3
$\begingroup$

I have already removed three outliers from the ANOVA and ANCOVA analysis to improve the models. I want to report the descriptive statistics and p-values.

First I removed these outliers, and then I applied the analysis. My question is, for publication, should I provide the information before or after removing the outliers?

$\endgroup$
  • 8
    $\begingroup$ "removed three outliers [...] to improve the models" - That sounds really sketchy at best. You should only remove outliers if you have strong reasons to believe systematic error in the data-collecting process. If you can't do that, try your hand on robust statistical estimation techniques to lessen the influence of outlier lookalikes. $\endgroup$ – Firebug Sep 24 '17 at 15:02
  • 2
    $\begingroup$ @ Firebug, thanks for the comment. I know the robust model is the better option and I really prefer to use the robust one but my boss forces me to remove the outliers!!! $\endgroup$ – joe Sep 24 '17 at 15:10
  • 3
    $\begingroup$ my boss forces me to remove the outliers. You have two courses or action here then: remove implausible data (data you know couldn't have happened, and you can justify it) and document it as well; and explain to your boss the rest of the outliers aren't really outliers, removing them is bad practice and, under full disclosure of the methods can even make the publication process harder. Here's a good question on the matter so it stays linked: Is it OK to remove outliers from data? $\endgroup$ – Firebug Sep 25 '17 at 1:10
10
$\begingroup$

In publications honesty is the best policy.

It's all right to remove outliers and to report descriptive statistics based on the reduced dataset. However in order for others to understand and appreciate what was done you have to state which samples were removed as outliers and your reasoning for why they were outliers.

If your justification of marking those samples as outliers is solid (for example - instrument malfunctioned when they were recorded) then others reading your paper will agree with you and thank you for removing those samples.

If there might be disputes about those samples being outliers then there is always a possibility of doing the calculations twice: one time with outliers present and another time with outliers removed. Then you can present numbers without outliers in the main text and include the numbers calculated with outliers in the supplement.


However it's hard to ignore one part of your question. You write:

removed three outliers ... to improve the models

Here you have to be extra careful. If a sample disagrees with your model it doesn't mean that the sample is outlier. The situation is like this:

  • there is a real phenomenon
  • you have data about that phenomenon
  • you have some imaginary model that you think explains the phenomenon

Now you find that some samples from the real world don't agree with your imaginary model. Would you say that removing those samples will improve your model? Likely not. What might be happening is your model not being a good representation of the real phenomenon. Then by removing these "outliers" you might be ignoring a part of the real world. And in that case removing them is dangerous.

Here is a nice illustration

$\endgroup$
3
$\begingroup$

It's a matter of taste, or what you think will be most helpful. Just be clear which you're presenting. If the outliers affect these statistics a great deal, particularly if you decided to remove them on that basis, it would make sense to include both.

$\endgroup$

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.