Hi everyone and thank you for you advice.
I'm running a two-way repeated-measure ANOVA, with two groups of subjects undergoing two different treatments (coded as: 0; 1) x 3 assessment times (coded as: 0;1;2. 0 is the baseline/pretreatment time). The dependent variable is a count variable, so I'm using a Poisson distribution.
The question is whether the two treatments produce a different change in the dependent variable at time 1 and 2 (compared to baseline).
I'm running the analysis with the data in long format using generalized linear models. My analytic strategy is, in order:
1) Looking for a general significance in the group X time interaction in the ANOVA table.
2) Looking in the coefficient table for the significance of group x time1 and group x time2. Considering the coding I did, they should indicate respectively the treatment differential effects at time 1 and 2 compared to baseline.
3) Making intra-group post-hoc comparisons among the three times (with sidak of bonferroni correction applied). This for both the two groups.
Do you think this is a correct strategy or you have something better to suggest?
Thank you,
Massimiliano
It is very likely that the data do not follow a Poisson distribution, so you will want to use a model that deals with overdispersion.
A Poisson regression model with overdispersion (or a generalised negative binomial regression) will allow you to cluster the observations within participant.
Rather than making post-hoc comparisons, why not test your experimental hypotheses directly using regression? All post-hoc tests are, to some extent, poor substitutes for a model that accurately reflects your hypothesis.
It is very likely that the data do not follow a Poisson distribution, so you will want to use a model that deals with overdispersion.
A Poisson regression model with overdispersion (or a generalised negative binomial regression) will allow you to cluster the observations within participant.
Rather than making post-hoc comparisons, why not test your experimental hypotheses directly using regression? All post-hoc tests are, to some extent, poor substitutes for a model that accurately reflects your hypothesis.
Ciao, you maybe should consider to use further attributes and run a configural frequency analysis of the prediction type (Ct0, A, B predict A,B-Ct1 A,B-Ct2) on a general linear background. Depends on the n=? - more promising with small ns than (fantasy-)distribution approaches. See my goose manifesto with useful links.
goffredo (wcbp2015)
http://www.biposuisse.ch/biomath
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA