# Dropping outliers based on “2.5 times the RMSE”

In Kahneman and Deaton (2010)$$^\dagger$$, the authors write the following:

This regression explains 37% of the variance, with a root mean square error (RMSE) of 0.67852. To eliminate outliers and implausible income reports, we dropped observations in which the absolute value of the difference between log income and its prediction exceeded 2.5 times the RMSE.

Is this common practice? What is the intuition behind doing so? It seems somewhat strange to define an outlier based upon a model which may not be well-specified in the first place. Shouldn't the determination of outliers be based on some theoretical grounds for what constitutes a plausible value, rather than how well your model predicts the real values?

$$\dagger$$: Daniel Kahneman, Angus Deaton (2010): High income improves evaluation of life but not emotional well-being. Proceedings of the National Academy of Sciences Sep 2010, 107 (38) 16489-16493; DOI: 10.1073/pnas.1011492107

• When you give a quote from a paper, always give a reference that includes the page number. – Reinstate Monica Jul 12 '19 at 1:26
• I can't say whether this is 'common practice', but I hope not. Automated removals of 'outliers' is fundamentally a bad idea. Maybe your model or removal criterion is not good, maybe there's something new going on (downturn beginning, fresh possibilities awakening) that you shouldn't ignore. // It's different if you can track a suspicious value to data entry error or equipment failure, or if the value is simply off-the-charts absurd (16'2" tall man, guy w/ 61 billable hours last Tuesday, 25min flight SFO-ORD). But not because it doesn't fit a model. I know a startup that went broke that way. – BruceET Jul 12 '19 at 1:38
• The statistical validity of this approach is reflected by the absurd number of decimals they report for the RMSE. – Frans Rodenburg Jul 12 '19 at 3:55
• This feels like a crude / heroic assumption solution to a question I asked a few months ago: stats.stackexchange.com/questions/390051/… – Adrian Jul 13 '19 at 0:37