Poisson Distribution Analysis in SPSS - Nonparametric, count, repeated-measures data

In relation to a recent post about what analyses to conduct for a data set, I am now asking a related question about the test to run in SPSS.

Background information on the data:

• Repeated measures design with count data
• Participants responded about the number of intrusions they had after exposure to two conditions.
• Data ranges from 0 to 4, with mostly zeroes, and thus is positively skewed.
• Thus, I will want to do an analysis for Poisson distributed data.

Question

What analysis would be ideal to carry out for the above data?

My take on issue

1. One idea was to carry out a generalised linear models analysis with a Poisson loglinear model.

• However, the issue here seems to be that there is only one dependent variable that can be inserted but I have two (see image below).

Image 1 - Generalised Linear Model with Poisson Loglinear Model - only one DV option

1. Another idea was to find the difference between the two variables, followed by dividing the answer by the square root of the number of values (in this case 2) to produce a composite score of the two variables.

Image 2 Image 2 - Univariate ANOVA - is it right for count, non-parametric data?

After I ran a univariate ANOVA which produced significant results (see image above). However, the problem with this idea is that:

• (a) I do not use a Poisson distribution.
• (b) If we consider the count data as being non-continuous as the data is non-normally distributed then an ANOVA should not be used.
• (c) In general I am not sure if producing a composite score so that I can run a univariate ANOVA is fine.

Update of information following original post

The DV is number of intrusions so yes you are right that it is actually only one DV I just realised meaning perhaps I didn't give a clear explanation (which I will fix in my original post after clarifying I understand). I had the DV of number of intrusions and each participant being exposed to a condition of low levels of information (low load) and high levels of information (high load), and then I wanted to see which condition was associated with the development of more intrusions. I coded the data into two separate columns for # of intrusions (low and high load).

2. Elaboration on the study design

• The study explored whether watching a film, while remembering more stimuli (high load), presented earlier before the film, led to less intrusions than remembering less stimuli (low load).
• The DV was number of intrusions after watching the film.
• The IVs were low and high load which were determined by the number of stimuli a person had to remember whilst watching a film.
• Thus low load (IV 1) should lead to more intrusions (DV) than high load (IV 2).
• All subjects were exposed to both conditions.
• I thought your dependent variable was just # intrusions. Are confusing dependent with independent (two binary variables for the experimental conditions?) – AdamO Aug 10 '15 at 19:34
• Can you rephrase the scientific question with specific attention to low/high load intrusions? Perhaps you are interested whether the intervention may decrease total overall intrusions and, secondarily, decrease the proportion of intrusions which are considered high load? – AdamO Aug 10 '15 at 19:54
• Trevor you need to edit your question with these details rather than having a long iterative comment sequence. We don't have time to comb over references on the field. The issue of how your measures relate to the hypothesis is still completely nebulous. Something needs to potentially increase/decrease as a result of the intervention... some combination of your two count measures. Take a look at the analysis wiki for some tips and pointers stats.stackexchange.com/tags/analysis/info – AdamO Aug 10 '15 at 20:32
• Okay AdamO, I have now done this - please let me know if I should change anything else. I will comment on your most recent point by adding the extra information into the post now. – Trevor Dubois Aug 10 '15 at 20:42
• If you have a repeated measures situation use Generalized Estimating Equations menu. It is the same procedure, basically, but with RM and some other additional options.Note however that there exist again only one field for the dependent variable. That means that you will have to restructure your "wide format" RM-dataset into "long format" leaving you only one dependent variable column. Use Restructure variables into cases menu for that job. – ttnphns Aug 11 '15 at 10:23