[S] Diff logit reg output b/w sas & Splus

Edward Malthouse (ecm@casbah.acns.nwu.edu)
Fri, 8 May 1998 09:06:36 -0500 (CDT)


I run Splus version 3.4 Rel 1 and SAS version 6.12 on an HP9000. I
am fitting logistic regressions using proc logistic in SAS and the
glm function in Splus. I get the same null deviance and deviance,
but the parameter estimates, standard errors, and p-values are
different. Has anyone else had this problem? Should I trust one
over the other?

Splus output
-------------------------------------------------------------------
> fit$null.deviance
[1] 945.4133
> deviance(fit)
[1] 792.6851
> fit$call
glm(formula = dvt ~ A + B + C + D + E + F + G + E:F, family = binomial,
data = gary, na.action = na.omit)
> summary(fit)$coefficients
Value Std. Error t value
(Intercept) -2.5270691 0.2063189 -12.248367
A 1.8408719 0.2019749 9.114358
B 1.4581433 0.2330308 6.257299
C 0.6755337 0.2689039 2.512175
D -1.4304647 0.5843410 -2.447996
E 0.7328473 0.2621257 2.795786
F 0.5814527 0.2379988 2.443091
G -0.4415661 0.2277794 -1.938570
E:F -0.7757377 0.3743964 -2.071969
> anova(fit, test="Chisq")
Analysis of Deviance Table

Binomial model

Response: dvt

Terms added sequentially (first to last)
Df Deviance Resid. Df Resid. Dev Pr(Chi)
NULL 998 945.4133
A 1 82.46387 997 862.9495 0.00000000
B 1 38.94716 996 824.0023 0.00000000
C 1 7.41460 995 816.5877 0.00646967
D 1 9.16642 994 807.4213 0.00246496
E 1 3.78895 993 803.6323 0.05159214
F 1 2.76857 992 800.8638 0.09613187
G 1 3.86951 991 796.9943 0.04917087
E:F 1 4.30913 990 792.6851 0.03790838

-------------------------------------------------------------------
SAS output
-------------------------------------------------------------------
Intercept
Intercept and
Criterion Only Covariates Chi-Square for Covariates

AIC 947.413 810.685 .
SC 952.320 854.846 .
-2 LOG L 945.413 792.685 152.728 with 8 DF (p=0.0001)
Score . . 166.560 with 8 DF (p=0.0001)

Analysis of Maximum Likelihood Estimates

Parameter Standard Wald Pr > Standardized Odds
Variable DF Estimate Error Chi-Square Chi-Square Estimate Ratio

INTERCPT 1 -2.5272 0.2073 148.5899 0.0001 . .
A 1 1.8410 0.2025 82.6674 0.0001 0.370551 6.303
B 1 1.4583 0.2335 38.9958 0.0001 0.256729 4.299
C 1 0.6755 0.2695 6.2829 0.0122 0.110839 1.965
D 1 -1.4382 0.6177 5.4202 0.0199 -0.205153 0.237
E 1 0.7329 0.2631 7.7595 0.0053 0.193257 2.081
F 1 0.5815 0.2390 5.9220 0.0150 0.159901 1.789
G 1 -0.4416 0.2287 3.7303 0.0534 -0.103627 0.643
EF 1 -0.7759 0.3756 4.2677 0.0388 -0.156959 0.460

-------------------------------------------------------------------
The parameter estimates are nearly the same, but differ in the third
or fourth decimal place. Likewise for the standard errors. The
P-values are alarmingly different. For example, predictor F is
almost highly significant in the SAS output (P=.0150), but not
significant in the Splus output (P=.0961).

The predictors are all dichotomous (0=no, 1=yes), the variable EF in
the SAS output is EF=E*F, and the data were imported into Splus with
the sas.get command.

The condition index of the design matrix is only 5.2.

Thank you for your help.

Dr. Edward C. Malthouse
Assistant Professor
Integrated Marketing Communications Department
Medill School of Journalism
1908 Sheridan Road
Evanston, IL 60208-1290
Tele: 847-467-3376
Fax: 847-491-5925

-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news