Research update

[jessica hekman]

The latest on my research here at Darwin’s Dogs. Folks, I’m really sorry about this one, but it is all statistics. So you may want to stop reading RIGHT NOW. You have been warned.

I’ve been working (as I’ve written about previously) on trying to find out how different breeds are different (or not) on our survey questions. I’m working with the first 110 questions only right now (they’re the ones the most people have answered).

Turns out, labs are pretty different from everyone else! On about 30 out of those 110 questions (so, a bit less than 1/3), labs have a statistical difference from other breeds. The problem is – we are asking a LOT of questions here (110 is a lot!). And “statistical significance” means “5% of those ‘different’ questions are just different by chance – but we don’t know which ones.”

I had already used a statistical technique to “adjust” the results to compensate for the large number of questions. (For those who are stats nerds – Benjamini-Hochberg.) But Elinor and I were still concerned that this wasn’t a real finding, that it was just by chance.

So last week I tried a different statistical approach – I ran permutations. I took the 217 labs I was using for this test, and the 217 comparison dogs. Then I took all of the 110 answers to questions for each dog, and shuffled them among dogs. A single set of 110 answers always stayed together, but which dog it was attached to was random. This should account for effects due to relationships between questions (for example, between two questions both getting at whether a dog is obsessed with food), but it should erase the division of groups that I’m actually testing for (labs vs other dogs). Then I re-ran the statistical test (for nerds – Wilcoxon) on this new group. In theory, it should have only statistically significant results due to chance (because the group of dogs is randomized). Then I did that again – 10,000 times. (We have fast computers; it took less than an hour.)

Then I looked at the results. How often, with this random shuffle, do I see “statistically significant” results? The answer is, usually about 5/110, maximum of 29/110. Which means that my 30/110 significant results with the real data set are starting to look like they might actually be real.

I always thought labs were total weirdos (I say this lovingly, having done retriever rescue for years) but it would be nice to be better able to describe exactly how!

[Jennifer]

So does this mean that Labs appear to show less behavioral diversity between individuals than other breeds? Or that Lab owners were more likely to choose the statistically least common answer to a question? Or something else?

[jessica hekman]

I compared the labs to a lot of other dogs from a variety of breeds – I didn’t do one on one breed comparisons with any other breeds besides German Shepherds (it’s hard to find any other breeds with enough dogs to do that!). So what this suggests is that labs have a personality profile – particular answers that are more likely for them than for purebred dogs in general. I expect the same to be true for other breeds (just, different answers).

So I don’t think it’s that labs are less behaviorally diverse than other breeds (although I haven’t tested that). They are still very behaviorally diverse! But as a breed they are more likely to like water and retrieving and meeting new people than other breeds (as a whole) are.

I hope that helps – let me know if not!

[Author]

Posts