2
$\begingroup$

I would like to ask for solution or advice on strange result that glmer.nb from lme4 generated when simulating using simR package. I’m working on longitudinal gut microbiome abundance data (23 patients, 2 cases (equal to time points), 482 bacteria), and they’re all count data. Here I will focus on only 2 bacteria among them, bacteria A and B.

    Bacteria_A <- structure(list(Individual = c(rep(c(26, 64, 1, 
                   35, 33, 30, 3, 24, 55, 46, 39, 34, 16, 49, 61, 
                   52, 28, 65, 62, 68, 74, 37, 67), each = 2)), 
    Case = c(3, 2, 3, 2, 2, 3, 3, 2, 3, 2, 2, 3, 2, 3, 3, 2, 3, 
    2, 3, 2, 3, 2, 2, 3, 2, 3, 2, 3, 2, 3, 3, 2, 3, 2, 3, 2, 2, 
    3, 2, 3, 2, 3, 2, 3, 3, 2), Abundance_value = c(18, 4, 10, 2,
    0, 0, 0, 0, 16, 1, 0, 0, 4, 16, 10, 18, 0, 0, 8, 7, 35, 16, 
    2, 22, 1, 6, 16, 9, 7, 12, 38, 32, 22, 4, 17, 13, 19, 20, 0, 
    6, 7, 13, 1, 22, 0, 0)),  class = "data.frame", row.names = 
    c(NA, 46L), .Names = c("Individual", "Case", 
      "Abundance_value"))
    Bacteria_B <- structure(list(Individual = c(rep(c(26, 64, 1, 
    35, 33, 30, 3, 24, 55, 46, 39, 34, 16, 49, 61, 52, 28, 65, 
    62, 68, 74, 37, 67), each = 2)), Case = c(3, 2, 3, 2, 2, 3, 
    3, 2, 3, 2, 2, 3, 2, 3, 3, 2, 3, 2, 3, 2, 3, 2, 2, 3, 2, 3, 
    2, 3, 2, 3, 3, 2, 3, 2, 3, 2, 2, 3, 2, 3, 2, 3, 2, 3, 3, 2), 
    Abundance_value = c(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 32, 0, 0, 0, 0, 0, 0, 
    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)),  class = "data.frame", 
    row.names = c(NA, 46L), .Names = c("Individual", "Case", 
    "Abundance_value"))

I tried both lmer and glmer.nb, and the code is as below.

    lmer(formula = rank(Abundance_value) ~ Case + 
      (1| Individual)) # I use rank() here because abundance    
                       # values aren’t normal-distributed
    glmer.nb(formula = Abundance_value ~ Case + (1| Individual)) 
            # I also tried negative binomial regression
    car::Anova(model, type=c("II"),  test.statistic=c("Chisq”)) 
       # Followed by Anova to get p value of “Case"

I found that compared to using lmer, glmer.nb gave much lower p values, and thus resulting in more significance of bacterial abundance changing over time. However, among the bacteria that showed significance, there were many having really low effect size (paired Cliff’s delta) between 2 cases. I’m wondering if glmer.nb is so sensitive that it can detect such low effect size changes. I did simulation by using the powerCurve function from simR package to check the power of glmer.nb.

    library(simR)
    m_NB <- glmer.nb(formula = Abundance_value ~ Case + 
         (1| Individual), REML = F)  # Fit the negative-binomial 
                                     # regression model
    m_NB_ext <- extend(m_NB, n=1000, along = "Individual”)   
     # Use the “extend" function in simR to increases sample size
    powerCurve(m_NB_ext, along = "Individual", nsim = 1000, 
              alpha = 0.1/N,  test = simr::fixed("Case", 
              method = "chisq”)) # Plot the simulated power at 
               # different sample size. The tested target here is 
               # whether “Case” (which is time point) can explain 
               # the variation of abundance value.

I plotted simulated power curve for different bacteria with Cliff’s delta between 2 cases being 0.609 (bacteria A) and 0.0435 (bacteria B). The power curve of 0.609 is smooth and looks normal, but the curve of 0.0435 looks really strange. It rose fast at the beginning, peaked at n=50, and then dropped all along the way to n = 600.

Bacteria A power curve: enter image description here

Bacteria B power curve: enter image description here

The power curve of bacteria B is really weird; nevertheless, it did reflect the truth that I got many low-effect size yet significant bacteria at n = 23 in my data set. I thought of some possible reasons and would like to ask for your advice.

  1. Is it due to sparsity? Most of my data are zero-inflated, and for bacteria A there were only 24% that the abundance equals to zero, while for bacteria B, there were 66%. Could the weird power curve being resulted from the high extent of zero-inflation? Is there any assumptions of data distribution for negative binomial regression using glmer.nb (or glmmTMB) that I wasn’t aware of?

  2. Maybe there is something wrong with my simR syntax?

Any advice or comments are appreciated!

$\endgroup$
9
  • $\begingroup$ For negative binomial, you can try cran.r-project.org/web/packages/glmmTMB/glmmTMB.pdf, and i noticed bacteria B has a lot of zeros.. as you pointed out, so in theory, negative binomial regression would estimate a really high dispersion (low theta) $\endgroup$
    – StupidWolf
    Apr 16 '20 at 14:05
  • $\begingroup$ Thank you @StupidWolf for your comments! However, I'm not using glm.nb() from MASS, instead, I'm using glmer.nb() from lme4 package. And the reason why I chose to use glmer.nb but not glmmTMB was because the simR function didn't support glmmTMB. Back to dispersion, I am not really sure what's the effect of high dispersion in negative binomial regression, could you please tell me more about it? Thank you! $\endgroup$ Apr 16 '20 at 14:23
  • $\begingroup$ sorry i did not realize they have the same name $\endgroup$
    – StupidWolf
    Apr 16 '20 at 14:24
  • $\begingroup$ @StupidWolf No worries! $\endgroup$ Apr 16 '20 at 14:28
  • $\begingroup$ Ok so when i fit using glmer.nb() for Bacteria_B, you can see one thing that is really weird, if you do summary(glmer.nb(formula = Abundance_value ~ Case + (1| Individual),data=Bacteria_B)) ; look under Family: Negative Binomial(165497.4) $\endgroup$
    – StupidWolf
    Apr 16 '20 at 14:28
0
$\begingroup$

As you correctly pointed out, there is a certain sparsity in your data, and especially for B, in the example dataset you provided:

    library(lme4)
    fit=glmer.nb(formula = Abundance_value ~ Case + 
          (1| Individual), data=Bacteria_B)

It throws some warnings and the fit simply doesn't make sense, especially the theta estimated:

    Generalized linear mixed model fit by maximum likelihood   
      (Laplace
      Approximation) [glmerMod]
     Family: Negative Binomial(165497.4)  ( log )
    Formula: Abundance_value ~ Case + (1 | Individual)
       Data: Bacteria_B

This explains why you get a low p-value, because you have a very high theta, meaning a very low parameter for the variance, for negative binomial the variance is parameterized as var = mean + (mean^2)/theta .

For the simulation, I guess it just doesn't make sense when the fit might be off. Conclusion is that if you have a lot of zeros or missing values, it might not be sensibly to fit a negative binomial. You can consider using a zero-inflated model (for example zero inflated poisson or zero inflated negative binomial. This is on the premise that there is enough non-zero data to fit.

$\endgroup$
4
  • $\begingroup$ Thank you for this detailed answer! One further question, is there a recommended upper threshold for theta? How should I know if the theta is too high? $\endgroup$ Apr 16 '20 at 16:35
  • $\begingroup$ It has to be relative to your mean, because var = mean + (mean^2)/theta ; in your example, your abundance mean is < 100 so 100^2 / 10000 means your extra variance is almost non existent... because for poisson distribution mean = variance $\endgroup$
    – StupidWolf
    Apr 16 '20 at 16:47
  • $\begingroup$ which doesn't make sense because we can see from the data, it is overdispersed and one way you can check is run a quasipoisson, ``` summary(glm(Abundance_value ~ Individual + Case,data=Bacteria_B,family="quasipoisson"))$dispersion ``` $\endgroup$
    – StupidWolf
    Apr 16 '20 at 16:47
  • 1
    $\begingroup$ Thank you once again, it's really helpful! :D $\endgroup$ Apr 16 '20 at 19:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.