Hi everybody, I want to determinate the asociation between change of paternity (risk factor, nominal) and the presence of pre-eclampsia (nominal) in a large sample of women. My covariate is intergenesic period (IGP) years between the actual and the last pregnancy) and it has three categories: short IGP (less than 2 years), normal IGP (2-4 years) and long IGP (more than 4 years). I choose this as covariate because an important proportion of women who have changed partenrs usually have long intergenesic period (more than 4 years). This variable is something difficult to control to me as if I only select women with short intergenesic period my sample will be reduced dramatically. As we know, long intergenesic period favors the onset of pre-eclampsia.
Any idea of what kind of analysis I could perform? Obviously, odds ratio is my best option in this case.
Dear José,
in general I would always prefer an explicit model-based analysis over stratified parametric or non-parametric tests for statistical significance (whatever this means).
In your case, if you have a binary response variable (presence/absence of eclampsia), the most straightforward choice is a logistic regression model: you can include in the model any effect (confounder) that you think influences the outcome (e.g. change of paternity, IGP etc ...).
You of course need to check which model best describes the data you have at hand, since some effects may not be estimable if your sample size is not sufficiently large or you have severely under/over represented classes in some variables.
Odds ratios can be derived directly from the logistic regression model: recall that logit(x) = log(p(x)/(1-p(x))) = log(odds), odds ratios can be obtained from exp(beta) [just pay attention to which contrast you're interested in when dealing with multinomial/polychotomous independente variables, as is your case].
Hope this helps!
Filippo
Dear José,
in general I would always prefer an explicit model-based analysis over stratified parametric or non-parametric tests for statistical significance (whatever this means).
In your case, if you have a binary response variable (presence/absence of eclampsia), the most straightforward choice is a logistic regression model: you can include in the model any effect (confounder) that you think influences the outcome (e.g. change of paternity, IGP etc ...).
You of course need to check which model best describes the data you have at hand, since some effects may not be estimable if your sample size is not sufficiently large or you have severely under/over represented classes in some variables.
Odds ratios can be derived directly from the logistic regression model: recall that logit(x) = log(p(x)/(1-p(x))) = log(odds), odds ratios can be obtained from exp(beta) [just pay attention to which contrast you're interested in when dealing with multinomial/polychotomous independente variables, as is your case].
Hope this helps!
Filippo
Dear Jose,
I will answer that you can include it as covariate but you but it will be a dummy variate in which case one of the category will become a reference when interpreting the result.
Kola
Another option is effect coding. The main practical difference between that and dummy coding is that there is no reference group--the effects are interpreted relative to the "center." How many categories do you have in your predictor?
José, do you have the actual IGP values (measured as a continuous variable)? Or was the data collected in the categories you indicated (short IGP = less than 2 years, normal IGP = 2-4 years, long IGP = more than 4 years)? By creating categories, you are imposing a step function (i.e., people on either side of a cut-point could have very different fitted values, despite having trivial differences in IGP, and people at the extreme ends of a category have the same fitted value, despite fairly substantial differences in IGP). If you have the actual values, you might consider including it as a continuous variable. You might also look at IGP2 (and possibly IGP3) to check for non-linearity of the relationship between IGP and the log-odds of the outcome. I'd also check for any extreme values of IGP, which might be influential. (If your software allows it, you could look at Cook's distance to check for influential points.)
HTH.
Hi Jose,
As explained by Kolawole, a nominal variable can be used as covariate but interpretation of the results need some reference range, i.e. what type of labeling of the original variable (or dummy )is used.
- Badges
- Science topic