ANOVA、GLM摘要及成对差异中意外NA值的问题求助

模型构建 #

我运行了以下GLM模型：

MM11<-glm(CR ~ Species + Season + Habitat, data=ds1)
MM12<-glm(CR ~ Species * Season + Habitat, data=ds1)
MM13<-glm(CR ~ Species + Season * Habitat, data=ds1)
MM14<-glm(CR ~ Species * Habitat + Season, data=ds1)
MM15<-glm(CR ~ Species * Habitat * Season, data=ds1)

模型筛选与异常结果 #

尝试用ANOVA筛选最优模型时出现报错：

anova(MM11, MM12, MM13, MM14, MM15, test="Chisq")

报错信息：

Error in if (dispersion == 1) Inf else min(resdf) : missing value where TRUE/FALSE needed

改用AIC筛选：

AIC(MM11, MM12, MM13, MM14, MM15)

结果显示MM15的AIC值最低，但对MM15单独运行Anova()时，结果全为NA，summary(MM15)也存在大量NA，计算边际均值时同样如此，其他模型无此问题。

得到的异常Anova结果：

Analysis of Deviance Table (Type II tests)

Response: CR
                       LR Chisq Df Pr(>Chisq)
Species                     NaN  3        NaN
Habitat                     NaN  2        NaN
Season                      NaN  1        NaN
Species:Habitat             NaN  4        NaN
Species:Season              NaN  3        NaN
Habitat:Season              NaN  1        NaN
Species:Habitat:Season      NaN  1        NaN

期望结果 #

期望得到类似如下的有效结果：

Analysis of Deviance Table (Type II tests)

Response: CR
                         LR Chisq Df Pr(>Chisq)    
Species                    45.423  3  7.522e-10 ***
Habitat                     6.434  2    0.04007 *  
Season                      2.275  1    0.13145    
Species:Habitat            26.451  5  7.294e-05 ***
Species:Season             2.5307  3   0.469771
Habitat:Season            11.0975  2   0.003892 **
Species:Habitat:Season     1.4956  1   0.23845

数据集 #

数据集ds1的dput输出：

structure(list(Season = c("Southwest monsoon", "Southwest monsoon",
"Southwest monsoon", "Southwest monsoon", "Southwest monsoon",
"Southwest monsoon", "Southwest monsoon", "Southwest monsoon",
"Winter", "Winter", "Winter", "Winter", "Winter", "Winter", "Winter",
"Winter", "Winter"), Habitat = c("Farmland", "Farmland", "Savannah",
"Savannah", "Savannah", "Savannah", "Scrubland", "Scrubland",
"Farmland", "Savannah", "Savannah", "Savannah", "Savannah", "Scrubland",
"Scrubland", "Scrubland", "Scrubland"), Species = structure(c(2L,
4L, 1L, 2L, 3L, 4L, 2L, 3L, 1L, 1L, 2L, 3L, 4L, 1L, 2L, 3L, 4L
), levels = c("1", "2", "3", "5"), class = "factor"), Capture = c(3,
1, 4, 15, 3, 1, 12, 7, 8, 1, 22, 3, 1, 1, 19, 9, 2), Trap = c(590L,
590L, 1761L, 1761L, 1761L, 1761L, 1197L, 1197L, 598L, 1703L,
1703L, 1703L, 1703L, 1134L, 1134L, 1134L, 1134L), CR = c(0.00508474576271186,
0.00169491525423729, 0.00227143668370244, 0.00851788756388416,
0.00170357751277683, 0.00056785917092561, 0.0100250626566416,
0.00584795321637427, 0.0133779264214047, 0.000587199060481503,
0.0129183793305931, 0.00176159718144451, 0.000587199060481503,
0.000881834215167548, 0.0167548500881834, 0.00793650793650794,
0.0017636684303351)), row.names = c(NA, -17L), class = "data.frame")

原因分析 #

1. 样本量远小于模型参数数：MM15是三因素全交互模型，需估计的参数总数约23个（含所有主效应、交互项），但数据集仅17行样本，样本量不足以支撑复杂模型，导致参数无法有效估计，出现NA。
2. 因素组合缺失：数据中存在大量Species-Habitat-Season组合缺失（例如西南季风季的农田仅包含物种2和5，缺少物种1和3），这些缺失组合导致全模型部分参数无对应数据支撑，无法计算。
3. ANOVA报错根源：MM15等模型存在奇异矩阵问题，导致离散度（dispersion）计算出现NA，触发anova()函数的条件判断报错。
4. 模型类型选择不当：响应变量CR是0-1之间的比例值，使用默认正态分布的GLM不符合数据分布特性，加剧了模型估计的不稳定性。

解决方法 #

1. 选择简约模型：放弃全交互模型MM15，从AIC结果中选择次优的、参数数与样本量匹配的模型（如MM12、MM13或MM14），这类模型复杂度更低，参数可被有效估计。
2. 修正模型类型：改用适合比例数据的模型：
   • 采用二项分布GLM，直接使用捕获数和未捕获数作为响应变量：

     MM11_binom <- glm(cbind(Capture, Trap - Capture) ~ Species + Season + Habitat, data=ds1, family=binomial)
     

   • 若捕获率存在大量0或1，可使用Beta回归（需加载betareg包）。
3. 优化数据：若条件允许，补充缺失的Species-Habitat-Season组合样本；若无法补充，需进一步简化模型结构。

内容的提问来源于stack exchange，提问作者Aditya Ganesh