Biomedical Engineering Reference
In-Depth Information
Laan et al. (2007) and van der Laan and Rose (2011)), for example. The
fluctuation sub-models described above thus become
Q
2;n
+
I(A(1) = 1;A(0) = a)
Q
2;n
()
=
expit
logit
;
g
0n
(a)g
1n
(a)
Q
1;n
+
I(A(0) = a)
g
0n
(a)
Q
1;n
()
=
expit
logit
;
expit
logit
Q
0;n
+
:
Q
0;n
()
=
One step of targeted minimum loss-based estimation results in revised
Q
a;
Q
2;n
(
a;
Q
2
,
Q
a;
Q
1;n
(
a;
Q
1
, and
Q
a;
estimates
2;n
=
2;n
) of
1;n
=
1;n
) of
0;n
=
Q
0;n
(
a;
0;n
) of Q
0
, where the optimal fluctuation parameters are given by
X
a;
L
2;a
( Q
2;n
())(O
i
) ;
2;n
=
argmin
i=1
X
L
1;a
( Q
1;n
();
Q
a;
a;
1;n
=
argmin
2;n
)(O
i
) ;
i=1
X
L
0;a
( Q
0;n
(); Q
a;
a;
0;n
=
argmin
1;n
)(O
i
) :
i=1
A clear advantage of the choice of loss function and fluctuation sub-model
above is that rather than requiring a possibly cumbersome use of general-
purpose optimization routines to obtain the optimizers above, widely avail-
able statistical software may be easily utilized instead. Indeed, the minimizer
a;
2;n
can be obtained as the estimated coecient in a logistic regression of the
binary outcome Y on covariate [g
0n
(a)g
1n
(a)]
1
with offset logit Q
2;n
, tted
on the subset of data for which A(1) = 1 and A(0) = a. Subsequently, the
minimizer
a;
1;n
is the estimated coecient in a logistic regression of the out-
come Q
a;
2;n
on covariate g
0n
(a)
1
with offset logit Q
1;n
, fitted on the subset of
data for which A(0) = a. Finally, the minimizer
a;
0;n
consists of the estimated
coecient in a logistic regression of the outcome Q
a;
1;n
on a constant predictor
with offset logit Q
0;n
, fitted on all available data.
In particular, it is easy to verify that
P
i=1
Q
a;
Q
a;
1
n
0;n
=
1;n
(O
i
), and there-
Search WWH ::
Custom Search