Skip to content
Prev 470 / 7419 Next

Teaching statistics to ecology undergraduates

Dear Graham,

11 hours is short - there's no mistaking. I teach (among other things) a 
6 day stats course for beginners, and find that I need the first 3 days 
to get the student to "think straight". I tried for a couple of years to 
teach "only" GLM, as you suggested. I "waste" one full day of explaining 
what a distribution is, what parameters of distributions are and on what 
ground to suspect data to be derived from a certain distribution. That 
would be at least 5 of your 11 hours. The next half day goes into 
explaining (and running examples) on likelihood and its maximisation. It 
is a good way to start, I find, and eventually students are very 
comfortable using glm rather than aov and friends.
Using only GLM is clear (and Ben Bolker's book sets the right tone, 
albeit at a much too high level for beginners). At the same time, the 
learning curve is VERY steep. 30% of the participants fall by the 
wayside. Is that acceptable? Maybe it is me, not the GLM.

However, I think you have to be very realistic about what you can 
achieve (and I have heard speaking highly of your courses, so I am sure 
you are doing the right things). Giving the students a "feeling" about 
what the idea of a "fit" is and what is behind comparisons of samples is 
rather independent of distributional assumptions and a very general 
point they can take away from a short course.
Also, as you said, visualising the data, getting a feeling for it, is SO 
important, particular when a student has little idea what to expect from 
an experiment/observation.

I my little 6 day course, I spend roughly 2 days on introducing R, 
distributions and (maximum) likelihood (half of this time the 
participants run examples). Another 2 days are devoted to multiple 
regression (going wildly through different distributions to make them 
comfortable with GLM) and issues such as collinearity and model 
selection. Then I throw in a day of design of experiments (randomised 
block, nested, split, survey design, stratification, sample size 
estimation) and run some simple (?) mixed models to illustrate the 
practical problems attached to DOE. The final two days we run largish 
examples (such as Harrell's Titanic data set), touch very superficially 
multivariate methods (PCA, CA and CCA) and end up with some 
miscellaneous issues such as randomisation and bootstrapping.

If I had to reduce it to 11 hours: Unless the students are likely to do 
experiments (which seems to have fallen out of funding), I would ditch 
DOE and focus on GLM plus a few sexy  but tricky examples. I love the 
Titanic study, because you can get the students to identify with the 
passengers. If that leads them to transfer their newly gained knowledge 
to the ecological work is a different question. If you additionally make 
the buy a good book (I always recommend Quinn & Keough, having myself 
been "raised" on Sokal and Rohlf and always hated it, because it never 
addressed my type of non-Gaussian problems) I think they should be set 
up for the next level.

I shall stop now (and prepare some stats course next week), otherwise I 
would also have a word to say about Crawley's approach, which I find 
enchanting and confusing.

Carsten
Graham Smith wrote: