draic.msm: Criteria for comparing two multi-state models with nested state spaces

Description

A modification of Akaike's information criterion, and a leave-one-out likelihood cross-validation criterion, for comparing the predictive ability of two multi-state models with nested state spaces. This is evaluated based on the restricted or aggregated data which the models have in common.

Usage

draic.msm(msm.full, msm.coarse, likelihood.only=FALSE,
         information=c("expected","observed"), tl=0.95)
drlcv.msm(msm.full, msm.coarse, tl=0.95, cores=NULL,
          verbose=TRUE,outfile=NULL)

Arguments

msm.full

Big model.

msm.coarse

Smaller model.

The two models must be fitted to the same datasets, except that the state space of the coarse model must be an aggregated version of the state space of the full model. That is, every state in the full dataset must correspo

likelihood.only

Don't calculate Hessians and trace term (DRAIC).

information

Use observed or expected information in the DRAIC trace term. Expected is the default, and much faster.

Width of symmetric tracking interval, by default 0.95 for a 95% interval.

cores

Number of processor cores to use in drlcv for cross-validation by parallel processing. Requires the doParallel package to be installed. If not specified, parallel processing is not used. If cores is set t

verbose

Print intermediate results of each iteration of cross-validation to the console while running. May not work with parallel processing.

outfile

Output file to print intermediate results of cross-validation. Useful to track execution speed when using parallel processing, where output to the console may not work.

Value

A list containing $D_{RAIC}$ (draic.msm) or $D_{RLCV}$ (drlcv.msm), its component terms, and tracking intervals.

Details

The difference in restricted AIC (Liquet and Commenges, 2011) is defined as

$$D_{RAIC} = l(\gamma_n |\mathbf{x}'' ) - l(\theta_n |\mathbf{x}'' ) + trace ( J(\theta_n |\mathbf{x}'')J(\theta_n |\mathbf{x})^{-1} - J(\gamma_n |\mathbf{x}'' )J(\gamma_n |\mathbf{x}' )^{-1})$$

where $\gamma$ and $\theta$ are the maximum likelihood estimates of the smaller and bigger models, fitted to the smaller and bigger data, respectively.

$l(\gamma_n |x'')$ represents the likelihood of the simpler model evaluated on the restricted data.

$l(\theta_n |x'')$ represents the likelihood of the complex model evaluated on the restricted data. This is a hidden Markov model, with a misclassification matrix and initial state occupancy probabilities as described by Thom et al (2014).

$J()$ are the corresponding (expected or observed, as specified by the user) information matrices.

$\mathbf{x}$ is the expanded data, to which the bigger model was originally fitted, and $\mathbf{x}'$ is the data to which the smaller model was originally fitted. $\mathbf{x}''$ is the restricted data which the two models have in common. $\mathbf{x}'' = \mathbf{x}'$ in this implementation, so the models are nested.

The difference in likelihood cross-validatory criteria (Liquet and Commenges, 2011) is defined as

$$D_{RLCV} = 1/n \sum_{i=1}^n \log( h_{X''}(x_i'' | \gamma_{-i}) / g_{X''}(x_i''| \theta_{-i}))$$

where $\gamma_{-i}$ and $\theta_{-i}$ are the maximum likelihood estimates from the smaller and bigger models fitted to datasets with subject $i$ left out, $g()$ and $h()$ are the densities of the corresponding models, and $x_i''$ is the restricted data from subject $i$.

Tracking intervals are analogous to confidence intervals, but not strictly the same, since the quantity which D_RAIC aims to estimate, the difference in expected Kullback-Leibler discrepancy for predicting a replicate dataset, depends on the sample size. See the references.

References

Thom, H. and Jackson, C. and Commenges, D. and Sharples, L. (2014) State-selection in multistate models with application to quality of life in psoriatic arthritis. Submitted.