应用于以两种不同语言编写的两个数据集（数字变量+分类变量）的lm函数的结果（一个用英文写，另一个用韩文写）是不同的。除分类变量外，数字变量完全相同。有什么可以解释结果的差异？
#data
df3 <- repmis::source_DropboxData("df3_v0.1.csv","gg30a74n4ew3zzg",header = TRUE)
#the one written in korean
out1<-lm(YD~SANJI+TAmin8+TMINup18do6+typ_rain6+DTD9,data=df3)
summary(out1)
#the one written in eng
df3$SANJI[df3$SANJI=="전북"]<-"JB"
df3$SANJI[df3$SANJI=="충북"]<-"CHB"
df3$SANJI[df3$SANJI=="경북"]<-"KB"
df3$SANJI[df3$SANJI=="전남"]<-"JN"
df3$SANJI2[df3$SANJI2=="고창"]<-"Gochang"
df3$SANJI2[df3$SANJI2=="괴산"]<-"Goesan"
df3$SANJI2[df3$SANJI2=="단양"]<-"Danyang"
df3$SANJI2[df3$SANJI2=="봉화"]<-"Fenghua"
df3$SANJI2[df3$SANJI2=="신안"]<-"Sinan"
df3$SANJI2[df3$SANJI2=="안동"]<-"Andong"
df3$SANJI2[df3$SANJI2=="영광"]<-"younggang"
df3$SANJI2[df3$SANJI2=="영양"]<-"youngyang"
df3$SANJI2[df3$SANJI2=="영주"]<-"youngju"
df3$SANJI2[df3$SANJI2=="예천"]<-"Yecheon"
df3$SANJI2[df3$SANJI2=="의성"]<-"Yusaeng"
df3$SANJI2[df3$SANJI2=="제천"]<-"Jechon"
df3$SANJI2[df3$SANJI2=="진안"]<-"Jinan"
df3$SANJI2[df3$SANJI2=="청송"]<-"Changsong"
df3$SANJI2[df3$SANJI2=="해남"]<-"Haenam"
out2<-lm(YD~SANJI+TAmin8+TMINup18do6+typ_rain6+DTD9,data=df3)
summary(out2)
#the one written in korean
#Call:
#lm(formula = YD ~ SANJI + TAmin8 + TMINup18do6 + typ_rain6 +
# DTD9, data = df3)
#Residuals:
# Min 1Q Median 3Q Max
#-98.836 -23.173 -2.261 22.626 111.367
#Coefficients:
# Estimate Std. Error t value Pr(>|t|)
#(Intercept) 970.33251 84.12479 11.534 < 2e-16 ***
#SANJI전남 -33.75664 12.53277 -2.693 0.008158 **
#SANJI전북 -44.17939 11.22274 -3.937 0.000144 ***
#SANJI충북 -44.09285 9.16736 -4.810 4.74e-06 ***
#TAmin8 -25.56618 3.36053 -7.608 9.37e-12 ***
#TMINup18do6 4.58052 0.96528 4.745 6.19e-06 ***
#typ_rain6 -0.19754 0.02862 -6.903 3.23e-10 ***
#DTD9 -16.15975 2.65128 -6.095 1.59e-08 ***
#---
#Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
#Residual standard error: 37.2 on 112 degrees of freedom
#Multiple R-squared: 0.58, Adjusted R-squared: 0.5538
#F-statistic: 22.1 on 7 and 112 DF, p-value: < 2.2e-16
#the one written in eng
#Call:
#lm(formula = YD ~ SANJI + TAmin8 + TMINup18do6 + typ_rain6 +
# DTD9, data = df3)
#Residuals:
# Min 1Q Median 3Q Max
#-98.836 -23.173 -2.261 22.626 111.367
#Coefficients:
# Estimate Std. Error t value Pr(>|t|)
#(Intercept) 926.23966 84.32621 10.984 < 2e-16 ***
#SANJIJB -0.08654 12.32752 -0.007 0.994
#SANJIJN 10.33620 13.09434 0.789 0.432
#SANJIKB 44.09285 9.16736 4.810 4.74e-06 ***
#TAmin8 -25.56618 3.36053 -7.608 9.37e-12 ***
#TMINup18do6 4.58052 0.96528 4.745 6.19e-06 ***
#typ_rain6 -0.19754 0.02862 -6.903 3.23e-10 ***
#DTD9 -16.15975 2.65128 -6.095 1.59e-08 ***
#---
#Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
#Residual standard error: 37.2 on 112 degrees of freedom
#Multiple R-squared: 0.58, Adjusted R-squared: 0.5538
#F-statistic: 22.1 on 7 and 112 DF, p-value: < 2.2e-16
答案 0 :(得分：7)
您的整体模型拟合是相同的，您只需要为您的因子（“SANJIJ”）提供不同的参考类。具有不同的参考水平也会影响您的截距，但不会改变您的连续协变量的估计值。
您可以使用relevel()强制使用特定的引用类（假设SANJIJ已经是一个因素）或使用levels=参数显式创建factor（），否则默认顺序按字母顺序排序在不同的语言中，级别可能不会以相同的方式排序。