Hi,
Do you mean you have very wide range of confidence interval for the estimate of coefficients in your logistic regression?
I would think of the problem of (quasi) complete separation in logistic.
Which software did you use?
- If you used R, it would have some warnings like: "Fitted probabilities numerically 0 or 1 occurred"
- If you used SAS, it would stated clearly that there is a problem of (quasi) complete separation and the parameter estimates should not be trusted.
If it is the problem of separation, you could look up on the internet to read more abut it, and how to deal with this problem.
Hope it help.
Hello,
You should also check your number of patients in this group. A small number of patients leads to a very large IC.
cordially
Thank you Ms. Thao, Mr. Bello and Prof. Stephen for your invaluable replies.
that may be as a result of large variance of estimates, possibly caused by presence of multicolinearity for OLS regression.
Answer from Prof. Joan Vaccaro !!
No, it sounds like an inflated value. This can be do to several causes. One could be having two variables that are highly correlated. Another reason could be outliers. Do some exploratory statistics such as box-plots to find if you have outliers. Another reason could be magnitude differences of the variables that you are comparing. How data is coded could cause an inflated odds ratio. You may have some variables that are continuous, but not normally distributed. For example, many biomarkers such as hemoglobin A1C, cholesterol, blood glucose, have outliers and once these are remove, still need to be transformed (usually by natural logarithm) in order to achieve linearity and normality. I suggest you check your independent variables using exploratory statistics. You can use the KS test for a single variable (non-parametric tests) to see if your data is normally distributed or not. Once you transform a variable, you will have to repeat the KS test to see if it is normally distributed. If you cannot achieve normality/linearity for a variable, then you may have to categorize it. Also check all independent variables that are continuous by correlation. Any correlation 0.8 or above may be considered too highly correlated and one should be dropped.
Is an odds ratio of 243200659 when comparing variables in logic regression normal?
- Badges
- Science topic
- Similar topics
- Statistics
- Survey