Hi all,
I've been trying to fit a mixed effects model and I've been having problems.
=>My aim:
to know whether states atributes, political parties and individual atributes
affect the electoral results of men and women candidates.
I use candidates as replications for states and for political parties.
=>Response: Percentage of valid votes casted to each individual.
(alternative response: electoral result (binomial: elected, non-elected)).
=>Function used to fit model: lme from nlme package.
=>Data Organization:
Data is unbalanced. 4946 observations. Each row in the data frame has info
on the individual candidates (educational level, marital status, sex, age,
party, etc) and on the states in which they run for election (literacy
rates, district magnitude, number of voters, urban population rates, sex
ratio, etc). Individual atributes vary within state and party%in%state.
State atributes vary between states. Though political parties are present
across states, I have reasons to believe that they behave differently in
each state.
Example:
State ID Literacy% #Voters/#Candidates
SexRatio Individual ID Sex Age
EducationalLevel Party Etc.
AC 84.5 8241.92
102.56 1 M 42
Undergraduate PT ...
AC 84.5 8241.92
102.56 2 F 35
Undergraduate PL ...
AC 84.5 8241.92
102.56 3 M 55
Undergraduate PMDB ...
DF 96.6 15593.38
89.64 4 M 40
PostGraduate PSDB ...
DF 96.6 15593.38
89.64 5 F 39
Undergraduate PSOL
RJ 96.0 15363.96
88.42 6 M 63
Undergraduate PT
RJ 96.0 15363.96
88.42 7 M 52
Undergraduate PMDB ...
=>Questions:
1) I've found that the random effects of the 'state' level have a very
low standard deviation.
>depfed.lme3<- lme(Votes~Sex, data=depfed.frm, random=~1|State/Party)
>summary (depfed.lme3)
(...)Random effects:
Formula: ~1 | State
(Intercept)
StdDev: 0.0001089504
This suggests that there is no significant advantage on using 'state' alone
as a grouping factor, I suppose. On the other hand, there is considerable
variation in the effects of parties inside states:
(...)Formula: ~1 | Party %in% State
(Intercept) Residual
StdDev: 0.9904834 1.002779
Is it possible to prevent R from calculating the random effects at the State
level and still calculate the effects of Parties inside States? It would
save degrees of freedom.
2) When I try to insert variables as random terms, the estimation
doesn't converge.
Ex: > depfed.lme4 <- lme(Votes~Sex, data=depfed.frm,
random=~Sex|State/Party)
>Erro em lme.formula(Votes ~ Sex, data = depfed.frm, random = ~Sex |
:
>nlminb problem, convergence error code = 1 message = iteration
limit reached without convergence (9)
> depfed.lme4 <- lme(Votes~Age, data=depfed.frm,
random=~Sex|State/Party)
>Erro em lme.formula(Votes ~ Age, data = depfed.frm, random = ~Sex |
:
>nlminb problem, convergence error code = 1 message = iteration limit
reached without convergence (9)
Am I doing anything wrong? How can I overcome this problem? Any
solutions using nlme or other packages are welcome.
3) I guess I should use State atributes as fixed effects and candidates
atributes both as fixed and random effects, any suggestion on this matter is
also welcome.
Thank you guys for your attention
Thiago Cortez
Rio de Janeiro, Brazil
[[alternative HTML version deleted]]
______________________________________________
R-help@r-project.org 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.
