Marginal treatment effects
Generate population-average marginal treatment effects. These are formed from population-average absolute predictions, so this function is a wrapper around predict.stan_nma().
marginal_effects( object, ..., mtype = c("difference", "ratio", "link"), all_contrasts = FALSE, trt_ref = NULL, probs = c(0.025, 0.25, 0.5, 0.75, 0.975), predictive_distribution = FALSE, summary = TRUE )
object: A stan_nma object created by nma().
...: Arguments passed to predict.stan_nma(), for example to specify the covariate distribution and baseline risk for a target population, e.g. newdata, baseline, and related arguments. For survival outcomes, type
can also be specified to determine the quantity from which to form a marginal effect. For example, type = "hazard" with mtype = "ratio"
produces marginal hazard ratios, type = "median" with mtype = "difference" produces marginal median survival time differences, and so on.
mtype: The type of marginal effect to construct from the average absolute effects, either "difference" (the default) for a difference of absolute effects such as a risk difference, "ratio" for a ratio of absolute effects such as a risk ratio, or "link" for a difference on the scale of the link function used in fitting the model such as a marginal log odds ratio.
all_contrasts: Logical, generate estimates for all contrasts (TRUE), or just the "basic" contrasts against the network reference treatment (FALSE)? Default FALSE.
trt_ref: Reference treatment to construct relative effects against, if all_contrasts = FALSE. By default, relative effects will be against the network reference treatment. Coerced to character string.
probs: Numeric vector of quantiles of interest to present in computed summary, default c(0.025, 0.25, 0.5, 0.75, 0.975)
predictive_distribution: Logical, when a random effects model has been fitted, should the predictive distribution for marginal effects in a new study be returned? Default FALSE.
summary: Logical, calculate posterior summaries? Default TRUE.
A nma_summary object if summary = TRUE, otherwise a list containing a 3D MCMC array of samples and (for regression models) a data frame of study information.
## Smoking cessation # Run smoking RE NMA example if not already available if (!exists("smk_fit_RE")) example("example_smk_re", run.donttest = TRUE) # Marginal risk difference in each study population in the network marginal_effects(smk_fit_RE, mtype = "difference") # Since there are no covariates in the model, the marginal and conditional # (log) odds ratios here coincide marginal_effects(smk_fit_RE, mtype = "link") relative_effects(smk_fit_RE) # Marginal risk differences in a population with 67 observed events out of # 566 individuals on No Intervention, corresponding to a Beta(67, 566 - 67) # distribution on the baseline probability of response (smk_rd_RE <- marginal_effects(smk_fit_RE, baseline = distr(qbeta, 67, 566 - 67), baseline_type = "response", mtype = "difference")) plot(smk_rd_RE) ## Plaque psoriasis ML-NMR # Run plaque psoriasis ML-NMR example if not already available if (!exists("pso_fit")) example("example_pso_mlnmr", run.donttest = TRUE) # Population-average marginal probit differences in each study in the network (pso_marg <- marginal_effects(pso_fit, mtype = "link")) plot(pso_marg, ref_line = c(0, 1)) # Population-average marginal probit differences in a new target population, # with means and SDs or proportions given by new_agd_int <- data.frame( bsa_mean = 0.6, bsa_sd = 0.3, prevsys = 0.1, psa = 0.2, weight_mean = 10, weight_sd = 1, durnpso_mean = 3, durnpso_sd = 1 ) # We need to add integration points to this data frame of new data # We use the weighted mean correlation matrix computed from the IPD studies new_agd_int <- add_integration(new_agd_int, durnpso = distr(qgamma, mean = durnpso_mean, sd = durnpso_sd), prevsys = distr(qbern, prob = prevsys), bsa = distr(qlogitnorm, mean = bsa_mean, sd = bsa_sd), weight = distr(qgamma, mean = weight_mean, sd = weight_sd), psa = distr(qbern, prob = psa), cor = pso_net$int_cor, n_int = 64) # Population-average marginal probit differences of achieving PASI 75 in this # target population, given a Normal(-1.75, 0.08^2) distribution on the # baseline probit-probability of response on Placebo (at the reference levels # of the covariates), are given by (pso_marg_new <- marginal_effects(pso_fit, mtype = "link", newdata = new_agd_int, baseline = distr(qnorm, -1.75, 0.08))) plot(pso_marg_new) ## Progression free survival with newly-diagnosed multiple myeloma # Run newly-diagnosed multiple myeloma example if not already available if (!exists("ndmm_fit")) example("example_ndmm", run.donttest = TRUE) # We can produce a range of marginal effects from models with survival # outcomes, specified with the mtype and type arguments. For example: # Marginal survival probability difference at 5 years, all contrasts marginal_effects(ndmm_fit, type = "survival", mtype = "difference", times = 5, all_contrasts = TRUE) # Marginal difference in RMST up to 5 years marginal_effects(ndmm_fit, type = "rmst", mtype = "difference", times = 5) # Marginal median survival time ratios marginal_effects(ndmm_fit, type = "median", mtype = "ratio") # Marginal log hazard ratios # With no covariates in the model, these are constant over time and study # populations, and are equal to the log hazard ratios from relative_effects() plot(marginal_effects(ndmm_fit, type = "hazard", mtype = "link"), # The hazard is infinite at t=0 in some studies, giving undefined logHRs at t=0 na.rm = TRUE) # The NDMM vignette demonstrates the production of time-varying marginal # hazard ratios from a ML-NMR model that includes covariates, see # `vignette("example_ndmm")` # Marginal survival difference over time plot(marginal_effects(ndmm_fit, type = "survival", mtype = "difference"))
Useful links