I need help clarify something. I had about 300 youths to investigate the association between perceptions and treatment adherence. I categorized age into three: children, young adolescents, and older adolescents. When I enter the data in SPSS and did multinomial logistic regression (three outcomes), there was no interaction detected, but when I stratified them by three age groups, the effect size as well as significance was very different among three groups. I just wonder why I could not detect the interaction. Is it possible that I did not have enough sample for the detection? Or am I talking about different things? Thank you for your help.


TS Contributor

If you could provide a bit more info on your study design:

  • list the independent variables and dependent variables,
  • exactly what you changed when you say "stratified," and
  • which two variables interacted / did not interact,
then we can help. Unfortunately right now it's not entirely clear.

Hi John, thank you. I recruited 300 children and adolescents (7-19 yeas old) from an existing cohort of parents whose children had this chronic illness. I mailed questionnaires to them. Items were psychosocial items and designed to measure their perceptions about certain things about their illness and others. The outcome was adhere, neutral, and poor adhere. I was going to use regular logistic using two outcome variables, adhere or poor adhere, but the advisor suggested that I use three. So, we chose to use multinomial logistic regression, but basically it is the same as regular one. So, it is a cross-sectional survey. I entered the data into SPSS and click interaction between age and all the perceptions variables, but no interaction was detected, but since I was not convinced that all the perception variables influence all the three groups equally towards poor adherence (or adherence), I entered each group separately into SPSS and did the the same anaysis. I compared all three analyses and found that OR of the significant variables (found in the initial analysis) was significant only in adolescents. None of the variables were significant predictor for poor adherence in children. But this should appear as an interaction in SPSS, I thought. Thank you, and I hope I made myself understood.
I forgot to mention that children were those aged 7-11, younger adolescents were those aged 12-14, and older adolescents were those aged 15-19. Thank you. I also should mention that I divided such because of Piaget's theory of cognitive develoment (12+ is an adolescent).


TS Contributor
I do recall that it's difficult to pick up interactions in logistic regression unless the sample size is huge. I would present the data both ways - one with the absence of an interaction, and then also show that certain effects are different within the age categories. It's possible that although there is a difference between the age groups, it wasn't "different enough" to show up as a significant interaction.