There is nothing like starting your Saturday morning with some statistics. Today, I’m going to talk about one of my favorite “recent” parasite ecology papers. It’s a 2010 Ecology Letters paper entitled “Detecting interspecific macroparasite interactions from ecological data: patterns and process.”
Here’s a question that people frequently ask in disease ecology: does parasite Species A interact with parasite Species B in/on the host? Do those species compete? If so, we might expect Species A to occur less frequently when Species B is present. Does one species facilitate the establishment of the other? If so, we might expect Species B to only be present when Species A is present. And so on.
(EDIT: Someone kindly pointed out in the comments that schistosomes aren’t intestinal parasites; they’re blood flukes. Oops.)
It is often difficult to approach this question experimentally. Even when we can experiment, it is important to see if the interaction occurs in the natural environment, too. Therefore, people often do observational studies where they collect a bunch of hosts and quantify infection with Species A and/or B. People typically use one of three statistical methods to analyze these data. In correlation analysis, they correlate the number of worms of Species A per host vs. the number of worms of Species B per host. In comparison analysis, they compare the number of worms of Species A per host when Species A is alone vs. the number when Species A is present with Species B. In generalized linear mixed modeling, they model roughly the same thing as in correlation analysis, except that additional variables like host age can be easily included, and the GLMMs can have non-normal error distributions.
Fenton et al. (2010) wanted to know if any of these methods actually work. So, here’s what they did. They generated “parasite data” with an agent based model in which “Species A” and “Species B” either had positive, neutral, or negative interactions. Then they evaluated those data with each of the three statistical techniques. They repeated that process a lot, and then calculated how many times each statistical technique got it right.
(I’m going to say this again because it’s awesome: they generated realistic parasite data with an agent based model!! That is so cool. I want to do that.)
It turns out that none of the methods were perfect, but GLMMs were the single best method. That’s probably because GLMMs can account for overdispersion, which is common in parasite data. (That’s a topic for a separate blog.) Additionally, GLMMs can handle adding host age to the statistical model, and that’s important because parasites typically accumulate with host age, and it might look like a positive relationship exists between Species A and B just because their accumulation curves are similar.
So, we should all start using GLMMs, right? Well, in this case, Fenton et al. (2010) suggested that a combination of approaches is actually the best option.
Does anyone else find this paper as awesome as I do?
Fenton, A., M. E. Viney, and J. Lello. 2010. Detecting interspecific macroparasite interactions from ecological data: patterns and process. Ecology Letters 13:606–15.
UPDATE: Since I wrote this, Dynamic Ecology posted about using fake data to test your ability to detect signals. Very relavent to this paper, and worth checking out!