vignettes/riskregression.Rmd
riskregression.Rmd
Let be a binary response, a binary exposure, and a vector of covariates.
In a common setting, the main interest lies in quantifying the treatment effect, , of on adjusting for the set of covariates, and often a standard approach is to use a Generalized Linear Model (GLM):
with link function , and , known vector functions of .
The canonical link (logit) leads to nice computational properties (logistic regression) and parameters with an odds-ratio interpretation. But ORs are not collapsible even under randomization. For example
When marginalizing we leave the class of logistic regression. This non-collapsibility makes it hard to interpret odds-ratios and to compare results from different studies
Relative risks (and risk differences) are collapsible and generally considered easier to interpret than odds-ratios. Richardson et al (JASA, 2017) proposed a regression model for a binary exposures which solves the computational problems and need for parameter contraints that are associated with using for example binomial regression with a log-link function (or identify link for the risk difference) to obtain such parameter estimates. In the following we consider the relative risk as the target parameter
Let , the idea is then to posit a linear model for , i.e.,
and a nuisance model for the odds-product
noting that these two parameters are variation independent as illustrated by the below L’Abbé plot.
p0 <- seq(0,1,length.out=100)
p1 <- function(p0,op) 1/(1+(op*(1-p0)/p0)^-1)
plot(0, type="n", xlim=c(0,1), ylim=c(0,1),
xlab="P(Y=1|A=0)", ylab="P(Y=1|A=1)", main="Constant odds product")
for (op in exp(seq(-6,6,by=.25))) lines(p0,p1(p0,op), col="lightblue")
p0 <- seq(0,1,length.out=100)
p1 <- function(p0,rr) rr*p0
plot(0, type="n", xlim=c(0,1), ylim=c(0,1),
xlab="P(Y=1|A=0)", ylab="P(Y=1|A=1)", main="Constant relative risk")
for (rr in exp(seq(-3,3,by=.25))) lines(p0,p1(p0,rr), col="lightblue")
Similarly, a model can be constructed for the risk-difference on the following scale
First the targeted
package is loaded
This automatically imports lava (CRAN) which we can use to simulate from the Relative-Risk Odds-Product (RR-OP) model.
m <- lava::lvm(a ~ x,
lp.target ~ 1,
lp.nuisance ~ x+z)
m <- lava::binomial.rr(m, response="y", exposure="a", target.model="lp.target", nuisance.model="lp.nuisance")
The lvm
call first defines the linear predictor for the
exposure to be of the form
and the linear predictors for the /target parameter/ (relative risk) and the /nuisance parameter/ (odds product) to be of the form
The covariates are by default assumed to be independent and standard
normal
,
but their distribution can easily be altered using the
lava::distribution
method.
The binomial.rr
function
args(lava::binomial.rr)
#> function (x, response, exposure, target.model, nuisance.model,
#> exposure.model = binomial.lvm(), ...)
#> NULL
then defines the link functions, i.e.,
with .
The risk-difference model with the RD parameter modeled on the
scale can be defined similarly using the binomial.rd
method
args(lava::binomial.rd)
#> function (x, response, exposure, target.model, nuisance.model,
#> exposure.model = binomial.lvm(), ...)
#> NULL
We can inspect the parameter names of the modeled
coef(m)
#> m1 m2 m3 p1 p2
#> "a" "lp.target" "lp.nuisance" "a~x" "lp.nuisance~x"
#> p3 p4
#> "lp.nuisance~z" "a~~a"
Here the intercepts of the model are simply given the same name as
the variables, such that
becomes a
, and the other regression coefficients are
labeled using the “~”-formula notation, e.g.,
becomes a~x
.
Intercepts are by default set to zero and regression parameters set
to one in the simulation. Hence to simulate from the model with
, we define the
parameter vector p
given by
p <- c('a'=-1, 'lp.target'=1, 'lp.nuisance'=-1, 'a~x'=2)
and then simulate from the model using the sim
method
d <- lava::sim(m, 1e4, p=p, seed=1)
head(d)
#> a x lp.target lp.nuisance z y
#> 1 0 -0.6264538 1 -2.4307854 -0.8043316 0
#> 2 0 0.1836433 1 -1.8728823 -1.0565257 0
#> 3 0 -0.8356286 1 -2.8710244 -1.0353958 0
#> 4 1 1.5952808 1 -0.5902796 -1.1855604 1
#> 5 0 0.3295078 1 -1.1709317 -0.5004395 1
#> 6 0 -0.8204684 1 -2.3454571 -0.5249887 0
Notice, that in this simulated data the target
parameter
has been set to lp.target =
1.
We start by fitting the model using the maximum likelihood estimator.
args(riskreg_mle)
#> function (y, a, x1, x2 = x1, weights = rep(1, length(y)), std.err = TRUE,
#> type = "rr", start = NULL, control = list(), ...)
#> NULL
The riskreg_mle
function takes vectors/matrices as input
arguments with the response y
, exposure a
,
target parameter design matrix x1
(i.e., the matrix
at the start of this text), and the nuisance model design matrix
x2
(odds product).
We first consider the case of a correctly specified model, hence we
do not consider any interactions with the exposure for the odds product
and simply let x1
be a vector of ones, whereas the design
matrix for the log-odds-product depends on both
and
x1 <- model.matrix(~1, d)
x2 <- model.matrix(~x+z, d)
fit1 <- with(d, riskreg_mle(y, a, x1, x2, type="rr"))
fit1
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 0.9512 0.03319 0.8862 1.0163 1.204e-180
#> odds-product:(Intercept) -1.0610 0.05199 -1.1629 -0.9591 1.377e-92
#> odds-product:x 1.0330 0.05944 0.9165 1.1495 1.230e-67
#> odds-product:z 1.0421 0.05285 0.9386 1.1457 1.523e-86
The parameters are presented in the same order as the columns of
x1
and x2
, hence the target parameter estimate
is in the first row
estimate(fit1, keep=1)
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 0.9512 0.03336 0.8858 1.017 7.159e-179
We next fit the model using a double robust estimator (DRE) which introduces a model for the exposure (propensity model). The double-robustness stems from the fact that the this estimator remains consistent in the union model where either the odds-product model or the propensity model is correctly specified. With both models correctly specified the estimator is efficient.
with(d, riskreg_fit(y, a, target=x1, nuisance=x2, propensity=x2, type="rr"))
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 0.9372 0.0339 0.8708 1.004 3.004e-168
The usual /formula/-syntax can be applied using the
riskreg
function. Here we illustrate the double-robustness
by using a wrong propensity model but a correct nuisance
paramter (odds-product) model:
riskreg(y~a, nuisance=~x+z, propensity=~z, data=d, type="rr")
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 0.9511 0.03333 0.8857 1.016 4.547e-179
Or vice-versa
riskreg(y~a, nuisance=~z, propensity=~x+z, data=d, type="rr")
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 0.9404 0.03727 0.8673 1.013 1.736e-140
whereas the MLE in this case yields a biased estimate of the relative risk:
fit2 <- with(d, riskreg_mle(y, a, x1=model.matrix(~1,d), x2=model.matrix(~z, d)))
estimate(fit2, keep=1)
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 1.243 0.02778 1.189 1.298 0
The more general model where
for a subset
of the covariates can be estimated using the target
argument:
fit <- riskreg(y~a, target=~x, nuisance=~x+z, data=d)
fit
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 0.95267 0.03365 0.88673 1.01862 2.361e-176
#> x -0.01078 0.03804 -0.08534 0.06378 7.769e-01
As expected we do not see any evidence of an effect of on the relative risk with the 95% confidence limits clearly overlapping zero.
Note, that when the propensity
argument is omitted as
above, the same design matrix is used for both the odds-product model
and the propensity model.
The syntax for fitting the risk-difference model is similar. To illustrate this we simulate some new data from this model
m2 <- lava::binomial.rd(m, response="y", exposure="a", target.model="lp.target", nuisance.model="lp.nuisance")
d2 <- lava::sim(m2, 1e4, p=p)
And we can then fit the DRE with the syntax
riskreg(y~a, nuisance=~x+z, data=d2, type="rd")
#> Estimate Std.Err 2.5% 97.5% P-value
#> (Intercept) 1.027 0.02081 0.9857 1.067 0
sessionInfo()
#> R version 4.5.1 (2025-06-13)
#> Platform: x86_64-pc-linux-gnu
#> Running under: Ubuntu 24.04.2 LTS
#>
#> Matrix products: default
#> BLAS: /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3
#> LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.26.so; LAPACK version 3.12.0
#>
#> locale:
#> [1] LC_CTYPE=C.UTF-8 LC_NUMERIC=C LC_TIME=C.UTF-8
#> [4] LC_COLLATE=C.UTF-8 LC_MONETARY=C.UTF-8 LC_MESSAGES=C.UTF-8
#> [7] LC_PAPER=C.UTF-8 LC_NAME=C LC_ADDRESS=C
#> [10] LC_TELEPHONE=C LC_MEASUREMENT=C.UTF-8 LC_IDENTIFICATION=C
#>
#> time zone: UTC
#> tzcode source: system (glibc)
#>
#> attached base packages:
#> [1] stats graphics grDevices utils datasets methods base
#>
#> other attached packages:
#> [1] targeted_0.6 lava_1.8.1
#>
#> loaded via a namespace (and not attached):
#> [1] Matrix_1.7-3 future.apply_1.20.0 jsonlite_2.0.0
#> [4] compiler_4.5.1 Rcpp_1.1.0 parallel_4.5.1
#> [7] jquerylib_0.1.4 globals_0.18.0 splines_4.5.1
#> [10] systemfonts_1.2.3 textshaping_1.0.1 yaml_2.3.10
#> [13] fastmap_1.2.0 lattice_0.22-7 R6_2.6.1
#> [16] knitr_1.50 htmlwidgets_1.6.4 future_1.67.0
#> [19] nloptr_2.2.1 desc_1.4.3 bslib_0.9.0
#> [22] rlang_1.1.6 cachem_1.1.0 xfun_0.52
#> [25] fs_1.6.6 sass_0.4.10 cli_3.6.5
#> [28] pkgdown_2.1.3 digest_0.6.37 grid_4.5.1
#> [31] mvtnorm_1.3-3 lifecycle_1.0.4 timereg_2.0.6
#> [34] pracma_2.4.4 evaluate_1.0.4 data.table_1.17.8
#> [37] numDeriv_2016.8-1.1 listenv_0.9.1 codetools_0.2-20
#> [40] ragg_1.4.0 survival_3.8-3 optimx_2025-4.9
#> [43] parallelly_1.45.1 rmarkdown_2.29 mets_1.3.6
#> [46] tools_4.5.1 htmltools_0.5.8.1