Let's say I perform multiple regression where y = income, x1 = educaiton, x2 = sex, and x3 = religion from 2003 to 2018, where the data is measured daily.

Is there any way to quantify an impact of a single day data (e.x. 2005-07-01) on the regression result?

  • 2
    $\begingroup$ If you want to quantify an impact of a single day data you can create a boolean feature that will be false for all days except you chosen one. $\endgroup$ Jul 19 '18 at 5:39
  • $\begingroup$ Maybe stating the obvious - this new feature would need to be included in the regression. $\endgroup$
    – bradS
    Jul 19 '18 at 8:15

It is hard to quantify exactly what a single point is doing, but you would benefit from plotting data to look for suspicious points and then measuring differences in statistics if you remove it.

Don't forget about the value of plotting data! Anscombe's quartet is a great reminder of how important this is. https://en.wikipedia.org/wiki/Anscombe%27s_quartet

enter image description here


I suppose you are trying to understand those data points in your sample that could impact your regression results to a large extent. These are called influential and high leverage observations. It is a practice to ensure that you do not have these sort of observations in your training sample as it would impact the final regression result significantly. The way to identify these observations is through the means of Cook's distance. You would need to further analyze those data points that high a high Cook's distance.

  • $\begingroup$ Hi, I know that people usually remove the outliers and high leverage observations from the dataset. However, in my dataset, those points actually have meanings and I do not want to remove them. Instead, I want to perform regression and identify those data points that have huge impact on the regression result. $\endgroup$
    – Jun Jang
    Jul 19 '18 at 11:56
  • $\begingroup$ What do you mean when you say ‘impact’? Do you mean how much the estimates change with and without that data point? $\endgroup$
    – Srikrishna
    Jul 19 '18 at 12:08
  • $\begingroup$ I dont have the precise definition of impact, but yes that would be one of the definitions. And actually, I am performing constrained regression, with coefficients being between 0 and 1, and the sum of the coefficients being 1. So I don't think I can use the usual Cook's distance. $\endgroup$
    – Jun Jang
    Jul 19 '18 at 12:19

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.