list object that contains matrix <code>theta</code> of elicitations and
character <code>link</code>, see <code><a rd-options="" href="/link/plotDesignPoint?package=indirect&version=0.2.1" data-mini-rdoc="indirect::plotDesignPoint">plotDesignPoint</a></code>

character, <code>moment</code>, <code>KL</code>, <code>SS</code>. Default is
<code>KL</code>.

fit.method

The default for <code>fit.method</code> is option <code>KL</code>. This option uses an 
objective function that minimises a discretised directed divergence from a 
cumulative distribution implied by raw elicited fractiles to a normal 
conditional mean prior for the linear predictor. An alterative method 
<code>moment</code> assigns the location parameter of the normal conditional mean 
prior to the elicited median on the linear predictor scale. The variance 
parameter is estimated as \(V = ((g(f_u) - g(f_l)/(qnorm(u) -
qnorm(l)))^2\), where \(l\) is the probability associated with the fractile
\(f_l\) that defines the lower bound for the central credible interval and 
\(u\) is the probability associated with the fractile \(f_u\) that
defines the upper bound for the central credible interval. This is also used
to initialise the optimisation for the <code>KL</code> method. Another optimsation 
method that minimises the sum of squares is also available as method 
<code>SS</code>. See the vignette for more details on the choice of objective 
function for <code>KL</code> and <code>SS</code>.

Functions are provided to facilitate prior elicitation for Bayesian generalised linear models using independent conditional means priors. The package supports the elicitation of multivariate normal priors for generalised linear models. The approach can be applied to indirect elicitation for a generalised linear model that is linear in the parameters. The package is designed such that the facilitator executes functions within the R console during the elicitation session to provide graphical and numerical feedback at each design point. Various methodologies for eliciting fractiles (equivalently, percentiles or quantiles) are supported, including versions of the approach of Hosack et al. (2017) <doi:10.1016/j.ress.2017.06.011>. For example, experts may be asked to provide central credible intervals that correspond to a certain probability. Or experts may be allowed to vary the probability allocated to the central credible interval for each design point. Additionally, a median may or may not be elicited.

Geoff Hosack

indirect

Elicitation of Independent Conditional Means Priors for
Generalised Linear Models

mV function

list object that contains matrix <code>theta</code> of elicitations and
character <code>link</code>, see <code><a rd-options='' href='plotDesignPoint'>plotDesignPoint</a></code>

mV: Helper function that translates elicited quantiles of target into independent conditional means normal prior for a defined inverse link function.

Description

Usage

Arguments

Value