Good morning everyone. These are the doubts I have. I have collected data from 2 groups of children. One is a group of children diagnosed with a disorder (data entered as 1). Another is a group of children who are healthy (data entered as 0). The groups are matched on age and gender. In both groups I have measured various continuous variables like child temperament (it has 5 dimensions like emotionality, sociability, etc) and parenting style (it has 3 dimensions- warmth, punitive, neglectful). Now my doubts are as follows:
1. Can I keep this objective (To estimate the risk of disorder in children with a specific type of parenting style and child temperament), and use binomial logistic regression to come up with odds ratio, and report whatever variable is significant as a risk? Is this the right approach? Or are there any other statistics to find risk.
2. I have two categorical variables (marital discord, and family history of mental illness) both are entered as 0,1. When I ran chi square test between the two groups, I found a significant difference between the groups. Do these two variables become confounding factors? Should I control for them when I do T test or regression analysis, which means I have to use Mancova instead of t-test? Or should I just report them as chi square test results and leave them out of the rest of the analysis?
This is for my thesis, and I appreciate any help I can receive.
I thank you for your valuable time and feedback.
Hello Sahithya,
Ideally, it would be good to match your cases on as many relevant variables as possible. Realistically, going beyond three or four becomes a real challenge unless your potential pool of study cases is very large.
To the specific questions:
1. Yes, LR could be a useful analytic technique for your study. To the extent that you successfully include nuisance variables as part of the data collection, at least you are able to invoke statistical control for them.
2. Finding that a couple of personal/family characteristics are related to group membership suggests that these, too, should be IVs in your study. I don't see that as implying mancova as a design, however.
Finally, you'll need to acknowledge that some of your variables, because you are most likely collecting them well after the group membership has been established (and not before), may or may not have a causal relationship to subsequent identification. Family history, not so much a temporal concern. Marital discord, quite possibly (since it might have arisen from, or been exacerbated by, the disorder).
Good luck with your work.
Hello Sahithya,
Ideally, it would be good to match your cases on as many relevant variables as possible. Realistically, going beyond three or four becomes a real challenge unless your potential pool of study cases is very large.
To the specific questions:
1. Yes, LR could be a useful analytic technique for your study. To the extent that you successfully include nuisance variables as part of the data collection, at least you are able to invoke statistical control for them.
2. Finding that a couple of personal/family characteristics are related to group membership suggests that these, too, should be IVs in your study. I don't see that as implying mancova as a design, however.
Finally, you'll need to acknowledge that some of your variables, because you are most likely collecting them well after the group membership has been established (and not before), may or may not have a causal relationship to subsequent identification. Family history, not so much a temporal concern. Marital discord, quite possibly (since it might have arisen from, or been exacerbated by, the disorder).
Good luck with your work.
You are quite likely on the right track as Prof. Morse indicates. You may be interested in a similar study in a different field. It is attached. Best wishes, David Booth PS it may be helpful if you can identify your goal: prediction or explanation. It appears to me to be prediction and so some of the variable selection methods in the attached may be helpful. In addition the logistic regression used in the attached would probably not be appropriate for you. Replace that by Multivariate multiple regression and you should be OK.A good link to start with may be:
https://www.bing.com/search?q=multivariate+lasso+regression+in+r&qs=RI&pq=lasso+in+multivariate&sk=RI1&sc=2-21&cvid=12FEE83183F04918BDB348DB49B28F7D&FORM=QBRE&sp=
BTW I suggest multivariate regression instead of Mancova because the regression coefficients will be the risk factors you want.
Good luck, David Booth PS another good link:
https://www.bing.com/search?q=Lasso+with+multivariate+regression&qs=n&form=QBRE&sp=-1&pq=lasso+with+multivariate+regression&sc=1-34&sk=&cvid=BDFF5223123641A98FBFB37288397018
Logistic regression analyses are good options. I give you an example of marital discord and how it affects young persons:
David Morse thank you so much for your valuable inputs and the time you are giving me. As suggested, I will go ahead with logistics regression. The only reason I am talking about t-test and MANCOVA here is because since it is a PhD thesis, I have to go about data analysis in a step by step method before I arrive at the regression. So I will have to first report, mean, SD for both the groups and say that these variables differ (or not differ) between the groups. Therefore, I was considering t-test. (I may not have to follow this procedure when I am publishing the paper). This is the reason I am asking if I can replace t-test with MANCOVA to report the differences between the two groups among various variables, while controlling for confounding variables (which t-test does not permit)?. Or shall I ignore the confounding variables (Marital discord and family history) in this analysis, go with t-test and use those confounding variables only during last regression. I am not sure of the right methodological way of doing this. Thank you once again.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression