This site makes extensive use of JavaScript.
Please enable JavaScript in your browser.
Live
PTR
10.2.7
PTR
10.2.6
Beta
Statistics question - anyone good at interpreting regression models?
Post Reply
Return to board index
Post by
Squishalot
Ok, statistical problem to deal with that I'm being stumped by. I've run two regressions which get 'interesting' results, and I'm not 100% sure how to interpret it. Here goes:
I've got 50 observations, and a number of variables associated with them. Key ones - Variable A and Variable B - can take a value of either 1 or 0.
Regression 1)
I take a subset of observations that meet the criteria Variable A = 1.
I run a linear regression on Y and Variable B and find that Variable B is significantly positive. The intercept is significantly negative.
Conclusion 1 - When Variable A = 1, then then Variable B is important.
Regression 2)
I take a subset of observations that meet criteria Variable B = 1.
I run a linear regression on Y and Variable A, and find that Variable A is significantly positive. The intercept is
not significantly different from zero
.
Conclusion 2 - When Variable B = 1, then Variable A is important.
Problem Question
Can I use these results to say that the value of Variable A determines the importance of Variable B, and that it's not just correlation?
I've got my own thoughts on whether I can or not, but I'd like to hear other ideas, and at 11pm at night, noone else I went to uni with is online right now :P So to you denizens of Asia, Europe and the US, what do you think?
Post by
314938
This post was from a user who has deleted their account.
Post by
Squishalot
Random is random. What can I say? I can put a WoW related spin on it and post it in WoW General instead, if you want :P
Post by
Skyfire
The answer to your question: No. I assume this is basic statistics? What matters are r and r^2. It is obvious that there is a strong correlation, but causation? You tread on dangerous ground when you're talking causation in statistics.
Post by
TheMediator
The answer to your question: No. I assume this is basic statistics? What matters are r and r^2. It is obvious that there is a strong correlation, but causation? You tread on dangerous ground when you're talking causation in statistics.
This post. You said you have 50 "observations" - you cannot draw causation from anything except proper experimentation, so if you have results from a study, you cannot say that one causes the other since you have not controlled confounding variables (a third outside variable could just have easily manipulated the two variables you're looking at).
Post by
Skyfire
The answer to your question: No. I assume this is basic statistics? What matters are r and r^2. It is obvious that there is a strong correlation, but causation? You tread on dangerous ground when you're talking causation in statistics.
This post. You said you have 50 "observations" - you cannot draw causation from anything except proper experimentation, so if you have results from a study, you cannot say that one causes the other since you have not controlled confounding variables (a third outside variable could just have easily manipulated the two variables you're looking at).
Indeed, there is always potential for confounding and hidden variables, among other things. Something I had forgotten about, tbh.
Post by
Squishalot
Mediator, you're a psych graduate, aren't you? ;)
I know that regression = correlation =/= causation. What I'm wondering is that if there's a significant correlation from one side with a significant default effect, and there's a significant correlation from the other side with no default effect, it suggests that Variable B is not important unless Variable A = 1.
It's not quite 'causation', because I'm not saying that Variable A or Variable B 'causes' the resulting dependent variable result. I'm asking if I can say anything about Variable B importance being
conditional
on Variable A.
It's an observational study, and a lot of the variables have been controlled for. (Cheap shots, both of you :P At least trust that I'm setting up my experiment properly!) Possible confounding effects have been identified, tested and have no significant impact - not to say that they have no impact, but that including them in the regression doesn't change the result.
r and r^2 aren't as important in this instance, because the study isn't to provide for an accurate forecasting tool, but to identify explanatory factors by finding significant coefficients of predictor variables. Having said that, all things considered, the dependent variable is related to stock market data, and the fact that I'm getting r^2s of 25%, I'm actually pretty happy :P
Anyway, after posting, it prompted me to run interaction tests, which suggested that there is a conditional significance. So all is good.
Post Reply
You are not logged in. Please
log in
to post a reply or
register
if you don't already have an account.