qreml_old: Quasi restricted maximum likelihood (qREML) algorithm, previous version

Description

This algorithm can be used very flexibly to fit statistical models that involve penalised splines or simple i.i.d. random effects, i.e. that have penalties of the form $$0.5 \sum_{i} \lambda_i b_i^T S_i b_i,$$ with smoothing parameters $\lambda_i$, coefficient vectors $b_i$, and fixed penalty matrices $S_i$.

The qREML algorithm is typically much faster than REML or marginal ML using the full Laplace approximation method, but may be slightly less accurate regarding the estimation of the penalty strength parameters.

qreml_old is the previous implementation, kept for reference. Use qreml, which estimates the same quantities with a better behaved outer iteration.

Under the hood, qreml_old uses the R package RTMB for automatic differentiation in the inner optimisation. The user has to specify the penalised negative log-likelihood function pnll structured as dictated by RTMB and use the penalty function to compute the quadratic-form penalty inside the likelihood.

Usage

qreml_old(
  pnll,
  par,
  dat,
  random,
  map = NULL,
  silent = 1,
  psname = "lambda",
  alpha = 0.3,
  smoothing = 1,
  maxiter = 100,
  tol = 1e-04,
  max_halve = 5,
  method = "BFGS",
  control = list(),
  conv_crit = "relchange",
  spHess = FALSE,
  joint_unc = FALSE,
  saveall = FALSE
)

Value

model object of class 'qremlModel'. This is a list containing:

...: everything that is reported inside pnll using RTMB::REPORT(). When using forward, tpm_g, etc., this may involve automatically reported objects.
obj: RTMB AD object containing the final conditional model fit
psname: final penalty strength parameter vector
all_psname: list of all penalty strength parameter vectors over the iterations
par: named estimated parameter list in the same structure as the initial par. Note that the name par is not fixed but depends on the original name of your par list.
relist_par: function to convert the estimated parameter vector to the estimated parameter list. This is useful for uncertainty quantification based on sampling from a multivariate normal distribution.
par_vec: estimated parameter vector
llk: unpenalised log-likelihood at the optimum
n_fixpar: number of fixed, i.e. unpenalised, parameters
edf: overall effective number of parameters
all_edf: list of effective number of parameters for each smooth
Hessian_condtional: final Hessian of the conditional penalised fit
obj_joint: if joint_unc = TRUE, joint RTMB object for joint uncertainty quantification in model and penalty parameters.

Arguments

pnll

penalised negative log-likelihood function that is structured as dictated by RTMB and uses the penalty function from LaMa to compute the penalty

Needs to be a function of the named list of initial parameters par only.

par

named list of initial parameters

The random effects/ spline coefficients can be vectors or matrices, the latter summarising several random effects of the same structure, each one being a row in the matrix.

dat

initial data list that contains the data used in the likelihood function, hyperparameters, and the initial penalty strength vector

If the initial penalty strength vector is not called lambda, the name it has in dat needs to be specified using the psname argument below. Its length needs to match the to the total number of random effects.

random

vector of names of the random effects/ penalised parameters in par

Caution: The ordering of random needs to match the order of the random effects passed to penalty inside the likelihood function.

map

optional map argument, containing factor vectors to indicate parameter sharing or fixing.

Needs to be a named list for a subset of fixed effect parameters or penalty strength parameters. For example, if the model has four penalty strength parameters, map[[psname]] could be factor(c(NA, 1, 1, 2)) to fix the first penalty strength parameter, estimate the second and third jointly, and estimate the fourth separately.

silent

integer silencing level: 0 corresponds to full printing of inner and outer iterations, 1 to printing of outer iterations only, and 2 to no printing.

psname

optional name given to the penalty strength parameter in dat. Defaults to "lambda".

alpha

optional hyperparamater for exponential smoothing of the penalty strengths.

For larger values smoother convergence is to be expected but the algorithm may need more iterations.

smoothing

optional scaling factor for the final penalty strength parameters

Increasing this beyond one will lead to a smoother final model. Can be an integer or a vector of length equal to the length of the penalty strength parameter.

maxiter

maximum number of iterations in the outer optimisation over the penalty strength parameters.

tol

Convergence tolerance for the penalty strength parameters.

max_halve

maximum number of step halvings per outer iteration.

After each update of the penalty strength parameters, the restricted log-likelihood is compared to its value in the previous iteration. If it decreased, the step (on the log scale) is repeatedly halved until it does not, or until max_halve halvings have been made. This makes the outer iteration monotone and hence more robust. Set to zero to switch step halving off.

method

optimisation method to be used by optim. Defaults to "BFGS", but might be changed to "L-BFGS-B" for high-dimensional settings.

control

list of control parameters for optim to use in the inner optimisation. Here, optim uses the BFGS method which cannot be changed.

We advise against changing the default values of reltol and maxit as this can decrease the accuracy of the Laplace approximation.

conv_crit

character, convergence criterion for the penalty strength parameters. Can be "relchange" (default) or "gradient".

spHess

logical, if TRUE, sparse AD Hessian is used in each outer iteration. If your Hessian is large and sparse (many cross derivatives are 0), this will speed up the computations a lot. If your Hessian is dense, this will slow down the computations slightly and might require significantly more memory.

joint_unc

logical, if TRUE, joint RTMB object is returned allowing for joint uncertainty quantification

saveall

logical, if TRUE, then all model objects from each iteration are saved in the final model object.

References

Koslik, J. O. (2024). Efficient smoothness selection for nonparametric Markov-switching models via quasi restricted maximum likelihood. arXiv preprint arXiv:2411.11498.

Examples

Run this code

data = trex[1:1000,] # subset

# initial parameter list
par = list(logmu = log(c(0.3, 2.5)), # step mean
           logsigma = log(c(0.3, 1.5)), # step sd
           beta0 = c(-2,-2), # state process intercept
           betaspline = matrix(rep(0, 18), nrow = 2)) # state process spline coefs
          
# data object with initial penalty strength lambda
dat = list(step = data$step, # step length
           tod = data$tod, # time of day covariate
           N = 2, # number of states
           lambda = rep(10,2)) # initial penalty strength

# building model matrices
modmat = modmat(~ s(tod, bs = "cp"), 
                       data = data.frame(tod = 1:24), 
                       knots = list(tod = c(0,24))) # wrapping points
dat$Z = modmat$Z # spline design matrix
dat$S = modmat$S # penalty matrix

# penalised negative log-likelihood function
pnll = function(par) {
  getAll(par, dat) # makes everything contained available without $
  Gamma = tpm_g(Z, cbind(beta0, betaspline), ad = TRUE) # transition probabilities
  delta = stationary_p(Gamma, t = tod[1] %% 24 + 1, ad = TRUE) # initial distribution
  mu = exp(logmu) # step mean
  sigma = exp(logsigma) # step sd
  # calculating all state-dependent densities
  allprobs = matrix(1, nrow = length(step), ncol = N)
  ind = which(!is.na(step)) # only for non-NA obs.
  for(j in 1:N) allprobs[ind,j] = dgamma2(step[ind],mu[j],sigma[j])
  -forward_g(delta, Gamma[,,tod], allprobs) +
      penalty(betaspline, S, lambda) # this does all the penalization work
}

# model fitting
mod = qreml_old(pnll, par, dat, random = "betaspline", silent = 2)