So many available options for regression . I am confused !!!!!!!!!!!

Hello All

Kindly have a look at the link below.. http://www.mathworks.in/help/stats/introduction-to-parametric-regression-analysis.html#btb52kr

It provides so various types of regression analysis. Can i know the internal working of all these ? Parametric / Supervised ???

How to choose suitable one for my work ?

In my case, I have entirely no idea regarding the relation between the regressor and response variables. (Linear or Polynomial ). M work has started from scratch with no previously available literature on my problem.

Which analysis will suit the best for my dataset ?

PLEASE HELP......

[Muthu Annamalai]

It would be nice to post questions without being dramatic; just saying.

Answer 1

[Image Analyst]

What does your data look like? Can you provide a sample of it to us? Then, is there some underlying theoretical model that you expect your noisy actual data to follow? Like there is a linear or quadratic relationship between some sets of data?

[Priya]

thanks a lot for your reply.. There is no previously stated relationship between my dataset. Being a beginner I am totally confused how to kick start from basic to complex???

I am extremely sorry I would not able to reveal the dataset, but I can definitely tell you the characteristics you of dataset you want to know.

Thanks a lot

[Image Analyst]

For example do you have to continuous variables, like time vs. tempaerature? Or do you have categorical variables, like names of cities, types of fruits, etc.? You can change your actual data or call it something other than what it is to keep your secrecy but you have to share something or else we have nothing to go on. How can we tell you a good regression method to use if we know absolutely nothing about your data?

[Priya]

Ok I got it..I am providing Sample dataset with 800 numerical entries for both Y and X variables. X variables are normalized values between 0 and 1, Y variables are logarithmic values.

Can i somehow attach excel sheet ?

[Image Analyst]

You can do

plot(X, Y);

and then hit alt-printscreen to copy the figure into the clipboard. Then go to http://snag.gy and type control-v to paste it in (upload it). Then tell us the URL it reports back so we can see your curve.

[Priya]

Thanks a lot.

Scatter Plot - http://snag.gy/mtpVj.jpg

Line Plot - http://snag.gy/Zwhf8.jpg

[Image Analyst]

And what sort of curve do you think should go through that? It looks nearly uncorrelated - almost just a bunch of random points. Do you think the points are tight enough to fit a line or quadratic through them? Sure you could, but would it be a useful/valuable estimate/prediction, or would it be essentially worthless because there's so much inherent randomness in your data?

[Priya]

Yeah !!!

If I somehow classify the dataset into different categories such that each class has approx 150 to 200 data points... If I make combination of such classes to further reduce the dataset in each class.

Any idea how shall I make fruitful combinations ?

[Image Analyst]

So you have a bunch of x and a bunch of y. Each y is associated with one x. Now what? Do you have a new x value and you'd like to predict a y value for that new x value? So that you want a model equation like y = m*x+b? I don't really know what you want to get from that data once you've measured it.

[Priya]

Actually I have around 600 X values. I know 790 is a random data and x and y may not fit into possible relationship.

However, I can classify data into catagories like A, B , C , D or double combinations like AB AC AD BC BD or Triple combinations like ABC, ABD, BCD to reduce the dataset number to approx 200 , 100, 50 respectively.

For each such classification I can perform simple and multiple linear regression analysis for 600 X variables.

However, i need to filter out best features among these 600 to perform multiple regression analysis.

Will this approach help me. Can you guide me to some fruitful direction ?

[Image Analyst]

So you don't want to preform prediction/estimation from a model, you want to classify your data set into 4 classes or more. Some of your data may belong to more than one class. But to do that you need to know what constitutes a class. From what I can see there are no clusters in your data so I'm not really sure what class A, B, C, or C would be. Do you have any training data that is already classfied into some known "ground truth" classes? Like an x of 30 and a y of 123 is known to be in class A?

[Priya]

Yeah. I can classify data according to 4 classification criteria, which are reasonable and make sense.

Regarding Training data, yes I can get training data from previous literature, but the dataset will be different and small.

Can't we make sense out of this data in any way. Because, generally most of the large datasets are quite random, out of which people have made some sense out of it by classification or maybe some other techniques.