But that is the question: Why is the variance of the difference in the
expected values too small if they are based on the formulas in Gary's notes?
Ben
Ben Goodrich
Graduate Student at Harvard University
Ph.D. Program in Government and Social Policy
-----Original Message-----
From: gov2001-l-bounces(a)lists.fas.harvard.edu [mailto:gov2001-l-
bounces(a)lists.fas.harvard.edu] On Behalf Of Olivia Lau
Sent: Thursday, December 09, 2004 1:08 PM
To: gov2001-l(a)lists.fas.harvard.edu
Subject: RE: [gov2001-l] setx question
If the variance is too narrow, then the confidence intervals won't have
appropriate coverage. So you need to use the difference in predicted
values, not the difference in expected values...
On Thu, 9 Dec 2004, Ben Goodrich wrote:
Yes, with an infinite number of draws, the point
estimates would be the
same
and the variances would be different. But why is
it necessary to go with
the
one with more variance? It isn't based on
the PATE / SATE distinction
because that just has to do with whether the dependent variable is
conditioned on.
Ben
Ben Goodrich
Graduate Student at Harvard University
Ph.D. Program in Government and Social Policy
www.people.fas.harvard.edu/~goodrich/
goodrich(a)fas.harvard.edu
> -----Original Message-----
> From: gov2001-l-bounces(a)lists.fas.harvard.edu [mailto:gov2001-l-
> bounces(a)lists.fas.harvard.edu] On Behalf Of Olivia Lau
> Sent: Thursday, December 09, 2004 12:37 PM
> To: gov2001-l(a)lists.fas.harvard.edu
> Subject: RE: [gov2001-l] setx question
>
> The point of the E(Y(1) - Y(0)) is that the variance of this term
differs
> from E(Y(1)) - E(Y(0)). The point estimate
should be the same.
>
> Olivia.
>
> On Thu, 9 Dec 2004, Ben Goodrich wrote:
>
> > But E "distributes" over addition and subtraction. I remember this
being
> in
> > Rice, although I do not have my copy with me. Irwin has a proof here
on
> the
> > last page:
www.people.fas.harvard.edu/~mirwin/stat110/Lecture/Day05.pdf
> It
> > is only multiplication where things are weird (E[ab] != E[a]E[b]).
> >
> > Either way, how do Gary's notes imply that we are supposed to take a
> draw
> > from a distribution?
> >
> > Ben
> >
> > Ben Goodrich
> > Graduate Student at Harvard University
> > Ph.D. Program in Government and Social Policy
> >
www.people.fas.harvard.edu/~goodrich/
> > goodrich(a)fas.harvard.edu
> >
> >
> > > -----Original Message-----
> > > From: gov2001-l-bounces(a)lists.fas.harvard.edu [mailto:gov2001-l-
> > > bounces(a)lists.fas.harvard.edu] On Behalf Of Olivia Lau
> > > Sent: Thursday, December 09, 2004 11:51 AM
> > > To: gov2001-l(a)lists.fas.harvard.edu
> > > Subject: Re: [gov2001-l] setx question
> > >
> > > The expectation is outside Y(1) - Y(0). So you want the expected
> value of
> > > the difference, not the difference of the expectations.
> > >
> > > ----- Original Message -----
> > > From: "Ben Goodrich" <goodrich(a)fas.harvard.edu>
> > > To: <gov2001-l(a)lists.fas.harvard.edu>
> > > Sent: Thursday, December 09, 2004 11:48 AM
> > > Subject: RE: [gov2001-l] setx question
> > >
> > >
> > > >> And for either model, you should be using the predicted values
to
> > > >> calculate the
ATE's.
> > > >
> > > > How does this square with the formula for ATE in Gary's notes
where
> > > >
> > > > ATE = (1/n)\Sigma E(Y(1) - Y(0))
> > > >
> > > > To me, that implies that we should be looking at the mean over
the
> > > > difference in expected values
rather than the mean over the
> predicted
> > > > values?
> > > >
> > > > Ben
> > > >
> > > > Ben Goodrich
> > > > Graduate Student at Harvard University
> > > > Ph.D. Program in Government and Social Policy
> > > >
www.people.fas.harvard.edu/~goodrich/
> > > > goodrich(a)fas.harvard.edu
> > > >
> > > >
> > > >> -----Original Message-----
> > > >> From: gov2001-l-bounces(a)lists.fas.harvard.edu [mailto:gov2001-
l-
> > > >>
bounces(a)lists.fas.harvard.edu] On Behalf Of Olivia Lau
> > > >> Sent: Thursday, December 09, 2004 11:29 AM
> > > >> To: gov2001-l(a)lists.fas.harvard.edu
> > > >> Subject: Re: [gov2001-l] setx question
> > > >>
> > > >> And for either model, you should be using the predicted values
to
> > > >> calculate
> > > >> the ATE's.
> > > >>
> > > >> ----- Original Message -----
> > > >> From: "Olivia Lau" <olau(a)fas.harvard.edu>
> > > >> To: <gov2001-l(a)lists.fas.harvard.edu>
> > > >> Sent: Thursday, December 09, 2004 11:17 AM
> > > >> Subject: Re: [gov2001-l] setx question
> > > >>
> > > >>
> > > >> > that's a great question, Mike.
> > > >> >
> > > >> > So here's the deal: Usually, the expected values will
differ
> across
> > > >> > columns only because the X observations are different. In
this
> case,
> > > >> the
> > > >> > X observations are all the same (intercept = 1 and treat =
1,
for
> > > >> > example), so each
column looks the same. The ev differ
across
> rows
> > > >> > because of the differences in the simulated betas.
> > > >> >
> > > >> > The predicted values are draws from the distribution defined
by
> the
> > > >> > corresponding value in the expected value matrix (and so
should
> be
> > > >> > different across rows and columns).
> > > >> >
> > > >> > ----- Original Message -----
> > > >> > From: "Michael Richard Kellermann"
<kellerm(a)fas.harvard.edu>
> > > >> > To: <gov2001-l(a)lists.fas.harvard.edu>
> > > >> > Sent: Thursday, December 09, 2004 11:05 AM
> > > >> > Subject: Re: [gov2001-l] setx question
> > > >> >
> > > >> >
> > > >> >>
> > > >> >> Here is the reason for my question: when I look at the
expected
> > > >> >> values
> > > >> >> generated by the sim command, using a normal model, they
are
the
> > > same
> > > >> in
> > > >> >> the univariate case:
> > > >> >>
> > > >> >> s.out.0$qi$ev[1:8, 1:8]
> > > >> >>
> > > >> >> [,1] [,2] [,3] [,4] [,5] [,6] [,7]
[,8]
> > > >> >> [1,] 0.4486 0.4486 0.4486 0.4486 0.4486 0.4486 0.4486
0.4486
> > > >> >> [2,] 0.4424 0.4424 0.4424 0.4424 0.4424 0.4424 0.4424
0.4424
> > > >> >> [3,] 0.4392 0.4392 0.4392 0.4392 0.4392 0.4392 0.4392
0.4392
> > > >> >> [4,] 0.4425 0.4425 0.4425 0.4425 0.4425 0.4425 0.4425
0.4425
> > > >> >> [5,] 0.4411 0.4411 0.4411 0.4411 0.4411 0.4411 0.4411
0.4411
> > > >> >> [6,] 0.4485 0.4485 0.4485 0.4485 0.4485 0.4485 0.4485
0.4485
> > > >> >> [7,] 0.4477 0.4477 0.4477 0.4477 0.4477 0.4477 0.4477
0.4477
> > > >> >> [8,] 0.4485 0.4485 0.4485 0.4485 0.4485 0.4485 0.4485
0.4485
> > > >> >>
> > > >> >> and then they are different for the predicted values:
> > > >> >>
> > > >> >> s.out.0$qi$pr[1:8, 1:8]
> > > >> >> [,1] [,2] [,3] [,4] [,5]
[,6]
[,7]
> > > >> [,8]
> > > >> >> [1,] 0.26540 0.69574 -0.30079 0.7086 0.095671
0.56353
> 0.5045
> > > >> 0.7559
> > > >> >> [2,] -0.29949 0.79896 1.81914 -0.2224 0.149480
0.07857
> 0.9538 -
> > > >> 0.7112
> > > >> >> [3,] -0.24355 0.54188 0.47502 1.1169 0.675997
0.88817
> 0.1099 -
> > > >> 0.1553
> > > >> >> [4,] -0.02701 -0.28023 0.75768 -0.2701 0.296251
0.86762
> 0.3645
> > > >> 0.5838
> > > >> >> [5,] 0.27580 0.46741 0.36535 0.4150 1.234477
0.40665
> 0.5650
> > > >> 0.9518
> > > >> >> [6,] 0.33215 0.81140 0.39242 0.7009 0.282338
0.97523
> 0.7354
> > > >> 0.1146
> > > >> >> [7,] -0.29423 0.09349 0.01257 1.0007 0.721501
0.12567
> 0.3155
> > > >> 0.0861
> > > >> >> [8,] 0.24989 0.28098 1.38814 0.6617 -0.007151
-0.19204
> 0.6433
> > > >> 0.5431
> > > >> >>
> > > >> >> So I guess my question more precisely is, does sim take
one
set
> of
> > > >> draws
> > > >> >> from the sampling distribution of the betas and reuse it
for
all
> of
> > > >> >> the
> > > >> >> sets of starting values?
> > > >> >>
> > > >> >> Also, for the normal model, should we be using the
predicted
> values
> > > or
> > > >> >> the
> > > >> >> expected values to calculate the ATE and ATT?
> > > >> >>
> > > >> >> Cheers,
> > > >> >> Mike
> > > >> >>
> > > >> >> On Thu, 9 Dec 2004, Olivia Lau wrote:
> > > >> >>
> > > >> >>> There's nothing random about setx(). The
randomness comes
in
> > > sim().
> > > >> >>> And my
> > > >> >>> understanding of how R handles seeds is that it
picks up
where
> it
> > > >> >>> left
> > > >> >>> off.
> > > >> >>> Hence,
> > > >> >>>
> > > >> >>> a <- rnorm(10)
> > > >> >>> b <- rnorm(10)
> > > >> >>> identical(a, b)
> > > >> >>>
> > > >> >>> should be false...
> > > >> >>>
> > > >> >>> ----- Original Message -----
> > > >> >>> From: "Michael Richard Kellermann"
<kellerm(a)fas.harvard.edu>
> > > >> >>> To:
<gov2001-l(a)lists.fas.harvard.edu>
> > > >> >>> Sent: Thursday, December 09, 2004 10:39 AM
> > > >> >>> Subject: Re: [gov2001-l] setx question
> > > >> >>>
> > > >> >>>
> > > >> >>> >
> > > >> >>> > Does Zelig start with the same random number
seed for
each
> vector
> > > >> >>> > of
> > > >> x
> > > >> >>> > values created by the setx command?
> > > >> >>> >
> > > >> >>> > Cheers,
> > > >> >>> > Mike
> > > >> >>> >
> > > >> >>> > On Thu, 9 Dec 2004, Olivia Lau wrote:
> > > >> >>> >
> > > >> >>> >> If you can't get Zelig to do what you
want (right now,
> before
> > > >> section
> > > >> >>> >> this
> > > >> >>> >> evening), I suggest that you manhandle the
output to get
it
> to
> > > >> >>> >> produce
> > > >> >>> >> what
> > > >> >>> >> you want...just a thought.
> > > >> >>> >>
> > > >> >>> >> ----- Original Message -----
> > > >> >>> >> From: "Andrew Eggers"
<aeggers(a)gmail.com>
> > > >> >>> >> To:
<gov2001-l(a)lists.fas.harvard.edu>
> > > >> >>> >> Sent: Thursday, December 09, 2004 10:23 AM
> > > >> >>> >> Subject: [gov2001-l] setx question
> > > >> >>> >>
> > > >> >>> >>
> > > >> >>> >> > Can anyone help me understand why my
setx function
isn't
> doing
> > > >> what
> > > >> >>> >> > I
> > > >> >>> >> > thought it would?
> > > >> >>> >> >
> > > >> >>> >> > I'm doing part 2a.
> > > >> >>> >> >
> > > >> >>> >> >> x.out0<- setx(z.out, fn= NULL,
phone =0)
> > > >> >>> >> >
> > > >> >>> >> >> x.out0[1:10,]
> > > >> >>> >> > (Intercept) phone
> > > >> >>> >> > 1 1 1
> > > >> >>> >> > 2 1 0
> > > >> >>> >> > 3 1 1
> > > >> >>> >> > 4 1 0
> > > >> >>> >> > 5 1 1
> > > >> >>> >> > 6 1 1
> > > >> >>> >> > 7 1 1
> > > >> >>> >> > 8 1 1
> > > >> >>> >> > 9 1 1
> > > >> >>> >> > 10 1 1
> > > >> >>> >> >
> > > >> >>> >> > I thought this would produce all
0's in the phone
column
> of
> > > >> x.out0,
> > > >> >>> >> > but clearly it didn't.
> > > >> >>> >> >
> > > >> >>> >> > Thank you!
> > > >> >>> >> >
> > > >> >>> >> > Andy
> > > >> >>> >> >
_______________________________________________
> > > >> >>> >> > gov2001-l mailing list
> > > >> >>> >> > gov2001-l(a)lists.fas.harvard.edu
> > > >> >>> >> >
http://lists.fas.harvard.edu/mailman/listinfo/gov2001- l
> >> >>> >> >
> >> >>> >>
> >> >>> >> _______________________________________________
> >> >>> >> gov2001-l mailing list
> >> >>> >> gov2001-l(a)lists.fas.harvard.edu
> >> >>> >>
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >> >>> >>
> >> >>> > _______________________________________________
> >> >>> > gov2001-l mailing list
> >> >>> > gov2001-l(a)lists.fas.harvard.edu
> >> >>> >
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >> >>> >
> >> >>>
> >> >>> _______________________________________________
> >> >>> gov2001-l mailing list
> >> >>> gov2001-l(a)lists.fas.harvard.edu
> >> >>>
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >> >>>
> >> >> _______________________________________________
> >> >> gov2001-l mailing list
> >> >> gov2001-l(a)lists.fas.harvard.edu
> >> >>
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >> >>
> >> >
> >> > _______________________________________________
> >> > gov2001-l mailing list
> >> > gov2001-l(a)lists.fas.harvard.edu
> >> >
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >> >
> >>
> >> _______________________________________________
> >> gov2001-l mailing list
> >> gov2001-l(a)lists.fas.harvard.edu
> >>
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >
> >
> > _______________________________________________
> > gov2001-l mailing list
> > gov2001-l(a)lists.fas.harvard.edu
> >
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
> >
>
> _______________________________________________
> gov2001-l mailing list
> gov2001-l(a)lists.fas.harvard.edu
>
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l