跨栏或零膨胀混合模型的置信区间

Question

I want to calculate CI in mixed models, zero inflated negative binomial and hurdle model. My code for hurdle model looks like this (x1, x2 continuous, x3 categorical):

m1 <- glmmTMB(count~x1+x2+x3+(1|year/class),
          data = bd, zi = ~x2+x3+(1|year/class), family = truncated_nbinom2,
          )

I used confint , and I got these results:

ci <- confint(m1,parm="beta_")
ci
                                          2.5 %       97.5 %     Estimate
cond.(Intercept)                   1.816255e-01  0.448860094  0.285524861
cond.x1                            9.045278e-01  0.972083366  0.937697401
cond.x2                            1.505770e+01 26.817439186 20.094998772
cond.x3high                        1.190972e+00  1.492335046  1.333164894
cond.x3low                         1.028147e+00  1.215828654  1.118056377
cond.x3reg                         1.135515e+00  1.385833853  1.254445909
class:year.cond.Std.Dev.(Intercept)2.256324e+00  2.662976154  2.441845815
year.cond.Std.Dev.(Intercept)      1.051889e+00  1.523719169  1.157153015
zi.(Intercept)                     1.234418e-04  0.001309705  0.000402085
zi.x2                              2.868578e-02  0.166378014  0.069084606
zi.x3high                          8.972025e-01  1.805832900  1.272869874

Am I calculating the intervals correctly? Why is there only one category in x3 for zi? If possible, I would also like to know if it's possible to plot these CIs.

Thanks!

Data looks like this:

     class id year count  x1      x2        x3 
956    5 3002 2002      3 15.6    47.9      high
957    5 4004 2002      3 14.3    47.9      low
958    5 6021 2002      3 14.2    47.9      high
959    4 2030 2002      3 10.5    46.3      high
960    4 2031 2002      3 15.3    46.3      high
961    4 2034 2002      3 15.2    46.3      reg

with x1 and x2 continuous, x3 three level categorical variable (factor)

Summary of the model:

summary(m1)
'giveCsparse' has been deprecated; setting 'repr = "T"' for you'giveCsparse' has been deprecated; setting 'repr = "T"' for you'giveCsparse' has been deprecated; setting 'repr = "T"' for you
 Family: truncated_nbinom2  ( log )
Formula:          count ~ x1 + x2 + x3 + (1 | year/class)
Zero inflation:          ~x2 + x3 + (1 | year/class)
Data: bd

     AIC      BIC   logLik deviance df.resid 
 37359.7  37479.7 -18663.8  37327.7    13323 

Random effects:

Conditional model:
 Groups    Name        Variance Std.Dev.
 class:year(Intercept) 0.79701  0.8928  
 year      (Intercept) 0.02131  0.1460  
Number of obs: 13339, groups:  class:year, 345; year, 15

Zero-inflation model:
 Groups    Name        Variance  Std.Dev. 
 dpto:year (Intercept) 1.024e+02 1.012e+01
 year      (Intercept) 7.842e-07 8.856e-04
Number of obs: 13339, groups:  class:year, 345; year, 15

Overdispersion parameter for truncated_nbinom2 family (): 1.02 

Conditional model:
               Estimate Std. Error z value Pr(>|z|)    
(Intercept)    -1.25343    0.23081  -5.431 5.62e-08 ***
x1             -0.06433    0.01837  -3.501 0.000464 ***
x2              3.00047    0.14724  20.378  < 2e-16 ***
x3high          0.28756    0.05755   4.997 5.82e-07 ***
x3low           0.11159    0.04277   2.609 0.009083 ** 
x3reg           0.22669    0.05082   4.461 8.17e-06 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Zero-inflation model:
               Estimate Std. Error z value Pr(>|z|)    
(Intercept)     -7.8188     0.6025 -12.977  < 2e-16 ***
x2              -2.6724     0.4484  -5.959 2.53e-09 ***
x3high           0.2413     0.1784   1.352  0.17635    
x3low           -0.1325     0.1134  -1.169  0.24258    
x3reg           -0.3806     0.1436  -2.651  0.00802 ** 
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

CI with broom.mixed

> broom.mixed::tidy(m1, effects="fixed", conf.int=TRUE)
# A tibble: 12 x 9
   effect component term           estimate std.error statistic  p.value conf.low conf.high
   <chr>  <chr>     <chr>             <dbl>     <dbl>     <dbl>    <dbl>    <dbl>     <dbl>
 1 fixed  cond      (Intercept)     -1.25      0.231      -5.43 5.62e- 8  -1.71     -0.801 
 2 fixed  cond      x1              -0.0643    0.0184     -3.50 4.64e- 4  -0.100    -0.0283
 3 fixed  cond      x2               3.00      0.147      20.4  2.60e-92   2.71      3.29  
 4 fixed  cond      x3high           0.288     0.0575      5.00 5.82e- 7   0.175     0.400 
 5 fixed  cond      x3low            0.112     0.0428      2.61 9.08e- 3   0.0278    0.195 
 6 fixed  cond      x3reg            0.227     0.0508      4.46 8.17e- 6   0.127     0.326 
 7 fixed  zi        (Intercept)     -9.88      1.32       -7.49 7.04e-14 -12.5      -7.30  
 8 fixed  zi        x1               0.214     0.120       1.79 7.38e- 2  -0.0206    0.448 
 9 fixed  zi        x2              -2.69      0.449      -6.00 2.01e- 9  -3.57     -1.81  
10 fixed  zi        x3high           0.232     0.178       1.30 1.93e- 1  -0.117     0.582 
11 fixed  zi        x3low           -0.135     0.113      -1.19 2.36e- 1  -0.357     0.0878
12 fixed  zi        x4reg           -0.382     0.144      -2.66 7.74e- 3  -0.664    -0.101 

Answer 1

tl;dr as far as I can tell this is a bug in confint.glmmTMB (and probably in the internal function glmmTMB:::getParms ). In the meantime, broom.mixed::tidy(m1, effects="fixed") should do what you want. (There's now a fix in progress in the development version on GitHub, should make it to CRAN sometime? soon...)

Reproducible example:

set up data

set.seed(101)
n <- 1e3
bd <- data.frame(
    year=factor(sample(2002:2018, size=n, replace=TRUE)),
    class=factor(sample(1:20, size=n, replace=TRUE)),
    x1 = rnorm(n),
    x2 = rnorm(n),
    x3 = factor(sample(c("low","reg","high"), size=n, replace=TRUE),
                levels=c("low","reg","high")),
    count = rnbinom(n, mu = 3, size=1))

fit

library(glmmTMB)
m1 <- glmmTMB(count~x1+x2+x3+(1|year/class),
          data = bd, zi = ~x2+x3+(1|year/class), family = truncated_nbinom2,
          )

confidence intervals

confint(m1, "beta_")  ## wrong/ incomplete
broom.mixed::tidy(m1, effects="fixed", conf.int=TRUE)  ## correct

You may want to think about which kind of confidence intervals you want:

• Wald CIs (default) are much faster to compute and are generally OK as long as (1) your data set is large and (2) you aren't estimating any parameters on the log/logit scale that are near the boundaries
• likelihood profile CIs are more accurate but much slower