Resolve an issue with starting betas being reset

[lhjohn]

@egillax and I started implementing #59 and came across a bug. During cross validation the beta values are reset to 0.0, while they should be reset to the starting betas, if provided.

PR has the following additions.

• Add local variable to hold starting coefficients.
• Reset betas during CV to starting coefficients, if provided.
• Add some getters to the interface useful for debugging

Code to reproduce initial problem providing useCrossValidation = TRUE, startingCoefficients and fixedCoefficients :

library(Cyclops)

set.seed(1)
nCovars <- 10

# simulate the data
simulant <- simulateCyclopsData(nstrata = 1,
                                nrows = 5000,
                                ncovars = nCovars,
                                model = "logistic")

targetData <- convertToCyclopsData(simulant$outcomes,
                                   simulant$covariates,
                                   modelType = "lr")

# fix all coefficients, but not the intercept
fixedCoef <- c(FALSE,
               rep(TRUE, nCovars))

# initialize a vector for starting coefficients to 0.5
startingCoef <- rep(0.5, nCovars+1)

# useCrossValidation = FALSE
fit <- fitCyclopsModel(targetData,
                       prior = createPrior("laplace",
                                           variance = 0.1,
                                           useCrossValidation = FALSE),
                       forceNewObject = TRUE,
                       fixedCoefficients = fixedCoef,
                       startingCoefficients = startingCoef)

# store coefficients
fitCoef <- coef(fit)

# print coefficients without cross validation: starting value is taken into account
fitCoef

# useCrossValidation = TRUE
fit2 <- fitCyclopsModel(targetData,
                        prior = createPrior("laplace",
                                            variance = 0.1,
                                            useCrossValidation = TRUE),
                        forceNewObject = TRUE,
                        fixedCoefficients = fixedCoef,
                        startingCoefficients = startingCoef)

# store coefficients
fit2Coef <- coef(fit2)

# print coefficients with cross validation: starting value is not taken into account
fit2Coef

[lhjohn]

@msuchard Could you have a look at this issue? This bug prevents us from implementing transfer learning using priors. I'd be happy to submit a revised version if you have some feedback.

[msuchard]

@lhjohn -- did this get fixed already?

[lhjohn]

@msuchard this problem persists. Could you run the code snippet on your end to show that when using cross validation the starting coefficients are ignored? This is caused by beta values resetting to 0.f for the folds. I could provide a more minimal solution if that helps.