0
$\begingroup$

I wonder whether anyone knows if I need to use a Poisson distribution when dealing with fixation count data from eye tracking?

I have currently used an ANOVA on fixation count data from an eye tracker and want to know whether this is a mistake.

$\endgroup$
6
  • $\begingroup$ What's the question? Your sentence isn't grammatically correct. $\endgroup$ Nov 11 '15 at 17:49
  • $\begingroup$ Apologies I was talking to Google which has a tendency to make me write crap. I have currently used an ANOVA on fixation count data from an eye tracker and want to know whether this is a mistake. In my original question substitute 'did' with'need' $\endgroup$ Nov 12 '15 at 19:53
  • $\begingroup$ You can edit your question directly by clicking the "edit" link under your question. $\endgroup$ Nov 12 '15 at 21:12
  • $\begingroup$ Please edit your additional explanation from your comment into your question (or something similar to it) $\endgroup$
    – Glen_b
    Jan 7 '17 at 22:35
  • $\begingroup$ @Glen_b, the OP hasn't been seen for over a year. $\endgroup$
    – gammer
    Jan 9 '17 at 1:23
1
$\begingroup$

Editing in information from your question and comment:

I have currently used an ANOVA on fixation count data from an eye tracker and want to know whether this is a mistake.

Do I need to use a poisson distribution when dealing with fixation count data from eye tracking?

The two do not correspond - for example it might be a mistake to perform ANOVA and yet it may not be that you need to use a Poisson model for your data. Indeed a Poisson-GLM with a mean model equivalent to ANOVA might be a poor model for such count data. I don't know this area so my concerns may not be an issue, but my first and most immediate worries would relate to independence ("eye tracking" suggests serial dependence may be an issue, e.g. if you're observing given subjects over time), and to heterogeneity of measurements within conditions (e.g. of having a mix of subjects with different means).

If you can assume independence the biggest problems with using ordinary ANOVA with count data will be heterogeneity of variance (e.g. if you do have Poisson-like conditional distributions then the variance will be proportional to the mean), and if some of the means are small, you'll likely be dealing with variables that are quite skew and variables that may only take a few distinct values, which can also impact the behaviour of your tests.

If the mean counts are never small, the heterogeneity of variance might be reasonably dealt with by a Welch-Satterthwaite type approach, but - as long as the dependence concern I mentioned isn't actually a problem - I'd be inclined to consider GLMs (and related models), whether Poisson models or something else.

[If you want practical advice you should identify clearly what your main questions of interest are, and explain something about how these data arise (we don't necessarily all know your area) and how you expect the counts to behave within the conditions (with particular mention of the issues I expressed concern about). Indeed, properly you should ask such questions of statisticians before collecting any data - among other things, suitable advice might profitably impact how you organize your experiment.]

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.