topic Re: Nonlinear regression in Statistical Procedures

Nonlinear regression

ilikesas — Sat, 12 Sep 2015 21:48:31 GMT

Hi,

I am trying to make a linear regression but my data is heteroskedastic, so will have to do a nonlinear regression.

Could you please tell me what sas procedure is good for nonlinear regressions, and which will also find the right model because I have several variables and don't even know if any should be logarithmic or exponential for example.

I tried searching the internet but only got more confused...

Thank you

Re: Nonlinear regression

PGStats — Sun, 13 Sep 2015 02:29:55 GMT

We need a lot more information about your data to suggest an appropriate statistical modeling approach. Nonlinear regression is not the most common way to deal with heteroskedasticity. Generally, if your data looks like this:

i.e. heteroskedastic because the standard deviation of the errors is not constant but proportional to the mean, then a logarithmic transformation of the dependent variable (y) is often a suitable approach.

Re: Nonlinear regression

ilikesas — Wed, 23 Sep 2015 22:01:00 GMT

Hi,

Thnak sfor the reply, I didn't answer for a while becasue I didn't knwo that I ha d areply, the new sas communities site doesn't send an email when there is a reply...

I tried transforming the y variable into log y (among others) but was still getting heteroskedasticity...

Please see what the y vs x graph and residual vs x graphs look like in the attached word doc (couldn't just copy and paste it here)

Thank you!

Re: Nonlinear regression

SteveDenham — Fri, 25 Sep 2015 14:37:44 GMT

I really don't see much evidence for heteroskedasticity, especially in the plot of car3 vs ln_mkt. The residual plot does show some more variablity at low values of the independent variable, but this may be more a function of high influence points.

Consider using either PROC ROBUSTREG or PROC QUANTREG to fit the data.

(On the other hand, I don't see much use of the independent variable in these plots as a predictor. That regression line is nearly as flat as the Texas Panhandle).

Steve Denham