When it comes to data analysis, there are hundreds of exciting approaches: simple summary statistics and hypothesis tests, various clustering methods, linear and nonlinear regression or classification techniques, neural networks of various types and depths, decision rules and frequent itemsets, feature extractors and dimension reductors, ensemble methods, bayesian approaches and graphical models, logic-based approaches and fuzzy stuff, ant colonies, genetic algorithms and other optimization methods, monte-carlo algorithms, sampling and density estimation, logic-based and graph methods. Don't even get me started on the numerous visualization techniques.

This sheer number of options is, however, both a blessing and a curse at the same time. In many practical situations just having those methods at your disposal may pose more problems than solutions. First you need to pick one of the approaches that might possibly fit your purpose. Then you will try to adapt it appropriately, spend several iterations torturing the data only to obtain very dubious first results, come to the conclusion that most probably you are doing something wrong, reconvince yourself that you need to try harder in that direction, spend some more iterations testing various parameter settings. Nothing works as you want it to, so you start everything from scratch with another method to find yourself obtaining new, even more dubious results, torturing the data even further, getting tired of that and finally settling on something "intermediately decent", which "probably makes sense", although you are not so sure any more and feel frustrated.

I guess life of a statistician was probably way simpler back in the days when you could run a couple of t-tests, or an F-test from a linear regression and call it a day. In fact, it seems that many experimental (e.g. wetlab) scientists still live in that kind of world, when it comes to analyzing their experimental results. The world of T-tests is cozy and safe. They don't get you frustrated. Unfortunately, t-tests can feel ad-hockish, because they force you to believe that something "is normally distributed". Also, in practice, they are mainly used to confirm the obvious rather than discover something new from the data. A simple scatterplot will most often be better than a t-test as an analysis method. Hence, I am not a big fan of T-tests. However, I do have my own favourite statistical method, which always feels cozy and safe, and never gets me frustrated. I tend to apply it whenever I see a chance. It is the Fisher exact test in the particular context of feature selection.

My appreciation of it stems from my background in bioinformatics and some experience with motif detection in particular. Suppose you have measured the DNA sequences for a bunch of genes. What can you do to learn something new about the sequence structure from that data? One of your best bets is to first *group *your sequences according to some known criteria. Suppose you know from previous experiments that some of the genes are cancer-related whereas others are not. As soon as you have specified those groups, you can start making observations like the following: "It seems that 10 out of my 20 cancer-related genes have the subsequence *GATGAG *in their DNA code. The same sequence is present in only 5 out of 100 non-cancer-related ones. *How probable* would it be to obtain similar counts of *GATGAG*, if the two groups were picked randomly?" If the probability to get those counts at random is very low, then obviously there is something fishy about *GATGAG *and cancer - perhaps they are related. To compute this probability you will need to use the hypergeometric distribution, and the resulting test (i.e. the question "how probable is this situation in a random split?") is known as the Fishers' exact test.

This simple logic (with a small addition of a multiple testing correction on top) has worked wonders for finding actually important short sequences on the DNA. Of course it is not limited to sequence search. One of our research group's most popular web tools uses the same approach to discover functional annotations, that are "significantly overrepresented" in a given group of genes. The same approach can be used to construct decision trees, and in pretty much any other "supervised learning" situation, where you have groups of objects and want to find binary features of those objects, associated with the groups.

Although in general the Fisher test is just *one *particular measure of association, it is, as I noted above, rather "cozy and comfortable". It does not force me to make any weird assumptions, there is no "ad-hoc" aspect to it, it is simple to compute and, most importantly, in my experience it nearly always produces "relevant" results.

A week ago me, Ilya and Alex happened to take part in a small data analysis hackathon, dedicated to the analysis of speech transcripts from the European Parliament. Somewhat analogously to DNA sequences, speeches can be grouped in various ways: you can group them by the speaker who gave them, by country, gender or political party of that speaker, by the month or year when the speech was given or by any combination of such groupings. The obvious "features" of a speech are *words, *which can be either present or not present in it. Once you view the problem this way the task of finding *group-specific words* becomes self-evident and the Fisher test is the natural solution to it. We implemented this idea and extracted "country-specific" and "time-specific" words from the speeches (other options were left out due to time constraints). As is usual the case with my favourite method, the obtained results look relevant, informative and, when shown in the form of a word cloud, fun. Check them out.

The complete source code of the analysis scripts and the visualization application is available on Github.

I've read that Fisher's test is usually accurate with "small" sample sizes, and that chi-square test is preferred otherwise. Not sure how small is "small", but the Talk of Europe data does not seem like that..

To better understand this claim I would need more details regarding the context in which it was made. Could you find the source?

In general, the question of "how improbable it is to get that K white balls from an urn with M white and N black balls" is meaningful no matter how large K, M or N is. The probability computation is indeed computationally harder for larger values, but it is an issue of computation, not statistical preference.

The Chi-square test, at the same time, is both harder to interpret ("the sum of squares between the observed and an expected distribution" - go figure) and, indeed, tends to be less stable for very small sample sizes.

The Fisher test is regularly and successfully used in bioinformatics on rather large datasets. The results we see on the ToE dataset are quite insightful as well.

Thank you for the clarification. I found that claim in many different contexts, e.g. http://www.biostathandbook.com/fishers.html and http://www.stat.purdue.edu/~tqin/system101/method/method_fisher_sas.htm

Interestingly, the latter says, "Fisher's exact test is particularly appropriate when dealing with small samples", I really cannot understand what makes it not very appropriate for larger data sets

Note that both of the links you mention only claim that Fisher test is *more precise* (i.e. it has more statistical power) for small samples. It does not imply that it is "inferior" or "inappropriate" for large samples.

It is just as good for large samples - the only "inferiority" is the computational expense - computing exact hypergeometric p-values for large k, n and m is tricky, and you would probably get qualitatively similar results using any other much simpler approach (one of them being Chi2).