Hi Daniel

it is a little bit tricky since I neither know the items nor what exactly you an "Impact on purchasing behaviour" (e.g. whether you more interested in what triggers a purchase or whether you're more interested in different consumer habits). Anyway, you might find that "factor analysis" (perhaps followed by cluster analysis) is a good approach. You could be able to differentiate between differentiate between types of buying (factors) and types of buyers (clusters). You find a quite good and short introduction to FA here:

https://www.theanalysisfactor.com/factor-analysis-1-introduction/

Hope this is a little helpful

Best, Alex

Alexander Ruser I have information about people total money spend on virtual items and then I have 15 variables from likert scale statements where people replied to series of statements about why they are buying these virtual items, there were simply 15 statements and people picked if they StronglyAgree Agree etc. that this factor influences their purchasing. From means of likert scale I can see what people think that influences them and I would like to know if people who picked for example StronglyAgree in one of the statements spend more money than the others plus if that statement is statistically significant. I would like to build some model, could simple linear model work? Thanks for your reply and I will check the factor analysis too

I am not sure what you mean by using factor analysis to convert your "Likert scale data into factor data with 5 levels." If the factor analysis shows that all 15 variables measure a single factor, then the simplest approach is to add together the 15 variables. If you want to get back to a score that ranges from 1 to 5, then divide that total by 15, which will give you a continuous variable instead of one with 5 "levels."

David L Morgan Thank you for your reply I really appreciate it Professor. "By Likert scale data into factor.." i ment that I converted their data type in RStudio to ordered factor (ordinal variable). I apologize that I did not specify better. There are 5 factors measured by the 15 variables (each by 3). So you would recommend to merge them to just five factors? I am sorry if my questions are stupid but I have never worked with likert scale data, so I am a bit lost.

If you have a hypothesis that your five factors are each measured by 3 variables, then the best way things is to test that via Confirmatory Factor Analysis and then evaluate the entire regression through Structural Equation Modeling.

However, since you are new to this kind of analysis, that approach is probably too advanced. Instead, I would assess the reliability of each scale using Cronbach's alpha. If they are all reliable (usual cut-off is .8 and above), then I would sum the items in each scale and then examine a correlation matrix among the five resulting scales. The issue here is multicollinearity, because if the scales are too highly intercorrelated, it will create problems in estimating the OLS coefficients.

If the scales are highly uncorrelated (say .6 and above), then you can sum the five sub scales into one total scale.

If I understood correctly, you have 3 questions in which the answers are in the Likert scale (1 - 5). These are basically 3 discrete variables. Since your dependent variable is continuous OLS is certainly applicable. You can indeed do a regression with these as factors. I usually prefer to convert them into indicator variables (set one of the levels, say 1, as your base label) and define an indicator for each of the other four levels (say, for example, for level 2 the indicator would be equal to 1 if the level is 2 and zero otherwise). Then, in your model, you will add those four indicators to capture the differences contributed by those levels relative to the base level. I am sure you can do this in other ways but this way gives you a lot of freedom and more importantly for me, I can interpret the meaning of those parameters. Of course, the most important thing is that the sign of the parameters and their magnitudes make sense (for this you may want to perform several hypothesis tests, not just of whether each parameter is significant but also whether the parameters for say levels 4 and 5 are statistically significant of each other). Without knowing anything else about your problem I cannot say much more. The only issue with this is that you end up with a model with a lot of variables and parameters (the model is still quite simple but many people would be intimidated by it). I hope this helps and that I am not adding more confusion.

Hi Daniel,

It seems that you’ve started working without properly defining variables under study n objective of your study. Let me first tell you few basics; data analysis technique to be used depends on the above two. So clearly define your objective of research, dependent n independent variables (number of variables as well as type of theirs- metric or non metric). Then I’ll be able to answer your query better

Also I want to mention that factor analysis (again two types EFA N CFA) is a statistical technique by which you can get few factors from statements of your questionnaire (basically dimension reduction).

Choosing or changing data type, in R studio or for that matter in any software, is different.

I know my answer might not be very clear to you in the first place but I would suggest you to go through https://m.youtube.com/watch?v=l2NkP-HJRPc and the next video in the series for better understanding.

Best Wishes !

Hi Daniel:

You might begin with an exploratory factor analysis. If you get a two (or more) factor solution, you need to do some serious thinking about the items that form factors regarding whether or not you want to include them as variables. This is a theoretical problem. You should not combine items that belong to different factors in the same scale.

Assuming you are now analyzing your first, or only, factor, you can either use an index or a latent variables. If you use an index, a Likert-type scale, you should look at the distribution of scores and transform your index, especially if it is a dependent variable (meaning it has one or more arrows coming into it), so that it is roughly normal. If you dichotomize all items, you can do a Guttman-scale or cumulative scale measure.

If you prefer to use a latent variable, use EQS, SAS, or some other structural equations modeling program to assess the fit of your items to a single latent variable. Statistics such as the root mean square residual and the goodness of fit index will give you insight into whether your items measure a single dimension. If you do this, use maximum likelihood estimation rather than least squares so you get t-tests for the significance of the item coefficients. Delete non-significant items, one by one, until all left are significant.

Good luck with your project, Best regards, Warren

Good Question