estimate: Estimate joint regression model for intervention costs and...

In uo-cmor/cea: Functions for Trial-Based Cost-Effectiveness Analysis

Description

Estimate a multivariate generalised linear model for the joint incremental costs and QALYs from a randomised trial (or other comparable data source). Models can be estimated as a Multivariate Covariance Generalized Linear Model (MCGLM; see Bonat & J\/orgensen 2016), using the mcglm package (Bonat 2018); a Multivariate Generalized Linear Mixed Model (MGLMM) via Penalized Quasi-Likelihood (Achana et al. 2021); or a Bayesian generalized multivariate model, using the brms package (B\/"urkner 2017, 2018).

Usage

estimate(
  QALYs,
  costs,
  treatment,
  covars,
  data,
  centre = NULL,
  cluster = NULL,
  family = NULL,
  prior = NULL,
  method = "mcglm",
  fixed = NULL,
  random = NULL,
  verbose = FALSE,
  control = NULL,
  ...
)

Arguments

QALYs, costs, treatment	Character strings naming the variables in data representing QALYs, costs, and treatment assignment, respectively.
covars	(optional) Character vector naming variables in data included as (baseline) covariates in the regression models.
data	A data frame (or object coercible by as.data.frame to a data frame) or a mids object containing the variables in the model.
centre	(optional) Character string naming a variable in data signifying centre membership in a multicentre trial.
cluster	(optional) Character vector naming a variable in data signifying cluster membership in a cluster-randomised trial. NOTE: At present, centre and cluster do exactly the same thing (estimate a mixed-effects model with centre/cluster random effects), and only one of these arguments should be specified.
family	(optional) A list of family functions specifying the distribution and link function for each outcome variable. Each element of the list should be a valid family value specified as for glm. If not specified, the default is to use gaussian() (i.e. OLS) for QALYs and Gamma("log") for Costs.
prior	For method = 'brms' only, specification of priors for Bayesian model. See brm and set_prior for details. If not specified, uses "normal(0, 5)" for fixed effects coefficients and brm defaults for all other parameters.
method	"mcglm", "mglmmPQL", or "brms", specifying which modelling approach to use.
fixed	An alternative way to specify the 'fixed effects' component of the model. Required if the RHS variables should differ for each equation or outcomes other than QALYs and Costs are to be estimated. If used, should be a (optionally named) list of two-sided model formulae specifying the equations for each outcome measure, and will override any values specified in QALYs, costs, and covars; the variable specified in treatment must be included in the RHS of each formula. If fixed does not have names, the LHS of each formula will be used as component names.
random	(optional) Specification of the random effects component of the model. Corresponds to the matrix_pred argument of mcglm and the random argument of lme; see the corresponding package help for details. If not specified, the default is to include centre/cluster random effects (intercepts) for each outcome (if applicable).
verbose	Whether to print messages from the model fitting function (default=FALSE).
control	A list of arguments to be passed to the fitting algorithm (argument control_algorithm of mcglm for method "mcglm"; control (returned via the function lmeControl) of lme for method "mglmmPQL").
...	Further optional arguments to be passed to the underlying model fitting function specified by method (mcglm/lme). See also the 'Details' section for additional considerations for specifying these arguments.

Details

The standard model specification provides QALYs, costs, treatment, and optionally covars and centre/cluster, to estimate a bivariate regression with common RHS variables and optional clustering by trial centre or clustered randomisation.

If a different specification is required (e.g., estimating additional outcomes beyond QALYs and Costs, or using different RHS variables in each equation), a custom model specification can be passed to fixed as a list of formulae. In this case, family should also be used to specify the GLM 'family' for each outcome.

Note the following details for model specification for each method:

mcglm:

• Model specification arguments linear_pred, matrix_pred, link, and variance are constructed via the estimate() arguments QALYs, costs, treatment, and covars; centre or cluster; and family (for which see below), and should not be passed directly. For custom model specifications, use the arguments fixed, random, and family.

• Instead of a standard GLM family specification, family can also be provided as a list where each element (corresponding to each model outcome) is a named list with elements "family" (a name for the custom family), "link", "variance" (passed as the corresponding arguments to mcglm), and "power" (the (starting) power parameter, currently only used for variance = "tweedie"). For the "tweedie" variance function, a power value 0 corresponds to a gaussian distribution, 1 to the (quasi-)Poisson distribution, 2 to the gamma distribution, and 3 to the inverse gaussian. (As in mcglm, the dafault is to fix the power parameter to its initial value.)

• The control_initial argument is derived by fitting separate GLM models to each outcome variable (as in the mcglm default) and cannot be directly set by the user. The initial power parameter can be set via family as descried above.

mglmmPQL:

• The lme argument method is renamed nlme_method to avoid a clash with the same-named argument to estimate.

• Formulae specified in random or weights (see below) can refer to the reserved name outvar, an indicator variable identifying the outcome variable to which an observation belongs in the transformed long-form (stacked) data used in the underlying lme estimation.

• By default, weights is set to nlme::varIdent(form = ~1 | outvar), to allow different variances for each outcome variable. See varIdent for details.

• max_iter can be used to specify the maximum number of PQL iterations. Use control =nlme::lmeControl(maxIter = <x>) to set the maximum number of iterations for the lme optimization algorithm.

Value

An object of class cea_estimate inheriting from mcglm.

References

Achana F, Gallacher D, Oppong R, et al. Multivariate generalized linear mixed-effects models for the analysis of clinical trial-based cost-effectiveness data. Med Decis Making 2021;41(6):667-84. https://doi.org/10.1177/0272989X211003880

Bonat WH. Multiple response variables regression models in R: The mcglm package. J Stat Soft 2018;84(4):1-30. https://doi.org/10.18637/jss.v084.i04

Bonat WH, Jørgensen B. Multivariate covariance generalized linear models. J Royal Stat Soc 2016;65:649-75. https://doi.org/10.1111/rssc.12145

Bürkner P-C. brms: An R package for Bayesian multilevel models using Stan. J Stat Soft 2017;80(1):1-28. https://doi.org/10.18637/jss.v080.i01

Bürkner P-C. Advanced Bayesian multilevel modeling with the R package brms. R Journal 2018;10(1):395-411. https://doi.org/10.32614/RJ-2018-017

uo-cmor/cea documentation built on Dec. 23, 2021, 2:01 p.m.