2

Thanks in advance.

I am new to mixed models and having several doubts about a mixed model (lme4's glmer, binomial) with multiple levels, measuring a proportion [0,1] in three time periods.

My data (without controls):

  • The dependent variable is the percentage of votes for a given party in a given city-year that were cast for women.
  • "year" Years 2008, 2012, 2016 as a factor variable.
  • "impeachment" A factor variable that groups the parties into a group that supported impeachment (PSDB and the right) and the parties that opposed impeachment (PT and the left) in 2016.
  • "pct_bolsonaro" A constant, continuous city-level variable that shows (latent) support for an anti-system candidate in a subsequent election.
  • my levels are the city "ibge7" and the party list "party.list", the latter of which contains one to three of the yearly city-level party lists.

I'm interested in how voting for female candidates in local city council elections changed in anti-system districts in the election year 2016, which occurred during the impeachment process of the female president.

Here is a Dropbox link to the data

My understanding is that I should nest the party.list level inside the city level. This, along with the fixed effects for year (of substantive interest) are the most appropriate for handling the repeated measures. Does that sound correct?

My problem arises in that I get a nesting error when I attempt to use the explicit nesting language (1 | ibge7 / party.list): "couldn't evaluate grouping factor ibge7:party.list..." . ( This post suggests I can nest the party.list level with separate random intercepts (as I did in the model below) as long as my party.list variable is unique within cities, which I fulfilled. However, this tutorial suggests using explicit nesting. Could the nesting error be related to the unbalanced nature of the yearly party.lists? Is separate random intercepts a valid nesting strategy since party.lists are unique to cities, and do readers agree I need to nest party.lists in cities based on my research question below?

I am also getting a convergence error (see below). Could this be related to the nesting issue?

Note: I weight the response variable following this advice

# Model
library(ggplot2)
library(lme4)
library(ggeffects)
library(see)
ver <- readRDS("data/gender_democracy.rds")

 conditional.intercepts <- glmer(Female / Party.Total ~  pct_bolsonaro * year * impeachment
      + (1 | ibge7) + (1 | party.list) , 
      weights = Party.Total, family = binomial, data = ver,
      control= glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 10000000)), nAGQ=1)
summary(conditional.intercepts)
plot(ggpredict(conditional.intercepts, term = c("year", "impeachment", "pct_bolsonaro")))


Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) ['glmerMod']
 Family: binomial  ( logit )
Formula: Female/Party.Total ~ pct_bolsonaro * year * impeachment + (1 |      ibge7) + (1 | party.list)
   Data: ver
Weights: Party.Total
Control: glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 1e+07))

      AIC       BIC    logLik  deviance  df.resid 
 32142242  32142570 -16071089  32142178    207439 

Scaled residuals: 
     Min       1Q   Median       3Q      Max 
-224.681   -4.557   -0.408    3.519  277.110 

Random effects:
 Groups     Name        Variance Std.Dev.
 party.list (Intercept) 11.5619  3.4003  
 ibge7      (Intercept)  0.2379  0.4878  
Number of obs: 207471, groups:  party.list, 98462; ibge7, 5568

Fixed effects:
                                                        Estimate Std. Error  z value Pr(>|z|)    
(Intercept)                                           -3.0965054  0.0345264  -89.685  < 2e-16 ***
pct_bolsonaro                                          0.6874152  0.1096435    6.270 3.62e-10 ***
year2012                                               0.0664325  0.0014046   47.296  < 2e-16 ***
year2016                                               0.1606648  0.0012625  127.260  < 2e-16 ***
impeachmentNo Federal Deputies                        -0.4166597  0.0454099   -9.176  < 2e-16 ***
impeachmentPSDB                                        0.3412179  0.0591027    5.773 7.77e-09 ***
impeachmentPT                                          0.5758914  0.0601784    9.570  < 2e-16 ***
impeachmentVoted to Impeach                           -0.0492452  0.0366724   -1.343  0.17932    
pct_bolsonaro:year2012                                 0.2002510  0.0082936   24.145  < 2e-16 ***
pct_bolsonaro:year2016                                 0.7151838  0.0074416   96.107  < 2e-16 ***
pct_bolsonaro:impeachmentNo Federal Deputies           1.3278641  0.1978885    6.710 1.94e-11 ***
pct_bolsonaro:impeachmentPSDB                          1.5146030  0.3291996    4.601 4.21e-06 ***
pct_bolsonaro:impeachmentPT                            0.7387725  0.2374355    3.111  0.00186 ** 
pct_bolsonaro:impeachmentVoted to Impeach              1.0411148  0.1178119    8.837  < 2e-16 ***
year2012:impeachmentNo Federal Deputies                0.2530214  0.0022913  110.426  < 2e-16 ***
year2016:impeachmentNo Federal Deputies                0.1681143  0.0021400   78.560  < 2e-16 ***
year2012:impeachmentPSDB                              -0.0108824  0.0020784   -5.236 1.64e-07 ***
year2016:impeachmentPSDB                               0.0006234  0.0018372    0.339  0.73437    
year2012:impeachmentPT                                 0.0307188  0.0019116   16.070  < 2e-16 ***
year2016:impeachmentPT                                -0.0051848  0.0018075   -2.868  0.00412 ** 
year2012:impeachmentVoted to Impeach                   0.0278516  0.0015123   18.417  < 2e-16 ***
year2016:impeachmentVoted to Impeach                  -0.0104830  0.0013640   -7.685 1.53e-14 ***
pct_bolsonaro:year2012:impeachmentNo Federal Deputies -0.6279720  0.0137620  -45.631  < 2e-16 ***
pct_bolsonaro:year2016:impeachmentNo Federal Deputies -1.1303625  0.0129276  -87.438  < 2e-16 ***
pct_bolsonaro:year2012:impeachmentPSDB                -0.5342600  0.0117564  -45.444  < 2e-16 ***
pct_bolsonaro:year2016:impeachmentPSDB                -1.0927577  0.0103405 -105.677  < 2e-16 ***
pct_bolsonaro:year2012:impeachmentPT                  -0.2603739  0.0111271  -23.400  < 2e-16 ***
pct_bolsonaro:year2016:impeachmentPT                  -0.1764549  0.0106061  -16.637  < 2e-16 ***
pct_bolsonaro:year2012:impeachmentVoted to Impeach    -0.1913510  0.0088472  -21.628  < 2e-16 ***
pct_bolsonaro:year2016:impeachmentVoted to Impeach    -0.6798329  0.0079621  -85.383  < 2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Correlation matrix not shown by default, as p = 30 > 12.
Use print(x, correlation=TRUE)  or
    vcov(x)        if you need it

convergence code: 0
Model failed to converge with max|grad| = 7.71506 (tol = 0.001, component 1)
Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?
Model is nearly unidentifiable: large eigenvalue ratio
 - Rescale variables?

Edit: The following shows my effort to nest party.list within municipalities(ibge7) while avoiding the error I described above. The level variable "party.list" is unique to the municipality but occurs up to three times per municipality (once per year). You can see that in the first municipality "1100015" the party list "1100015DEM" occurs in both 2008 and 2012; however, "1100015PPS" occurs in 2008 but not in 2012. SIGLA_PARTIDO is just the name of the political party, so I used that with ibge7 to create a unique municipality-list factor (party.list):

> head(ver %>% select(year, ibge7, SIGLA_PARTIDO, party.list), n=20)
# A tibble: 20 x 4
   year    ibge7 SIGLA_PARTIDO party.list    
   <fct>   <dbl> <fct>         <fct>         
 1 2008  1100015 DEM           1100015DEM    
 2 2008  1100015 PC do B       1100015PC do B
 3 2008  1100015 PDT           1100015PDT    
 4 2008  1100015 PMDB          1100015PMDB   
 5 2008  1100015 PPS           1100015PPS    
 6 2008  1100015 PR            1100015PR     
 7 2008  1100015 PSB           1100015PSB    
 8 2008  1100015 PSDB          1100015PSDB   
 9 2008  1100015 PSDC          1100015PSDC   
10 2008  1100015 PSL           1100015PSL    
11 2008  1100015 PT            1100015PT     
12 2008  1100015 PTB           1100015PTB    
13 2008  1100015 PTN           1100015PTN    
14 2008  1100015 PV            1100015PV     
15 2012  1100015 DEM           1100015DEM    
16 2012  1100015 PC do B       1100015PC do B
17 2012  1100015 PDT           1100015PDT    
18 2012  1100015 PMDB          1100015PMDB   
19 2012  1100015 PP            1100015PP     
20 2012  1100015 PR            1100015PR
Robert Long
  • 53,316
  • 10
  • 84
  • 148
dcoy
  • 70
  • 7

1 Answers1

3

Try the GLMMAdaptive package. This package also allows for more quadrature points during the optimisation.

Your random effects are (1 | ibge7) + (1 | party.list) which could be crossed or nested random effects depending on how they are coded. Can you provide some detail about these factors ?

Robert Long
  • 53,316
  • 10
  • 84
  • 148
  • I added an edit to the end to the end of the post that shows how I coded the lowest level. It seems like you are saying it is possible to nest party.list in ibge7 with the current code, which is what I intended to do. Does the edit allow you to further comment? Does this seem appropriate in combination with the fixed effects for years? Thanks for your response. – dcoy Dec 15 '19 at 04:27
  • Regarding the choice of logit: I fixed a typo describing my dependent variable from (0,1) to [0,1]. This renders beta regression unusable without a transformation, correct? My understanding is that a logit link is fine for continuous, binomial [0,1] modeling as long as I expect the response to follow a logistic curve, which I do. There are many zero values in the response (party lists with no female candidates); however, due to a known phenomenon of placing "fake" female candidates in party lists who approach zero votes, a logistic curve with no hurdle makes sense, I think. – dcoy Dec 15 '19 at 04:51
  • Here are two links on logit for proportions: [first](https://stats.stackexchange.com/questions/189115/fitting-a-binomial-glmm-glmer-to-a-response-variable-that-is-a-proportion-or-f) [second](https://stats.stackexchange.com/questions/87956/how-to-apply-binomial-glmm-glmer-to-percentages-rather-than-yes-no-counts?noredirect=1&lq=1) – dcoy Dec 15 '19 at 04:55
  • Yes, that looks like `party.list` is nested in `ibge7`. I assume you have multiple measures within each level of `party.list` ? I have also removed the objection to the binomial. – Robert Long Dec 16 '19 at 18:42
  • Yes there are up to three measures of `party.list` in each `ibge7` (one per year). Thanks for that response. That is my biggest concern, and your response helped me realize I could confirm it by ensuring `(1 | ibge7) + (1 | party.list)` is equal to `(1 | ibge7) + (1 | ibge7:SIGLA_PARTIDO)`. – dcoy Dec 19 '19 at 20:53