mfp2Obtains predictions from an mfp2 object.
# S3 method for mfp2
predict(
object,
newdata = NULL,
type = NULL,
terms = NULL,
terms_seq = c("equidistant", "data"),
alpha = 0.05,
ref = NULL,
strata = NULL,
newoffset = NULL,
nseq = 100,
...
)For any type other than "terms" the output conforms to the output
of predict.glm() or predict.coxph().
If type = "terms" or type = "contrasts", then a named list with entries
for each variable requested in terms (excluding those not present in the
final model).
Each entry is a data.frame with the following columns:
variable: variable values on original scale (without shifting).
variable_pre: variable with pre-transformation applied, i.e. shifted, and centered as required.
value: partial linear predictor or contrast (depending on type).
se: standard error of partial linear predictor or contrast.
lower: lower limit of confidence interval.
upper: upper limit of confidence interval.
a fitted object of class mfp2.
optionally, a matrix with column names in which to look for
variables with which to predict. If provided, the variables are internally
shifted using the shifting values stored in object. See mfp2() for
further details.
the type of prediction required. The default is on the scale of
the linear predictors. See predict.glm() or predict.coxph() for details.
In case type = "terms", see the Section on Terms prediction. In case
type = "contrasts", see the Section on Contrasts.
a character vector of variable names specifying for which
variables term or contrast predictions are desired. Only used in case
type = "terms" or type = "contrasts". If NULL (the default) then all
selected variables in the final model will be used. In any case, only
variables used in the final model are used, even if more variable names are
passed.
a character string specifying how the range of variable
values for term predictions are handled. The default equidistant computes
the range of the data range and generates an equidistant sequence of
100 points from the minimum to the maximum values of shifted values to
properly show the functional form estimated in the final model.
The option data uses the observed data values directly, but these may not
adequately reflect the functional form of the data, especially when extreme
values or influential points are present.
significance level used for computing confidence intervals in terms prediction.
a named list of reference values used when type = "contrasts".
Note that any variable requested in terms, but not having an entry in this
list (or if the entry is NULL) then the mean value of shifted data
(or minimum for binary variables) will be used as reference. Values should be
specified on the original scale of the variable since the program will
internally scale it using the scaling factors obtained from
find_scale_factor(). By default, this function uses the means
(for continuous variables) and minimum (for binary variables) as
reference values.
stratum levels used for predictions.
A vector of offsets used for predictions. This parameter is important when newdata is supplied. The offsets are directly added to the linear predictor without any transformations.
Integer specifying how many values to generate when
terms_seq = "equidistant". Default is 100.
further arguments passed to predict.glm() or predict.coxph().
If type = "terms", this function computes the partial linear predictors
for each variable included in the final model. Unlike predict.glm() and
predict.coxph(), this function accounts for the fact that a single variable
may be represented by multiple transformed terms.
For a variable modeled using a first-degree fractional polynomial (FP1),
the partial predictor is given by \(\hat{\eta}_j = \hat{\beta}_0 + x_j^* \hat{\beta}_j\), where
\(x_j^*\) is the transformed variable (centered if center = TRUE).
For a second-degree fractional polynomial (FP2), the partial predictor
takes the form \(\hat{\eta}_j = \hat{\beta}_0 + x_{j1}^* \hat{\beta}_{j1} + x_{j2}^* \hat{\beta}_{j2}\),
where \(x_{j1}^*\) and \(x_{j2}^*\) are the two transformed components
of the original variable (again, centered if center = TRUE).
This functionality is particularly useful for visualizing the functional
relationship of a continuous variable, or for assessing model fit when
residuals are included. See also fracplot().
If type = "contrasts", this function computes contrasts relative to a
specified reference value for the \(j\)th variable (e.g., age = 50). Let
\(x_j\) denote the values of the \(j\)th variable in newdata, and
\(x_j^{\text{ref}}\) the reference value. The contrast is defined as the
difference between the partial linear predictor evaluated at the transformed
(and centered, if center = TRUE) value \(x_j\), and that evaluated at the
transformed reference value \(x_j^{(\text{ref}})\),
i.e., \(f(x_j^*) - f(x_j^{*(\text{ref})})\).
For a first-degree fractional polynomial (FP1), the partial predictor is: $$\hat{f}(x_j^*) = \hat{\beta}_0 + x_j^* \hat{\beta}_j$$ and the contrast is: $$\hat{f}(x_j^*) - \hat{f}(x_j^{*(\text{ref})}) = x_j^* \hat{\beta}_j - x_j^{*(\text{ref})} \hat{\beta}_j$$
For a second-degree fractional polynomial (FP2), the partial predictor is: $$\hat{f}(x_j^*) = \hat{\beta}_0 + x_{j1}^* \hat{\beta}_{j1} + x_{j2}^* \hat{\beta}_{j2}$$ and the contrast is: $$\hat{f}(x_j^*) - \hat{f}(x_j^{*(\text{ref})}) = x_{j1}^* \hat{\beta}_{j1} + x_{j2}^* \hat{\beta}_{j2} - x_{j1}^{*(\text{ref})} \hat{\beta}_{j1} - x_{j2}^{*(\text{ref})} \hat{\beta}_{j2}$$
where \(x_j^*\), \(x_{j1}^*\), and \(x_{j2}^*\) are the transformed (and centered, if applicable) components of the \(j\)th variable, and the \(\hat{\beta}\) terms are the corresponding model estimates
The reference value \(x_j^{(\text{ref})}\) is first shifted using the
same shifting factor estimated from the training data, then transformed using
the estimated fractional polynomial (FP) powers, and finally centered
(if center = TRUE) using the mean of the transformed (and shifted) values
of \(x_j\) in the training data—ensuring full consistency with the fitted model.
If ref = NULL, the function uses the mean of the shifted \(x_j\) as the
reference value when \(x_j\) is continuous, or the minimum of \(x_j\)
(typically 0) when \(x_j\) is binary. This provides a natural and interpretable
baseline in the absence of a user-specified reference.
The fitted partial predictors are centered at the reference point, meaning the contrast at that point is zero. Correspondingly, confidence intervals at the reference value have zero width, reflecting no contrast.
This functionality is especially useful for comparing the effect of a variable relative to a meaningful baseline, such as clinically relevant value.
To prepare the newdata for prediction, this function applies any
necessary shifting based on factors obtained from the training data.
It is important to note that if the shifting factors estimated from the
training data are not sufficiently large, variables in newdata may end up
being non-positive, which can cause prediction errors when non-linear
functional forms such as logarithms are used. In such cases, the function
issues a warning.
The next step involves transforming the data using the selected
fractional polynomial (FP) powers. After transformation, variables are
centered if center was set to TRUE in mfp2(). Once transformation
(and centering) is complete, the transformed data is passed
to either predict.glm() or predict.coxph(), depending on the model
family used, provided that type is neither terms nor contrasts (see the
section handling terms and contrasts for details).
# Gaussian model
data("prostate")
x = as.matrix(prostate[,2:8])
y = as.numeric(prostate$lpsa)
# default interface
fit1 = mfp2(x, y, verbose = FALSE)
predict(fit1) # make predictions
Run the code above in your browser using DataLab