当我运行模型 lme4 时,我得到了这个错误信息,但仍然可以得到摘要。我的数据集如下:分类响应是 object_case(binary: nominative (marked as 1) and partitive (0));随机效应因子为verb_lemma;解释变量是数字(单数或复数)、生命(有生命、无生命或人)、顺序(VO 或 OV)、时态(现在或过去)、结果性(是或否)和有界性(是或否)。 我的代码如下:
m2 = glmer(object_case ~ number + animacy + order + tense + resultativity +
boundedness + (1|verb_lemma), data = my_dataset, family=binomial(link = "logit"))
但是,我收到了这样的警告
Warning message:
In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Model failed to converge with max|grad| = 1.26418 (tol = 0.002, component 1)
但我仍然可以得到摘要
> summary(m2)
Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) ['glmerMod']
Family: binomial ( logit )
Formula: object_case ~ Arv + Elusus + Järjekord + Aeg + piiritletus_s + piiritletus_o + (1 | verb_lemma)
Data: lra
AIC BIC logLik deviance df.resid
148.7 183.3 -65.4 130.7 336
Scaled residuals:
Min 1Q Median 3Q Max
-0.71743 -0.00001 0.00004 0.00723 0.75621
Random effects:
Groups Name Variance Std.Dev.
verb_lemma (Intercept) 218.5 14.78
Number of obs: 345, groups: verb_lemma, 171
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -16.411 8.946 -1.834 0.0666 .
ArvMitmus -1.223 1.405 -0.871 0.3839
ElususEluta -9.782 7.993 -1.224 0.2210
ElususInimene -12.344 8.615 -1.433 0.1519
JärjekordVO -1.281 1.853 -0.691 0.4893
AegOlevik -0.375 1.530 -0.245 0.8064
piiritletus_sPiiritlematud 21.882 5.074 4.312 1.61e-05 ***
piiritletus_oPiiritlematud 15.440 2.768 5.577 2.44e-08 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Correlation of Fixed Effects:
(Intr) ArvMtm ElssEl ElssIn JrjkVO AgOlvk prtlts_sP
ArvMitmus -0.210
ElususEluta -0.760 0.227
ElususInimn -0.735 0.188 0.979
JärjekordVO -0.402 0.154 0.441 0.436
AegOlevik -0.380 -0.119 0.272 0.295 0.070
prtlts_sPrt -0.307 -0.160 -0.324 -0.335 -0.321 0.156
prtlts_Prtl -0.139 0.033 -0.395 -0.432 0.005 -0.095 0.565
optimizer (Nelder_Mead) convergence code: 0 (OK)
Model failed to converge with max|grad| = 1.18052 (tol = 0.002, component 1)
Warning message:
In abbreviate(rn, minlength = 6) : abbreviate不适用于非ASCII字元
> lra <- read.csv(file.choose())
> m2 = glmer(object_case ~ number + animacy + order + tense + resultativity +
+ boundedness + (1|verb_lemma), data = lra, family=binomial(link = "logit"))
Warning message:
In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Model failed to converge with max|grad| = 1.26418 (tol = 0.002, component 1)
> summary(m2)
Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) ['glmerMod']
Family: binomial ( logit )
Formula: object_case ~ number + animacy + order + tense + resultativity + boundedness + (1 | verb_lemma)
Data: lra
AIC BIC logLik deviance df.resid
148.4 183.0 -65.2 130.4 336
Scaled residuals:
Min 1Q Median 3Q Max
-0.73592 0.00000 0.00005 0.00844 0.77798
Random effects:
Groups Name Variance Std.Dev.
verb_lemma (Intercept) 221.6 14.89
Number of obs: 345, groups: verb_lemma, 171
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -26.1970 9.7317 -2.692 0.0071 **
numbersingular 1.1154 1.3319 0.837 0.4023
animacyeluta -2.6822 5.0778 -0.528 0.5973
animacyinimene -5.0652 5.4291 -0.933 0.3508
orderVO -0.8986 1.5268 -0.589 0.5562
tensepresent -0.2413 1.4865 -0.162 0.8710
resultativityyes 22.8050 5.0525 4.514 6.37e-06 ***
boundednessyes 15.6020 2.6845 5.812 6.18e-09 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Correlation of Fixed Effects:
(Intr) nmbrsn anmcyl anmcyn ordrVO tnsprs rslttv
numbersnglr -0.105
animacyelut -0.791 -0.072
animacyinmn -0.739 -0.022 0.947
orderVO -0.111 -0.121 0.057 0.068
tensepresnt -0.331 0.134 0.262 0.304 0.043
reslttvtyys -0.872 0.148 0.459 0.402 -0.094 0.253
bounddnssys -0.568 -0.037 0.114 0.023 0.127 -0.034 0.651
optimizer (Nelder_Mead) convergence code: 0 (OK)
Model failed to converge with max|grad| = 1.26418 (tol = 0.002, component 1)
我的主要问题是
In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Model failed to converge with max|grad| = 1.26418 (tol = 0.002, component 1)
这可能更像是一个 CrossValidated 问题,但这里的问题几乎可以肯定是基线水平中主格(1-结果)结果的流行率非常低,正如一个模型中的截距估计 -16 和 - 26 在另一个中,以及一些其他参数的相应大值。这很可能是 (准)完全分离 的情况,您可以阅读 CrossValidated 或其他地方。一般来说,逻辑回归中绝对值 >10 的参数估计是这种现象的危险信号(前提是您的预测变量是分类变量或合理缩放的数值变量)。
glmmTMB
中运行相同的模型并使用 diagnose()
函数可能会给你一个类似的答案。?lme4::troubleshooting
和网络上的其他地方所示,对您是否应该关注收敛警告的蛮力检查是使用 allFit()
并查看使用各种不同优化器的结果是否与您的目的足够相似.