Hi,
I have a multi-state model that I would like to estimate using the 'mstate'
package - but I am not sure how best to approach it and was wondering
if anyone could provide some insight for me.
Basically, I have a group of kids who have been randomized to one of 2
treatments, A or B. If they do well on either of these, they
are discharged (A -->D or B--> D). If they do poorly, kids who were receiving
treatment A are moved onto treatment B (A-->B), but kids
on B can't go to A, they go straight into ICU (A-->C). If the kids who went
from A-->B do well they are discharged (A-->B-->D), and if they
do poorly they go to ICU (A-->B-->C), and then finally after ICU they are
discharged (C-->D). There are also a small number of kids who were sent
straight to ICU (so...in fact there are 3 initial states). So the transition
matrix looks like:
A B C D
A NA 1 NA 1
B NA NA 1 1
C NA NA NA 1
D NA NA NA NA
However, I am not sure whether I can do this in 'mstate', as there are three
initial states? The program requires that we start with
a wide data set, in which there is an event indicator for each state, and a
time of entry into each state. If they don't enter a state, then
the time for that state is set to last follow-up (page 4 of
https://www.jstatsoft.org/article/view/v038i07 ).
So, in the case where a kid starts in A, and goes to D; the data would look
like:
A.status A.time B.status B.time C.Status C.Time D.Status
D.time
1 0 0 Final 0
Final 1 Final
And I think this is OK in terms of what is specified for B, because technically
the kid is at risk of transitioning into B
until they are discharged.
But what about a kid who starts in B and goes to D? The corresponding data
would possibly be?:
A.status A.time B.status B.time C.Status C.Time D.Status
D.time
0 Final 1 0 0
Final 1 Final
However, this to me doesn't look right, as technically they are never at risk
of going to A if they started in B.
I tried setting the A.time to 0 to reflect the fact that they are never at risk
of going back to A, but the numbers I got
from the model (events$model) were incorrect - basically all the events were
going to A, and no one was in B.
And for the kids that started in C...similarly.
So, I was wondering, would it be valid to create a new initial state (e.g. P =
pre-treatment), from which the child
then transitions immediately (at say, time = 0.1) to either A or B (or C). So
the matrix would be:
P A B C D
P NA 1 1 1 NA
A NA NA 1 NA 1
B NA NA NA 1 1
C NA NA NA NA 1
D NA NA NA NA NA
And then the data for someone who went from A--D would be
A.status A.time B.status B.time C.Status C.Time D.Status
D.time
1 0.1 0 0.1 0
Final 1 Final
And then the data for someone who went from B--D would be
A.status A.time B.status B.time C.Status C.Time
D.Status D.time
0 0.1 1 0 .1 0
Final 1 Final
When I code the model like this, then the number of events I get from the model
is correct.
However, I am not sure whether adding this extra initial state is a valid
option?
Thanks in advance for anyone who can help me out,
Lucy Leigh
mstate: An R Package for the Analysis of Competing Risks
...<https://www.jstatsoft.org/article/view/v038i07>
www.jstatsoft.org
Authors: Liesbeth C. de Wreede, Marta Fiocco, Hein Putter: Title: mstate: An R
Package for the Analysis of Competing Risks and Multi-State Models
[[alternative HTML version deleted]]
______________________________________________
R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
