Are you using linear regression with a dichotomous DV? Binary logistic regression would surely be a better choice? Assumptions of constant error variance and normal error distributions will be breached.
I'm performing multiple regression for an international relations thesis regarding various predictor variables and the outcome of intervention in civil wars. My ability to interpret the analysis is limited. I used RegrissIt add-on for excel, and dummy coded my variables to dichotomous ones and ran the macro for several different combinations of variables. For example, I used independent variables of biased (whether the intervention was supporting one side vice neutral) and multilateral with the dependent variable of success. I got the following from the equation:
R sq. 0.147, adj. R sq. 0.093
Coefficients:
Constant .525
Biased -.541
Multilateral .464, all with p <0.05
From what I've read, it would seem to me, when biased is constant, if multilateral=1, then there is a 46.4% chance success=1. Am I understanding that correctly? What exactly does it mean for biased to be constant? Does that mean if biased is .525, then multilateral would have a 46.4% effect on success. What does that mean for a binary variable? How then do I interpret adjusted R sq.? Is that saying if both biased and multilateral=1, then there is a 9.3% chance success=1? If you could clear this up for me I'd greatly appreciate it.
Are you using linear regression with a dichotomous DV? Binary logistic regression would surely be a better choice? Assumptions of constant error variance and normal error distributions will be breached.
Matt aka CB | twitter.com/matthewmatix
Let denote the dependent variable then you're linear model is:
Insert the value 1 for "multilateral" to get:
Insert the value 0 for "multilateral" to get:
now take the difference between the equations to get:
for any - constant - value of biased. So if biased is kept constant then a change in multilateral from 0 to 1 increases the probability of y=1 by 46.4 percentage points.
In a linear probability model not all calculated probabilities for the sample are necessarily between 0 and 1 (one drawback of a linear probability model).
datorres85 (06-19-2016)
Thanks, that helps. So, including some of my other variables into the regression, I do get a >1 value for Pr(y=1). How do I interpret this? I can't say the likelihood of success is 143%. What about the R sq. value?
Last edited by datorres85; 06-19-2016 at 06:49 PM.
Maybe start by just having a google around to find a resource that explains it in a way that makes sense to you? To compute, I'm not sure you can use Excel - Excel isn't really a great program for statistical analysis - but almost any dedicated statistical analysis program or statistical computing language has this as an option (SPSS, SAS, R, etc.)
Matt aka CB | twitter.com/matthewmatix
Tweet |