I want to find what is the impact of race on poverty controlling for income, unemployment, and gender. Do I have to apply hierarchical regression analysis?
Abu, that sounds like a fairly standard regression analysis, so no, you don't need hierarchies and can just do a normal OLS regression, with poverty as your dependent variable and income, unemployment, gender and race as independent variables. However, if your data are hierarchically structured (eg you have people nested in places, or data over time with occasions nested in people), then you would need to do something more hierarchical. I'm not an SPSS user but either way should be pretty easy to do - google should tell you the answer if you can't work it out yourself.
I agree with Frei, on the other hand yiu could also first decide if poverty is a categorical or continuous variable. This might help determine the precise model structure you might need to specify in addition to the already suggested OLS regression analysis.
Chiming in on some of the answers. "Hierarchical" has (at least) two meanings in linear/multiple regression analysis. The one Andrew is referring to is sometimes called a mixed model -- where you have sources of variance at more than one level. I agree with Andrew in that it doesn't sound like that's your situation.
I *think* you're asking about what Frei was talking about, which refers to entering predictors in a sequential fashion. I rarely use that technique, preferring to enter all predictors simultaneously. In that case, the test of significance for each predictor is as if it was entered last in a hierarchical approach.
However, I'd also add that income and unemployment are virtually guaranteed to be highly correlated, so including both as predictors is going to result in inflated standard errors, so you'll want to interpret with caution.
I'm not an SPSS user, so I'm not sure exactly what you mean. "Stepwise" is sometimes used in the same sense as "hierarchical," but often means an automated selection procedure for predictors.
The latter is *never* recommended for drawing substantive conclusions -- it's been shown conclusively to be extremely sensitive to minor variations in the data, and the results can not be trusted. About the only use it has, to my mind, is if you have a small sample and are concerned about using more df than you have to for covariates, only then might you use maximum-R^2 form of automated selection.
HI, This may help. A reference that may help re: Hierarchical Modelling; "The Role of Conceptual Frameworks in Epidemiologic Analysis: A Hierarchical Approach". Victoria Cesar et al.International Journal of Epidemiology. Vol 26 No 1.