# Regression with an unknown dependent variable - estimating “likelihood” to do something

This probably seems like a really strange question, but let me try to explain what I want to do; hopefully it will make sense.

I have a data set with a couple dozen variables, such as age, level of education, self-assessed (via a Likert scale) measures of technical ability, experience, willingness to share personal information, etc etc. Using this data, I want to estimate "likelihood of participation" in using a not-yet-developed social network. Since the site doesn't exist yet, I don't have a simple dependent variable with a "1" or "0" indicating whether they've used it or not, so I can't take the usual approach with regression and estimate what impact each variable had on their participation.

Is there some way to "create" a dependent variable that models "likely vs. not likely" (or something like that), and do the regression against that? I'm really lost ... would be open to any ideas.

"Data! Data! Data! I can't make bricks without clay."

              Sherlock Holmes


Sorry, you just can't do it. If you make up some measure that combines some of your variables, then guess what: those exact variables will be predictive of it.

I agree with @Aniko that you won't be able to "predict" anything without an outcome. Now @Andy suggestion makes sense, provided you find a users database sharing similar characteristics. As an example of related studies, I guess you might find interesting google hits on users' characteristics in studies on Twitter, Facebook or other social networks or community driven sites. Here is a couple of references that I just found: Comparing community structure to characteristics in online collegiate social networks (Traud et al., arXiv:0809.0690); Social Computing Privacy Concerns: Antecedents & Effects (Nov and Wattal, CHI 2009).

Aside from users' characteristics, willingness to particpate or contribute to a social network is also a function of its characteristics (number of members, contact frequency, content quality, etc.).

Finally, you still can do some kind of exploratory analysis on your data set. Specifically, a cluster analysis would help to highlight groups of individuals sharing similar response profiles. Another approach would be to use Multiple correspondence analysis to uncover potentially interesting patterns in your data. Those features might be related to already published results in a further step. Obviously, you can't regress an outcome that is still to be observed from this, but you will certainly get a better idea of how structured your data are.

The simplest approach would be to do a literature search and see if someone else developed a model with your variables, and then generate predictions based on those prior models. Another possibility would be to make your own model from another data source that does have the required information, and then use that to generate predictions.

Obviously you can't do much with your data besides generate predictions with my suggestion, although that may be all you want from the context of your question.