pmcalibration
library(pmcalibration)
# simulate some data for vignette
set.seed(2345)
dat <- sim_dat(1000, a1 = -3, a3 = .3)
# show the first 3 columns (col 4 is the true linear predictor/LP)
head(dat[-4])
#> x1 x2 y
#> 1 -1.19142464 -0.9245914 0
#> 2 0.54930055 -1.0019698 0
#> 3 -0.06240514 1.5438665 0
#> 4 0.26544150 0.1632147 0
#> 5 -0.23459751 -1.2009388 0
#> 6 -0.99727160 -1.1899600 0
We have data with a binary outcome, y
, and two
‘predictor’ variables, x1
and x2
. Suppose we
have an existing model for predicting y
from
x1
and x2
that is as follows
p(y = 1) = plogis( -3 + 1*x1 + 1*x2 )
To externally validate this model on this new data we need to calculate the predicted probabilities. We’ll also extract the observed outcomes.
First we can check ‘calibration-in-the-large’ via the calibration intercept and slope.
logistic_cal(y = y, p = p)
#> Logistic calibration intercept and slope:
#>
#> Estimate Std. Error z value Pr(>|z|) lower upper
#> Calibration Intercept -0.044 0.12 -0.36 0.7223 -0.29 0.19
#> Calibration Slope 1.336 0.13 2.62 0.0088 1.10 1.60
#>
#> z-value for calibration slope is relative to slope = 1.
#> lower and upper are the bounds of 95% profile confidence intervals.
The calibration-intercept suggests no particular bias with a point estimate not far off zero. The calibration slope suggests that predicted probabilities are too extreme. However, this logistic calibration enforces a linear relationship between logit transformed probabilities and the log odds of y = 1.
Below we use pmcalibration
to fit a flexible calibration
curve, allowing for a non-linear relationship between predicted and
actual probabilities. This assesses ‘moderate calibration’ according to
the hierarchy of Van
Calster et al. (2016).
In the example below, we fit a calibration curve using
mgcv::gam
via a penalized thin plate regression spline (see
?mgcv::tprs
). pmcalibration
calculates various
metrics from the absolute difference between the predicted probability
and the actual probability (as estimated by the calibration curve). In
this case 95% confidence intervals for these metrics are calculated via
simulation based inference.
(cc <- pmcalibration(y = y, p = p,
smooth = "gam", bs = "tp",
k = 10, transf="logit",
ci = "sim", method="REML"))
#> Calibration metrics based on a calibration curve estimated for a binary outcome via a generalized additive model (see ?mgcv::s) using logit transformed predicted probabilities.
#>
#> Estimate lower upper
#> Eavg 0.020 0.0087 0.032
#> E50 0.013 0.0031 0.023
#> E90 0.024 0.0111 0.062
#> Emax 0.211 0.0700 0.305
#> ECI 0.140 0.0263 0.368
#>
#> 95% confidence intervals calculated via simulation based inference with 1000 replicates.
The printed metrics can be interpreted as follows:
Eavg
suggests that the average difference between
prediction and actual probability of the outcome is 0.02 (or 2%) with a
95% CI of [0.009, 0.031].E50
is the median difference between prediction and
observed probability (inferred from calibration curve). 50% of
differences are 0.013 or smaller.E90
is the 90th percentile difference. 90% of
differences are 0.024 or smaller.Emax
is the largest observed difference between
predicted and observed probability. The model can be off by up to 0.21,
with a broad confidence interval.ECI
is the average squared difference between predicted
and observed probabilities (multiplied by 100). See Van Hoorde et
al. (2015).A quick and simple plot of the calibration curve, and 95% confidence
interval, can be obtained via plot
.
Or one could use get_cc
to extract data for plotting
with method of your choice. The plot below also shows the distribution
of predicted probabilities.
library(ggplot2)
pcc <- get_cc(cc)
ggplot(pcc, aes(x = p, y = p_c, ymin=lower, ymax=upper)) +
geom_abline(intercept = 0, slope = 1, lty=2) +
geom_line() +
geom_ribbon(alpha = 1/2, fill="lightblue") +
coord_cartesian(xlim=c(0,1), ylim=c(0,1)) +
labs(x = "Predicted", y = "Estimated") +
theme_bw(base_size = 14) +
geom_histogram(data = data.frame(p = p), aes(x=p, y=after_stat(density)*.01),
binwidth = .001, inherit.aes = F, alpha=1/2)
The model in its current form very slightly underestimates risk at low levels of predicted risk and then overestimates risk at predicted probabilities of over 0.4.
The results above can be compared with rms::val.prob
.
Note that this uses lowess(p, y, iter=0)
to fit a
non-linear (nonparametric) calibration curve. This calibration curve
suggests that the overestimation at high levels of predicted risk is
even more extreme that that suggested by gam
calibration
curve above. This is particularly evident in the estimate of
Emax
(0.35 vs 0.21).
library(rms)
#> Loading required package: Hmisc
#>
#> Attaching package: 'Hmisc'
#> The following objects are masked from 'package:base':
#>
#> format.pval, units
val.prob(p = p, y = y)
#> Dxy C (ROC) R2 D D:Chi-sq
#> 0.725449678 0.862724839 0.343241418 0.162057146 163.057146265
#> D:p U U:Chi-sq U:p Q
#> NA 0.005910946 7.910945691 0.019149612 0.156146201
#> Brier Intercept Slope Emax E90
#> 0.057594793 0.528823286 1.336388397 0.350214394 0.029680412
#> Eavg S:z S:p
#> 0.020095036 -1.684470284 0.092090818
Note also that the calibration intercept reported by
rms::val.prob
comes from the same logistic regression as
that used to estimate the calibration slope. In
logistic_cal
the calibration intercept is estimated via a
glm
with logit transformed predicted probabilities included
as an offset term (i.e., with slope fixed to 1 - see, e.g., Van Calster et al.,
2016). The calibration slope is estimated via a separate
glm
.
The code below produces a calibration curve, and associated metrics, for a time-to-event outcome.
library(simsurv)
library(survival)
# simulate some data
n <- 2000
X <- data.frame(id = seq(n), x1 = rnorm(n), x2 = rnorm(n))
X$x3 <- X$x1*X$x2 # interaction
b <- c("x1" = -.2, "x2" = -.2, "x3" = .1)
d <- simsurv(dist = "weibull", lambdas = .01, gammas = 1.5, x = X, betas = b, seed = 246)
mean(d$eventtime)
#> [1] 19.59251
median(d$eventtime)
#> [1] 16.53999
mean(d$status) # no censoring
#> [1] 1
d <- cbind(d, X[,-1])
head(d)
#> id eventtime status x1 x2 x3
#> 1 1 6.59752 1 -1.8030953 -0.9757324 1.7593385
#> 2 2 28.84754 1 0.8369289 -1.6105987 -1.3479566
#> 3 3 14.84345 1 -0.4916106 0.8366313 -0.4112968
#> 4 4 32.89247 1 2.0750555 -0.1059696 -0.2198928
#> 5 5 14.29186 1 -0.4198547 -1.9941478 0.8372523
#> 6 6 11.85858 1 -0.5349368 0.2494071 -0.1334170
# split into development and validation
ddev <- d[1:1000, ]
dval <- d[1001:2000, ]
# fit a cox model
cph <- coxph(Surv(eventtime, status) ~ x1 + x2, data = ddev)
# predicted probability of event at time = 15
p = 1 - exp(-predict(cph, type="expected", newdata = data.frame(eventtime=15, status=1, x1=dval$x1, x2=dval$x2)))
y <- with(dval, Surv(eventtime, status))
# calibration curve at time = 15
(cc <- pmcalibration(y = y, p = p, smooth = "rcs", nk = 5, ci = "pw", time = 15))
#> Calibration metrics based on a calibration curve estimated for a time-to-event outcome (time = 15) via a restricted cubic spline (see ?rms::rcs) using complementary log-log transformed predicted probabilities.
#>
#> Estimate
#> Eavg 0.035
#> E50 0.029
#> E90 0.070
#> Emax 0.201
#> ECI 0.217
# pointwise standard errors for plot but no CI for metrics
# 'boot' CIs are also available for time to event outcomes
Compare to rms::val.surv
, which with the arguments
specified below uses polspline::hare
to fit a calibration
curve. Note val.surv
uses probability of surviving until
time = u not probability of event occurring by time = u.