[R] help interpreting a model summary

zozio32 remy.pascal at gmail.com
Tue Sep 21 15:58:02 CEST 2010 


David Winsemius wrote:
> 
> 
> On Sep 19, 2010, at 5:59 PM, zozio32 wrote:
> 
>>
>> Thanks for you're long answer.
>> I have to say, I am not fully sure of what you're meaning  
>> everywhere. As I
>> said, I am merely following a recipe book, and when things depart  
>> from it I
>> am a bit lost.
>>
>> I'll try to answer to each of your paragraphs:
>> 3:  I was not wanting to include 3-way interactions, but that's the  
>> only
>> way I found to include a 2 way interaction in my piecewise linear  
>> model.
> 
> Yes. That makes sense to me.
> 
>> I could obviously include only angleNoise*reflection, but I thought  
>> that was
>> not very consistent with the fact that reflection variable was split  
>> in 2.
>> I could may be define the point of separation, and then create 2  
>> separate
>> models in the form "lm(weightedDiff ~ angleNoise*reflection)".
> 
> Yes, I see that you did, and that may be helpful in understanding the  
> estimates. The question I would ask you is why you are putting a  
> breakpoint in a physical model. Unless there is a phase change or some  
> discontinuity in effects at that point, I think the breakpoint looks  
> artificial. Is "refelction" somehow physically connected with the  
> breakpoint? Some sort of acoustic timing phenomenom? Or reflected wave  
> effect in a hydraulic model?
> 
> 
 I am generating waves in a wave tank and I have some of it bouncing back
from the tank wall to mess up my measures, and i want to quantify this
effect. "weightedDiff" is a measure of the difference between the target
spectra and what I am measuring.   For that I have generated virtual wave
elevation with different level of reflection and I am analysing the results
of my waves measurement method.
Basically, I can observe a strong curvature in my data "weightedDiff" as a
function of "reflection". Now, I can try to model this curvature by a
"square" term, a linear piece wise linear model or may be a log model in the
form y = a*log(x)+b.    I don't like the "square" model as it will not go to
+inf with reflection -> +inf.
So I've tried the second option with what I thought was not too bad results
...
I think I'll investigate the log option now. i have to say that speaking to
someone definitely clears up my mind on this. And R is not the cup of tea of
people in my department.
I am not really thinking that there is a breaking point, but that the
influence of the "angleNoise" perturbations (level of uncertainty in the
direction of propagation of my waves) is negligible when the reflection get
too high. If I could identify this point, or a king of limit between 2 zones
( one with reflection low enough that "angleNoise" as to be taken into
account, one with reflection so high that "angleNoise" do not matter any
more) that will be helpful.




> 
>> I merely
>> thought that my formulation was a way to combine them together.
>> Basically, i am expecting both parameters to degrade my signal, but  
>> I'll not
>> be surprised if passed a certain level of reflection, having noise  
>> or not in
>> my angles is not really relevant, hence the interaction parameter. The
>> piecewise linear model is a way to take into account the curvature  
>> in the
>> data that I can observe on a straight scatter plot.
>>
>> 1:  Thanks for the first part, i think I can make sens of it. ;)
>> I guess I can ignore this parameter in that case.  By the way, which  
>> type of
>> Anova you refering to: creating a factor with "high" and "low" level  
>> of
>> interation, and fitting the interation between angleNoise and this new
>> factor?
> 
> If you took a model with all of the data and fit first a model without  
> the break point and one with the breakpoint and then looked at the  
> output of anova(model1) and annova(model2) the difference in deviance  
> across the two models is distributed (asymptotically anyway) as a chi- 
> square statistic with the difference in number of degrees of freedom.  
> That's a much better basis for deciding whether the addition of the  
> term is "statistically significant".
> 
yeah, I did that and there is definitely no match for the linear model
without the break or square term, or anything. I just need to model this
curvature somehow to get a decent model here.  As I am using virtual wave
elevation, I have over 600 observations point, so the anova test results are
not ambiguous at all.


>>
>> 2: first, i was mislead by the meaning of this factor. i only  
>> encounter the
>> version were it's "TRUE", not "FALSE" which is the difference.
>> I think I also use "important" in a wrong way. I should have used
>> "significant" instead.
> 
> You had specified a model that that terms with both (reflection >=  
> Break[xMin]) and (reflection < Break[xMin]) and the lm program  threw  
> away all the levels with reflection >= Break[xMin]. If you had only  
> specified the the model with only reflection >= Break[xMin] you would  
> have gotten an identical model as far as predictions were concerned,  
> but the signs would have been reversed for any of the levels with the  
> inequality term in them.
> 
>> After, i have to admit that I am lost  when you're talking about  
>> models with
>> reversed inequality...
>>
>> 4. Not much to say here, i knew they were pointless but the results  
>> from my
>> formulation of the model. I don't thnik there is a need to removed  
>> them no?
> 
> Not really. And I did not mean to say they were meaningless... as long  
> as there is some physical meaning that motivated the exercise. I was  
> chiding you for making constructions that you did not know how to  
> interpret. If this were a consulting session I would have stopped you  
> at an early point and tried to understand what the physical system  
> "looked like", ie., what the measurements represented, and where the  
> current state of knowledge about them stands.
>>
>>
>> -- 
>> View this message in context:
>> http://r.789695.n4.nabble.com/help-interpreting-a-model-summary-tp2546161p2546292.html
>> Sent from the R help mailing list archive at Nabble.com.
>>
>> ______________________________________________
>> R-help at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-help
>> PLEASE do read the posting guide
>> http://www.R-project.org/posting-guide.html
>> and provide commented, minimal, self-contained, reproducible code.
> 
> David Winsemius, MD
> West Hartford, CT
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 
> 

-- 
View this message in context: http://r.789695.n4.nabble.com/help-interpreting-a-model-summary-tp2546161p2548675.html
Sent from the R help mailing list archive at Nabble.com.

More information about the R-help mailing list