multinma (version 0.1.3)

plot.nma_dic: Plots of model fit diagnostics

Description

The plot() method for nma_dic objects produced by dic() produces several useful diagnostic plots for checking model fit and model comparison. Further detail on these plots and their interpretation is given by TSD2;textualmultinma.

Usage

# S3 method for nma_dic
plot(
  x,
  y,
  ...,
  show_uncertainty = TRUE,
  stat = "pointinterval",
  orientation = c("vertical", "horizontal", "x", "y")
)

Arguments

x

A nma_dic object

y

(Optional) A second nma_dic object, to produce "dev-dev" plots for model comparison.

...

Additional arguments passed on to other methods

show_uncertainty

Logical, show uncertainty with a ggdist plot stat? Default TRUE.

stat

Character string specifying the ggdist plot stat to use if show_uncertainty = TRUE, default "pointinterval". If y is provided, currently only "pointinterval" is supported.

orientation

Whether the ggdist geom is drawn horizontally ("horizontal") or vertically ("vertical"). Only used for residual deviance plots, default "vertical".

Value

A ggplot object.

Details

When a single nma_dic object is given, a plot of the residual deviance contribution for each data point is produced. For a good fitting model, each data point is expected to have a residual deviance of 1; larger values indicate data points that are fit poorly by the model.

When two nma_dic objects are given, a "dev-dev" plot comparing the residual deviance contributions under each model is produced. Data points with residual deviance contributions lying on the line of equality are fit equally well under either model. Data points lying below the line of equality indicate better fit under the second model (y); conversely, data points lying above the line of equality indicate better fit under the first model (x). A common use case is to compare a standard consistency model (fitted using nma() with consistency = "consistency") with an unrelated mean effects (UME) inconsistency model (fitted using nma() with consistency = "ume"), to check for potential inconsistency.

See TSD2;textualmultinma for further details.

References

Examples

# NOT RUN {
## Smoking cessation
# Set up network of smoking cessation data
head(smoking)

smk_net <- set_agd_arm(smoking,
                       study = studyn,
                       trt = trtc,
                       r = r,
                       n = n,
                       trt_ref = "No intervention")

# Print details
smk_net

# }
# NOT RUN {
# Fitting a fixed effect model
smk_fit_FE <- nma(smk_net,
              trt_effects = "fixed",
              prior_intercept = normal(scale = 100),
              prior_trt = normal(scale = 100))

smk_fit_FE
# }
# NOT RUN {
# }
# NOT RUN {
# Fitting a random effects model
smk_fit_RE <- nma(smk_net,
                  trt_effects = "random",
                  prior_intercept = normal(scale = 100),
                  prior_trt = normal(scale = 100),
                  prior_het = normal(scale = 5))

smk_fit_RE
# }
# NOT RUN {
# }
# NOT RUN {
# Compare DIC of FE and RE models
(smk_dic_FE <- dic(smk_fit_FE))
(smk_dic_RE <- dic(smk_fit_RE))   # substantially better fit

# Plot residual deviance contributions under RE model
plot(smk_dic_RE)

# Changing the plot stat used
plot(smk_dic_RE, stat = "interval", orientation = "horizontal")

# Further customisation is possible using ggplot commands
# For example, highlighting data points with residual deviance above a certain threshold
plot(smk_dic_RE) +
  ggplot2::aes(colour = ifelse(..y.. > 1.5, "darkorange", "black")) +
  ggplot2::scale_colour_identity()

# Or by posterior probability, for example here a central probability of 0.6
# corresponds to a lower tail probability of (1 - 0.6)/2 = 0.2
plot(smk_dic_RE, .width = c(0.6, 0.95)) +
  ggplot2::aes(colour = ifelse(..ymin.. > 1, "darkorange", "black")) +
  ggplot2::scale_colour_identity()

# Check for inconsistency using UME model
# }
# NOT RUN {
# Fitting an unrelated mean effects (inconsistency) model
smk_fit_RE_UME <- nma(smk_net,
                      consistency = "ume",
                      trt_effects = "random",
                      prior_intercept = normal(scale = 100),
                      prior_trt = normal(scale = 100),
                      prior_het = normal(scale = 5))

smk_fit_RE_UME
# }
# NOT RUN {
# }
# NOT RUN {
# Compare DIC
smk_dic_RE
(smk_dic_RE_UME <- dic(smk_fit_RE_UME))  # no difference in fit

# Compare residual deviance contributions with a "dev-dev" plot
plot(smk_dic_RE, smk_dic_RE_UME)

# By default the dev-dev plot can be a little cluttered
# Hiding the credible intervals
plot(smk_dic_RE, smk_dic_RE_UME, show_uncertainty = FALSE)

# Changing transparency
plot(smk_dic_RE, smk_dic_RE_UME, point_alpha = 0.5, interval_alpha = 0.1)
# }