#asreml #error asr_multinomial() Errors in GLM.

Invalid Binomial/mltinomial proportion ,2.000000/1.000000 in record 559Error in asreml(str ~ 1, random = ~Mum, family = asr_binomial(), subset = Spacing == : Errors in GLM.

原因是响应变量不是二项分布，而是有多个分类。这是要用将响应变量转因子后，用asr_multinomial()。
林元震老师教材的第二版中10.4.2.5（p485）中提到泊松分布，但所用的响应变量str应该是干直度的分级，属于多元分类变量而非泊松分布，实际的运行效果如下：

1. 泊松分布

bm_asr <- asreml(str~1,
                 random = ~Mum,
                 family = asr_poisson(),
                 subset = Spacing==3,
                 data = dfm2)

# Model fitted using the sigma parameterization.
# ASReml4 Beta-release 4.0.0.9005 Tue Jul 11 09:16:32 2019
# Poisson; Log   Mu=exp(XB) V=Mu; 
# Note: The LogLik value is unsuitable for comparing GLM models
# 
#           LogLik        Sigma2     DF     wall    cpu
#  1      -89.0416           1.0    558 09:16:32    0.0 (1 restrained)
#  2      -66.0837           1.0    558 09:16:32    0.0
#  3      -71.2797           1.0    558 09:16:32    0.0
#  4      -71.4648           1.0    558 09:16:32    0.0
#  5      -71.5269           1.0    558 09:16:32    0.0
#  6      -71.5594           1.0    558 09:16:32    0.0
#  7      -71.5616           1.0    558 09:16:32    0.0
#  8      -71.5617           1.0    558 09:16:32    0.0
# Deviance from GLM fit:   802.73
# Variance heterogenity factor (Deviance/df):    1.44
# (assuming 558 degrees of freedom)

summary(bm_asr)$varcomp
#           component  std.error  z.ratio bound %ch
# Mum      0.01318554 0.01032939 1.276507     P   0
# units(R) 1.00000000         NA       NA     F   0

plot(bm_asr)

h2 <- vpredict(bm_asr, h2~4*V1/(V1+V2))
#    h2 h2.se 
# 0.052 0.040 

泊松分布残差图

2. 多项分类

dfm2$str <- dfm2$str %>% factor
bm_asr <- asreml(str~trait, #注意这里是trait，不是1
                 random = ~Mum,
                 family = asr_multinomial(),
                 subset = Spacing==3,
                 data = dfm2)
summary(bm_asr)$varcomp

# Model fitted using the sigma parameterization.
# ASReml4 Beta-release 4.0.0.9005 Tue Jul 11 09:14:16 2019
# Cum. Multinomial; Logit  P=1/(1+exp(-XB)); 
# Note: The LogLik value is unsuitable for comparing GLM models
# 
#           LogLik        Sigma2     DF     wall    cpu
#  1     -1388.772           1.0   2790 09:14:16    0.0
#  2     -1391.806           1.0   2790 09:14:16    0.0
#  3     -1393.413           1.0   2790 09:14:16    0.0
#  4     -1394.429           1.0   2790 09:14:16    0.0
#  5     -1395.058           1.0   2790 09:14:16    0.0
#  6     -1395.444           1.0   2790 09:14:16    0.0
#  7     -1395.913           1.0   2790 09:14:16    0.0
#  8     -1396.005           1.0   2790 09:14:16    0.0
#  9     -1396.034           1.0   2790 09:14:16    0.0
# 10     -1396.041           1.0   2790 09:14:16    0.0
# 11     -1396.043           1.0   2790 09:14:16    0.0
# Deviance from GLM fit:  1986.22
# Variance heterogenity factor (Deviance/df):    0.71
# (assuming 2790 degrees of freedom)

summary(bm_asr)$varcomp
#                        component  std.error   z.ratio bound %ch
# Mum                   0.04835465 0.06831176 0.7078524     P   0
# units:trait(R)        1.00000000         NA        NA     F   0
# units:trait!trait_0:0 1.00000000         NA        NA     F   0
# units:trait!trait_1:0 0.00000000         NA        NA     F  NA
# units:trait!trait_1:1 1.00000000         NA        NA     F   0
# units:trait!trait_2:0 0.00000000         NA        NA     F  NA
# units:trait!trait_2:1 0.00000000         NA        NA     F  NA
# units:trait!trait_2:2 1.00000000         NA        NA     F   0
# units:trait!trait_3:0 0.00000000         NA        NA     F  NA
# units:trait!trait_3:1 0.00000000         NA        NA     F  NA
# units:trait!trait_3:2 0.00000000         NA        NA     F  NA
# units:trait!trait_3:3 1.00000000         NA        NA     F   0
# units:trait!trait_4:0 0.00000000         NA        NA     F  NA
# units:trait!trait_4:1 0.00000000         NA        NA     F  NA
# units:trait!trait_4:2 0.00000000         NA        NA     F  NA
# units:trait!trait_4:3 0.00000000         NA        NA     F  NA
# units:trait!trait_4:4 1.00000000         NA        NA     F   0

plot(bm_asr)
# 出这个图时，会报错Error in log(y[nz]) : non-numeric argument to mathematical function，
# 在原函数plot.asreml中，将rr <- resid(x, type = res, spatial = spatial)改成rr <- resid(x, spatial = spatial)

h2 <- vpredict(bm_asr, h2~4*V1/(V1+V2*3.29))
#    h2 h2.se 
# 0.058 0.081
# 遗传力估计值和上面模型是近似的，但标准误差得比较大。

多项分类残差图

这个图怎么解读呢？？？

---

最新的版本残差图是这样的：

image.png

也需是前面修改plot的代码导致残差值提取错误导致的。