nma.fit: Assess Model Fit

In audrey-b/BUGSnet: Bayesian network meta-analyses in compliance with best practice and reporting guidelines

Assess Model Fit

Description

Computes the Deviance Information Criteria (DIC) and produces a leverage plot (as in the NICE Technical Support Document 2) for a given model. These can be used to assess and compare the fit of different models (i.e fixed vs random effects, consistency vs inconsistency). nma.fit also produces a plot comparing the leverage of each data point against their contribution to the total posterior deviance. Points lying outside the purple dotted line are generally identified as contributing to the model's poor fit. Points with high leverage are influencial i.e. they have a stong influence on the estimates.

Usage

nma.fit(nma, plot.pD = TRUE, plot.DIC = TRUE, plot.Dres = TRUE, ...)

Arguments

nma	A BUGSnetRun object produced by running nma.run().
plot.pD	Whether to include pD on the plot. Default is TRUE.
plot.DIC	Whether to include DIC on the plot. Default is TRUE.
plot.Dres	Whether to include Dres on the plot. Default is TRUE.
...	Graphical arguments such as main=, ylab=, and xlab= may be passed as in plot(). These arguments will only effect the leverage plot.

Value

DIC - A number indicating the Deviance Information Criteria. The DIC is calculated as the sum of Dres and pD. A larger DIC is indicative of a worse model fit.

leverage - A vector with one value per study arm indicating the leverage of each data point (study arm). Leverage is defined as pmdev minus the deviance at the posterior mean of the fitted values.

w - A vector with one value per study arm. The magnitude of w represents the data point's contribution to the posterior mean deviance of the model and is simply the square root of pmdev. The sign indicates whether the data point is being over (negative sign) or under (positive sign) estimated by the model and is calculated as the sign of the difference of the observed outcome minus the predicted outcome.

pmdev - A vector with one value per study arm representing the posterior mean residual deviance for each data point (study arm).

Dres - The posterior mean of the residual deviance.

pD - The effective number of parameters, calculated as the sum of the leverages.

See Also

nma.run

Examples

data(diabetes.sim)

diabetes.slr <- data.prep(
  arm.data = diabetes.sim, 
  varname.t = "Treatment", 
  varname.s = "Study"
)

#Random effects, consistency model.
#Binomial family, cloglog link. This implies that the scale will be the Hazard Ratio.
diabetes.re.c <- nma.model(
  data = diabetes.slr,
  outcome = "diabetes", 
  N = "n",
  reference = "Placebo",
  family = "binomial",
  link = "cloglog",
  effects = "random",
  type = "consistency",
  time = "followup"
)
 
diabetes.re.c.res <- nma.run(
  model = diabetes.re.c,
  n.adapt = 100,
  n.burnin = 0,
  n.iter = 100
)

#Fixed effects, consistency model.
#Binomial family, cloglog link. This implies that the scale will be the Hazard Ratio.
diabetes.fe.c <- nma.model(
  data = diabetes.slr,
  outcome = "diabetes", 
  N = "n",
  reference = "Placebo",
  family = "binomial",
  link = "cloglog",
  effects = "fixed",
  type="consistency",
  time="followup"
)
 
diabetes.fe.c.res <- nma.run(
  model = diabetes.fe.c,
  n.adapt = 100,
  n.burnin = 0,
  n.iter = 100
)  

#Compare fixed vs random effects via leverage plots and DIC 
par(mfrow=c(1,2))
nma.fit(diabetes.fe.c.res, main = "Fixed Effects Model")
nma.fit(diabetes.re.c.res, main= "Random Effects Model")

audrey-b/BUGSnet documentation built on Feb. 2, 2025, 5:10 p.m.