0
$\begingroup$

I am recording animal behaviour in four different conditions but in the same environment. So my data are several values (the number of which vary) representing the frequency of a single behaviour for each of the four conditions.

Previously people have used a one-way ANOVA to analyse this data, however, I have realised that the data are not actually normally distributed. So I thought of using a Kruskall Wallis test, but the data are also of repeated measures. Then I thought of using a Friedman test but the data groups (conditions) have unequal numbers of values...

If anyone knows of a test I can use for this data I would be very grateful for some help.

Improve this question
$\endgroup$
6
  • $\begingroup$ Maybe I don't understand your question properly, but couldn't you just use a normal chi^2 test? $\endgroup$ Commented Nov 8, 2012 at 14:53
  • $\begingroup$ Hi and thanks for the quick reply. I can run a chi square test on the average values - can you run a chi square on averaged values? Or do you mean I can run a chi square regardless of group size etc? Thanks, Roddy. $\endgroup$ Commented Nov 8, 2012 at 14:56
  • $\begingroup$ I thought it might help if I provide an example; for one of my data sets in condition 1 the values are: 0.36, 1.49, 0.02, 0, 0, 0.19, 0, in condition 2 the values are: 0, 2.32, 0, 1.03, 0, 0, 0, 0, 0 in condition 3 the values are: 0.49, 3.33, 4.54, 4.022, 6.70, 4.45, 7.52, 3.37, 0.53, 0, 0.78, 3.57 and the values in condition 4 are: 0, 0, 0, 0, 0, 0, 1.66, 6.9, 0.52, 5.67, 0.52, 0, 0. Obviously in condition 3 the behaviour was much more frequent but I can't find a statistical test to show this! $\endgroup$ Commented Nov 8, 2012 at 15:11
  • $\begingroup$ How about running ANOVA on rank-normalized data? See en.wikipedia.org/wiki/ANOVA_on_ranks $\endgroup$ Commented Nov 8, 2012 at 15:37
  • $\begingroup$ I have just tried ranking and two groups are still showing as statistically significant under a Kolmogorov-Smirnov test... I should mention that I'm using SPSSv19 $\endgroup$ Commented Nov 8, 2012 at 15:56

1 Answer 1

Reset to default
1
$\begingroup$

Some possibilities (though I still am not sure of the actual details):

If the response is a frequency then you generally want to use a binomial or poisson outcome for counts, but your sample data is not integers, is it counts divided by time or something similar? If so you could still use a poisson regression model to analyze the data (generalized linear models, glm). If there is concern about the dependence in the data then generalized estimating equations (gee) or generalized mixed effects models (gmm) extend the poisson regression this way.

A flexible non-parametric method is permutation testing (many of the other non-parametric methods are special cases of permutation testing).

I don't remember enough about SPSS to give more specific suggestions.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Hi and thanks for your suggestions. SPSS doesn't do permutation testing, but, the GLM sounds like something that could work for me. My values are frequencies (Hz). I can run a generalized linear model but in SPSS I am a bit overwhelmed by settings. I have set it up so I am running a poisson loglinear model, the response (DV) is my frequency values, the predictor is what condition (1 2 3 or 4). But SPSS gives me a couple of other choices - one tab is labelled 'model' and another is labelled 'EM means' I don't know what I'm supposed to do here... Thanks again for any help, Roddy. $\endgroup$ Commented Nov 9, 2012 at 15:20

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.