predict.tramME {tramME}

Predict method for tramME objects

Description

Evaluates the _conditional_ distribution implied by a tramME model, given by a set of covariates and random effects on a selected scale.

Usage

## S3 method for class 'tramME'
predict(
  object,
  newdata = model.frame(object),
  ranef = NULL,
  fix_smooth = TRUE,
  type = c("lp", "trafo", "distribution", "logdistribution", "survivor", "logsurvivor",
    "density", "logdensity", "hazard", "loghazard", "cumhazard", "logcumhazard", "odds",
    "logodds", "quantile"),
  ...
)

Arguments

object

A tramME object.

newdata

an optional data frame of observations

ranef

Random effects it can be a ranef.tramME object, a named list, an unnamed list, NULL or the word "zero". See Details.

fix_smooth

If FALSE, the random effects coefficients of the smooth terms are refitted to newdata. It's probably not what you want to do.

type

The scale on which the predictions are evaluated:

• lp: Linear predictor (Xb + Zg). For more information, see Details.

• trafo: The prediction evaluated on the scale of the transformation function.

• (log)distribution: The prediction evaluated on the scale of the conditional (log-)CDF.

• (log)survivor: The prediction evaluated on the scale of the (conditional) (log-)survivor function.

• (log)density: The prediction evaluated on the scale of the conditional (log-)PDF.

• (log)hazard: The prediction evaluated on the (log-)hazard scale.

• (log)cumhazard: The prediction evaluated on the (log-)cumulative hazard scale.

• (log)odds: The prediction evaluated on the (log-)odds scale.

• quantile: Return the quantiles of the conditional outcome distribution corresponding to newdata. For more information, see Details.

...

Additional arguments, passed to predict.mlt.

Details

When newdata contains values of the response variable, prediction is only done for those values. In this case, if random effects vector (ranef) is not supplied by the user, the function predicts the random effects from the model using newdata.

When no response values are supplied in newdata, the prediction is done on a grid of values for each line of the dataset (see predict.mlt for information on how to control the setup of this grid). In this case, the user has to specify the vector of random effects to avoid ambiguities.

The linear predictor (type = "lp") equals to the shift terms plus the random effects terms _without the baseline transfromation function_.

The linear predictor (type = "lp") and the conditional quantile function (type = "quantile") are special in that they do not return results evaluated on a grid, even when the response variable in newdata is missing. The probabilities for the evaluation of the quantile function can be supplied with the prob argument of predict.mlt.

In the case of type = "quantile", when the some of the requested conditonal quantiles fall outside of the support of the response distribution (specified when the model was set up), the inversion of the CDF cannot be done exactly and tramME returns censored values.

ranef can be different objects based on what we want to calculate and what the other inputs are. If ranef is a ranef.tramME, we assume that it contains the full set of random effects, but not the penalized coefficients of the smooth terms. In this case fix_smooth must be TRUE. If ranef is a named vector, we are fixing the supplied random effects (and penalized coefficients) and predict the rest from newdata (fix_smooth may also be used in this case). In this case, the random effects are identified with the same naming convention as in object$param$gamma.

If ranef is an unnamed vector, the function expects the full set of necessary random effects (with or without penalized coefficients, depending on fix_smooth). If ranef = NULL (the default), all random effects and optionally penalized parameters (although this is not recommended) are predicted from newdata. Finally, if ranef is equal to "zero", a vector of zeros with the right size is used.

Value

A numeric vector/matrix of the predicted values (depending on the inputs) or a response object, when the some of the requested conditonal quantiles fall outside of the support of the response distribution specified when the model was set up (only can occur with type = "quantile").

Examples

data("sleepstudy", package = "lme4")
fit <- BoxCoxME(Reaction ~ Days + (Days | Subject), data = sleepstudy)
predict(fit, type = "trafo") ## evaluate on the transformation function scale
nd <- sleepstudy
nd$Reaction <- NULL
pr <- predict(fit, newdata = nd, ranef = ranef(fit), type = "distribution",
              K = 100)