I am studying animal cognition and I have data about animals' latency to approach a cue and animals' distance from the cue in a test. Data are not normally distributed and when I do scatterplots having latencies in one axis and distances in the other, there are extreme values in both variables or in just one variable, depending on which sample I select. I decided to look for statistical techniques to detect these outliers in a proper and non subjective way. For the moment I found:
- MAD (absolute deviation around the mean, with conservative approach), that not only removes the outliers that I can spot from the graph, but also other values that are more clustered with the rest. Since it seems very aggressive even with conservative threshold, I decided to apply it only when I can actually see that there are gaps in the scatterplot, otherwise I would have just a few datapoints left.
- boxplot, which is just a way to see how my data are distributed, but I found a page that says that values outside whiskers are outliers https://medium.com/@agarwal.vishal819/outlier-detection-with-boxplots-1b6757fafa21. This method seems to detect datapoints that are really far from the rest of datapoints more often than the MAD method.
The first method is documented in a paper "Detecting outliers: Do not use standard deviation around the mean, use absolute deviation around the median" and has been used in some research areas, such as psychology. The second method is not documented, but it seems more straightforward and it seems to work better. Are you aware of studies in the field of animal cognition (with measures of latencies and distances) that needed to remove outliers and what they did?
You might find the statistical ideas in the attached file of some interest. Best, David Booth any stat package will do this now days.
Lots on RTs being close-ish to log-normal in many circumstances. Also, presumably the two are related. A scatter plot on the logged data might be useful, but first, why do you want to detect outliers? Just to identify some cases for further investigation, or are you trying to identify cases that are not part of the population of interest? I assume that you aren't thinking of just throwing them out.
Hello Alessandra,
You've already gotten some good advice, but please permit me to add a couple of points that you may find germane:
1. The Boxplot, if applied to normal distributions, would classify as "outside" cases approximately 1% of the distribution (0.007).
2. Why (as in the article title you cite) is the median (as measure of location) so commonly recommended? First, it's the only measure of location about which probability statements can be made, regardless of the shape of the distribution. Second, if the distribution is not symmetric (often the case!) it's more robust against non-normality triggering some other attribute check. For example, if Levene's test for variance equality is applied to non-normal data, it's more likely to trigger a false positive for lack of homogeneity than if one computes deviation from sample/group/cell medians.
Good luck with your work.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques