问题描述
我有一个逻辑回归模型，我想绘制 ROC 曲线。所有变量都有一些缺失数据。总结如下：
X<-cbind(outcome,var1,var2)
summary(X)
# outcome var1 var2
# Min. :0.0000 Min. : 0.100 Min. : 65.1
# 1st Qu.:0.0000 1st Qu.: 0.600 1st Qu.: 91.9
# Median :0.0000 Median : 1.000 Median :101.0
# Mean :0.2643 Mean : 2.421 Mean :110.3
# 3rd Qu.:1.0000 3rd Qu.: 2.200 3rd Qu.:114.5
# Max. :1.0000 Max. :34.800 Max. :388.4
# NA's :165 NA's :80 NA's :30
该模型似乎有效：
model <- glm(outcome~var1+var2,family=binomial)
summary(model)
# Call:
# glm(formula = outcome ~ var1 + var2,family = binomial)
#
# Deviance Residuals:
# Min 1Q Median 3Q Max
# -1.63470 -0.67079 -0.56255 0.01727 2.07577
#
# Coefficients:
# Estimate Std. Error z value Pr(>|z|)
# (Intercept) -3.652208 0.973013 -3.754 0.000174 ***
# var1 0.386811 0.147054 2.630 0.008528 **
# var2 0.016165 0.008075 2.002 0.045316 *
# ---
# Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
#
# (dispersion parameter for binomial family taken to be 1)
#
# Null deviance: 135.91 on 117 degrees of freedom
# Residual deviance: 108.84 on 115 degrees of freedom
# (187 observations deleted due to missingness)
# AIC: 114.84
#
# Number of Fisher Scoring iterations: 6
但是当我尝试计算 ROC 曲线时，出现错误：
library(pROC)
roc(model)
# Error in roc.default(model) : No valid data provided.
我认为可能是由于缺少数据，我尝试添加 na.action = na.exclude 选项，但问题仍然存在：
model2 <- glm(outcome~var1+var2,family=binomial,na.action = na.exclude)
roc(model2)
# Error in roc.default(model2) : No valid data provided.
我也尝试过使用 lrm 而不是 glm，但仍然不起作用：
model.lrm<-lrm(outcome~var1+var2,options(na.action="na.delete"),x=TRUE,y=TRUE)
model.lrm
# Frequencies of Missing Values Due to Each Variable
# outcome var1 var2
# 165 80 30
#
# Logistic Regression Model
#
# lrm(formula = outcome ~ var1 + var2,data = options(na.action = "na.delete"),# x = TRUE,y = TRUE)
#
#
# Model Likelihood discrimination Rank discrim.
# Ratio Test Indexes Indexes
# Obs 118 LR chi2 27.07 R2 0.300 C 0.782
# 0 87 d.f. 2 g 1.377 Dxy 0.565
# 1 31 Pr(> chi2) <0.0001 gr 3.964 gamma 0.565
# max |deriv| 7e-05 gp 0.189 tau-a 0.221
# Brier 0.150
# Coef S.E. Wald Z Pr(>|Z|)
# Intercept -3.6522 0.9730 -3.75 0.0002
# var1 0.3868 0.1471 2.63 0.0085
# var2 0.0162 0.0081 2.00 0.0453
#
roc(model.lrm)
# Error in roc.default(model.lrm) : No valid data provided.
以下是前 20 个观察结果：
> dput(head (dati[,c(2,3,4)],20))
structure(list(outcome = c(NA,1,NA,0),var1 = c(NA,0.3,0.5,1.5,4.5,2,2.2,0.7,0.3),var2 = c(117,84,90,91,113,88,108,178,100,86,95,92,111,103,81,95)),row.names = c(NA,-20L),class = c("tbl_df","tbl","data.frame"))
有什么问题？
解决方法
ROC 曲线不是建立在模型上，而是建立在从模型得出的预测上。因此，您需要使用 predict 函数来获得对数据的预测。它看起来像这样：
predictions <- predict(model)
然后您可以使用这些调用 roc 函数：
roc(outcome,predictions)
缺失值将被自动忽略。
如果您使用的是测试集，这将变得简单且非常相似：
test_predictions <- predict(model,newdata = test_data)
roc(test_data$outcome,test_predictions)
我找到了一个修改代码的解决方案，如下所示：
roc(outcome,as.vector(fitted.values(model)),plot=TRUE)