I am using generalized linear models with one response variable and 6 predictors (1 covariate and 5 factors). I want to assess the effect of smoking on my response variable. When I split my participants in sex groups and then in three smoking groups (smokers, ex-smokers, non-smokers) I find weaker associations than when I add smoking in the model (as a factor since it is a categorical variable) and split the participants in sex groups.

Could anyone please explain why is there this difference and what would be the best practice in this model?

Thank you!

ps: when I split in three smoking groups, the associations are similar per smoking group, they are just weaker.

  • $\begingroup$ What is your sample size? It could be that that the sample size is not very large and now you have essentially created 6 groups. The effects may be real but different by gender. What are the 5 factors? Keep in mind that 3 genders with three smoking groups makes 3x2=6 groups. $\endgroup$ Commented Jul 23, 2012 at 10:40
  • $\begingroup$ I have 32,484 (20,008 women and 14,619 men) people. But when I run the models, it says that only about 5,ooo are included (probably because of missing values in some covariates). My dependent variable is body mass index and the factors are alcohol drinking frequency, alcohol amount, economic status, index of deprivation, physical activity level. $\endgroup$
    – Vasia
    Commented Jul 23, 2012 at 10:45
  • 3
    $\begingroup$ The fact that most of your observations have missing data is worrisome. $\endgroup$
    – Peter Flom
    Commented Jul 23, 2012 at 11:21
  • 1
    $\begingroup$ Shouldn't 20,008 women + 14,619 men = 34,627 people, instead of 32,484? $\endgroup$ Commented Jul 23, 2012 at 13:49
  • 1
    $\begingroup$ It sounds like you tried your model 2 ways, entering smoking first or entering sex first, and found 2 different patterns of results. Is this the basis for your question? $\endgroup$ Commented Jul 23, 2012 at 14:02

2 Answers 2


After your comments I think the sample size is not a big issue. But like Peter I would be concerned about the missing data and understanding why so many samples have missing information. If your software won't let you fit some data because if missing covariates see which covariates are missing.

If it is just a couple that cause you to lose so much data drop them and then fit the model. Now many observations will enter the model because they will not have missing covariates. Maybe based on what you observed smoking doesn't really have much effect and any relationship you might see with a model that relate smoking to BMI where smoking is the only covariate may be because it is the smoker that drink that tend to be obese. I have seen many a smoker that is pretty thin.

Compare a model that includes smoking, alcohol and an interaction term between the two and also look at a model with alcohol alone. If the more detailed model is not doing much better fit to the same data then maybe it is okay to drop the smoking groups.

But before taking any of these recommendation get to understand better why covariates are missing in so many cases. Maybe you can looked at the dropped cases and compare the demographics (gender, age etc) for the fitted sample with what it is for the missing data. There could be bias due to covariate imbalance. Vance Berger's book which I have previously cited here on CV could help you with that.


If I am following you, you are doing one of two things: 1) Controlling for what you are trying to analyze. In the non-smoking group, there can be no association between smoking and BMI (or anything else). 2) Comparing the three groups (non-smokers, ex-smokers, current smokers) somehow, after separating them. I am not sure how you would then find associations: At most you could find different levels of BMI.

  • $\begingroup$ My goal is to find whether there are associations between alcohol and BMI (body mass index). I have to control though for potential confounding factors (such as economic status, index of deprivation, physical activity level, age) and one of them is smoking. Apart from alcohol and the confounders that I always add, when I put smoking as a factor my associations are similar compared to when I do not put smoking. When I separate in smoking groups (and ran the generalized linear models with BMI as the dependent variable and all the rest as factors and covariates) I find much weaker associations. $\endgroup$
    – Vasia
    Commented Jul 23, 2012 at 12:38
  • $\begingroup$ So, I am getting very confused about what is the best practice. I hope that makes more sense. $\endgroup$
    – Vasia
    Commented Jul 23, 2012 at 12:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.