Description

Meta-Analytic-Predictive (MAP) analysis for generalized linear models suitable for normal, binary, or Poisson data. Model specification and overall syntax follows mainly glm conventions.

Usage

Arguments

Details

The meta-analytic-predictive (MAP) approach derives a prior from historical data using a hierarchical model. The statistical model is formulated as a generalized linear mixed model for binary, normal (with fixed σ) and Poisson endpoints:

y_ih|θ_ih ~ f(y_ih | θ_ih)

Here, i=1,…,N is the index for observations, and h=1,…,H is the index for the grouping (usually studies). The model assumes the linear predictor for a transformed mean as

g(θ_ih; x_ih,β) = x_ih β + ε_h

with x_ih being the row vector of k covariates for observation i. The variance component is assumed by default normal

ε_h ~ N(0,τ^2), h=1,...,H

Lastly, the Bayesian implementation assumes independent normal priors for the k regression coefficients and a prior for the between-group standard deviation τ (see taud.dist for available distributions).

The MAP prior will then be derived from the above model as the conditional distribution of θ_* given the available data and the vector of covariates x_* defining the overall intercept

θ_*| x_*,y .

A simple and common case arises for one observation (summary statistic) per trial. For a normal endpoint, the model then simplifies to the standard normal-normal hierarchical model. In the above notation, i=h=1,…,H and

y_h|θ_h ~ N(θ_h,s_h^2)

θ_h = μ + ε_h

ε_h ~ N(0,τ^2),

where the more common μ is used for the only (intercept) parameter β_1. Since there are no covariates, the MAP prior is simply Pr(θ_* | y_1,…,y_H).

The hierarchical model is a compromise between the two extreme cases of full pooling (τ=0, full borrowing, no discounting) and no pooling (τ=∞, no borrowing, stratification). The information content of the historical data grows with H (number of historical data items) indefinitely for full pooling whereas no information is gained in a stratified analysis. For a fixed τ, the maximum effective sample size of the MAP prior is n_∞ (H->∞), which for a normal endpoint with fixed σ is

n_∞ = (τ^2/σ^2)^-1

(Neuenschwander et al., 2010). Hence, the ratio τ/σ limits the amount of information a MAP prior is equivalent to. This allows for a classification of τ values in relation to σ, which is crucial to define a prior P_τ. The following classification is useful in a clinical trial setting:

The above formula for n_∞ assumes a known τ. This is unrealistic as the between-trial heterogeneity parameter is often not well estimable, in particular if the number of trials is small (H small). The above table helps to specify a prior distribution for τ appropriate for the given context which defines the crucial parameter σ. For binary and Poisson endpoints, normal approximations can be used to determine σ. See examples below for concrete cases.

The design matrix X is defined by the formula for the linear predictor and is always of the form response ~ predictor | grouping, which follows glm conventions. The syntax has been extended to include a specification of the grouping (for example study) factor of the data with a horizontal bar, |. The bar separates the optionally specified grouping level, i.e. in the binary endpoint case cbind(r, n-r) ~ 1 | study. By default it is assumed that each row corresponds to an individual group (for which an individual parameter is estimated). Specifics for the different endpoints are:

normal

family=gaussian assumes an identity link function. The response should be given as matrix with two columns with the first column being the observed mean value y_ih and the second column the standard error se_ih (of the mean). Additionally, it is recommended to specify with the weight argument the number of units which contributed to the (mean) measurement y_ih. This information is used to estimate σ.

binary

family=binomial assumes a logit link function. The response must be given as two-column matrix with number of responders r (first column) and non-responders n-r (second column).

Poisson

family=poisson assumes a log link function. The response is a vector of counts. The total exposure times can be specified by an offset, which will be linearly added to the linear predictor. The offset can be given as part of the formula, y ~ 1 + offset(log(exposure)) or as the offset argument to gMAP. Note that the exposure unit must be given as log-offset.

Value

The function returns a S3 object of type gMAP. See the methods section below for applicable functions to query the object.

Methods (by generic)

• print: displays a summary of the gMAP analysis.

• fitted: returns the quantiles of the posterior shrinkage estimates for each data item used during the analysis of the given gMAP object.

• coef: returns the quantiles of the predictive distribution. User can choose with type if the result is on the response or the link scale.

• as.matrix: extracts the posterior sample of the model.

• summary: returns the summaries of a gMAP. analysis. Output is a gMAPsummary object, which is a list containing

  tau

  posterior summary of the heterogeneity standard deviation

  beta

  posterior summary of the regression coefficients

  theta.pred

  summary of the predictive distribution (given in dependence on the type argument either on response or link scale)

  theta

  posterior summary of the mean estimate (also depends on the type argument)

Differential Discounting

The above model assumes the same between-group standard deviation τ, which implies that the data are equally relevant. This assumption can be relaxed to more than one τ. That is,

ε_h ~ N(0,τ_s(h)^2)

where s(h) assignes group h to one of S between-group heterogeneity strata.

For example, in a situation with two randomized and four observational studies, one may want to assume τ_1 (for trials 1 and 2) and τ_2 (for trials 3-6) for the between-trial standard deviations of the control means. More heterogeneity (less relevance) for the observational studies can then be expressed by appropriate priors for τ_1 and τ_2. In this case, S=2 and the strata assignments (see tau.strata argument) would be s(1)=s(2)=1, s(3)=…=s(6)=2.

Prior Specification

The prior distribution for the regression coefficients β is normal.

• If a single number is given, then this is used as the standard deviation and the default mean of 0 is used.

• If a vector is given, it must be of the same length as number of covariates defined and is used as standard deviation.

• If a matrix with a single row is given, its first row will be used as mean and the second row will be used as standard deviation for all regression coefficients.

• Lastly, a two-column matrix (mean and standard deviation columns) with as many columns as regression coefficients can be given.

It is recommended to always specify a beta.prior. Per default a mean of 0 is set. The standard deviation is set to 2 for the binary case, to 100 * sd(y) for the normal case and to sd(log(y + 0.5 + offset)) for the Poisson case.

For the between-trial heterogeniety τ prior, a dispersion parameter must always be given for each exchangeability stratum. For the different tau.prior distributions, two parameters are needed out of which one is set to a default value if applicable:

For a prior distribution with a default location parameter, a vector of length equal to the number of exchangability strata can be given. Otherwise, a two-column matrix with as many rows as exchangability strata must be given, except for a single τ stratum, for which a vector of length two defines the parameters a and b.

Random seed

The MAP analysis is performed using Markov-Chain-Monte-Carlo (MCMC) in rstan. MCMC is a stochastic algorithm. To obtain exactly reproducible results you must use the set.seed function before calling gMAP. See RBesT overview page for global options on setting further MCMC simulation parameters.

References

Neuenschwander B, Capkun-Niggli G, Branson M, Spiegelhalter DJ. Summarizing historical information on controls in clinical trials. Clin Trials. 2010; 7(1):5-18

Schmidli H, Gsteiger S, Roychoudhury S, O'Hagan A, Spiegelhalter D, Neuenschwander B. Robust meta-analytic-predictive priors in clinical trials with historical control information. Biometrics 2014;70(4):1023-1032.

See Also

plot.gMAP, forest_plot, automixfit, predict.gMAP

Examples