2

I am performing a MANOVA and when I use split files I get significant interactions. However if I don´t split the file I do not get any significant interactions.

I am interested in knowing how attention to visual stimulus is affected under two different conditions. But I am specifically interested on how this happens in people that have different scores on resilience and anxiety. Therefore I have recoded resilience and anxiety scores into high and low, and I have split the file in order to run the MANOVA. Is that allowed?

mpiktas
  • 35,099
Natalia Giraldo Holguín
  • 47
  • 1
    What are your dependent variables? What do you mean you split the file? – StevenP Jul 24 '15 at 13:55
  • My dependant variable is attention latency means (from dot probe task- time latency to congruent stimulus trial and time latency to incongruent stimulus trial) my independant variable is condition. I split the file before running the analisis to separate it by cases. Therefore the output shows comparison between possible combinations between high-low resilience and high-low anxiety. – Natalia Giraldo Holguín Jul 24 '15 at 14:08
  • 1
    So you have two measurements of attention latency (congruent and incogruent) for every participant? What do you mean you split the file to separate it by cases? Did you split your sample in four sub-sets for every low-high anxiety-resilience combination? Sorry for the questions I am trying to understand what you are after – StevenP Jul 24 '15 at 14:17
  • 1
    I split the file by resilience and anxiety to observe the interaction between attention latency and condition by groups (high resilience- low anxiety, low resilience- high anxiety and so on) sorry I am not being clear enough.. – Natalia Giraldo Holguín Jul 24 '15 at 14:51
  • I do have two measurements of attention latency for every participant. – Natalia Giraldo Holguín Jul 24 '15 at 14:52
  • You don't need to split your dataset, you should just examine the interaction Anxiety*Resilience on each of the dependent variables. If it's significant you could run a post-hoc test to see which of the four possible combinations are significantly associated with lower/higher mean attention latency scores. – StevenP Jul 24 '15 at 15:10

1 Answers1

1

So basically you are performing a MANOVA to examine mean differences in two types of attention latency across two stress and two resilience conditions, as well as their interactions.

To answer your question, is this allowed? Yes, it's allowed. Is it the best possible way to address your problem? Probably not!

There is a certain reason why you run a MANOVA and not two between-subjects 2-way-ANOVAs. It's the same reason why you shouldn't brake down your sample in four subsets according to the interaction level your subjects belong to. This is reason is minimizing Type I error. For every additional test you run, you introduce a new source of Type I error.

Instead, run a MANOVA as you suggest and interpret the main effects of stress, resilience and their interaction. SHould you get a significant p value for the interaction term run post-hoc tests to identify which combination of conditions differ significantly from each other with respect to your dependent variables. Keep in mind that even if the p value of the main effect of the interaction term is significant (p<.05) there is a chance your post-hoc test might not identify any of the conditions as differing significantly from each other with respect to the mean latency values. Should this be the case, it will coincide with your observation of no significant findings after splitting your sample in 4 sub-sets.

StevenP
  • 524
  • I don't get any significance when I run the Manova with the condition and resilience and anxiety scores as between subject factors. But if I run the analysis only with condition as between subjects factor (spliting the file) I do get significant interactions between condition and attention larency for high resilience and low anxiety. I was wondering if I was allowed to run the analysis like that. – Natalia Giraldo Holguín Jul 24 '15 at 18:19
  • 1
    Can you check what's happenning with your degrees of freedom (df)? I suspect that in your MANOVA model they are lower than the ones in the single 2-way ANOVAs. It could very well be that when you run the MANOVA you run it only for those with complete data on both dependent variables. – StevenP Jul 27 '15 at 09:07