Hi all, I am running into a snag creating a path analysis model using ordinal and multinomial logistic regression. I am using scaled scores as predictors of either an ordinal (OnlineSatisfaction, a scaled score that was an average of 6 survey items, value of 1-5) or categorical outcome variable (EnrollmentDecision: no enroll, enroll online, enroll on campus). The predictors are also scaled scores that are an average of the Likert scores of 5 items and 6 items respectively. The problem comes in that when I try to use the predictors in logistic regression, the crosstab of the variables indicates that over 50% of the cells end up with expected frequencies less than 5. Suggestions?
Why are you using logistic regression for ordinal or multiple-categorical data? You should be using ordinal logit or multi-logit models. That may not solve your problem, but at least you'll be using the correct statistical procedures.
I thought that the logit model was just another name for logistic regression? How do I run a logit model in SPSS?
It is the same. That is not the issue. The issue is that you have ordinal (or nominal data) and you are using a model designed for binary variables.
Apologies, I am still missing something. When you say the model is designed for binary variables, are you inferring the path model? I was under the impression that since I have two exogenous variables (one ordinal, and one nominal) that I would run an ordinal logistic regression and then a multinomial logistic regression (after checking to ensure that there are no independent/dependent combinations with 0 to ensure convergence) and check for model fit and significance before entering my data into a path model. Does path modeling only handle multiple regression or binary logistic regression exogenous variables? I am confused what the issue is. Thank you for clarifying.
You wrote: "The problem comes in that when I try to use the predictors in logistic regression, the crosstab of the variables indicates that over 50% of the cells end up with expected frequencies less than 5."
Given that you stated before that, that you have ordinal and categorical variables, I don't understand how you decided to use logistic regression or how that is being used. This is the thread I am trying to follow.
Apologies for not being clearer, I was planning on using ordinal and multinomial logistic regression, not one designed for binary outcomes. Also, by reducing my Likert responses from a 5 point to 3 point scale, I was able to eliminate 0 cell sizes in my sample for the main exogenous variable (categorical), but I still have cell sizes of 0 when comparing my predictors with the other exogenous (ordinal).
It sounds to me like you have a sample size problems here. You either may have more categories/variables than subjects within cells, or else you have cells with small samples and everyone has the same result. As a general rule of thumb, there should be no less than 7 observations per variable (this is obviously dependent om the type of variable and the distribution of values within each variable and the relationship between the variable and outcome). In any case, you are probably very low here on sample size, or very high on the number of variables. This is just a guess, but that's what it sounds like
Hi there, I also have the same problem and I think Ariel Linden;s last comment is correct. if is very difficult to run multivariate logistics regression with a small sample size or many covariates (predictors) or few predictors but with many categories because it will result in zero cell or low cell. One solution i think is to minimize the number of category in both predictors and dependent variables. (in logistics regression 2 for dependent variable and dichotomous covariates or continuous covariates if it has normal distribution (very unlikely). If you cannot use continuous variables (because of abnormality distribution), then divide groups based on published cut-off score or the median of that variable in your data. You also can reduce the number of covariats by looking at the frequency 2x2 table. if frequency of outcome observation is very similar between sub groups of predictor (for example, percentage of disease between male and female is 49% and 51%, and insignificant p >0.05), then consider removing gender from your multivairate analysis.
Hope this help, I think it is simple but quite time consuming..
- Badges
- Science topic