Hi Soko,
We just fixed the nnet size default bug, so the next time that MatchIt
rebuilds, the size default will get pulled through to change the default.
The fact that extraneous arguments to nnet don't get a warning message is
due to the way nnet.formula internally handles (...) arguments. For
example:
nnet.formula(treat~educ,data=lalonde,node=3,size=4,soko=4)
doesn't yield a warning. This is something peculiar to the nnet package, so
just be careful in passing arguments when model="nnet."
Dan
----- Original Message -----
From: "Daniel E. Ho" <daniel.ho(a)yale.edu>
To: "S Setoguchi" <soko(a)kuhp.kyoto-u.ac.jp>
Cc: "Kosuke Imai" <kimai(a)Princeton.Edu>du>; "Elizabeth Stuart"
<EStuart(a)Mathematica-Mpr.com>
Sent: Monday, March 07, 2005 12:26 AM
Subject: Re: [matchit] question on pscore matching using nnet or cart (fwd)
You're right. This is a bug in distance, which
should be pulling through
any (...) arguments to nnet but actually isn't. The rpart implementation
works this way:
ff<-matchit(treat~educ,data=lalonde,model="cart",cp=0.03)
ff$assign$call
rpart(formula = treat ~ educ, data = lalonde, cp = 0.03)
Sorry for the confusion, but we need to fix the nnet portion.
Dan
----- Original Message -----
From: "S Setoguchi" <soko(a)kuhp.kyoto-u.ac.jp>
To: "'Daniel E. Ho'" <daniel.ho(a)yale.edu>
Sent: Monday, March 07, 2005 12:09 AM
Subject: RE: [matchit] question on pscore matching using nnet or cart
(fwd)
I see what was wrong.
You used the term 'node' in your previous email
> >
foo<-matchit(treat~educ,model="nnet",data=lalonde,node=4)
But now you are using 'size' in this reply.
That was my question. I got confused with the term. I thought it was
'size'
but you used 'node' once in your previous
email.
I see you are using 'size' this time.
Soko
-----Original Message-----
From: Daniel E. Ho [mailto:daniel.ho@yale.edu]
Sent: Monday, March 07, 2005 12:09 AM
To: S Setoguchi; matchit(a)latte.harvard.edu
Subject: Re: [matchit] question on pscore matching using nnet or cart
(fwd)
I'm not exactly sure what you are trying, but the problem might be with
nnet
itself.
nnet uses some sort of random seed so you will get different answers with
each neural net that you run. For example:
nnet(treat~educ,size=4,data=lalonde)
#
weights: 13
initial value 295.270340
final value 185.000000
converged
a 1-4-1 network with 13 weights
inputs: educ
output(s): treat
options were -
nnet(treat~educ,size=4,data=lalonde)
#
weights: 13
initial value 130.405931
iter 10 value 128.240413
iter 20 value 128.174714
iter 30 value 127.232077
iter 40 value 126.821942
final value 126.755841
converged
a 1-4-1 network with 13 weights
inputs: educ
output(s): treat
options were -
If you're running into a different problem, please feel free to send along
a
few lines to replicate the problem.
Thanks,
Dan
----- Original Message -----
From: "S Setoguchi" <soko(a)kuhp.kyoto-u.ac.jp>
To: "'Daniel E. Ho'" <daniel.ho(a)yale.edu>du>;
<matchit(a)latte.harvard.edu>
Sent: Sunday, March 06, 2005 11:47 PM
Subject: RE: [matchit] question on pscore matching using nnet or cart
(fwd)
Hi
I was playing with psmatch for nnet.
I realized that an argument to change size of nodes in the original
nnet
is
'size' but not 'node'.
I tried both in matchit and both run but give me different results.
Can you explain why this is??
-----Original Message-----
From: Daniel E. Ho [mailto:daniel.ho@yale.edu]
Sent: Sunday, March 06, 2005 10:48 PM
To: S Setoguchi
Subject: Re: [matchit] question on pscore matching using nnet or cart
(fwd)
Great. For some reason I wasn't getting the matchit@latte emails, so
I'm just going to resend my answer to the listserv in case others
might find
it
helpful.
Dan
----- Original Message -----
From: "S Setoguchi" <soko(a)kuhp.kyoto-u.ac.jp>
To: "'Daniel E. Ho'" <daniel.ho(a)yale.edu>du>;
"'Elizabeth Stuart'"
<stuart(a)stat.harvard.edu>
Cc: "'Kosuke Imai'" <kimai(a)Princeton.Edu>du>;
<ssetoguc(a)hsph.harvard.edu>
Sent: Sunday, March 06, 2005 9:32 PM
Subject: RE: [matchit] question on pscore matching using nnet or cart
(fwd)
Thanks. It makes sense.
Soko
-----Original Message-----
From: Daniel E. Ho [mailto:daniel.ho@yale.edu]
Sent: Sunday, March 06, 2005 8:46 PM
To: Elizabeth Stuart
Cc: Kosuke Imai; ssetoguc(a)hsph.harvard.edu
Subject: Re: [matchit] question on pscore matching using nnet or
cart
(fwd)
Weird, for some reason I'm not getting any of these matchit
messages.
Yes, all of the usual tweaks can be used with nnet and cart.
For example:
data(lalonde)
foo<-matchit(treat~educ,model="nnet",data=lalonde)
foo<-matchit(treat~educ,model="nnet",data=lalonde,node=4)
foo<-matchit(treat~educ,model="cart",data=lalonde)
foo<-matchit(treat~educ,model="cart",data=lalonde,cp=0.02)
I hope that makes sense!
Dan
----- Original Message -----
From: "Elizabeth Stuart" <stuart(a)stat.harvard.edu>
To: "Daniel E. Ho" <daniel.ho(a)yale.edu>
Cc: "Kosuke Imai" <kimai(a)Princeton.Edu>
Sent: Sunday, March 06, 2005 8:35 PM
Subject: [matchit] question on pscore matching using nnet or cart
(fwd)
Hey Dan,
I seem to remember that you played around with nnet and cart a
little. My memory is that any of the nnet or cart options can be
changed by just adding that to the options in the matchit command
(that's what the "..." in the matchit command is for). Is that
right?
It was good talking to you tonight. Good luck with your
preparations!
liz
---------- Forwarded message ----------
Date: Sun, 6 Mar 2005 13:59:25 -0500
From: Soko Setoguchi <ssetoguc(a)hsph.harvard.edu>
To: matchit(a)latte.harvard.edu
Subject: [matchit] question on pscore matching using nnet or cart
Hi
I have questions on propensity score matching using nnet or cart.
For nnet, I know the default set by matchit is node=3 but how do
you
change it
to other values.
Similarly for cart, I assume that matchit uses default setting of
rpart.
How
do you change the setting of rapart for example
cp, xval,
minbucket etc...
Thank you
Soko
Soko Setoguchi
Harvard School of Public Health
Soko Setoguchi MD MPH ScM
Doctor of Public Health Candidate
Department of Epidemiology
Harvard School of Public Health
-
MatchIt mailing list served by Harvard-MIT Data Center List
Address: matchit(a)latte.harvard.edu
Subscribe/Unsubscribe:
http://lists.hmdc.harvard.edu/?info=matchit
MatchIt Software and Documentation:
http://gking.harvard.edu/matchit/
-
MatchIt mailing list served by Harvard-MIT Data Center
List Address: matchit(a)latte.harvard.edu
Subscribe/Unsubscribe:
http://lists.hmdc.harvard.edu/?info=matchit
MatchIt Software and Documentation:
http://gking.harvard.edu/matchit/