1
$\begingroup$

I have the marker data of 32 patients for eight different markers. WHat needs to be done here is to predict the type of marker which is suitable for the disease control.

I used Disease control as the dependent variable and markers as the independent variable and did a multiple regression analysis. And from the results of which I did not see a linear trend in the data and the below figure gives an idea of how the data looks like and it is very clear that a model cannot be fit using linear methods.

I would like to focus on the non-linear trends in the data.

Will doing a support vector regression help in this case? Iam using R to do the regression.

enter image description here

$\endgroup$
  • $\begingroup$ Why do you want to do a regression method? Is it a requirement? Or do you just want to get something meaningful from the data? Have you tried validating all the assumptions that linear regression makes? If you dont have to use linear regression, it looks to me like you're better off fitting a distribution to your data. Also, what does disease control mean and why use it as the dependent variable? If you used the markers as the independent variable, why are you getting some negative values? $\endgroup$ – markovchain Mar 19 '14 at 14:41
  • $\begingroup$ @markovchain : The aim is to predict a model as to which of the marker is best suited to control a particular disease. So, the regression analysis was done. In this picture,I used a simple linear model, use only one marker at a time just to see how the data looks like. Do you think a non-linear regression would help? $\endgroup$ – user1805343 Mar 19 '14 at 15:07
  • 2
    $\begingroup$ I think you might want to rephrase your question. As stated it seems like your focus is on possible non-linear trends in the data and how linear models may be modified to facilitate these. But your real issue is lack of data and how to avoid overfitting, while at the same time modeling these weak predictors. To my knowledge SVMs do have built in regularization so may be useful, but so would any model with regularization. $\endgroup$ – charles Mar 19 '14 at 17:06
2
$\begingroup$

You could use polynomials or, better yet, splines to address non-linearity. Harrell's RMS package makes this straigh forward and has convenient features to interpret results from splines.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ I've always splines, but fractional polynomials were popular at one stage and I think there is an R package for them. But....you don't have a lot of data, nothing is going to solve that or fit the data well as a result $\endgroup$ – charles Mar 19 '14 at 16:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.