This is a transference of the ergm-private" "control_llik" branch to the public version ergm 4.0. #337
Conversation
…the public version ergm 4.0. The branch consolidates control arguments to the likelihood functions used in ergm.estimate into a single list variable called "control.llik". This simplifies the calls to the functions. The branch also adds a simple hook in "ergm.estimate" for the Kpenalty likelihood functions. These are required for the "ergm'tapered" package to interoperate with "ergm". Also fixed a few documentation typos in "ergm".
|
To make life easy, I created a branch of ergm 4.0 and manually made the edits from the private "control_llik" branch. As it turned out it was necessary as ergm 4.0 had evolved so that automatic merging would have been a mess. |
|
Thanks, Mark! I'll take a look the week after Sunbelt. |
…llik # Conflicts: # man/ergm-terms.Rd
These required extensive algebra and computation. However, they dramatically improve the computational speed and accuracy of the Kpenalty method.
| @@ -48,6 +48,7 @@ Suggests: | |||
| tergm (>= 4.0.0), | |||
| ergm.count (>= 4.0), | |||
| ergm.userterms (>= 3.10.0), | |||
| ergm.tapered (>= 1.0.0), | |||
There was a problem hiding this comment.
Since ergm.tapered is not on CRAN, we also need to add a Remotes: directive.
That said, we might want to keep it away from requirements to avoid a chicken-and-egg problem the first time we release.
| @@ -30,36 +30,69 @@ | |||
| #' \eqn{q}-vectors of estimating function values. | |||
| #' @keywords internal | |||
| #' @export | |||
| ergm.estfun <- function(stats, theta, model, ...){ | |||
| ergm.estfun <- function(stats, theta, model, exclude=NULL, ...){ | |||
There was a problem hiding this comment.
I am not sure whether this is a good idea in its current form:
- It doesn't seem to do anything that subsetting the function's output wouldn't already. Both matrices and
mcmc.listobjects already support column subsetting. If it did the subsetting prior to the estimating function calculation, that might save computing time, but it doesn't do that, as far as I can tell. - Matching columns by name can be very unreliable, because in
ergmyou can have duplicate coefficient names, at least at the moment.
There was a problem hiding this comment.
Finally replying to this. I tried the subsetting but can not get it to work as there are also places like ergm.bridge where it is needed. In the end I just left it as an option as it is cleaner and more flexible.
| @@ -57,6 +57,12 @@ ergm.stocapprox <- function(init, nw, model, | |||
| stats <- model$nw.stats - NVL(model$target.stats,model$nw.stats) | |||
| target.stats <- model$target.stats | |||
| }) | |||
|
|
|||
| control.llik <- list(MCMLE.metric=control$MCMLE.metric, MCMLE.trustregion=control$SA.trustregion, | |||
There was a problem hiding this comment.
Do we want to implement a control.ergm.llik.<METRIC>() family of functions so that the user could potentially specify different parameters for each metric and get argument completion etc.?
There was a problem hiding this comment.
Maybe latter. At the moment the list is small and has not changed for a long while (until ergm.tapered/Kpenalty. At some point it may be worth adding if/when the list grows
| MCMLE.metric=c("lognormal", "logtaylor", | ||
| "Median.Likelihood", | ||
| "EF.Likelihood", "naive"), | ||
| "EF.Likelihood", "Kpenalty", "naive"), |
There was a problem hiding this comment.
Point of clarification: is Kpenalty an alternative way to estimate the likelihood ratio, or is it estimating the penalised likelihood?
There was a problem hiding this comment.
Kpenalty estimates the penalized likelihood. Explicitly, is computed the log-likelihood ratio (using lognormal) and then the kurtosis estimates to add to the log-likelihood ratio as the penalty. Otherwise, it is close to the standard lognormal version. The gradient is computed (mostly) exactly.
Added a MPLE.save.xmat control argument to return the design matrix from the MPLE fit (if any). The current and retained default is to remove it.
control$MCMC.esteq.exclude.statistics allows network statistics to not be included in the estimating equations. There equations determine most convergence diagnostics. For some models there are some terms that are not zero at convergence and should be removed. For example, this supports tapering models which may have terms with non-specific means. The estimating equations are used in a number of places and the code follows those places. Specifically, it is a vector of strings being the names of the statistics to be excluded from the estimating equations.
It now passes R CMD check
statnet-janitor
force-pushed
the
master
branch
from
12d1476
to
951da42
Compare
The branch consolidates control arguments to the likelihood functions used in ergm.estimate into a single list variable called "control.llik".
This simplifies the calls to the functions.
The branch also adds a simple hook in "ergm.estimate" for the Kpenalty likelihood functions. These are required for the "ergm'tapered" package to interoperate with "ergm".
Also fixed a few documentation typos in "ergm".