Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

RE: st: Dropping the largest and smallest 1% of observations

From   "Nick Cox" <>
To   <>
Subject   RE: st: Dropping the largest and smallest 1% of observations
Date   Thu, 13 Feb 2003 12:00:40 -0000

> > I want to cut off extraordinarily large and small values 
> from variables
> > used in regression analysis.
> > What is the easiest way to drop the largest and smallest 
> 1% of observations
> > from variables in STATA-7?

Ronan Conroy

> These values are potentially very informative. You can try 
> other approaches
> such as 
> - median regression
> - intreg 
> - robust regression
> Very large and very small values can indicate problems with 
> measurement.
> -intreg- can be used to specify that these values are not 
> known precisely
> but are bigger/smaller than some threshold.
> Robust regression is useful to confirm that substantive 
> conclusions from
> your analysis are not being 'driven' by influential observations.
> I hate discarding data. These strange values are trying to tell us
> something. We ignore them at our peril. I am analysing some 
> microbiology
> data at the moment. There is a tradition of discarding any 
> measurements
> where there were so many bugs that the plate was 
> unreadable. You can imagine
> the havoc that this has played with results!

I'd echo this strongly. Two other points: 

1. identifying points extreme within univariate 
distributions is not guaranteed to identify 
multivariate outliers. -hadimvo- is one command 
in this area. 

2. tagging outliers and comparing results 
with and without is another relatively simple strategy. 

*   For searches and help try:

© Copyright 1996–2019 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index