ANOVA or Chi Square for Nominal Non Parametric Data

[DeeD]

Dear TalkStat Forum members,

I was hoping to get some advice over whether I am using the appropriate test for data that I have.

The notes of 70 patients were analysed and 2 variables (documentation of postive symptoms yes/no and negative symptoms yes/no) were noted on 4 occasions.

An intervention was performed (teaching session) and then a second cycle was performed looking at a different set of 35 patient notes.

• From the table of results it is clear that positive symptoms were recorded more frequently than negative symptoms. Would a Chi Square test be the most appropriate to show that this difference was significant.

• What is the best test to see if the intervention was significant? My research suggests a Chi Square test or a form of ANOVA, but would it be best to use the mean frequency (positive symptoms yes/no and negative symptoms yes/no) on each occasion?

• And finally other general factors were recorded Drug use (Yes/No), Illness duration (Less than 5 years or over 5 years) and Level of Person documenting (5 possible outcomes). Would the best solution be a series of Chi-Square analyses to see if this had an effect? Again would it be best to use the mean of each occasion

I have tried to look at previous forums threads and looked through the internet and hopefully am on the right track. But confirmation/alternative advice would be gratefully appreciated.

Many Thanks for any help/advice offered
Dee

 

[noetsi]

Chi square requires comparing two variables and I don't see two variables here (that is a variable predicting another variable). What is your two variables you are actually comparing? I suspect you could use a test of proportion if you wanted to compare how often something occured in one situation to another.

Repeated measure anova is a common way to compare change over time. Your dependent variable has to be interval and it is not clear to me what your dependent variable is. If it is frequency that something occurs than I would thing that works (although there are many ways to test this). You can use ANOVA to test a percent of something. You can not use it to test something coded yes no (as a dependent variable).

If you have a series of predictors you don't want to run a series of one predictor test. You want to put them all in the model at one time. If your dependent variable is interval than ANOVA, or possibly linear regression should work. If it has two levels than logistic regression is probably your best bet. If it has three or more levels than ordered logistic regression is probably best.

 

[DeeD]

Thanks for the advice Noetsi!

Taking your advice on board. I assessed the data again. I have tried to simplify what I need to do. In essence what I would like to do is: (Not sure if I need to ask this on a new thread though!)

Compare the percentage of positive symptoms documented and against the percentage of negative symptoms documented firstly within the same group. And then comparing it between two sample groups. (Before and after a teaching session on documentation - they are two completely different sample groups with different numbers of patients)

And then compare what effect various variable had on the percentage of symptoms documentated. (Grade of staff (nominal), Length of stay (interval))

Would a test of proportion and ordinal linear regression still be appropriate.

Many Thanks again
Sorry for essentially changing the question!
DeeD