I need to analyze some data obtained from "in vivo" software testing. Unfortunately I’m a SW engineer, not a statistician.
The dataset contains observations for a number of messages sent between 2 nodes (actually, sent and returned).
There are two independent possible outcomes: 1. “message was flagged” and 2. “message lost integrity”. (I’m also interested in analyzing “1 OR 2” and “1 AND 2.”)
For each message, I have:
a) Binary, overlapping data, describing system features activated or not (for instance, the system may have been setup with the “total ordering” feature only, or “total ordering and encryption X” features, or none, etc.). There’s a set of 10 such features, activated or not, for each message. These parameters are completely independent from each other.
b) One parameter of categorical data: in which of 5 locations was the message analyzed.
c) Finally, I have four parameters for message category (from automated semantics detection), which might be set to “yes”, “no” or “couldn’t tell”. For instance: (urgent, yes) (coordination, couldn’t tell) (critical, yes) (illicit, couldn’t tell).
I’m trying to find out if there are combinations of these parameters that make it more likely for the possible outcomes to occur. For instance, activation of features 3 and 4 at location X for messages categorized as urgent and non-coordination predicts that messages loose integrity.
I’ve looked at some techniques and already tried chi-squared for independence for each variable. This approach, however, can’t be right: I need combinations of parameters that explain the outcomes. How should I go about getting such thing?
