Forty years later, it's still hard to improve on William Kruskal's February1960 Technometrics paper, "Some Remarks on Wild Observations". Since permission was granted to reproduce it in whole or in part, here it is in its entirety.
Some Remarks on Wild Observations *
William H. Kruskal**
The University of Chicago
* This work was sponsored by the Army, Navy and Air Force through the Joint Services Advisory Committee for Research Groups in Applied Mathematics and Statistics by Contract No. N6qri4)2035. Reproduction in whole or in part is permitted for any purpose of the United States Government.
** With generous suggestions from LJ Savage, HV Roberts, KA Browalee, and F Mosteller.
Editor's Note: At the 1959 meetings of the American Statistical Association held in Washington D.C., Messrs. F. J. Anscombe and C. Daniel presented papers on the detection and rejection of 'outliers', that is, observations thought to be maverick or unusual. These papers and their discussion will appear in the next issue of Technometrics. The following comments of Dr. Kruskal are another indication of the present interest of statisticians in this important problem.
The purpose of these remarks is to set down some non-technical thoughts on apparently wild or outlying observations. These thoughts are by no means novel, but do not seem to have been gathered in one convenient place.
1. Whatever use is or is not made of apparently wild observations in a statistical analysis, it is very important to say something about such observations in any but the most summary report. At least a statement of how many observations were excluded from the formal analysis, and why, should be given. It is much better to state their values and to do alternative analyses using all or some of them.
2. However, it is a dangerous oversimplification to discuss apparently wild observations in terms of inclusion in, or exclusion from, a more or less conventional formal analysis. An apparently wild (or otherwise anomalous) observation is a signal that says: "Here is something from which we may learn a lesson, perhaps of a kind not anticipated beforehand, and perhaps more important than the main object of the study." Examples of such serendipity have been frequently discussed--one of the most popular is Fleming's recognition of the virtue of penicillium.
3. Suppose that an apparently wild observation is really known to have come from an anomalous (and perhaps infrequent) causal pattern. Should we include or exclude it in our formal statistics? Should we perhaps change the structure of our formal statistics?
Much depends on what we are after and the nature of our material. For example, suppose that the observations are five determinations of the percent of chemical A in a mixture, and that one of the observations is badly out of line. A check of equipment shows that the out of line observation stemmed from an equipment miscalibration that was present only for the one observation.
If the magnitude of the miscalibration is known, we can probably correct for it; but suppose it is not known? If the goal of the experiment is only that of estimating the per cent of A in the mixture, it would be very natural simply to omit the wild observation. If the goal of the experiment is mainly, or even partly, that of investigating the method of measuring the per cent of A (say in anticipation of setting up a routine procedure to be based on one measurement per batch), then it may be very important to keep the wild observation in. Clearly, in this latter instance, the wild observation tells us something about the frequency and magnitude of serious errors in the method. The kind of lesson mentioned in 2 above often refers to methods of sampling, measurement, and data reduction, instead of to the underlying physical phenomenon.
The mode of formal analysis, with a known anomalous observation kept in, should often be different from a traditional means-and-standard deviations analysis, and it might well be divided into several parts. In the above very simple example, we might come out with at least two summaries: (1) the mean of the four good observations, perhaps with a plus-or-minus attached, as an estimate of the per cent of A in the particular batch of mixture at hand, and (2) a statement that serious calibration shifts are not unlikely and should be investigated further. In other situations, nonparametric methods might be useful. In still others, analyses that suppose the observations come from a mixture of two populations may be appropriate.
The sort of distinction mentioned above has arisen in connection with military equipment. Suppose that 50 bombs are dropped at a target, that a few go wildly astray, that the fins of these wild bombs are observed to have come loose in flight, and that their wildness is unquestionably the result of loose fins. If we are concerned with the accuracy of the whole bombing system, we certainly should not forget these wild bombs. But if our interest is in the accuracy of the bombsight, the wild bombs are irrelevant.
4. It may be useful to classify different degrees of knowledge about an apparently wild observation in the following way:
a. We may know, even before an observation, that it is likely to be wild, or at any rate that it will be the consequence of a variant causal pattern. For example, we may see the bomb's fins tear loose before it has fallen very far from the plane. Or we may know that a delicate measuring instrument has been jarred during its use.
b. We may be able to know, after an observation is observed to be apparently outlying, that it was the result of a variant causal pattern. For example, we may check a laboratory notebook and see that some procedure was poorly carried out, or we may ask the bombardier whether he remembers a particular bomb's wobbling badly in flight. The great danger here, of course, is that it is easy after the fact to bias one's memory or approach, knowing that the observation seemed wild. In complex measurement situations we may often find something a bit out of line for almost any observation.
c. There may be no evidence of a variant causal pattern aside from the observations themselves. This is perhaps the most difficult case, and the one that has given rise to various rules of thumb for rejecting observations.
Like most empirical classifications, this one is not perfectly sharp. Some cases, for example, may lie between b and c. Nevertheless, I feel that it is a useful trichotomy.
5. In case c above, I know of no satisfactory approaches. The classical approach is to create a test statistic, chosen so as to be sensitive to the kind of wildness envisaged, to generate its distribution under some sort of hypothesis of nonwildness, and then to 'reject' (or treat differently) an observation if the test statistic for it comes out improbably large under the hypothesis of nonwildness. A more detailed approach that has sometimes been used is to suppose that wildness is a consequence of some definite kind of statistical structure--usually a mixture of normal distributions--and to try to find a mode of analysis well articulated with this structure.
My own practice in this sort of situation is to carry out an analysis both with and without the suspect observations. If the broad conclusions of the two analyses are quite different, I should view any conclusions from the experiment with very great caution.
6. The following references form a selected brief list that can, I hope, lead the interested reader to most of the relevant literature.