如何修复零膨胀泊松回归的错误

Question

I ran zero-inflated Poisson regression with package pscl and came across a same error with this post

However, since I know there is a separate process for excess zeros indicated by z , does it still make sense to just run Poisson as a solution (Poisson results are just fine)? Is there an alternative way to fix this problem for ZIP? I also tried zero-inflated negative binomial regression but it got the same error. Thanks.

Call:
zeroinfl(formula = y ~ x1 + x2 + x3 + x4 + x5 + x6 + x7 + x8 + x9 | z, data = df)

Pearson residuals:
     Min       1Q   Median       3Q      Max 
-2.48465 -0.06156 -0.06126 -0.06091  5.57840 

Count model coefficients (poisson with log link):
                 Estimate Std. Error z value Pr(>|z|)
(Intercept)      3.547e+00         NA      NA       NA
x1              -3.251e-02         NA      NA       NA
x2               6.290e-03         NA      NA       NA
x3               8.867e-01         NA      NA       NA
x4               1.432e-01         NA      NA       NA
x5               2.705e-01         NA      NA       NA
x6              -8.223e-10         NA      NA       NA
x7              -7.218e-02         NA      NA       NA
x8               3.322e-02         NA      NA       NA
x9              -2.072e-01         NA      NA       NA

Zero-inflation model coefficients (binomial with logit link):
            Estimate Std. Error z value Pr(>|z|)
(Intercept)    5.531         NA      NA       NA
z              158.108         NA      NA       NA
Error in if (getOption("show.signif.stars") & any(rbind(x$coefficients$count,  : 
  missing value where TRUE/FALSE needed

Answer 1

It's hard to answer this without a reproducible example, but I'll offer a couple of observations (too long for a comment):

• pscl 's default behaviour is to use the same formula for both the zero-inflated and the count (conditional) part of the model. Unless you have an extremely large data set, you are very likely to have trouble fitting a 10-parameter model (intercept + 9 covariates) to both the count and zero-inflation aspects of the data. (A reasonable rule of thumb is that you should have 20 times as many observations as parameters, so that's a minimum of 400 observations -- and that rule is probably conservative for estimating zero-inflation.)
• one of your parameter estimates ( x6 ) is approximately zero, suggesting that you don't have enough variation in your data to estimate that parameter (or that there is some other issue with this covariate, eg you have an extreme outlier in this dimension). This could easily mess up the standard errors etc. for your whole model.

General advice:

• plot your data
• find a reasonably complex model that you can actually fit by splitting the difference between an over-complicated model that breaks and an over-simplified model that misses important phenomena, as in this figure:

Uriarte and Yackulic, Ecological Applications, 19(3), 2009, pp. 592–596