r - glmer : Error in eval(expr, envir, enclos) 中的错误:找不到有效的起始 > 值:请指定一些

Question

我正在尝试运行 glmer 模型:

这是我要运行的模型

md$count<-as.integer(md$count)
Model1 <- glmer (count ~ distance_nest_tag + ep_nest  + (1|tag_ID), 
              family = gaussian (link="log"),           
              data=md )

• count 是 0 到 19 之间的整数，表示访问次数。
• distance_nest_tag 是一个以米为单位的连续变量
• ep_nest 是一个二进制变量 ( 1, 0)

随机效应项只是巢穴和个体的数量。

我也尝试过这个

Model1 <- glmer (count ~ distance_nest_tag + ep_nest + (1|tag_ID), 
              family = gaussian (link="log"),
               start = coef (lm(md$count ~ md$distance_nest_tag + md$ep_nest)),                 
              data=md )

但我总是明白这一点

> Error in eval(expr, envir, enclos) :    cannot find valid starting
> values: please specify some

有什么想法吗？？？

dput



       Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) ['glmerMod']
 Family: binomial  ( logit )
Formula: EP0 ~ counts + dist_mean_hour + nest_reproductive_period + X.vistors +  
    (1 | nest_ID) + (1 | tag_ring_ID) + (1 | distance_nesttag_nest) +      (1 | date)
   Data: only_visits_encounternet
     AIC      BIC   logLik deviance df.resid 
 34.7643  81.1238  -7.3822  14.7643      752 
Random effects:
 Groups                Name        Std.Dev.
 date                  (Intercept)  8.507  
 distance_nesttag_nest (Intercept) 68.095  
 tag_ring_ID           (Intercept) 64.491  
 nest_ID               (Intercept)  2.251  
Number of obs: 762, groups:  date, 96; distance_nesttag_nest, 35; tag_ring_ID, 20; nest_ID, 19
Fixed Effects:
              (Intercept)                     counts             dist_mean_hour  
                -13.08952                   -1.80786                   -0.04528  
nest_reproductive_periodI  nest_reproductive_periodP                  X.vistors  
                 -4.04944                   -1.56264                    0.06601  
convergence code 0; 1 optimizer warnings; 0 lme4 warnings 
> dput(R)
structure(list(nest_ID = c(6L, 6L, 6L, 6L, 6L, 6L, 6L, 6L), nest_reproductive_period = structure(c(3L, 
1L, 2L, 3L, 1L, 2L, 3L, 1L), .Label = c("F", "I", "P"), class = "factor"), 
    tag_ID = c(23L, 23L, 23L, 4L, 4L, 4L, 5L, 5L), ring_tag_id = structure(c(3L, 
    3L, 3L, 1L, 1L, 1L, 2L, 2L), .Label = c("AU34180", "AU70442", 
    "BD12273"), class = "factor"), age_tag = c(5L, 5L, 5L, 4L, 
    4L, 4L, NA, NA), p3_tag = c(590L, 590L, 590L, 585L, 585L, 
    585L, 590L, 590L), weight_tag = c(1770L, 1770L, 1770L, 1843L, 
    1843L, 1843L, 1856L, 1856L), next_box_tag = c(42L, 42L, 42L, 
    56L, 56L, 56L, 0L, 0L), dist_mean = c(NA, NA, NA, NA, NA, 
    NA, NA, NA), count = c(3L, 2L, 1L, 0L, 0L, 0L, 0L, 0L), comb = structure(c(2L, 
    2L, 2L, 3L, 3L, 3L, 1L, 1L), .Label = c("6_0", "6_42", "6_56"
    ), class = "factor"), distance_nest_tag = c(0.189813484, 
    0.189813484, 0.189813484, 0.649465717, 0.649465717, 0.649465717, 
    NA, NA), epp_male_comb = c(NA, NA, NA, NA, NA, NA, NA, NA
    ), epp_male = c(NA, NA, NA, NA, NA, NA, NA, NA), nest_epmale = c(NA, 
    NA, NA, NA, NA, NA, NA, NA), neighbour = c(0L, 0L, 0L, 0L, 
    0L, 0L, 0L, 0L), ep_nest = c(0L, 0L, 0L, 0L, 0L, 0L, 0L, 
    0L), roost_nextbox_tag = c(8L, 8L, 8L, 335L, 335L, 335L, 
    332L, 332L), nestroost = structure(c(3L, 3L, 3L, 2L, 2L, 
    2L, 1L, 1L), .Label = c("6_332", "6_335", "6_8"), class = "factor"), 
    neighbours_roost = c(1L, 1L, 1L, 0L, 0L, 0L, 0L, 0L)), .Names = c("nest_ID", 
"nest_reproductive_period", "tag_ID", "ring_tag_id", "age_tag", 
"p3_tag", "weight_tag", "next_box_tag", "dist_mean", "count", 
"comb", "distance_nest_tag", "epp_male_comb", "epp_male", "nest_epmale", 
"neighbour", "ep_nest", "roost_nextbox_tag", "nestroost", "neighbours_roost"
), class = "data.frame", row.names = c(NA, -8L))

编辑

这是一个可重现的小型数据集示例，其中响应变量中包含 0 以及分组变量。

dat = structure(list(y = c(0, 13.0988072077744, 1.53920266020577, 12.1207857178524, 
33.9470080593601, 0, 0, 3.46572339150589, 1.05917038733605, 14.295924854232, 
11.2930736701237, 8.1866351980716, 0.0106620794860646, 0.731016675010324, 
5.99548577982932, 0, 0, 3.15404516097624, 7.62275500199758, 0.604545763926581, 
0, 2.37143378704786, 2.39386320579797, 0.800569675164297), x = c(0, 
0, 0, 0, 0, 0, 2.43684896701947, 0.808418724797666, 0.672910050840583, 
0, 0.763154948372394, 1.44573423753027, 0.96113385074772, 0, 
0.498556550480425, 2.43977373047965, 1.55665618954226, 0.88557694968069, 
4.12758995011915, 3.16827587767271, 2.55354765986558, 3.99502024875255, 
1.71053826174466, 3.59887218330055), tag = c(1, 1, 1, 1, 1, 1, 
2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 5, 5, 5, 5, 5, 5)), .Names = c("y", 
"x", "tag"), row.names = c(NA, -24L), class = "data.frame")

mod2 = glmer(y ~ x + (1|tag),
           data = dat, family = gaussian(link = log), 
           start = list(theta = 1, coef(lm(y ~ x, data = dat))))

Answer 1

tl;dr 要实现此功能，您需要将 mustart 参数设置为合理的值；报告了 mgcv 的类似问题 here 。在当前版本的 glmer 中，您只需要包含

mustart=pmax(dd$count,1e-3)

在 glmer 调用中。

解释:GLM(迭代重新加权最小二乘法)中通常使用的迭代算法需要设置一个起点，不仅为参数设置起点，还为预测的响应值设置起点。高斯模型的默认设置是将预测值设置为等于观测值；如果观察到的值是特定链接函数的不可能的预测(平均值)值(例如，在对数转换时零值变为负无穷大)，那么这将破坏东西(即使观察到的值)值是合法的(即，即使平均值被限制为正，我们也可以观察到零值)。因此，gaussian(link="log")<的 $initialize 组件 家庭项目是:

if (is.null(etastart) && is.null(start) && is.null(mustart) && 
    ((family$link == "inverse" && any(y == 0)) || (family$link == 
        "log" && any(y <= 0)))) 
    stop("cannot find valid starting values: please specify some")
mustart <- y

所以我们真正想做的是设置mustart，这样所有的值都是正数。

接下来的问题是这是否是正确的统计方法。这并不疯狂；尝试泊松模型、基于 log(1+x) 的简单线性模型或序数模型也可能是合理的……具体取决于您对驱动访问的过程的看法以及您观察到的计数分布/条件分布...