zeroinfl模型-警告消息：在sqrt（diag（对象$vcov））中：生成了NAN

我试图运行零膨胀负二项式，但在检查模型时，我遇到了一个“NaNs生成的”警告，它阻止我看到结果。以下是一些模拟数据，它是我实际数据的简化版本-我的实际数据中每个物种有更多的观测值+更多的物种：

df1 <- data.frame(species = c("Rufl","Rufl","Rufl","Rufl","Assp","Assp","Assp","Assp","Elre", "Elre","Elre", "Elre","Soca","Soca","Soca","Soca"),
                  state = c("warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient"),
                  p_eaten = c(0, 0, 3, 0, 0, 1, 15, 0, 20, 0, 0, 2, 0, 3, 87, 0))

---

df1对于这样一个复杂的模型来说，这是非常有限的数据，但是如果对数据帧的版本执行
xtabs
，您会发现其中一个引用类别的计数为零。如果交换
state
变量的级别，NA将消失，尽管仍存在一些较大的标准错误

xtabs(p_eaten~ state + species, data=df1)
         species
state     Assp Elre Rufl Soca
  ambient    1    2    0    3
  warmed    15   20    3   87

未清理的控制台输出如下所示：

df1 <- data.frame(species = c("Rufl","Rufl","Rufl","Rufl","Assp","Assp","Assp","Assp","Elre", "Elre","Elre", "Elre","Soca","Soca","Soca","Soca"),
+                   state = factor(c("warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient","warmed","ambient"), levels=c("warmed","ambient")),
+                   p_eaten = c(0, 0, 3, 0, 0, 1, 15, 0, 20, 0, 0, 2, 0, 3, 87, 0))
> xtabs(p_eaten~ state + species, data=df1)
         species
state     Assp Elre Rufl Soca
  warmed    15   20    3   87
  ambient    1    2    0    3

df1xtabs（p_~state+species，data=df1）
种
国家法律援助委员会
温暖的1520387
环境1 2 0 3

尝试：

> library(pscl)
> mod1 <- zeroinfl(p_eaten ~ state * species,
+                      dist = 'negbin',
+                      data = df1)
> summary(mod1)

Call:
zeroinfl(formula = p_eaten ~ state * species, data = df1, dist = "negbin")

Pearson residuals:
     Min       1Q   Median       3Q      Max 
-0.98868 -0.80384 -0.00342  0.80387  0.98872 

Count model coefficients (negbin with log link):
                           Estimate Std. Error z value Pr(>|z|)    
(Intercept)               2.708e+00  2.582e-01  10.488  < 2e-16 ***
stateambient             -3.401e+00  1.033e+00  -3.292 0.000994 ***
speciesElre               2.877e-01  3.416e-01   0.842 0.399623    
speciesRufl              -1.671e+00  6.874e-01  -2.431 0.015068 *  
speciesSoca               1.758e+00  2.796e-01   6.288 3.22e-10 ***
stateambient:speciesElre  8.714e-01  1.400e+00   0.622 0.533627    
stateambient:speciesRufl -3.972e-04         NA      NA       NA    
stateambient:speciesSoca -2.763e-02  1.218e+00  -0.023 0.981906    
Log(theta)                1.501e+01  1.848e+02   0.081 0.935234    

Zero-inflation model coefficients (binomial with logit link):
                           Estimate Std. Error z value Pr(>|z|)
(Intercept)              -1.327e-04  1.414e+00   0.000    1.000
stateambient             -8.538e+00  1.206e+02  -0.071    0.944
speciesElre               1.379e-04  2.000e+00   0.000    1.000
speciesRufl              -1.267e-01  2.083e+00  -0.061    0.952
speciesSoca               1.757e-04  2.000e+00   0.000    1.000
stateambient:speciesElre  8.016e+00  1.206e+02   0.066    0.947
stateambient:speciesRufl  1.757e+01         NA      NA       NA
stateambient:speciesSoca  8.411e+00  1.206e+02   0.070    0.944
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 

Theta = 3316847.5216 
Number of iterations in BFGS optimization: 41 
Log-likelihood: -21.87 on 17 Df
Warning message:
In sqrt(diag(object$vcov)) : NaNs produced

>库（pscl）
>mod1概要（mod1）
电话：
zeroinfl（公式=p_~state*物种，数据=df1，dist=“negbin”）
皮尔逊残差：
最小1季度中值3季度最大值
-0.98868 -0.80384 -0.00342  0.80387  0.98872 
计算模型系数（带有日志链接的negbin）：
估计标准误差z值Pr（>z）
（截距）2.708e+00 2.582e-01 10.488<2e-16***
州环境温度-3.401e+00 1.033e+00-3.2920.000994***
规格ELRE 2.877e-01 3.416e-01 0.842 0.399623
规格RUFL-1.671e+00 6.874e-01-2.431 0.015068*
规格SOCA 1.758e+00 2.796e-01 6.288 3.22e-10***
州环境：规范ELRE 8.714e-01 1.400e+00 0.622 0.533627
州环境：物种RUFL-3.972e-04 NA
州环境：规格SOCA-2.763e-02 1.218e+00-0.023 0.981906
对数（θ）1.501e+01 1.848e+02 0.081 0.935234
零通胀模型系数（带有logit链接的二项式）：
估计标准误差z值Pr（>z）
（截距）-1.327e-04 1.414e+00 0.000 1.000
州环境温度-8.538e+00 1.206e+02-0.071 0.944
规格ELRE 1.379e-04 2.000e+00 0.000 1.000
规格RUFL-1.267e-01 2.083e+00-0.061 0.952
规格SOCA 1.757e-04 2.000e+00 0.000 1.000
州环境：规格ELRE 8.016e+00 1.206e+02 0.066 0.947
州环境：规格RUFL 1.757e+01 NA
州环境：规范SOCA 8.411e+00 1.206e+02 0.070 0.944
---
签名。代码：0'***'0.001'***'0.01'*'0.05'.'0.1''1
θ=3316847.5216
BFGS优化中的迭代次数：41
对数可能性：-17 Df时为21.87
警告信息：
在sqrt中（diag（对象$vcov））：生成的NAN

方差-协方差矩阵应为正定矩阵，负值可能会妨碍反演工作。当CrossValidated.com上提出类似问题时，Ben Bolker建议使用brglm2版本的glm：

> library(brglm2)
> summary(m1 <- glm(p_eaten~ state * species, data=df1,
+             family=poisson,
+             method="brglmFit"))

Call:
glm(formula = p_eaten ~ state * species, family = poisson, data = df1, 
    method = "brglmFit")

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-9.3541  -1.8708  -0.7071   0.8567   5.7542  

Coefficients:
                         Estimate Std. Error z value Pr(>|z|)    
(Intercept)                2.0477     0.2540   8.062 7.52e-16 ***
stateambient              -2.3354     0.8551  -2.731  0.00631 ** 
speciesElre                0.2796     0.3366   0.831  0.40619    
speciesRufl               -1.4881     0.5918  -2.514  0.01192 *  
speciesSoca                1.7308     0.2756   6.281 3.37e-10 ***
stateambient:speciesElre   0.2312     1.0863   0.213  0.83142    
stateambient:speciesRufl   0.3895     1.7369   0.224  0.82258    
stateambient:speciesSoca  -0.8835     1.0141  -0.871  0.38362    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for poisson family taken to be 1)

    Null deviance: 443.21  on 15  degrees of freedom
Residual deviance: 183.08  on  8  degrees of freedom
AIC: 225.38

Number of Fisher Scoring iterations: 1

>库（brglm2）
>总结（m1 | z |）
（截距）2.0477 0.2540 8.0627.52e-16***
州环境-2.3354 0.8551-2.731 0.00631**
规格0.2796 0.3366 0.831 0.40619
规格RUFL-1.4881 0.5918-2.514 0.01192*
规格SOCA 1.7308 0.2756.281 3.37e-10***
州环境：speciesElre 0.2312 1.0863 0.213 0.83142
州环境：speciesRufl 0.3895 1.7369 0.224 0.82258
州环境：规格SOCA-0.8835 1.0141-0.871 0.38362
---
签名。代码：0'***'0.001'***'0.01'*'0.05'.'0.1''1
（泊松族的色散参数取1）
零偏差：在15个自由度上为443.21
剩余偏差：8自由度时为183.08
AIC:225.38
Fisher评分迭代次数：1

事实上，互动的必要性似乎值得怀疑。随着相互作用项的去除，偏差的变化很小。很难知道完整的数据集是否也会出现than mich。
这很可能是一种情况，但如果没有完整的数据集，则无法确定

当类别全部为零或全部为非零时，可能会发生这种情况。在您上面给出的示例中：

with（df1，表（物种，状态，p_=0））

表明在物种==“Rufl”、状态==“ambient”和
p
==0为
FALSE
的情况下，无观察结果；换句话说，对于这些因素的组合，所有的观察结果都是零。因此，涉及与该状态进行比较的任何系数将具有较大的参数值（即abs（β）>>1）；理论上它应该是无限的，但通常在10到30之间（取决于数值方法放弃的地方）。这些系数要么具有大得离谱的标准误差和（Wald）置信区间，要么（如您的情况）具有
NaN
值

本说明适用于
speciesRufl
和
statewarmed:speciesRufl
的零膨胀系数。计数模型系数不大，但仍有
NaN
标准误差，我认为，因为它们的不确定性与零通胀系数的不确定性有关

计算模型系数（带有日志链接的negbin）：
估计标准误差z值Pr（>z）
（截距）-0.69312 1.00016-0.693 0.488307
...
种类RUFL-1.69427楠楠
零通胀模型系数（带有logit链接的二项式）：
估计标准误差z值Pr（>z）
...
品种17.560楠楠楠楠楠

对此你能做些什么


忽略这个问题。对于受影响的系数，您将无法获得可靠的标准误差（以及p值、置信区间等）
> library(brglm2)
> summary(m1 <- glm(p_eaten~ state * species, data=df1,
+             family=poisson,
+             method="brglmFit"))

Call:
glm(formula = p_eaten ~ state * species, family = poisson, data = df1, 
    method = "brglmFit")

Deviance Residuals: 
    Min       1Q   Median       3Q      Max  
-9.3541  -1.8708  -0.7071   0.8567   5.7542  

Coefficients:
                         Estimate Std. Error z value Pr(>|z|)    
(Intercept)                2.0477     0.2540   8.062 7.52e-16 ***
stateambient              -2.3354     0.8551  -2.731  0.00631 ** 
speciesElre                0.2796     0.3366   0.831  0.40619    
speciesRufl               -1.4881     0.5918  -2.514  0.01192 *  
speciesSoca                1.7308     0.2756   6.281 3.37e-10 ***
stateambient:speciesElre   0.2312     1.0863   0.213  0.83142    
stateambient:speciesRufl   0.3895     1.7369   0.224  0.82258    
stateambient:speciesSoca  -0.8835     1.0141  -0.871  0.38362    
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for poisson family taken to be 1)

    Null deviance: 443.21  on 15  degrees of freedom
Residual deviance: 183.08  on  8  degrees of freedom
AIC: 225.38

Number of Fisher Scoring iterations: 1