predict.glmmTMB function

prediction

## S3 method for class 'glmmTMB'
predict(
  object,
  newdata = NULL,
  newparams = NULL,
  se.fit = FALSE,
  cov.fit = FALSE,
  re.form = NULL,
  allow.new.levels = FALSE,
  type = c("link", "response", "conditional", "zprob", "zlink", "disp", "latent"),
  zitype = NULL,
  na.action = na.pass,
  fast = NULL,
  debug = FALSE,
  aggregate = NULL,
  do.bias.correct = FALSE,
  bias.correct.control = list(sd = TRUE),
  ...
)

Arguments

• object: a glmmTMB object

• newdata: new data for prediction

• newparams: new parameters for prediction

• se.fit: return the standard errors of the predicted values?

• cov.fit: return the covariance matrix of the predicted values?

• re.form: NULL to specify individual-level predictions; ~0 or NA to specify population-level predictions (i.e., setting all random effects to zero)

• allow.new.levels: allow previously unobserved levels in random-effects variables? see details.

• type: Denoting $mu$ as the mean of the conditional distribution and p as the zero-inflation probability, the possible choices are:

  • "link": the linear predictor of the conditional model, or equivalently the conditional mean on the scale of the link function (this equivalence does not hold for truncated distributions, where the link-scaled value is not adjusted for the effect of truncation on the mean; to get the corrected value of the conditional mean on the linear predictor scale, use family(m)$linkfun(predict(m, type = "conditional")))
  • "response": expected value; this is $mu*(1-p)$ for zero-inflated models and mu otherwise
  • "conditional": mean of the conditional response; mu for all models (i.e., synonymous with "response" in the absence of zero-inflation
  • "zprob": the probability of a structural zero (returns 0 for non-zero-inflated models)
  • "zlink": predicted zero-inflation probability on the scale of the logit link function (returns -Inf for non-zero-inflated models)
  • "disp": dispersion parameter, however it is defined for that particular family (as described in sigma.glmmTMB)
  • "latent": return latent variables

• zitype: deprecated: formerly used to specify type of zero-inflation probability. Now synonymous with type

• na.action: how to handle missing values in newdata (see na.action); the default (na.pass) is to predict NA

• fast: predict without expanding memory (default is TRUE if newdata and newparams are NULL and population-level prediction is not being done)

• debug: (logical) return the TMBStruc object that will be used internally for debugging?

• aggregate: (optional factor vector) sum the elements with matching factor levels

• do.bias.correct: (logical) should aggregated predictions use Taylor expanded estimate of nonlinear contribution of random effects (see details)

• bias.correct.control: a list sent to TMB's function sdreport(). See documentation there.

• ...: unused - for method compatibility

Details

• To compute population-level predictions for a given grouping variable (i.e., setting all random effects for that grouping variable to zero), set the grouping variable values to NA. Finer-scale control of conditioning (e.g. allowing variation among groups in intercepts but not slopes when predicting from a random-slopes model) is not currently possible.
• Prediction of new random effect levels is possible as long as the model specification (fixed effects and parameters) is kept constant. However, to ensure intentional usage, a warning is triggered if allow.new.levels=FALSE (the default).
• Prediction using "data-dependent bases" (variables whose scaling or transformation depends on the original data, e.g. poly, ns, or poly) should work properly; however, users are advised to check results extra-carefully when using such variables. Models with different versions of the same data-dependent basis type in different components (e.g. formula= y ~ poly(x,3), dispformula= ~poly(x,2)) will probably not produce correct predictions.
• Bias corrected predictions are based on the method described in Thorson J.T. & Kristensen (2016). These should be checked carefully by the user and are not extensively tested.

Examples

data(sleepstudy,package="lme4")
g0 <- glmmTMB(Reaction~Days+(Days|Subject),sleepstudy)
predict(g0, sleepstudy)
## Predict new Subject
nd <- sleepstudy[1,]
nd$Subject <- "new"
predict(g0, newdata=nd, allow.new.levels=TRUE)
## population-level prediction
nd_pop <- data.frame(Days=unique(sleepstudy$Days),
                     Subject=NA)
predict(g0, newdata=nd_pop)
## return latent variables (BLUPs/conditional modes/etc. ) with standard errors
##  (actually conditional standard deviations)
predict(g0, type = "latent", se.fit = TRUE)

References

Thorson J.T. & Kristensen K. (2016) Implementing a generic method for bias correction in statistical models using random effects, with spatial and population dynamics examples. Fish. Res. 175, 66-74.