Date: Sun, 5 Dec 1999 05:53:53 -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: Presentation of logistic and linear regression
Content-Type: text/plain; charset=us-ascii
I've had recently to present logistic regression results to a lay
audience, and can understand Staffan's plight. I chose to explain the
gist of the procedure in verbal form, made a reference to the
classification table an the pseudo R2 as measures of the analysis
goodness of fit, and then discussed the most significant log reg
coefficients in terms of "how much is the impact of an increase in X on
the probability of Y, other things being equal?" I evaluated the effects
of one independent variable for people with average values in other
independent variables.
For this I generated graphs of the logistic regression equation
obtained, keeping all other indep variables at their meamn values, and
showing the relation between p(Y) and X in the manner chosen by
Herrnstein and Murray to present their results in The Bell Curve.
These graphs are not a standard SPSS result: I used Excel (one column
for X, and another for the value of the log reg equation for varying
values of X while keeping other independent variables constant at their
means).
Below are some specific observations on Staffan's message.
Staffan Lindberg wrote:
>
> I have done some logistic regressions with approximately a dozen
> independent variables in each, all binary coded except age that has 4
> levels. David Nichols and John Hendrickx kindly explained the difference
> between simple and indicator contrasts. Now I have a new headache.
> Are the results in SPSS (I am
> running 6.1.4) univariate or multivariate ?
If you have more than one independent variable, they are
multivariate in the usual sense of the word. But your dependent variable
must be a dichotomous one. In later versions, SPSS introduced logistic
regression for dependent variables with more than two categories.
>Are both necessary ?
I do not think this question has any sense. Univariate log reg is used
when the problem involves only one independent variable, and
multivariate when there are more, as with ordinary linear regression.
>Should
> confidence levels for Exp(b) be presented ?
I guess so. Exp(b) is the log odds ratio, and their CI may be useful to
grasp the decisiveness of the conclusions reached.
>Is there some form of commonly
> accepted table layout analogous to an ANOVA table ?
No.
> These questions also pertain to some linear regressions I made. In this
> context I also wonder what is commonly presented, the R Square or the
> adjusted R Square or both ?
I ordinarily present both (sometimes adding a brief explanation of the
"adjusted" coefficient). For lay audiences, it may help to state that
both R2 reflect the proportion of variance in the dependent variable
that is explained by the other variables in the equation.
Remember, though, that log reg can yield deceptively high R2 values
when you have many variables relative to the number of cases. The
"actual" number of variables includes k-1 dummy variables for every
categorical independent variable having k categories.
Hector Maletta
Universidad del Salvador
Buenos Aires, Argentina
| 
