If you are trying to calculate the adjusted OR for a study, you will likely need to run a multiple logistic regression model.
At present, I am working on a meta-analysis that observes the effect of physical activity for a disease (i.e. cancer).
For each study we have used adjusted (by age and sex) odds ratios (ORs) and for studies that do not present ORs, we have requested data from respective authors.
For one study, I have calculated the unadjusted OR.
I have no experience performing this and would like some support on how to adjust for age (continuous or categorical), gender (male, female) and smoking status (never, former, present)
I expect that it will be challenging but I would appreciate any instructions or advice on how to go about this.
Last edited by jle; 05-13-2015 at 07:08 PM.
If you are trying to calculate the adjusted OR for a study, you will likely need to run a multiple logistic regression model.
Stop cowardice, ban guns!
Thanks hlsmith, I've run the analysis and it looks good for the meta-analysis.
For studies that present unadjusted ORs - is there a way to adjust these figures for age and sex without raw data?
I have read some literature about external adjustments, however they are hard to decipher. Can summary statistics (proportions) of age and gender be used to adjust reported ORs?
Last edited by jle; 05-13-2015 at 10:03 PM.
Was age formatted as categorical or continuous in the other studies?
Stop cowardice, ban guns!
This tends to vary by study, for example one has categorised age (e.g. 45-49, 50-54, 55-59) and another has reported the mean and standard deviation for age (continuous).
Looking pretty doubt full. How was PA formatted?
Stop cowardice, ban guns!
That's understandable.
I was hoping there was a way around this. Any ideas (apart from requesting data)?
Physical activity is dichotomous (i.e. no physical activity, physical activity)
If you had a bunch of binary variables and counts for each unique combination of the groups you could do it. You could pull it off if you had groups with > 2 groups but it gets fairly tedious.
If you are not familiar with logistic regression how are you going to potentially pull off a model with random intercepts and slopes?
Stop cowardice, ban guns!
I have good experience with regression models in SPSS.
I try to learn different techniques online but there isn't a straightforward tutorial I can follow.
The main problem has been adjusting ORs without access to raw data.
If I were to have data for the frequency of persons by age group, gender, physical activity and disease outcome - adjustment is possible?
Can it be just be 2 groups or does it need to be greater than 2?
If all variables were categorical, and you had the counts for unique permutations, then you could run logistic with weights (e.g., 10 individuals: smoke = 1, PA = 1, age cat = 1, gender = 1; 8 individuals; smoke = 1 PA = 1 age cat = 1; gender = 0,..., 5 individuals smoke = 0, PA = 0 age cat = 0; gender = 0.
Say you had 4 binary groups like above (plus the binary dependent variable), you would need to know the counts for 2^5 groups (or the 32 unique permutations). And that is with binary variables. You could do it with non-binary variables as well, I just can figure out the number of permutations in my head.
Stop cowardice, ban guns!
Thanks hlsmith!
I really appreciate your input.
I'll have to see if I can extract any more data from supplement papers before speaking to statistician.
Tweet |