Date: Thu, 24 Feb 2000 20:08:31 -0300
Reply-To: hmaletta@overnet.com.ar
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: "Hector E. Maletta" <hmaletta@OVERNET.COM.AR>
Subject: Re: multiple regression
Content-Type: text/plain; charset=us-ascii
Hector Musacchio wrote:
> When I calculate bivariate correlations, the variables A and B
> are correlated with an r=0.44 (positive). In the multivariate
> regression, adding another 3 or 4 variables (and method enter),
> results that A is an independent predictor of B, but with a
> NEGATIVE beta !!
> Is this possible? Thanks in advance.
Of course it's possible. The positive bivariate correlation is totally
explained by the other variables in the second equation. Controlling for
those other variables, as you did with your multiple regression
equation, the relation between B and A has a negative sign.
Once I found a similar situation in a study of Latin American women. It
was approximately as follows. The dependent variable was fertility
(number of children), the independent variables were education attained
by women, household income, and labor force participation (whether or
not the woman was employed or seeking work). Lower income, lower
education and no employment all were positively related to fertility
(bivariate correlations were all negative). Controlling for education
and employment, the net effect of income was surprisingly positive,
meaning that for women with the same education and labor participation
values, the number of children increased as income grew higher. In
particular, for women with no college education that did not work
outside their home, higher income meant more children. The overall
negative bivariate correlation of income and children was completely
explained by education and labor market participation. Many similar
situations appear in texts such as Hubert Blalock's Causal Inferences in
Nonexperimental Research. In short, the sign of a correlation can be
changed, or the correlation reduced to zero, by introducing control
variables.
Hector Maletta
Universidad del Salvador
Buenos Aires, Argentina
| 
