Hello,
For my dissertation project, I want to look at the relationship between two predictor variables with a dependent variable.
Independent variables- social media activity use (Likert scale, ORDINAL DATA) and personality traits (will get a score for each of the five personality traits, continuous data)
Dependent variable- depression scores (score data, continuous)
My query is what is the most appropriate regression analysis to use. Ideally, I would like to sed a multiple linear regression. The only issue I’m having is that one of my INDEPENDENT/PREDICTOR variables is ORDINAL and my other predictor variable and dependent variable are numerical/continous data.
Does anyone have any advice on what regression to use?
Thank you!
An ordinal scale can be coded numerically and treated as continuous on average; however, as the measure is not interval, I would use sensitivity analysis to compare multiple coding schemes.
For example, the ordinal scale: poor, good, and excellent might be codes as (1) 1,2,3 or (2) 1,2,4 or (3) 1,2,8 and them comparing the significance of the contributing factors.
For additional inside to the relationship that exist among these variables:
(1) Consider each 'independent variable' as a function of the ordinal measure coded by dummy variables
(2) Compare and contrast the means of the 'independent variable' by score on the 'independent variable'
(3) Consider the relationship between the 'independent variables' without a 'dependent variable' using non-response analysis (a subset of implicit regression)
Do you have any information about the relations between the independent variables and dependent variable. If you don't have use the nonparametric regression.
A lot depends on how you measured your variables. In particular, there is a major difference between a single Likert-scored item (which is ordinal) and a set of items that are combined into a Likert scale (which can be treated as interval, if it has a high level of internal consistency, i.e., reliability as typically assessed by Cronbach's alpha).
Common practice is to treat such predictors as continuous. This might be OK - particularly for validated scales summing over many items. There are various alternatives:
1) Models that treat predictors as ordinal are not straightforward, but do exist (e.g., dominance based methods). Probably not recommended for everyday use as they are difficult to run and interpret.
2) Empirical logistical transformation - this could be seen as a kind of halfway option between ordinal and continuous.
3) Ordinal coding of dummy variables (e.g., staircase coding). Inefficient but would work (it would just eat up df very rapidly - especially with interactions).
4) Treat the predictor as monotonic. This is effectively modelling the variable as a 0/1 dummy variable with fixed or variable break points. This is a very new approach and probably the most interesting as it is genuinely ordinal, pretty efficient and relatively easy to interpret. As far as I know this is only easily run in the R package brms. See: https://psyarxiv.com/9qkhj/ and https://cran.r-project.org/web/packages/brms/vignettes/brms_monotonic.html
- Badges
- Science topic
- Similar topics
- Social Science
- Media
- Social Media