2

I did multiple regression analysis.

There are 2 independence variables and one dependence variable.

Because there are heteroscedasticity problem, I did log transport about dependence variable.

All of the variable types are numeric.

""reg3<-lm(log(h3+1)~e3+i3,data=h3.train)"" (my code to fit the data)

I check the residual plots to see the model is well fitted.

I know the basic rules of interpreting residual plot in regression analysis.

But these 3 residual plots below are pretty ambiguous for me.

I want you to help me to interpret the plots and what should I do for next.

Here are the plots.

enter image description here

I think there some rules at the first graph.

I can see both small and big one line in the graph.

I don't think it is randomly distributed.

enter image description here

It looks the dots are randomly distributed.

But I can see there is a limit at the bottom line.

I mean at the under left. I can see there is a line.

enter image description here

I think this is ok. It seems that the residual spots are randomly distributed. But I doubt it could have same problem with the second graph. It is just difference of degrees.

Thank you for your help in advance. Sincerely.

Doramph
  • 21
  • 1
  • Would you please post a link to the raw data? – James Phillips Oct 24 '19 at 10:53
  • 1
    The pattern in the residual plot suggests that you might be using count data for your response variable, in which case there are better models than linear regression. – Ben Oct 24 '19 at 13:09
  • 1
    @Ben Yes, the response variable is count data. I also fitted Poisson regression model. But the prediction accuracy of linear regression prediction is a bit better than Poisson regression. That's why I try to use linear regression. And It would be my pleasure if you teach me what kind of test I have to do after fitting Poisson regression model. I just know to check overdispersion about it. – Doramph Oct 25 '19 at 00:22
  • 1
    @JamesPhillips I am having trouble to upload raw data. I don't know how to do that. Let me figure it out. – Doramph Oct 25 '19 at 00:49
  • @Doramph: Poisson regression sucks; you should almost never use it (see [here](https://stats.stackexchange.com/questions/392591/)). I would suggest you start with a negative binomial regression, which allows both the mean and variance to be fit to the data. This model often gives a good fit to count data. – Ben Oct 25 '19 at 03:44
  • @Ben Thanks for your advice. I will try a negative binomial regression. – Doramph Oct 25 '19 at 06:31

0 Answers0