# Beginner to prediction/statistics: Where do I start?

I sincerely apologize if there is another thread already that will answer this question. I'm so incredibly out of my league here that I don't even know what keywords to search for :-).

I'm a computer programmer by trade, and while I have a basic background in math, statistics was never really my cup of tea.

I currently work at a school and just finished developing a basic set of tools to help automatically collect and analyze data on our student's behaviors (this is a school for children with autism and other disabilities). So, we have a couple of year's worth of data for things like: given Billy, how frequently did he have Aggressions, Self-Injurious Behaviors, Drop, etc. Probably 6 - 10 "inputs" (I think that's the correct term) per student. We'll be adding more as well in the future.

What I'm curious about is this: Are there any beginning tutorials out there that might show me some interesting things to do with this data (besides just graphing it?) For example, it would be interesting to be able to predict when Billy is likely to have a long string of aggressions given that these x other factors have been increasing lately. Or, there is an increasing trend of this behavior which is way out of whack with its previous values, that should raise a big red flag.

I've been doing some basic Googling and this seems to be in the realm of "Statistical Data Mining"—some brief tutorials were found on Andrew Moore's site, but these just aren't detailed enough for me to really learn anything.

I realize that this is akin to someone walking into Stack Overflow and saying "Hey, tell me how to write the next Facebook." So, if these are the sorts of things that I can only do with years and years of statistical experience, just let me know and I'll be on my way. However, I also know that while someone couldn't walk into SO and write the next Facebook in a few weeks, we could probably point them in the right direction to create a basic site for their dad's business, even if it would be a pretty basic site. Likewise, I'm not looking to create a genius AI capable of predicting student behavior down to the millisecond; rather, I'm just curious if there's any low-hanging fruit that a guy like me could pick up in a few weeks or months of diligent reading that might make for some interesting uses of this new data we've unlocked.

I'm open to online tutorials, books, textbooks, videos, open source programs and libraries, etc.

-
You seem to want to do forecasting using time-series or longitudinal data; presumably modelling, not data mining. You have data for each student at multiple time points. Are these data collected at regular time intervals? What are your predictors like (the "inputs")? Are the outcome variables aggressions, self-injurious behaviours counts? Do they account for a given time period? Is drop binary (yes/no)? –  GaBorgulya Apr 28 '11 at 20:49
Well, I believe that doing forecasting is probably the more advanced of the two things I'd like to do. I'd be happy to start with just some basica anomaly detection outside of just doing standard deviation. To answer your questions, the data is collected at various time intervals although it is usually aggregated into uniform intervals (e.g. hourly, daily, etc.) Right now the predictors are the behaviors being tracked (aggression, self-injurious behaviors, etc.), although we will be adding more in the future along the lines of sleep, meals eaten, special incidents, medical information... –  Riley Dutton Apr 28 '11 at 21:00
I guess in theory I was wondering if the inputs and the outputs can be similar (e.g. can more aggressions predict more self injurious behaviors)? Drop is actually duration, but could be simplified down to frequency as well for these purposes. Again, I'm not at this point as concerned with our specific situation, I Just wanted to give a basic example to make sure I'm on the right track. I'd rather learn how to do this in general with some basic examples then hopefully become more proficient and apply it to our exact circumstances. –  Riley Dutton Apr 28 '11 at 21:02
You may want to take a look at a somewhat similar question: stats.stackexchange.com/questions/242/… –  GaBorgulya Apr 28 '11 at 22:43

"anomaly detection" -- This is usually called detection of outliers. You can find many references via googling.

"can more aggressions predict more self injurious behaviors" -- You can try one of the basic things: correlation between different variables or features (you called them "inputs").

The data will be easier to analyze if they will be in the format, say, "number of xxx incidents per week", i.e., if your variables will be measured on the same timescale.

-

It sounds like you have some really wonderful data to work with! One person suggested trying analyses in R, and that's definitely a powerful option. With your background in programming, it may be well-suited for you. I personally prefer a program like SPSS, which is built specifically user-friendly(ish) analysis of social science data. If you're new to the program, I'd suggest Julie Pallant's "SPSS Survival Manual," which has basic how-to instructions for most common analyses.

Regardless of the software, it sounds like using correlations, regressions, and some time-series work could help you investigate your variables. If it seems overwhelming to learn all the stats in a short period of time, I might suggest advertising the fact that you have data to work with. I'm certain that psychology undergrads or grad students at a nearby university would jump at the chance to help you do analysis and possibly publish any useful results. Best of luck!

-
Thanks for the great answer. Definitely gives me some specific terms to start Googling to hopefully start making heads and tails of all of this. –  Riley Dutton Apr 29 '11 at 13:25

One of my colleagues at StatSoft put this series of videos together.

http://www.statsoft.com/textbook/data-mining-techniques/