Logistic regression or DFA?

[Merlion]

Hi!

I am trying to get my masters thesis done. I am taking a statistics course right now, but I need to get started on things sooner rather than later.

Basically, I am looking at a treatment study data. Subjects underwent randomization to two conditions. Sample size is 153. I am trying to identify predictors of treatment outcome. I previously attempted to run logistic regression on the data, but I think I need to double check things before I move ahead. My questions are:

1. Can I do this in SPSS or AMOS? I worked with SPSS before, but not AMOS.
Should I use structural equation modelling?
2. My outcome variable is presence or absence of diagnosis at posttreatment, but also reduction in symtoms severity from pre to post. I am still not clear if logistic regression is the correct approach for diagnosis/no diagnosis as the outcome?
Another student I work with suggested DFA? Which technique is better?

For the reduction in symptoms from pre-post (measured based on self-report), I read a paper where the authors used reliable change indeces (RCs). How does one obtain RCs? I see from that articles that there is an interval to guide if RC is statistically significant. Then it looks as if I need to figure it if change is clinically significant? Some other studies regressed the predictors on the raw outcome score, which does not seem appropriate. How does one go about to regressing the predictors on the change indices? How will I split sample in improved/not improved based on RCs?

I am trying to figure this out on my own. I am taking a stats class now, but it will take some time to learn the answers to these and other questions, but I need to make progress on the thesis, so any help would be welcome.

Thank you.

 

[TheAnalysisFactor]

Hi!

I am trying to get my masters thesis done. I am taking a statistics course right now, but I need to get started on things sooner rather than later.

Basically, I am looking at a treatment study data. Subjects underwent randomization to two conditions. Sample size is 153. I am trying to identify predictors of treatment outcome. I previously attempted to run logistic regression on the data, but I think I need to double check things before I move ahead. My questions are:

1. Can I do this in SPSS or AMOS? I worked with SPSS before, but not AMOS.
Should I use structural equation modelling?

You can do a logistic regression in SPSS, but not AMOS. Nothing you've said indicates SEM is appropriate.

2. My outcome variable is presence or absence of diagnosis at posttreatment, but also reduction in symtoms severity from pre to post. I am still not clear if logistic regression is the correct approach for diagnosis/no diagnosis as the outcome?
Another student I work with suggested DFA? Which technique is better?

Neither is better. Just appropriate for different situations.

A few things to keep in mind:
DFA requires that all the variables that make up the discriminant function have multivariate normality. That means no categorical predictors, like treatment.

Logistic Regression will also allow you to include interactions between predictors. DFA will not.

For the reduction in symptoms from pre-post (measured based on self-report), I read a paper where the authors used reliable change indeces (RCs). How does one obtain RCs? I see from that articles that there is an interval to guide if RC is statistically significant. Then it looks as if I need to figure it if change is clinically significant? Some other studies regressed the predictors on the raw outcome score, which does not seem appropriate. How does one go about to regressing the predictors on the change indices? How will I split sample in improved/not improved based on RCs?

I am trying to figure this out on my own. I am taking a stats class now, but it will take some time to learn the answers to these and other questions, but I need to make progress on the thesis, so any help would be welcome.

Thank you.

I don't know anything about RCs. It might be unique to your field. Many fields have statistics that they use frequently that no one else has ever heard of.

Karen

 

[Merlion]

If I may ask additinal questions:

1. On some variables I have more than 10% missing data. I understand that I can inpute missing data using AMOS. Do you know of any link where it explains how to do this?
2. I understand that I need to check for multicollinearity. Do you have any resource that explains how this is done (preferably a detailed one) ? I read this as a start to get an idea
http://cscu.cornell.edu/news/statnews/stnews65.pdf
3. Do I need to check for power? I understand that if I include too many predictors, the power decreases. I was told that I do not have power to study moderation effects. My sample has 153 subjects-half in each condition. Would power be an issue if I consider the whole sample, or the tx conditions separate? Both conditions have been designed to test if changes in an active component (different in the 2 conditions) mediate treatment outcome, so I don't have any reasons to expect difference between two. Both would lead to change by through different mechanisms.
4. Are there any other preliminary analyses that I should do?

I know these are basic questions, but any help is welcome.

 

[TheAnalysisFactor]

If I may ask additinal questions:

1. On some variables I have more than 10% missing data. I understand that I can inpute missing data using AMOS. Do you know of any link where it explains how to do this?

I don't, but I wrote this handout a few years ago. I have been hesitant to make it publicly available, because I haven't checked it against the new version of AMOS. But it might help. http://www.analysisfactor.com/resources/AMOS-handout.pdf.

But just to clarify, AMOS doesn't impute missing data. It's even better. It uses Maximum Likelihood to work around the missing values. It's too difficult to explain here, but it works as well as Multiple Imputation without the work of setting up the imputation model.

But, unless this has changed with newer versions, you're limited to normal-errors regression models. No logistic regression.

2. I understand that I need to check for multicollinearity. Do you have any resource that explains how this is done (preferably a detailed one) ? I read this as a start to get an idea
http://cscu.cornell.edu/news/statnews/stnews65.pdf

That's a good article. If you go to http://www.analysisfactor.com/learning/teletraining1.html, you can get a recording of a teleseminar I recently did on multicollinearity. There is also a handout, which includes a list of resources.

3. Do I need to check for power? I understand that if I include too many predictors, the power decreases. I was told that I do not have power to study moderation effects. My sample has 153 subjects-half in each condition. Would power be an issue if I consider the whole sample, or the tx conditions separate? Both conditions have been designed to test if changes in an active component (different in the 2 conditions) mediate treatment outcome, so I don't have any reasons to expect difference between two. Both would lead to change by through different mechanisms.

You don't really need to check for power. If your results aren't significant, then you don't have enough power. That sounds like a decent sample size, but it depends on your model as a whole.

4. Are there any other preliminary analyses that I should do?

I know these are basic questions, but any help is welcome.

It's always good to do descriptives, but beyond that, I don't know enough about your study to recommend specifics.

Karen