Update: Model Selection and Fieldwork in Maine

The summer session is almost at an end, and the past few weeks have been absolutely packed! I have fully incorporated the milkweed data that we took in the field last month into a new set of models. We now have 4.5 years of data tracking thousands of individual milkweed stems from sprouting/germination until reproduction and dieback at the end of the season each year. This is enough data to make robust statistical models for the vital rates of the population (survival, growth, reproduction), which is what I have been working on. The process of deciding which type of model is the best fit and most suitable for our purposes is a challenging one that takes a surprising amount of time and thought. This stage in the modeling process is often referred to as “model selection” (and validation), and is very heavy on the statistics. Various techniques are used to quantify and visualize how and where a model is performing well and where it may be lacking. A huge part of this involves keeping track of which models can actually be compared with one another in any sort of a meaningful way. This was one of the issue I needed to solve in my analysis. Specifically, we are working with a class of models called Generalized Linear Mixed Models (GLMMs), which do not perform the same way in tests designed for standard linear models. In order to get a more accurate way of assessing the quality of our models, I needed to find and implement some new statistical tools that would make these types of models comparable for us. The below plot is an example of what I used to visually assess the new models, and was generated by the DHARMa package in R. If it looks even remotely interesting to you, check out it’s vignette here.DHARMa Residual Plot

The other exciting development recently (that is actually only indirectly related to my project) is that I had the opportunity to spend the past two weeks in Milo, Maine doing fieldwork as part of a collaborative study on population and community ecology of American Chestnut trees. We took very similar demographic data on these trees as we did for milkweed last month, which is going to be used in a similar way, as well as data related to chestnut blight infection, which has been responsible for the death of nearly a continent’s worth of ecologically and economically valuable trees since it’s introduction around the turn of the century. While it was not directly part of my project, I learned a ton about tree/forest ecology, field methods in forestry, and got to meet several collaborating ecologists that do fascinating work (including Nate Lichti at Purdue, and Mike Steele at Wilkes University).


P.S. I also got to use a giant slingshot in the name of science!