I work with a new youth outreach program that's trying to reduce incidents of STDs in our city. So far we've been using a naive approach that target everyone equally, but this has been stretching us very thin with little results... However, it turns out we have access to a lot of data that could be quite valuable (I think). Unfortunately none of us really know how to use it - yet.
We have several datasets that I'd like to analyze, but there are three particularly promising ones:
- Sparse demographic data: Race, gender, age, and some academic data (truant days, GPA, etc). The parents can opt-in so we either have all or none of this data for a student (~80% opt-in). What data we have is completely accurate.
- Self-reported lifestyle survey: Scale questions (1-5) on different lifestyle choices. Roughly 100 questions per student covering everything from sexual activity to drug use, collected over multiple sessions. We have this data for every student in our program but it is self-reported. This is collected by a parent organization so we also have data for students that haven't even entered the program yet (!).
- STD results: Yes/No/Not-Reported indicating whether the student tested positive for some STD in the past 6 months. The answer is accurate if the parent opt to provide results, but a sizable portion declined to report (~35%).
All datasets have a common identifying label so we can track an anonymous student's results across the data. Roughly 2000 students have yes/no STD answers and appear in both the other datasets.
Given a large background dataset, how can we predict which background attributes correlate with test results? I've done some research but this is a new field for me.
Right now my thought is to join all the data into one big file, and then build two matrixes representing "Answers for students that tested positive" and "Answers for students that tested negative".
Then I was planning to use PCA to determine which questions/demographics are most important for that population. Hopefully the output will give us the most important features (eg frequent drug use as a junior or senior with no job means high probability for STDs) so we can pre-emptively target those populations.
Does this sound like a reasonable approach? Are there better methods of analysis?
What sort of literature should I be looking for if I want to learn more about this kind of analysis?