Hi there,
I am planning to run some classification models and found that several of my 53 predictors are count variables with a 5-6 extreme outlying values (out of N=1,500)
My questions
1) I tend to eliminate those but would like to know how outliers in predictors can have any effect on the modeling (perhaps affect stabilty, variance, convergence?)
2) Are there systematic approaches how to detect outliers *in count variables*. I find that difficult as e.g. poisson distributions would not enable it to conceptualize outliers at all (as the vlaues run to infinity theoretically).
3) Should I remove or trim the outlying values? And if I should trim, what would be the approach?
4) Am I hysterical and do I make a fuzz about nothing? :D
Best,
Holger
I think U can delete them (5~6), and test the model's classification performance firstly. And secondly, train model using all variables again and see the results. Compare these two results, and discuss their differences or significances in your research background or study problem.
Hi there,
thanks for the comments. I winsorized the variables to the 98% percentile (95% was too harsh). Doing both (raw and winsorized) and comparing the results sounds reasonable.
Best,
Holger
Hello,
As these are only a few values out of 1500 (Depending on the minority class proportion) - I think what the other posters have suggested and what you are trying sounds fine to me.
There are many ways to deal with this issue - in my previous work where the effect of outliers was significant, I used models that are robust to outliers, e.g. a mixture binomial model - which reduces the influence of the outliers on the estimated model parameters -Article Combining Allergen Components Improves the Accuracy of Peanu...
Regards
Hi Fatma,
thanks for your response. If you consider the IQR that means that using boxplots (that depict them via the whiskers) may be a reasonable easy basis for detecting the outliers?
All the best,
Holger
- Badges
- Science topic