Interesting results, how to deal with these interaction effects?

[Peter Flom]

Good morning

I am dealing with a clinical data set; the DV is a measure of blood flow reduction, measured in percent.
The IVs are in three sets: Risk factors (especially hypertension and heart failure, each measured as a dichotomy), demographics, and medicines (especially one called ACEARB). ACEARB is used to treat hypertension and some other similar problems.

My client came to me because he had confusing results - the medicine had negative effects on the DV.

Some exploration revealed the following

A model with HTN, CHF, and ACEARB shows significant and fairly large effects of all three, but with ACEARB in the wrong direction, with R^2 = .18

A model with these three, plus two way interactions between the risk factors and the drugs shows highly sig. effects of the risk factors, but no other sig. results (using type III SS). But the parameter estimates indicate that ACEARB negates about half the effect of CHF, but makes HTN slightly worse.

Those who have read my posts know I am no fan of p-values and sig. levels, but I am wondering how to present these to my client, and to the larger community, which will probably say that, since the interactions are not sig., they should be excluded.

I am also curious if others have seen similar findings

Thanks in advance

Peter

Peter L. Flom, PhD
Statistical Consultant
Website: www DOT peterflomconsulting DOT com
Writing; http://www.associatedcontent.com/user/582880/peter_flom.html
Twitter: @peterflom

[BXC (Bendix Carstensen)]

First: How was the data collected?
If not randomized you may just have confounding by indication??
Bendix Carstensen
_______________________________________________

Bendix Carstensen
Senior Statistician
Steno Diabetes Center
Niels Steensens Vej 2-4
DK-2820 Gentofte
Denmark
+45 44 43 87 38 (direct)
+45 30 75 87 38 (mobile)
b...@steno.dk http://www.biostat.ku.dk/~bxc
www.steno.dk

[Peter Flom]

There have been good responses already, here are some responses

Kevin Spratt suggested that a percentage might need to be transformed - absolutely right,
but I wasn't clear enough, the DV is actually a change in percentage terms, and it ranges - 45 to + 30,
so a transformation isn't needed.

He also suggested CHAID - an excellent idea. I can do this in R (I don't have SAS data miner)

He also asked about the sample size - N = 200. But very few people have certain combinations of the different
risk factors and different drugs. e.g. only 3 people without HTN are taking ACEARB, and very few people have CHF but not HTN.

Bendix Carstenson asked how the data was collected - I am investigating this

Tracy Clegg (off list) suggested the possibility of collinearity - this was my thought, as well, but the condition indexes and VIFs did not indicate a problem.

Thanks to all!

[Brett Magill]

Hi Peter,

Just another question...what are the individual correlations like
among the IV's?

There isn't really any reason to expect a regression/ancova
coefficient in any particular direction unless you can explain the
intercorrelation patterns among all the IV's and understand that it is
the partial IV/DV relationship that is represented by the
coefficients.

Brett

[Jeff Allard]

Out of curiosity - how would CHAID (or any decision tree) address the problem?

2009/10/22 Peter Flom <peterflom...@mindspring.com>

[mcap]

It's a room full of statisticians!!

I think you should step back and consider things clinically for a
minute. Please provide some more detail on your abbreviations...

DV - Ductus Venosus? Distribution volume?

There may be a rather simple clinical explanation for what you are
seeing but we need a little more detail on your dep var. What was it
exactly, and how was it measured? Also, who were these patients?

There may be a variable that wasn't measured that may play a major
role here. But you can look to clinical considerations to justify
your interactions. There are a lot of data on those meds in cardiac
conditions. There are also cases where reduction of blood pressure
can make things worse. Patients with CHF and low blood pressure for
example, don't do as well.

Marc

> > Twitter:   @peterflom- Hide quoted text -
>
> - Show quoted text -

[Peter Flom]

mcap <mca...@yahoo.com> Marc wrote

>
>
>It's a room full of statisticians!!
>

Yeah. It is.

>I think you should step back and consider things clinically for a
>minute. Please provide some more detail on your abbreviations...
>
>DV - Ductus Venosus? Distribution volume?
>

Dependent variable. Which is a measure of change in blood flow. It's a percentage, and ranges from -30 to + 20 or so. The change was after application and removal of a blood pressure cuff. Per my colleague, it is a good thing for there to be large decreases. And ACEARB ought to have good effects. He says it's a well known effect.

The sample was, again per my colleague, a random sample of patients in a major medical center.

>There may be a rather simple clinical explanation for what you are
>seeing but we need a little more detail on your dep var. What was it
>exactly, and how was it measured? Also, who were these patients?
>

See above

>There may be a variable that wasn't measured that may play a major
>role here. But you can look to clinical considerations to justify
>your interactions. There are a lot of data on those meds in cardiac
>conditions. There are also cases where reduction of blood pressure
>can make things worse. Patients with CHF and low blood pressure for
>example, don't do as well.
>

I will keep looking, but so far, things look quite odd, clinically.

Thanks!

[Frank Harrell]

The laws of arithmetic are violated any time you use % change as a
dependent variable. A simple way to see this is that an increase of
100% is balanced by a decrease of 50%. Percent changes may only be
computed on group summary measures, not per-patient measures.

I also find that it is disappointingly uncommon for analysts to check
the assumptions of any change scores (using for example Bland-Altman
plots). Huge assumptions are made about the proper transformations of
the variables.

Also it is mandatory to include the baseline variable as a covariate
any time change is analyzed.

See http://biostat.mc.vanderbilt.edu/ManuscriptChecklist for details.

A far better approach is to use the follow-up value as the dependent
variable. You can even use a semi-parametric approach such as the
proportional odds model so as to not make strong distributional
assumptions.

Frank Harrell

On Oct 22, 10:07 pm, Peter Flom <peterflomconsult...@mindspring.com>
wrote:

[Peter Flom]

Thanks for this, Frank.

I will investigate a model with these adjustments when I am back at that job, on Monday

Peter

[Bruce Weaver]

On Oct 23, 9:26 am, Frank Harrell <f.harr...@vanderbilt.edu> wrote:
> The laws of arithmetic are violated any time you use % change as a
> dependent variable.  A simple way to see this is that an increase of
> 100% is balanced by a decrease of 50%.  Percent changes may only be
> computed on group summary measures, not per-patient measures.

Peter, here are a couple articles that might help you argue Frank's
point about % change with the investigators.

http://www.jstor.org/stable/2983064
http://www.biomedcentral.com/1471-2288/1/6

If you don't have JSTOR access, I can send you the Kronmal article--
just drop me a line.

--
Bruce Weaver
bwe...@lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM."

[Chris Everyman]

Peter, another article about % change:

Törnqvist, L., Vartia, P., & Vartia, Y. O. (1985). How should relative
changes be measured? The American Statistician, Vol. 39 (Iss. 1), pp.
43-46.

Best wishes,

Chris