I'm certainly no expert in hypothesis tests, but could a
Kolmogorov-Smirnov test be useful to deal with the below?
Best,
Dan
PS I think the "Matching" package implements the KS test.
On Sun, 16 Apr 2006, Abby Williamson wrote:
Thanks, Gary and Ian for this helpful information.
We'll try to digest this
and follow up if we continue to have trouble - our authors say they are
using "standard p-weights," as we thought we were doing, but our results are
still off.
Also, we are wondering if anyone is aware of a "difference in modes" test
similar to a difference in means. Our paper emphasizes a change in the mode
number of close confidants that respondents named between two waves of the
GSS (from a mode of 3 in 1985 to a mode of 0 in 2004). Those who commented
on our replication wondered whether there was any way to know whether this
change was more than just random. Justin recommended a useful simulation
technique, but before we go down that road, I wanted to check whether any
standard test existed.
Many thanks,
Abby
-----Original Message-----
From: gov2001-l-bounces(a)lists.fas.harvard.edu
[mailto:gov2001-l-bounces@lists.fas.harvard.edu]On Behalf Of Gary King
Sent: Sunday, April 16, 2006 8:31 PM
To: gov2001-l(a)lists.fas.harvard.edu
Subject: Re: [gov2001-l] p-weights
I don't know what p-weights are, but a little known fact is that weights
are almost always used wrong in the literature. if you're interested in a
descriptive quantity (such as the average age in the population), then you
certainly should take the weighted average,
sum(age_i * weight_i)/sum(weight_i). But suppose you're interested in the
causal effect of some variable. Does that mean if you're running a
regression you should use the weight option in whatever stat'l program
you're running? no way. if its regression, then the only reason to
weight is heteroskedasticity, not sampling schemes. you definitely want
to control for the sampling process (the probabilty that unit i made it
into your sample, since if the unexplained part is correlated with Y you
can have bias), but you do this by adding the weights (or the vars that
make up the weights) as control variables in the regression; you don't say
weight=whatever. this is completely counterintuitive of course, but if
you think about what assumption of regression you're violating by not
running WLS instead of LS say i think you'll understand that the answer is
none. yet, if you don't control for the process of getting into the
sample, you can get into big trouble.
Gary
On Sun, 16 Apr 2006, Abby Williamson wrote:
Can anyone recommend a good source of information
on how to use p-weights?
Likewise, if anyone has experience properly using the weights for the 2004
GSS, my paper group would love to hear your insights.
Many thanks,
Abby
-----Original Message-----
From: gov2001-l-bounces(a)lists.fas.harvard.edu
[mailto:gov2001-l-bounces@lists.fas.harvard.edu]On Behalf Of Dan Hopkins
Sent: Sunday, April 16, 2006 7:51 PM
To: gov2001-l(a)lists.fas.harvard.edu
Subject: Re: [gov2001-l] problem (e): y_i =0 when y_i^* >0?
Hi Yuki and Everyone,
This is a good question--and could be defined either way, depending
whether we want the ultimate probability to be given by CDF(y*) or
1-CDF(Y*). With thanks to Ian and Gary, we've updated the slides to be
consistent with "Unifying Political Methodology"--and yes, to put y* on
the left-hand side, switching the observation mechanism from that given in
the problem set.
Best,
Dan
----
Ph.D. Student
Department of Government
Harvard University
Tutor, Currier House
dhopkins(a)fas.harvard.edu
http://www.danhopkins.org
On Sun, 16 Apr 2006, Yuki Takagi wrote:
> Hi all,
>
> The question says that "y_i = 1 when y_i^* > 0 and 0 otherwise." Could
you
tell me
if it is correct?
I feel that it might be "y_i = 0 when y_i^* > 0 and 1 otherwise" rather
than
"y_i = 1 when y_i^* > 0 and 0
otherwise" in order to approximates the
probit
model because the observation mechanism for the
probit model is is y_i =
0
if y_i^* > \tau and 1 otherwise where y_i is
observed realization , y_i^*
is
unobserved variable, and \tau is the threshold
parameter.
Thanks,
Yuki
_______________________________________________
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