdm is both boring and and not available if we want to get this on CRAN.

(Topic for next lab meeting?)

Which means they aren't directly relatable to the outputs of dmultinom from base R.

This is not a problem for fitting the models, so we may simply need to document this difference rather than add the extra calculation.