I have a SPSS dataset in which I detected some significant outliers. The outliers were detected by boxplot and 5% trimmed mean. How do I deal with these outliers before doing linear regression? Should I remove them altogether or should I replace them with something else?
Hi Sharmistha,
I would run the regression with all the data and check residual plots. The one of interest in this particular case is the Residuals vs Leverage plot: http://data.library.virginia.edu/diagnostic-plots/
https://stats.stackexchange.com/questions/58141/interpreting-plot-lm
If the outliers are influential - high leverage and high residual I would remove them and rerun the regression.
Anyway I would check the differences in the coefficients in the two models (with and without outliers), if they are minor I would keep the all data model, if they are huge I would keep the model with the outliers omitted and report why and how I chose to remove certain data points.
EDIT: if it appears the residuals have a trend perhaps you should investigate non linear relationships as well.
Regards,
Milan
Hi Sharmistha,
I would run the regression with all the data and check residual plots. The one of interest in this particular case is the Residuals vs Leverage plot: http://data.library.virginia.edu/diagnostic-plots/
https://stats.stackexchange.com/questions/58141/interpreting-plot-lm
If the outliers are influential - high leverage and high residual I would remove them and rerun the regression.
Anyway I would check the differences in the coefficients in the two models (with and without outliers), if they are minor I would keep the all data model, if they are huge I would keep the model with the outliers omitted and report why and how I chose to remove certain data points.
EDIT: if it appears the residuals have a trend perhaps you should investigate non linear relationships as well.
Regards,
Milan
Do not deal with outliers. Just accept them as a natural member of your dataset
Article TO DETERMINE SKEWNESS, MEAN AND DEVIATION WITH A NEW APPROAC...
Dear Sharmistha,
I agree with Milan and understand the point made by Guven. All I would add is there are two reasons to remove outliers:
Dealing with outliers has no statistical meaning as for a normally distributed data with expect extreme values of both size of the tails. I suggest you first look how significant is the difference between your 5% trimmed mean and mean. If not significant then go ahead because your extreme values does not influence that much
I have a question: Is there any difference between parametric and non-parametric values to remove outliers? Is it really necessary to remove? Thanks in advance...
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS