Bob Green
2007-Mar-26 20:19 UTC
[R] fitted probabilities in multinomial logistic regression are identical for each level
I was hoping for some advice regarding possible explanations for the
fitted probability values I obtained for a multinomial logistic
regression. The analysis aims to predict whether Capgras delusions
(present/absent) are associated with group (ABH, SV, homicide; values
= 1,2,3,), controlling for previous violence. What has me puzzled is
that for each combination the fitted probabilities are identical. I
haven't seen this in the worked examples I have come across and was
interested to know if this is a problem or what might be the cause
for this. I ran an analysis with another independent variable and
obtained a similar pattern.
Any assistance with this is appreciated
Bob Green
> predictors <- expand.grid(group=1:3, in.acute.danger =
c("y","n"),
violent.convictions = c("y","n"))
> p.fit <- predict(mod.multacute, predictors, type='probs')
> p.fit
1 2 3
1 0.4615070 0.3077061 0.2307869
2 0.4615070 0.3077061 0.2307869
3 0.4615070 0.3077061 0.2307869
4 0.7741997 0.1290310 0.0967693
5 0.7741997 0.1290310 0.0967693
6 0.7741997 0.1290310 0.0967693
7 0.4230927 0.3846055 0.1923017
8 0.4230927 0.3846055 0.1923017
9 0.4230927 0.3846055 0.1923017
10 0.7058783 0.1647063 0.1294153
11 0.7058783 0.1647063 0.1294153
12 0.7058783 0.1647063 0.1294153
> mod.multacute <- multinom(group ~ in.acute.danger *
violent.convictions, data = kc, na.action = na.omit)
# weights: 15 (8 variable)
initial value 170.284905
iter 10 value 131.016050
final value 130.993722
converged
> summary(mod.multacute, cor=F, Wald=T)
Call:
multinom(formula = group ~ in.acute.danger * violent.convictions,
data = kc, na.action = na.omit)
Coefficients:
(Intercept) in.acute.dangery violent.convictionsy
in.acute.dangery:violent.convictionsy
2 -1.455279 1.3599055 -0.3364982
0.02651913
3 -1.696416 0.9078901 -0.3830842
0.47860722
Std. Errors:
(Intercept) in.acute.dangery violent.convictionsy
in.acute.dangery:violent.convictionsy
2 0.2968082 0.5282077 0.6162498
0.9936493
3 0.3279838 0.6312569 0.6946869
1.1284891
Value/SE (Wald statistics):
(Intercept) in.acute.dangery violent.convictionsy
in.acute.dangery:violent.convictionsy
2 -4.903094 2.574566 -0.5460419
0.02668862
3 -5.172256 1.438226 -0.5514486
0.42411327
Residual Deviance: 261.9874
AIC: 277.9874
> Anova (mod.multacute)
Anova Table (Type II tests)
Response: group
LR Chisq Df Pr(>Chisq)
in.acute.danger 10.9335 2 0.004225 **
violent.convictions 0.5957 2 0.742430
in.acute.danger:violent.convictions 0.1895 2 0.909600
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05
'.' 0.1 ' ' 1
John Fox
2007-Mar-26 21:30 UTC
[R] fitted probabilities in multinomial logistic regression are identical for each level
Dear Bob, If I understand correctly what you've done, the "newdata" that you're using to get predicted values includes the three values of the response variable, which are irrelevant to the predictions and cause each prediction to be repeated three times. I hope that this helps, John On Tue, 27 Mar 2007 06:19:12 +1000 Bob Green <bgreen at dyson.brisnet.org.au> wrote:> > I was hoping for some advice regarding possible explanations for the > fitted probability values I obtained for a multinomial logistic > regression. The analysis aims to predict whether Capgras delusions > (present/absent) are associated with group (ABH, SV, homicide; values > > = 1,2,3,), controlling for previous violence. What has me puzzled is > that for each combination the fitted probabilities are identical. I > haven't seen this in the worked examples I have come across and was > interested to know if this is a problem or what might be the cause > for this. I ran an analysis with another independent variable and > obtained a similar pattern. > > Any assistance with this is appreciated > > Bob Green > > > predictors <- expand.grid(group=1:3, in.acute.danger = c("y","n"), > > violent.convictions = c("y","n")) > > p.fit <- predict(mod.multacute, predictors, type='probs') > > p.fit > 1 2 3 > 1 0.4615070 0.3077061 0.2307869 > 2 0.4615070 0.3077061 0.2307869 > 3 0.4615070 0.3077061 0.2307869 > 4 0.7741997 0.1290310 0.0967693 > 5 0.7741997 0.1290310 0.0967693 > 6 0.7741997 0.1290310 0.0967693 > 7 0.4230927 0.3846055 0.1923017 > 8 0.4230927 0.3846055 0.1923017 > 9 0.4230927 0.3846055 0.1923017 > 10 0.7058783 0.1647063 0.1294153 > 11 0.7058783 0.1647063 0.1294153 > 12 0.7058783 0.1647063 0.1294153 > > > > mod.multacute <- multinom(group ~ in.acute.danger * > violent.convictions, data = kc, na.action = na.omit) > # weights: 15 (8 variable) > initial value 170.284905 > iter 10 value 131.016050 > final value 130.993722 > converged > > summary(mod.multacute, cor=F, Wald=T) > Call: > multinom(formula = group ~ in.acute.danger * violent.convictions, > data = kc, na.action = na.omit) > > Coefficients: > (Intercept) in.acute.dangery violent.convictionsy > in.acute.dangery:violent.convictionsy > 2 -1.455279 1.3599055 -0.3364982 > 0.02651913 > 3 -1.696416 0.9078901 -0.3830842 > 0.47860722 > > Std. Errors: > (Intercept) in.acute.dangery violent.convictionsy > in.acute.dangery:violent.convictionsy > 2 0.2968082 0.5282077 0.6162498 > 0.9936493 > 3 0.3279838 0.6312569 0.6946869 > 1.1284891 > > Value/SE (Wald statistics): > (Intercept) in.acute.dangery violent.convictionsy > in.acute.dangery:violent.convictionsy > 2 -4.903094 2.574566 -0.5460419 > 0.02668862 > 3 -5.172256 1.438226 -0.5514486 > 0.42411327 > > Residual Deviance: 261.9874 > AIC: 277.9874 > > Anova (mod.multacute) > Anova Table (Type II tests) > > Response: group > LR Chisq Df Pr(>Chisq) > in.acute.danger 10.9335 2 0.004225 ** > violent.convictions 0.5957 2 0.742430 > in.acute.danger:violent.convictions 0.1895 2 0.909600 > --- > Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 > > ______________________________________________ > R-help at stat.math.ethz.ch mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide > http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code.-------------------------------- John Fox Department of Sociology McMaster University Hamilton, Ontario, Canada http://socserv.mcmaster.ca/jfox/