Get simulations from a model object
sim_qi.Rdsim_qi() is a function to simulate quantities of interest
from a regression model
Arguments
- mod
a model object
- nsim
number of simulations to be run, defaults to 1,000
- newdata
A data frame with a hypothetical prediction grid. If absent, defaults to the model frame.
- original_scale
logical, defaults to TRUE. If TRUE, the ensuing simulations are returned on their original scale. If FALSE, the ensuing simulations are transformed to a more practical/intuitive quantity that for now is the simulated probability. This argument is ignored in the context of simulations on the linear model.
- return_newdata
logical, defaults to FALSE. If TRUE, the output returns additional columns corresponding with the inputs provided to
newdata. This may facilitate easier transformation along with greater clarity as to what the simulations correspond.
Value
sim_qi() returns a data frame (as a tibble) with the
quantities of interest and identifying information about the particular
simulation number. For linear models, or simple generalized linear models
where the dependent variable is either "there" or "not there", the quantity
of interest returned is a single column (called y). For models where the
underlying estimation of the dependent variable is, for lack of a better
term, "multiple" (e.g. ordinal models with the basic proportional odds)
assumption), the columns returned correspond with the number of distinct
values of the outcome. For example, an ordinal model where there are five
unique values of the dependent variable will return columns y1, y2, y3,
y4, and y5.
Details
Specifying a variable in newdata with the exact same name as the
dependent variable (e.g. mpg in the simple example provided in this
documentation file) is necessary for matrix multiplication purposes. If you
set return_newdata to TRUE, you should not interpret the column matching
the name of the dependent variable as communicating the kind of information
you want from this function. That particular column is just a simple
placeholder you need for matrix multiplication. The information you want will
always be in a column (or columns) named (or starting with) y.
This function builds in an implicit assumption that your dependent variable
in the regression model is not called y.
For ordinal models, I recommend setting original_scale to be FALSE. The
function, underneath the hood, is actually calculating things on the level of
the probability. It's just transforming back to a logit or a probit, if that
is what you say you want.