actuaryzhang / cplm Goto Github PK

Hi,
Great library; thanks for putting it together. Having an odd problem with running CPGLMM. I'm doing a few thousand regressions in a loop (apply isn't optimal for my use case). I don't expect all of them to converge, and have it set up in a tryCatch loop to handle this. However, I've found that once 1 regression fails, all those following it fail as well, even if when run in isolation they don't. If I start the loop at at different value (ie i=8 instead of i=1), regressions that previously failed do not...Any idea why this would be the case? I'm curious if using cpglmm changes some environmental feature that does not get shifted back when a regression completes?

Image 1 shows what happens if I start the loop at "regression 1," image 2 shows what happens if I start the loop at "regression 2."

Any help you can provide would be greatly appreciated.
Thanks,
Braden

See also our e-mail to you, from Jan.24, 2023.
The <Rdefines.h> header has become "semi - deprecated",
and so in the Matrix package we have stopped using. It contains a very simple macro GET_SLOT() which you have been using.

A back compatible change such that {cplm} continues to install fine with the upcoming Matrix 1.5-4 is

1 file changed, 5 insertions(+), 1 deletion(-)
src/common.h | 6 +++++-

modified   src/common.h
@@ -18,7 +18,11 @@
 #include <Rinternals.h>
 #include <Rmath.h>
 #include <R_ext/Lapack.h>        /* for BLAS and Lapack related */
-#include "Matrix.h"		 /* for cholmod functions and S4 structures (GET_SLOT)*/
+#include "Matrix.h"		 /* for cholmod functions and S4 structures */
+#ifndef GET_SLOT
+   // in Rdefines.h (formerly in Matrix.h) :
+#  define GET_SLOT(x, what)	R_do_slot(x, what)
+#endif
 #ifndef FCONE
 # define FCONE
 #endif

fitted(model) and predict(model, type='response') are supposed to return the same set of values. But it is not the case for the cplm model objects.

library(cplm)

da <- subset(AutoClaim, IN_YY == 1)
da <- transform(da, CLM_AMT = CLM_AMT / 1000)

P1 <- cpglm(CLM_AMT ~ 1, data = da, offset = log(NPOLICY))

# Compare the values of fitted and predicted with type response specified.

all.equal(fitted(P1), predict(P1, type='response'), check.attributes = FALSE)
# Mean relative difference: 0.5924107

# The values are different!
# Plot and see:

plot(fitted(P1))
plot(predict(P1, type='response'))

# But if added the `newdata` argument, it's now same.

all.equal(fitted(P1), predict(P1, newdata=da, type='response'), check.attributes = FALSE)
# TRUE

# Compare to other models, there is no such problem?
# I checked with `glm::glm()`, `MASS::glm.nb()`, and `tweedie()` with `glm::glm()`.

Hello,

Under R 4.1.0, I'm using cplm::cpglm() within a tryCatch.
The control argument is understanding things like max.iter = X, but fails to accepts trace = 0 (at least my understanding is that it should remove any verbose print in the console).

Code below:

fit <- tryCatch(
  expr = {
    cplm::cpglm(
      formula = formula, 
      data = data,
      control = list(trace = 0)
    )
  },
  error = function(cond) {
    # Stuff
  },
  warning = function(cond) {
    # Stuff
  }
)

Console

Call:
cplm::cpglm(formula = formula, data = data, control = list(trace = 0))
Deviance Residuals:
Min 1Q Median 3Q Max
-3.0237 -2.6788 -0.5744 0.3147 5.9825
Estimate Std. Error t value Pr(>|t|)
(Intercept) -9.3936 0.3071 -30.589 <2e-16 ***
Group_A_vs_Control 0.2241 0.4094 0.547 0.586
Group_B_vs_Control 0.1434 0.4179 0.343 0.732
Signif. codes: 0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.’ 0.1 ‘ ’ 1
Estimated dispersion parameter: 3.7397
Estimated index parameter: 1.3619
Residual deviance: 400.3 on 86 degrees of freedom
AIC: 441.87
Number of Fisher Scoring iterations: 5

Any tips on how to suppress the verbose output?

I had a user ask about emmeans support for cplm objects. I found that the terms method does not work in the namespace of my package. (See also rvlenth/emmeans#116) As explained in an answer on Stack Overflow -- https://stackoverflow.com/questions/56739673/flaky-s4-method-dispatch -- there exists an S3 generic for terms but there is no S3 method for cplm or cpglm objects. The answerer points to documentation recommending that such S3 methods be added, and how to do it. In particular, add and export functions like this:

terms.cplm <- function(x, ...) terms(x, ...)
terms.cpglm <- function(x, ...) terms(x, ...)

To test it, try something like:

m <- cpglm(...)
stats::terms(m)

I think you need to do this for all S4 methods you provide for which there also exists an S3 generic.

In emmeans, I opted to create S4 objects, but use S3 methods exclusively, no S4 ones, for exactly these kinds of reasons.