Hi, I have 3 groups of data. SGA: 95% birth weight. 50 SGA and LGA cases were selected from a cohort, and matched to 50 AGA cases according to ethnicity. Hormones levels were measured. I can't find any significant associations within each group. Can I run a multiple regression with these 150 cases altogether with hormones as independent variables to predict birth weight? How can I test it is reasonable to do that?
I don't see anything wrong with running multiple regression in principle. But in this particular case, trying out different methods after the first test didn't give the result you were hoping for could be considered data fishing / p-hacking.
Dear Shutan, If I understood your question, why not?
The worst case, I guess, is you may end up failing to find significant predictors if some other important independent variables (ethnicity, for instance, or any other that I'm not aware of - age o mother, number of previous birth's, just guessing...) play a major role in the outcome and are not included in your regression model. Then you would have a "large" amount of non-explained (residual) variance that, among other issues, will make difficult to establish statistical significance for the predictors included in your model.
You are assuming that the weight at birth varies linearly with the hormone level, either positively or negatively correlated. Is there the possibility of not being so? (for instance reaching a plateau?)
Just curious, how many hormones (level's) you have in your data set?
Hope it helped.
Hi Shutan. Yes. As your dependent is in ordered categories, once you have merged all together, you need an ordered logit model. Then simply add your other variables in on the right hand side, including hormones, and run the model. Good luck. Cheers, Marc
Thank you for all your replies. The key issue is that someone criticises doing regression on a grouped design. Of course, the models had been controlled for maternal age, ethnicity, socioeconomic status, smoking and drinking habits, maternal BMI, infant sex and gestation at delivery. Maternal placental growth hormone, IGFs and IGFBPs were measured. It's a merely stats question. I found significant correlations in these 150 cases, but couldn't find any significance within each group (probably due to the smaller sample size). The paper is ready to submit. I'm not sure everyone is happy with the stats method I used.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression