Oversampling in logistic regression

[sachin01663]

Hi ,

Can someone please suggest when we do oversampling ,do we do this on main dataset and then we spilt it into Training and Validation (so both are over sampled)? Or we first make Training and Validation datasets nd then oversample Training dataset and perform validation on non-oversampled Validation dataset? Thanks.

[PGStats]

Oversampling for what purpose? Bootstrap?

[sachin01663]

Hi,

I am working on a data with very low event rate around 0.8% however the observations are quite high around 2 million. I am thinking to take 50-50 events and non events to build a better logistic regression model. Usually when I do logistic regression, I split my data into validation and training datasets. Build model on Training and validate on validation.

However in this case, where I am taking lower proportion of non events, I am not sure how to split Training and Validation datasets. Do I do oversampling/undersampling before split data into Training and Validation or I only do this on Training dataset and Validation data stays as it is?

I hope it is clear.
Thanks Sachin

[PGStats · Accepted solution]

I don't see the need for unequal probability sub-sampling for the validation data. By keeping the original event probability in the validation dataset you could also check that your weighting in the training data was done properly.

Note : I edited the topic title to attract attention from others to this question.

[sachin01663]

Okay. I was wondering if the event rate is too low to build model on Training dataset (hence we did sampling), is it sufficient to do validation without oversampling ?

Thanks for editing the title.