The Internet is flooded with explanations related to P-values (To P or not to P?) and problems around multiple testing (the more tests you do, the more significant results you get even if everything is random). I don’t bring anything new, I just wrap it for myself and add a few examples that hopefully make some concepts clearer. I will talk about Type I and Type II error rate, the power of the test, disadvantages of multiple tests, how it looks like when we use the correction for multiple tests (not many significant results left) and what else we could do with that (some global analysis?). Read further if you also think about these things.

P-values

Several decades ago, ecologists were strongly arguing that it is important to start using rigorous statistical approaches and standards in ecology; that was the era when ecology studies were based mostly on natural history observations instead of systematically sampled data analysed by rigorous statistics. Meanwhile, ecology grew into a standard scientific discipline, and some start to argue that P-values are often being misused and misunderstood, and should be banned from ecological publications (as it happens in medicine or psychology). Putting this discussion aside, P-values are still representing a useful tool telling us how much evidence we have against the null hypothesis, which we are testing.

P-value is the probability of obtaining the value of test statistic equal or more extreme than the observed one, even if the null hypothesis is true. The null hypothesis is usually formulated as “there is no relationship between the variables” (or there is some fixed relationship, from which we are testing the difference). For example, if I relate species composition to some environmental factor using constrained ordination, P-value is the probability of obtaining the test statistic (R², pseudo-F-value) equal or higher than the one we truly calculated, even if the environment is truly not related to species composition (null hypothesis is true). If the P-value is sufficiently low (e.g. less than 5%, as is the common standard, proposed originally by Ronald Fisher), we may conclude that it is quite unlikely that the result could be obtained at a situation that the null hypothesis is true and state that the observed value is “significant”. That 5% (or in general alpha level) is the probability of Type I error, and we report such result as “significant” at given alpha threshold (e.g. P < 0.05). Some examples. First, let’s generate a data for which we are sure that the null hypothesis is true: we generate two random variables, independently from each other; then we calculate the correlation between them (Pearson’s r) and test it using parametric t-test (all this is done by ‘’cor.test’’ function in R):

1
2
3

x.rand <- rnorm (100)
y.rand <- rnorm (100)
cor.test (x.rand, y.rand)$p.value

We obtain some P-value (e.g. if you include ”set.seed (1234)” before the script above, the P-value should be P = 0.8020633, i.e. not significant at alpha = 0.05). Yet, even if these variables are randomly generated, occasionally we still can obtain a “significant” result with P-value smaller than alpha, with a probability of alpha (try it – you can e.g. ‘’set.seed (1257)’’, and you obtain P-value = 0.001678019, well significant at 0.05 level).

Now, the second situation – we generate data, for which we are sure that the null hypothesis is not true; variable x is still generated randomly, but variable y is the function of x, with added random noise. The same – calculate correlation and test it:

1
2
3

x.dep <- rnorm (100)
y.dep <- 0.2 * x.dep + rnorm (100)
cor.test (x.dep, y.dep)$p.value

If in the code above you ‘’set.seed (12345)’’, you got P-value = 0.001676154, well below the 0.05 threshold, so you would conclude that the value is significant at P < 0.05. But of course, it may also happen that the result is not significant, even it should be (e.g. with ‘’set.seed (12355)’’, you obtain P-value = 0.06517154, which would already not be significant at the 0.05 significance level). To summarize, see the possible outcomes in the table:

The left column shows the situation that the null hypothesis is, in reality, true (our first example with randomly generated variables). If I still reject it, I conduct the Type I error or a false positive. The probability that this will happen is set by me and equals to alpha level at which I decided to reject the null hypothesis (alpha = 0.05 or 5% in this case). If I did not reject it, I conduct the true negative (with the probability of 1 – alpha).

The right column shows the situation that the null hypothesis is in reality not true (our second example, where y is dependent on x because I enforced this dependence). If I reject the null hypothesis, I did correct decision (true positive), if I did not reject it, I conducted the Type II error or a false negative). The probability of Type II error is beta, and the probability of true positive is 1 – beta, also called the power of the test. Beta and the power of the test depends on several factors – test itself, how powerful it is (some are more and some less powerful), and also the data – sample size, and amount of variance. In our case, the data are quite noisy and I sample only 100 samples, so it may be not so easy to find the relationship which is there (this will improve if we increase the sample size, e.g. to 1000).

Now, in reality, I do not know whether the null hypothesis is true or not (that is why I want to test it), and I need to collect the data and use the statistic to get an idea. Both Type I and Type II errors are problematic – it means that I rejected something I should not (Type I error) or did not reject something I should (Type II error). The probability of Type I error is selected by me (here 0.05), while the probability of Type II error is dependent on three factors: the sample size (larger sample size means lower Type II error), the selected value of Type I error (lower alpha means higher beta), and the true difference of the data from the null hypothesis, i.e. true effect size (the higher is the difference, the lower is Type II error rate – it’s easier to reject the null and conduct true positive). This means that decreasing our choice of Type I error rate (alpha) increases the probability of Type II error (beta) and decreases the power of the test (1 – beta). Not an easy life; best solution – sample more to get higher sample size, which will reduce the probability of Type II error (of course, not always possible).

Both the effect size and the number of samples in the dataset are also influencing the P-value (it decreases with both the effect size and the number of samples), so in case of a large dataset, we are likely to get strongly significant results even if the true effect (e.g. R²) is pretty small. Some argue that datasets with more than several hundreds of samples the P-value already does not represent a useful tool for the decision, because virtually anything could get highly significant. Also, the threshold for rejecting the null hypothesis (often set as P < 0.05) should not be taken too seriously. It does not mean that at P = 0.048 we have evidence against the null hypothesis, but at P = 0.052 we already do not. Better is to consider P-value as a continuous measure, and use those thresholds (P < 0.05, P < 0.01, P < 0.001) only as a recommendation to which results we should pay more attention to. The other extreme, i.e. interpreting everything, would return us several decades back, to the era when interpretations of ecological studies were based on feelings and intuition rather than on the results of statistical analyses.

Multiple testing issue

Sometimes the study includes an analysis which is based on testing more than one null hypothesis. Imagine example: you want to know whether plant functional traits (e.g. leaf area, SLA) are related to some of the environmental factors (e.g. elevation, soil P). We will keep technical details of how this can be done aside, just imagine that you have ten different traits and ten different environmental variables, and you collect 100 samples (e.g. vegetation plots) in which you do all these measurements. Then, you take each of the ten traits and regress it against each of the ten environmental variables, test the significance, and interpret significant results. What is the problem?

Let’s do the same exercise as above. First, simulate the situation when all null hypotheses are true (variables are randomly and independently generated). R code gets more complex, so I deposited it in the Github repository here. What I do: I generate two random variables from the normal distribution, each having 100 values, calculated regression of y on x, and tested the significance (by parametric F-test in this case; note that if I calculate Pearson’s r correlation, the P-value would be identical). I repeat this 100 times and highlighted significant results by adding the red regression line into the scatterplot.

You see that I got a few significant results, although I should not – the null hypothesis is truly correct, I generated data in this way. These are false positives, and their probability is equal to the significance level alpha I set up (here 0.05, i.e. 5%). You can see that if I repeat the same analysis again, I got again some significant hits, just this time in different combinations of traits and environmental variables.

In fact, I can calculate the probability that, for given number of tests and given alpha value, I will get at least one significant result (with P < alpha), even if the null hypothesis is true: p_{{at least one significant result}} = 1 – (1 – alpha)^m, where m is the number of tests. In case of 100 tests, the probability of obtaining at least one result significant at P < 0.05 is 1 – (1 – 0.05)^100 = 99.4; I can be almost sure that I get at least one! Now, the second scenario, when all 100 relationships are generated with null hypothesis not true – I created variables which are dependent on each other. The same as above applies – I test each of them, and if significant, I add the red regression line. In this case, I coloured the background each scatter by light blue, to indicate that the variables are truly not independent.

The number of hits is high, but not all combinations are significant. Those which are not are false negatives, and those which are significant are true positive; the probability of true positive equals to the power of the test, in this case, something around 50% (How do I know that? Repeat the analysis many times, e.g. 1000 times, to see the proportion of true positives; this gives you the power of the test, from which you can calculate Type II error rate as beta = 1 – power). If you repeat the same thing, again quite high proportion of relationships will be significant (true positives), but some not even they should (false negatives).

In reality, we don’t know for which combination of variables the null hypothesis is true and for which it is not. Let’s make a third scenario, in which we mix both options – some of the relationships are generated from independent variables (null hypothesis is true), some from related variables (null hypothesis is not true). For simplicity, the 10 relationships in the first row are those with null hypothesis not true, and the rest (90) with null hypothesis true.

The figure shows the whole spectrum of situations: true positives (blue background with red regression line), true negatives (white background, no regression line), false positive (white background, regression line) and false negatives (blue background, no regression line). The problem is that, in reality, we do not know which of the relationships have the blue background (null hypothesis is not true), so when we do the interpretation one by one, we may well interpret true positives (correct!) as false positives (not correct!).

How to deal with this? One way is to adjust P-values of individual tests to correct for multiple testing issue. The simplest is the Bonferroni correction, which multiplies each P-value by the number of tests done (P_adj = P * m) and then evaluates P_adj instead of the original P for the significance. This helps to control for the overall (or family-wise) Type I error rate, but at the same time, it quite drastically reduces the power of the test, and you are unlikely to get false positives, but you are quite unlikely to get true positives. There are better methods then Bonferroni (e.g. FDR, false discovery rate, or Holm’s correction), but the drastic reduction of test power is true for all of them.

If we increase the power of the test by increasing the sample size, this problem will be less serious, but this is not always possible. Below, I increased the number of samples in each comparison from 100 to 500.

This looks great! No false positive, and only two false negatives. However, in real data, this is quite unlikely to happen, because you won’t have the power of the test more than 99% as in this case.

Seems that multiple testing strategy is not the best way to go – either you suffer from too many false positives, or you correct and get (almost) nothing significant. If possible, opt for other types of analysis, some which conduct a single global test instead of the series of pairwise ones, or combine these two strategies together. A well-known example of such an approach is one-way ANOVA and post-hoc comparison. ANOVA tests whether the mean of at least one of the groups deviates from the overall mean; if yes, then one can conduct a pairwise post-hoc comparison of groups to identify which of the groups are those different (still, this pairwise comparison is not a simple series of t-tests, but includes an implicit correction for multiple testing). In the case of trying to find the relationship between a set of two variable types (e.g. traits and environment as in our example), you may opt for constrained ordination, which is in fact multivariate regression (many traits regressed on many environmental variables, more precisely CWM-RDA in this case, but details are beyond this text). If the ordination is significant, then go for pairwise regressions to see which of the variables are truly related (still, to use some of the corrections for multiple testing is advisable).

You may ask: so if I measure things, relate them to each other and find a significant relationship, does it mean that these variables, in fact, may have no real meaning, they could be in fact just random variables? The answer is yes, it may happen that this result is just a false positive. The way to make sure that the relationship between those variables (e.g. given trait and environment) truly exists is to replicate the study, best in a different location. Or, research the literature to see whether ever someone detected the same relationship in a similar context. It’s unlikely that the same combination of variables would be a false positive in many studies (you see above that false positives occur, but in the new replication it occurs in different variable combination). If you repeat the study again (and again and again) and you get the same relationship significant (or you see multiple papers reporting the same thing), that is strong evidence that it may truly mean something.

Where to go next?

A nice readable text on this topic is the chapter The p value and the base rate fallacy in the book Statistics Done Wrong by Alex Reinhart (see the comics there!). Also, a few years ago there was an inspiring series of forum papers in Ecology, started by the paper of Paul A. Murtaugh entitled In defense of P values (Murtaugh 2014), followed up by several others and wrapped together again by Murtaugh. Also, there is recently an interesting suggestion that we should redefine statistical significance and lower the alpha threshold from 0.05 to 0.005, which will greatly reduce the probability of reporting false positives in scientific studies and increase reproducibility. As it goes, one of the first comments (from many) to this paper is entitled “Remove, rather than redefine, statistical significance”…

R script for all calculations in this post is in Github repository: https://gist.github.com/zdealveindy/39feabd7c7b4f3d780d2b8c5232456a1

Ecological drift is one of the four main processes maintaining species coexistence in ecological community, along with selection, dispersal and speciation (Vellend 2010, 2016). To me, it seems that from these four processes, the effect of ecological drift on diversity and species composition of ecological communities is the least intuitive. Here I come with a few interactive illustrations and videos showing how the ecological drift works and what it does.

I separated this post into two parts, forming a logical sequence. Here, in part 1, I focus purely on the random drift, without any other process in action. Destiny of all simulated communities is the same: eventually, only one species will prevail, and all other will go extinct, it is just a matter of time when. In part 2, I will also add dispersal and create a virtual landscape in which it matters who is your neighbour and how far your offspring can disperse. Then, I will also add selection, making species not ecologically equal and letting them search for their ecological optima. Combination of drift with dispersal and selection will be the most closely resembling the community assembly in real ecological communities, although being still very far from the situation in any real community. And hopefully, it will convince you that random drift is a process which can get more important than you would think.

But let’s start slowly.

Island with 16 trees of two species: how the ecological drift works

Imagine a small, tiny island – so tiny, that only 16 trees could squeeze in, not a single one more. The island is surrounded by sea, far from any mainland, perfectly isolated. In our example, we start with eight individuals of each species (red and black, see the figure below). A single process is happening: at any given moment, one individual dies and empties its place, and another individual reproduces, and its offspring fills the vacant place. The selection of individual which dies and which reproduces is random; it does not matter to which species the individual belongs or where on the island it occurs. See the illustration below.

Random drift on remote island with 16 trees of two species. Yellow circle indicates species which has been selected to reproduce and whose offspring fills the vacant place.

What happens after several cycles of die-reproduce-replace? Watch below.

The population is small, so even a tiny fluctuations in units of individuals, in the end, lead to the extinction of one species (in this case the red one). If both species have an equal proportion of individuals at the beginning, the probability to extinct is equal for both since there is no difference between species and between individuals (species are ecologically equal, meaning that they have no fitness differences – they have the same probability to reproduce and survive to the next generation).

The same thing can be drawn with the x-axis as time and y-axis as the individuals, with different species displayed by different colour. Because the position of individuals in the space of island does not play any role in this example, we can stack the individuals of the same species (the same colour) together to see overall trends of species abundances in time:

Trend of abundances of two species in time, influenced by random drift. Less than 100 generations were needed in this case for one species go extinct from the island.

Now, let’s rerun the same scenario several times (nine times in this case), to see what happens:

Random drift on island with 16 individuals and two species, running for 300 generations.

Note that I increased the number of generations from 100 to 300 in this multiple run example, to show that for most of the replicates this is enough time for one species to disappear from the island. Result in this case: in eight cases, one species extinct (in four cases it was red, in four black), and in one case still both species co-exist on the island. Note that the first panel (top left) is the same as the single figure above (with 300 instead of 100 generations).

We can play further with this visualisation. How about a bigger island, say with 100 individuals?

Random drift on island with 100 individuals and two species, running for 300 generations.

Not enough generations for any of the two species to extinct. But if we increase considerably the time (into 5000 generations), things happen:

Random drift on island with 100 individuals and two species, running for 5000 generations.

The size of the population decides how fast the random drift will cause one of the species go extinct; more individuals need more time. But anyway, it is just a matter of time when the species will “drift away” of the community.

Now, how about more species? If the community of 100 individuals is cut into ten species, the population of each species becomes rather small, prone to extinction.

Random drift on island with 100 individuals and 10 species, running for 300 generations.

Even if the time is short (300 generations), in three out of nine replicates, one species extincted from the community (check the text in the titles of each plot: Start: indicates how many species were in the community at the start of the simulation, end: counts how many were in the end). If we take the same communities and increase the time to 5000 generations, again, things happen:

Random drift on island with 100 individuals and 10 species, running for 5000 generations. First 300 generations in each panel are identical with previous figure.

After 5000 generations, between one to three species left from what was community of 10 species at the beginning. A little hard to imagine that this figure is showing the same communities as the previous one, just “much later”, right? See the animation below (whcih focuses on the first panel (A) from the nine):

From the rainbow flag at the initial situation (after 100 generations, all species still have rather a similar number of individuals), after 5000 generations, all but one species eventually extinct.

Forest fragment with ~5000 trees and 74 species: getting community parameters closer to something real

Ok, so let’s start to get more real. First, the populations of organisms are usually much larger than 100 individuals. Also, different species have different relative abundance, some are common, and some are rare (this is called species abundance distribution, which is usually highly uneven). Obviously, more common species seems to have a lower probability of extinction compared to rare species. To get more realistic parameters of our island community, I borrowed the data from the real forest in Taiwan, 25-ha large forest permanent plot in Lienhuachih (LHC). As in other forest dynamics plots established following the standards of Smithsonian Institute, all tree individuals with DBH (diameter in breast height) larger than 1 cm are measured, determined to species, their exact location is recorded, and each is permanently marked. Altogether, during the census of LHC plot in 2005 researchers found 153,255 woody individuals of 144 species. To make things simple here, I will consider only individuals with DBH > 20 cm; we get 5449 stems of 74 species. Their absolute abundances are rather variable: two the most abundant species, Engelhardia roxburghiana and Schefflera octophylla, have 599 and 532 individuals each, respectively, while seven species have each only one individual. The two figures below illustrate the flavor of these data:

Distribution of individual trees (DBH > 20 cm) in the 25ha plot of Lienhuachih forest dynamics plot, Taiwan. The size of the circle is proportional to DHB of individual trees, and different color means different species.

Liehnuachih forest dynamics plot. Photo credit: David Zelený.

Let’s put these data into the random drift visualisation, directly with 5000 generations:

Random drift in forest with 5449 trees, 74 species and realistic species abundance distribution, running for 5000 generations.

Seven species went extinct (mostly the rarest ones), but otherwise nothing too much is happening in such short time for such a big community. Note the new visualisation tool here: the left panel is the same type of figure you have seen already above (x-axis is time, y-axis is absolute number of individuals for each species), while the right panel shows species abundance distribution at the beginning (gray bars) and the end of 5000 generations (colorful transparent bars). In the barplot, species are sorted from the most to the least abundant at the start of the simulation (i.e. the tallest bar is Engelhardia roxburghiana with 599 individuals, the second tallest is Schefflera octophylla with 532 individuals).

The figures below show the same community, progressing further (50,000, 500,000 and 5,000,000 generations):

Random drift in forest with 5449 trees, 74 species and realistic species abundance distribution, running for ~50,000 generations.

Random drift in forest with 5449 trees, 74 species and realistic species abundance distribution, running for ~500,000 generations.

Random drift in forest with 5449 trees, 74 species and realistic species abundance distribution, running for ~5,000,000 generations.

The result is as expected; rare species became rather quickly extinct, and few dominant species survived. The expected probability that species will extinct is negatively related to its initial abundance, being highest for rare species. But that is “expected probability”; what we see in this simulation is one of many possibilities how the random drift changes species composition in time. In the figure with 500,000 generations, you can see that dark blue species, which has initially been rather rare, became quite abundant (almost 200 individuals); eventually, however, it still extinct. Video below shows the whole sequence between 5,000 and 5,000,000 generations with steps of 10,000 generations:

Play with random drift: interactive application

Time to try things by yourself. Below is an interactive application, in which you can play with random drift by yourself. By default, it creates a community of 10 species, 100 individuals and 1000 generations, and displays abundances of individual species by colours of the rainbow. Feel free to modify different parameters; for example, set two species, 16 individuals and 100 generations to get the same scenario as displayed in the first figure of this website. Whenever you modify some parameter, the whole simulation is recalculated with new randomly generated values. If you want to stick to some version, fix the seed of the pseudorandom generator (set seed to YES); this will return the same generated community. You can calculate more than one (four or nine) scenarios in parallel.

Feel free to play! (click here to open the app in a new windows)

Technical details: how I made it

Everything on this website was done in R, using a relatively simple random.drift function. Originally I created this function as an exercise with students in my R for Ecologists class; it has been a simple function with two species from which one will eventually drift away, and the purpose was for students to practice how to do a loop and write a function. Eventually, I found it interesting to add an option to have more species and include more arguments to improve flexibility. See the definition of the function on my AnaDat-R website. Videos are not fully R-made – I used R to generate the sequence of *.png images but merged them into a video using Window’s Movie Maker. Hope in the future this can be replaced by some function in ”magick” package, which is a recent fresh import of ImageMagick into R (but seems not having all the functionality ImageMagick has). Yeh and GIF animations were done with package ”animate”, function ”saveGIF”; there are some shortcomings in this package, for example, while I can modify the width and height of the figure, I cannot modify the resolution and pointsize argument; or maybe it just needs to study more.

Additionally to the R code, I also made Shiny application for the random.drift function, see the previous section. The background calculation is run in R on Shiny servers, and if you choose larger values, it may take a while to display the outcome. I got quite fascinated by the possibility to create such interactive apps; it is actually quite simple. More will for sure come in future.

References

• Vellend, M. 2010. Conceptual synthesis in community ecology. The Quarterly Review of Biology 85: 183–206.
• Vellend, M. 2016. The Theory of Ecological Communities. Princeton University Press.

Since I started to teach multivariate methods in community ecology and additionally to chapters about ordination I also included the chapter about diversity, I found it confusing that there are two “Simpson’s indices,” one for similarity, and one for diversity. Indeed, each index has a completely different meaning and use, and I may well be the only one confused by them, but for a while, I had an impression that both indices are somehow related and maybe could be derived from each other. I did some search to clarify it for myself, and it turned out that although both indices have been published in roughly the same time (in the ’40s), each one is attributed to different Mr Simspon and appeared in quite a different context.

A bit of theory first. Simpson’s similarity index is used to calculate the similarity between a pair of community samples, to quantify whether their species composition is similar (they share most or all the species) or different. It comes in the company with other, similar indices, like Jaccard and Sørensen, which do the same job but with slightly different logic. If we have two samples, there will be some number of species shared between both, and some species occurring only in the first or only in the second sample. Jaccard or Sørensen divides the number of species both samples share by the number of all species occurring in both samples together (and Sørensen cares about shared species more than Jaccard, so it multiplies it twice in both numerator and denominator). The problem occurs if one of the samples has much higher species richness than the other. The number of shared species (in the numerator) may be the same, but the high richness of one of the samples will increase the denominator and decrease the overall index value, making the index losing the resolution power. Here comes the Simpson’s similarity index, which has the same nominator (number of shared species), but in the denominator, it takes only the number of species occurring in the smaller sample. Its use is recommended in a situation when there are large differences in species richness between the samples, and eventually, it turned out that this index is also the one quantifying the nestedness of two community samples (if the first sample contains the same species as the other one plus some more, the second sample is perfectly nested within the first one, and the value of the Simpson’s similarity index is equal to one).

Simpson’s diversity index is a completely different story. It aims to quantify the diversity of a single community sample, while considering both species richness (the number of species recorded in the sample) and also the relative species abundances (some species within the sample can be dominant, represented by a high number of individuals/higher biomass/larger cover, while some other are less common or rare). Two communities may have the same species richness (i.e. number of species), but if they differ by relative species abundance (i.e. evenness), their overall diversity will be different. If we have two forests, both with ten species (richness), but first with all species represented by the same number of individuals (perfectly even), while the second with one or two dominant trees (highly uneven), the first will intuitively (and also functionally) be more diverse – a walk through such forest will give you a feeling of diversity since you keep meeting different trees. In contrast, the forest dominated by a single or few species will feel species-poor (rare species are also there, so richness is the same, but they are not likely to be encountered as frequently as and also they are not likely to have an important ecological function in the community). Simpson’s diversity index, when calculating the diversity, is weighing the importance of species richness and relative species abundances. It comes in the company of his famous sibling, Shannon index, which does the same thing in a slightly different way. While Simpson’s index cares more about relative abundances, the Shannon index cares more about species richness; or, put in another way, the importance of rare species decreases in order species richness > Shannon index > Simpson index.

Now about the name. Both indices share the author’s surname, but in each case, it was a different Mr Simspon. Simpson’s similarity index was introduced by an American palaeontologist George Gaylord Simpson (1902-1984) to solve the situation often encountered in his practice with fossil samples, where some samples are much richer than the others, meaning that conventional similarity indices (like Jaccard and Sørensen) do not work properly. In 1943, G.G. Simpson wrote a paper “Mammals and the nature of continents” and published it in American Journal of Science; later, in 1960, he published (in the same journal) paper “Notes on the measurement of faunal resemblance”, where he also provided the formula for his index. The index has been rediscovered by Jack J. Lennon et al. in 2001 as βsim (often called beta sim; I suspect “sim” stems from “Simpson”). Note that Lennon et al. (2001) does not talk about “similarity”, but instead they used Simpson’s formula to calculate mean pairwise dissimilarities between the focal sample and the other samples; in Koleff et al. (2003), however, the beta sim is already considered as a pairwise dissimilarity measure based on the original Simpson’s index (1 – Simpson).

Simpson’s diversity index (also called concentration index) was published by British statistician Edward Hugh Simpson (born 1922, living in Oxfordshire), who is mostly famous for his formulation of Simpson’s paradox. E.H.Simpson published the index in the 1949’s Nature’s paper entitled “Measurement of diversity”. It quantifies the probability that two individuals took independently (without replacement) from the sample will be of the same species. With increasing diversity the index actually decreases (two individuals have the highest chance to be of the same species if the sample has only one species), so more often the one complement (1-Simpson) or reciprocal (1/Simpson) values are used, also known as Gini-Simpson index or Hurlbert’s PIE (probability of interspecific encounter, Hurlbert 1971).

So here we go – no reason to confuse them any more!

References

• Hurlbert, S.H. 1971. The nonconcept of species diversity: A critique and alternative parameters. Ecology, 52: 577–586.
• Koleff, P. Gaston, K.J. & Lennon, J.J. 2003. Measuring beta diversity for presence-absence data. Journal of Animal Ecology, 72: 367-382.
• Lennon, J.J., Koleff, P., Greenwood, J.J.D. & Gaston, K.J. 2001. The geographical structure of British bird distributions: diversity, spatial turnover and scale. Journal of Animal Ecology, 70: 966-979.
• Simpson, E.H. 1949 Measurement of diversity. Nature, 163: 688.
• Simpson, G.G. 1943. Mammals and the nature of continents. American Journal of Science, 241: 1-31.
• Simpson, G.G. 1960. Notes on the measurement of faunal resemblance. American Journal of Science, 258-A: 300-311.