16 Dec
2004
16 Dec
'04
9:33 a.m.
On plots of DFBETAs: How do we decide what deserves a plot, and what
does not? Seems like we could just do a DFBETA plot matrix and plot the
effect of each independent variable on the other. Is this shotgun
approach correct, or is there a more systematic way of approaching the
problem?
Wearily,
Andy
16 Dec
16 Dec
10:04 a.m.
Hi Andy-
You'll want to look at the dfbetas associated with independent variable.
There are two ways to contstruct these. One version of the plot simply
contains the dfbetas for one regression coefficient. (an "index plot") The
other plot contains dfbetas for one coefficient on the x axis, dfbetas for another
coefficient on the y axis. The latter type of plot contains the same
dfbeta values but does allow you to make half the number of graphs.
I would recommend plotting dfbetas for each of the regression coefficients
for this problem. Each of these various types of outlier diagnostics is
getting at something slightly different...
Alison
On Thu, 16 Dec 2004, Andy Harris wrote:
On plots of DFBETAs: How do we decide what deserves a
plot, and what
does not? Seems like we could just do a DFBETA plot matrix and plot the
effect of each independent variable on the other. Is this shotgun
approach correct, or is there a more systematic way of approaching the
problem?
Wearily,
Andy
_______________________________________________
gov1000-list mailing list
gov1000-list(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov1000-list
10:13 a.m.
Hi Andy,
Several options here. In what follows I'm assuming that the output
from my fitted model is in lm.out. and that I've created a matrix of
DFBETAS called using
dfb <- dfbetas(lm.out)
The first option is to construct k index plots (one for each column of
dfb) where k is the number of estimated coefficients:
plot(dfb[,1])
plot(dfb[,2])
...
plot(dfb[,k])
This will tell you which observation numbers tend to exert a large
influence on each coefficient. With a lot of data points you may want
to identify points interactively with identify()
You could also look at all pairwise scatterplots (either by hand, or
as you suggest with a scatterplot matrix). The scatterplot matrix is
much easier to do (1 line of code) but the identify function doesn't
work with the R pairs() function. Doing the plots by hand is a lot
more work but it does allow you to use identify(). The bivariate plots
give you information about which observations have a large joint
influence on 2 coefficients of interest.
With a lot of covariates it is typically easiest to look at index
plots.
Hope this helps.
Best,
Kevin
------------------------------------------------------
Kevin Quinn
Assistant Professor
Department of Government and
Center for Basic Research in the Social Sciences
34 Kirkland Street
Harvard University
Cambridge, MA 02138
On Thu, 16 Dec 2004, Andy Harris wrote:
On plots of DFBETAs: How do we decide what deserves a
plot, and what
does not? Seems like we could just do a DFBETA plot matrix and plot the
effect of each independent variable on the other. Is this shotgun
approach correct, or is there a more systematic way of approaching the
problem?
Wearily,
Andy
_______________________________________________
gov1000-list mailing list
gov1000-list(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov1000-list
12:21 p.m.
Hi Kevin,
The index plots make sense, but how should be systematically evaluate
the points? In each plot (there will be 10 on them, including the
intercept), there are over 50 points over the benchmark. We don't get
an easily useable output of those points that we can compare across all
10 plots, so, as far as I can tell, the only way to do that is to copy
the CL output that we get after using identify, coercing it into a
vector after inserting commas between each of the identified points. Is
there an efficient way to do this? Or am I trying to be too systematic?
Do we just need to show that there is a good amount of high leverage
points?
andy
On Dec 16, 2004, at 10:13 AM, Kevin Quinn wrote:
Hi Andy,
Several options here. In what follows I'm assuming that the output
from my fitted model is in lm.out. and that I've created a matrix of
DFBETAS called using
dfb <- dfbetas(lm.out)
The first option is to construct k index plots (one for each column of
dfb) where k is the number of estimated coefficients:
plot(dfb[,1])
plot(dfb[,2])
...
plot(dfb[,k])
This will tell you which observation numbers tend to exert a large
influence on each coefficient. With a lot of data points you may want
to identify points interactively with identify()
You could also look at all pairwise scatterplots (either by hand, or
as you suggest with a scatterplot matrix). The scatterplot matrix is
much easier to do (1 line of code) but the identify function doesn't
work with the R pairs() function. Doing the plots by hand is a lot
more work but it does allow you to use identify(). The bivariate plots
give you information about which observations have a large joint
influence on 2 coefficients of interest.
With a lot of covariates it is typically easiest to look at index
plots.
Hope this helps.
Best,
Kevin
------------------------------------------------------
Kevin Quinn
Assistant Professor
Department of Government and
Center for Basic Research in the Social Sciences
34 Kirkland Street
Harvard University
Cambridge, MA 02138
On Thu, 16 Dec 2004, Andy Harris wrote:
On plots of DFBETAs: How do we decide what
deserves a plot, and what
does not? Seems like we could just do a DFBETA plot matrix and plot
the
effect of each independent variable on the other. Is this shotgun
approach correct, or is there a more systematic way of approaching the
problem?
Wearily,
Andy
_______________________________________________
gov1000-list mailing list
gov1000-list(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov1000-list
1:18 p.m.
Andy,
I think the point with the dfbeta plots is just to say that there are a lot of
outlying observations/leverage points being shown, which essentially
demonstrates that the dfbeta is not such a good diagnostic tool. That is the
whole point of 1(b) I think ...just that these diagnostics show us that there
are a lot of unusual observations that are likely to be causing problems for
our OLS coefficients. The existence of outliers, leverage points and also the
fact that we have non-constant error variance is showing us that we should be
suspicious of our OLS estimates.... so we need robust regression (thiis is my
understanding)!
For dfbeta I didn't do the intercept (I don't think thats as important as the
coefficients, I just did index plots of each coefficient). I only identified
several of the wayward points, mentioning that there were more ... I think
thats the main idea.
Amy
Quoting Andy Harris <jaharris(a)fas.harvard.edu>du>:
Hi Kevin,
The index plots make sense, but how should be systematically evaluate
the points? In each plot (there will be 10 on them, including the
intercept), there are over 50 points over the benchmark. We don't get
an easily useable output of those points that we can compare across all
10 plots, so, as far as I can tell, the only way to do that is to copy
the CL output that we get after using identify, coercing it into a
vector after inserting commas between each of the identified points. Is
there an efficient way to do this? Or am I trying to be too systematic?
Do we just need to show that there is a good amount of high leverage
points?
andy
On Dec 16, 2004, at 10:13 AM, Kevin Quinn wrote:
Hi Andy,
Several options here. In what follows I'm assuming that the output
from my fitted model is in lm.out. and that I've created a matrix of
DFBETAS called using
dfb <- dfbetas(lm.out)
The first option is to construct k index plots (one for each column of
dfb) where k is the number of estimated coefficients:
plot(dfb[,1])
plot(dfb[,2])
...
plot(dfb[,k])
This will tell you which observation numbers tend to exert a large
influence on each coefficient. With a lot of data points you may want
to identify points interactively with identify()
You could also look at all pairwise scatterplots (either by hand, or
as you suggest with a scatterplot matrix). The scatterplot matrix is
much easier to do (1 line of code) but the identify function doesn't
work with the R pairs() function. Doing the plots by hand is a lot
more work but it does allow you to use identify(). The bivariate plots
give you information about which observations have a large joint
influence on 2 coefficients of interest.
With a lot of covariates it is typically easiest to look at index
plots.
Hope this helps.
Best,
Kevin
------------------------------------------------------
Kevin Quinn
Assistant Professor
Department of Government and
Center for Basic Research in the Social Sciences
34 Kirkland Street
Harvard University
Cambridge, MA 02138
On Thu, 16 Dec 2004, Andy Harris wrote:
_______________________________________________
gov1000-list mailing list
gov1000-list(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov1000-list
On plots of DFBETAs: How do we decide what
deserves a plot, and what
does not? Seems like we could just do a DFBETA plot matrix and plot
the
effect of each independent variable on the other. Is this shotgun
approach correct, or is there a more systematic way of approaching the
problem?
Wearily,
Andy
_______________________________________________
gov1000-list mailing list
gov1000-list(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov1000-list
1:35 p.m.
Hi Andy,
The index plots make sense, but how should be systematically evaluate
the points? In each plot (there will be 10 on them, including the
intercept), there are over 50 points over the benchmark. We don't get
an easily useable output of those points that we can compare across all
10 plots, so, as far as I can tell, the only way to do that is to copy
the CL output that we get after using identify, coercing it into a
vector after inserting commas between each of the identified points. Is
there an efficient way to do this? Or am I trying to be too systematic?
Yes, there is an easier way to do this. recall that you can select
elements of a matrix using the logical operators in R. for instance,
dfb[dfb[,1]>c, 1]
will select all of the elements in the first column of dfb that are
greater than some constant c.
Hope this helps.
Best,
Kevin
Do we just need to show that there is a good amount of
high leverage
points?
andy
On Dec 16, 2004, at 10:13 AM, Kevin Quinn wrote:
Hi Andy,
Several options here. In what follows I'm assuming that the output
from my fitted model is in lm.out. and that I've created a matrix of
DFBETAS called using
dfb <- dfbetas(lm.out)
The first option is to construct k index plots (one for each column of
dfb) where k is the number of estimated coefficients:
plot(dfb[,1])
plot(dfb[,2])
...
plot(dfb[,k])
This will tell you which observation numbers tend to exert a large
influence on each coefficient. With a lot of data points you may want
to identify points interactively with identify()
You could also look at all pairwise scatterplots (either by hand, or
as you suggest with a scatterplot matrix). The scatterplot matrix is
much easier to do (1 line of code) but the identify function doesn't
work with the R pairs() function. Doing the plots by hand is a lot
more work but it does allow you to use identify(). The bivariate plots
give you information about which observations have a large joint
influence on 2 coefficients of interest.
With a lot of covariates it is typically easiest to look at index
plots.
Hope this helps.
Best,
Kevin
------------------------------------------------------
Kevin Quinn
Assistant Professor
Department of Government and
Center for Basic Research in the Social Sciences
34 Kirkland Street
Harvard University
Cambridge, MA 02138
On Thu, 16 Dec 2004, Andy Harris wrote:
On plots of DFBETAs: How do we decide what
deserves a plot, and what
does not? Seems like we could just do a DFBETA plot matrix and plot
the
effect of each independent variable on the other. Is this shotgun
approach correct, or is there a more systematic way of approaching the
problem?
Wearily,
Andy
_______________________________________________
gov1000-list mailing list
gov1000-list(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov1000-list
7095
days inactive
7095
days old
gov1000-list@lists.fas.harvard.edu
5 comments
4 participants
participants (4)
-
Alison Elizabeth Post -
Amy Louise Catalinac -
Andy Harris -
Kevin Quinn