tests/testthat/test-estimation-hessian.R

# Reconstruct marginal covariance matrices independently from their free
# parameters, then differentiate the scalar Gaussian likelihood directly.
independent_profile_objective_2year <- function(theta, fit, side, persistence) {
  years <- 1L + as.numeric(fit$X[,2])
  n <- length(fit$y); q <- ncol(fit$Z)
  Z <- as.matrix(fit$Z); X <- as.matrix(fit$X)
  G <- matrix(0,q,q); cursor <- 0L; col <- 0L
  if(side=='G'){
    R <- diag(theta[years]);cursor <- fit$num.year
    G[seq_len(fit$num.student),seq_len(fit$num.student)] <- diag(theta[cursor+1L],fit$num.student)
    cursor <- cursor+1L;col <- fit$num.student
  }else{
    Ri <- matrix(0,2,2);Ri[lower.tri(Ri,diag=TRUE)]<-theta[1:3]
    Ri<-Ri+t(Ri)-diag(diag(Ri));R<-kronecker(diag(fit$num.student),Ri);cursor<-3L
  }
  for(g in seq_len(fit$num.year)){
    k<-fit$Kg[g];S<-matrix(0,k,k);nk<-k*(k+1)/2
    S[lower.tri(S,diag=TRUE)]<-theta[cursor+seq_len(nk)]
    S<-S+t(S)-diag(diag(S),nrow=k);cursor<-cursor+nk
    for(j in seq_len(fit$num.teach[g])){
      ids<-col+seq_len(k);G[ids,ids]<-S;col<-col+k
    }
  }
  if(persistence=='VP'){
    a<-theta[cursor+1L]
    idx <- which(years==2L)
    cols <- seq_len(fit$num.teach[1])
    zz<-Z[idx,cols,drop=FALSE];zz[zz!=0]<-a;Z[idx,cols]<-zz
  }
  V<-R+Z%*%G%*%t(Z); Vi<-solve(V)
  xv<-crossprod(X,Vi%*%X)
  beta<-solve(xv,crossprod(X,Vi%*%fit$y));r<-fit$y-X%*%beta
  val<-0.5*(n*log(2*pi)+as.numeric(determinant(V,TRUE)$modulus)+sum(r*(Vi%*%r)))
  if(fit$REML)val<-val-0.5*ncol(X)*log(2*pi)+0.5*as.numeric(determinant(xv,TRUE)$modulus)
  val
}

test_that("profile information matches direct ML and REML likelihood Hessians", {
  d <- vam_fixture(24L, 2L)
  for (side in c("R", "G")) {
    models <- if (side == "G") "GP" else c("GP", "rGP", "CP", "VP", "ZP")
    for (model in models) for (reml in c(FALSE, TRUE)) {
      fit <- suppressWarnings(GPvam(d, ~0+as.factor(year)+x, student.side = side,
          persistence = model, REML = reml, max.iter.EM = 12, hessian = TRUE,
          hes.method = "richardson", verbose = FALSE))
      objective <- function(theta) independent_profile_objective_2year(theta,
          fit, side, model)
      expect_equal(objective(fit$covariance_parameters), -fit$loglik,
                   tolerance = 1e-8, info = paste(side, model, reml))
      expected <- numDeriv::hessian(objective, fit$covariance_parameters)
      expect_equal(as.matrix(fit$Hessian), expected, tolerance = 2e-5,
                   ignore_attr = TRUE, info = paste(side, model, reml))
    }
  }
})

test_that('Hessian calculations respect response units in both methods', {
  d <- vam_fixture(24L, 2L)
  small <- d
  small$y <- d$y / 1000
  for (side in c('R', 'G')) {
    models <- if (side == 'G') 'GP' else c('GP', 'CP', 'VP')
    for (model in models) for (reml in c(FALSE, TRUE)) {
      for (method in c('simple', 'richardson')) {
        fit <- suppressWarnings(GPvam(d, ~0+as.factor(year)+x,
            persistence = model, student.side = side, REML = reml,
            max.iter.EM = 8, hessian = TRUE, hes.method = method, verbose = FALSE))
        scaled <- suppressWarnings(GPvam(small, ~0+as.factor(year)+x,
            persistence = model, student.side = side, REML = reml,
            max.iter.EM = 8, hessian = TRUE, hes.method = method, verbose = FALSE))
        factors <- rep(1e-6, nrow(fit$Hessian))
        if (model == 'VP') factors[length(factors)] <- 1
        expect_true(all(is.finite(scaled$Hessian)))
        expect_equal(scaled$Hessian * outer(factors, factors), fit$Hessian,
                     tolerance = 2e-4, ignore_attr = TRUE,
                     info = paste(side, model, reml, method))
      }
    }
  }
})

Try the GPvam package in your browser

Any scripts or data that you put into this service are public.

GPvam documentation built on Oct. 23, 2026, 1:06 a.m.