I am the author of -fractileplot-. You should
look at the concurrent thread on -mlowess-.
I doubt that you really want -fractileplot-
unless you know that you do. For arbitrary
predictors you want -mrunning-, -mlowess-
or the much more comprehensive -gam-.
Use -search- or failing that -findit- to
search for locations.
You are right that -fractileplot- (and
also -mrunning- and -mlowess-) do not
address the issue of modelling interactions.
That is not their job at all. The user
should come to each command with a list
of variables. It's the user's choice what
those variables should be, and none of these
programs knows or cares how the variables
processed were defined, and whether some
are original variables and some defined
as interactions.
I doubt that the results of any of these
programs will depend on quite what
scales you use to define interactions.
Finally, note that you should not read
too much into the existence of these programs.
Their authors intend them as exploratory
graphics tools, no less and not much more.
In some ways they are not very Stataish at all.
If you are looking for a rich source of modelling tools in this
terrain, S-Plus and R will provide much happier hunting
for you than Stata.
By the way
multivariate, nonparametric, kernel-based smoothing
is one thing and
curve fitting
is another, at least in my book.
Nick
[email protected]
[email protected] (reminds me of
"No, Groucho is not my real name.
I'm just breaking it in for a friend.") wrote
> after searching statalist and the web, i concluded
> that fractileplot is the only Stata machinery readily
> available for performing multivariate, nonparametric,
> kernel-based smoothing (curve fitting). so my many
> thanks to the author of this extremely useful tool,
> and if anyone is aware of any other relevant Stata
> tools please advise.
>
> my question pertains to how to best use fractileplot
> to capture possible interactive effects among the
> right hand side variables. from reading the help file
> and going through the ado file, it looks to me like
> fractileplot assumes that the effect of each rhs
> variable is strictly additive. please correct me if
> that is not accurate.
>
> i am trying to use fractileplot while allowing for
> interaction effects, and i would welcome comments and
> suggestions as to my proposed back-of-the-envelope
> recipe below. for simplicity, lets consider the
> bivariate case, i.e., what would normally by estimated
> as "fractileplot y x1 x2".
>
> to capture interactive effects, i suggest doing the
> following:
>
> . egen std_x1 = std(x1)
> . egen std_x2 = std(x2)
> . gen x12 = std_x1 * std_x2
> . fractileplot x1 x2 x12
>
> i dont have a well articulated explanation for why i
> standardized x1 and x2 prior to the creation of the
> interactive variable, just a vague notion that this
> would fit best with the whole concept of
> distributional smoothing.
>
> some basic questions and concerns that i have around
> this include:
> - is an interactive term a reasonable approach to
> expanding beyond additive marginals?
> - if so, does the standardization of x1 and x2 prior
> to their interaction make sense or does it carry
> hidden dangers?
> - any other suggestions for how to allow for
> interactive effects?
> - any other easy recipes for multivariate
> beyond-additive nonparametric smoothing using existing
> Stata commands and ado's?
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
