Greeting, I am designing a survey to test 7 independent variables (IV's) impact on 2 dependent variables (DV's)
My draft questionnaire has variables measured on Likert scale which would result in ordinal data.
I have been advised to consider converting the DV's to dicotomous variables (would recommend/would not recommend) and (would use the supplier again/would not use the supplier again)
My question is would binary logistical regression be appopriate where the IV data is ordinal and the DV data is dicotomous?
If I stuck to keeping all variables as ordinal, would ordinal logistic regression be appropriate?
many thanks for any assistance
Binary Logistic Regression is used when your DV is dicotomous, irrespective of what scale your IVs are. So in BLR you can have either binary or ordinal or continuous IV (or combination of these).
Ordinal Regression is used when your DV (not IV) is in ordinal scale
Logistic regression can handle ordinal data as dependents or independents. The number of categories doesn't matter. However when a binary value is used for dependent and indeoendent, it will.give a nice odds ratio to work with. When dependent (or independent) is linear and other is binary, it gives a t value. If it has more categories it gives ANOVA which gives less information. Whwn both a linear it gives a correlation coefficient.
thanks for the feedback. As I am still designing my study I am in the fortunate position of being able to design my DV's as ordinal or catagorical. Are there any particular advantages of binary logistic regression over standard multiple regression other than less stringent requirements for normal distribution of the DV data? After digging around I was shocked to find out how may cases are required for BLR. If I have 7 IV's the best case scenario is 150 cases needed. But this could go up dramatically if the the responses to the DV are not 50/50 yes/no. for example I calculate that if either yes or no responses are say 20% with the other response 80% then I need 350 cases which will be unacheivable.
Linear scale data as raw number always have more precision, than when they are categorized or ranked. If we are dealing with linear scale variables, linear (+/- multiple) regression is always precise than logistic regression. However when it comes to ordinal or categorical data, only logistic regression is possible.
If your data items don't fall on a standard normal distribution, Spearman's rank correlation is an option. (https://en.wikipedia.org/wiki/Spearman%27s_rank_correlation_coefficient)
I am a little late to the party but since you are in the design phase let me suggest that you consult the attached notes. Whether a variable is a DV or IV is generally not a choice of the researcher but a consequence of the research question. Best, David Booth PS read these closely. you on a very slippery slope here as some of your comments show.
Hello,
I'm not sure if this fits the original question. I am trying to model a 10 point likert scale variable and wonder about "measures" to compare the fit of a linear, binomial (dichotomize my DV) or ordinal regression.
The variable goes from never justifiable, 2, 3, … to always justifiable and is highly skewed to the left so that the existing literature dichotomized it into a 4 or 2 point variable.
The results from binary probit regression and linear regression are roughly equal but the ordinal regression is not. (I know about the assumptions of proportional odds in the latter case.)
Would you advice power analysis in this point or are there other methods?
Thanks in advance. Best, Luise
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R