ANCOVA with dichotomous dependent variable - How?

#1
Hi everyone! I am a new subscriber with a little talent for statistics and I study developmental psychology :)

I need to perform and ANCOVA on a dichotomous dependent variable with SPSS. I’d like to compare 3 groups to see whether they differ in terms of rates for Conduct Disorder (CD), controlling for the effects of my covariates.

DV: conduct disorder – binary categorical variable 1: yes, 0: no.
IV – categorical grouping variable, 3 levels: Irritability
Group 0: children with no irritability
Group 1: children with either one subtype of irritability
Group 2: children with both subtypes of irritability.
Covariates: Age, Social Economical Status, IQ and Gender.

Throughout my reading, I have the vague idea that I am supposed to use the logistic regression, after dummy coding my IV. I tried to run it and I have specified the categorical nature of my IV on SPSS logistic regression, using then the automatic dummy coding. In my output I have two odds ratios, one is associated to my dummy coded group 1 compared to the baseline (group 0) and the other one is associated to my dummy coded group 2 still compared to the baseline (group 0).
They are both high and significant, meaning that having both or either one type of irritability is a risk factor for CD compared to children who do not display any symptoms.
Now my questions are the following: is this the correct procedure so far? Is there a way to compare my group 1 and my group 2 between them to see if having either one type of irritability or both is associated to a greater odds ratio for conduct disorder?

I hope this is clear, thank you so much in advance.
 

CowboyBear

Super Moderator
#2
Your procedure sounds sensible so far, but questions:

  • Why does Group 1 includes both subtypes of irritability (why not a group for each)?
  • How did you measure irritability levels? If this was measured with a continuous scale, it might be better not to collapse participants into groups

In terms of comparing groups 1 and 2 there are a couple of things you could do:
  • Change the indicator/reference level to Group 1
  • Or look at the confidence intervals for the ORs in your existing output - if they don't overlap there's clearly a significant difference (although if they do overlap that doesn't necessarily imply the difference isn't significant)
 
#3
Hi!

I am investigating tonic and phasic irritability. I'd like to know if having both tonic and phasic irritability is worst than having either tonic or phasic, also compared to children who had none (in term of CD diagnosis). Tonic and phasic irritability are categorical variables and binary coded.

Thank you for your suggestions, I have tried to change the baseline condition I am comparing my groups to and it looks like it makes sense. I have used my "either" group as a baseline and looked at the odds ratio associated with my "both" group.
However, I was reviewing my stats and I was concerned also because my results are not symmetrical. I know from my descriptive statistics that the rate of CD diagnosis are 3%, 7% and 10% for my "no irritability", "either one" and "both" group, respectively. Running the logistic regression using the "no irritability" group as a baseline, I have a sig. odds ratio associated with my "either" group but no sig. odds ratio associated with my "both" group (still compared to the baseline).
How is it possible?

Thank you
Silvia.
 
#4
I think I found the solution. Those percentages refer to the rate of CD diagnosis within group, without taking into account the demographic factors that dropped those estimate because of missing data. Indeed, if I run the logistic regression without covariates the odds ratios are symmetrical. I have checked the frequency of CD diagnosis per group considering the demographic factors and it is as following:
none: 3.2% (vs. 3.5%)
either: 9.6% (vs. 7.1% - demographic dropped the subjects in this group but not the cases)
both: 8.7% (vs. 10.1%)

that is why the difference is significant between either-none gropus and not between both-none groups. Anyway, I know at least I am using the right statistics.
Thank you.