R： 回归分析案例

案例所用数据集bschool0.txt。

1. 读数据

一般我们用read.table()函数直接读入文件。如果是csv文件，可以用read.cvs()读入，内部参数设置都是一样的。其中路径的表达注意不要用反斜杠而要用斜杠/表示，header取T/F，表示是否第一行为列名，sep表示文件的分隔方式，这里为逗号。于是有

x <- read.table("E:/R/bschool0.txt",header=T,sep=",")

2. 数据初步认识

查看维度

dim(x)  #,60个观测9个变量

>(60,9)

查看前五行

head(x)

>                      FiveYearGain FiveYearGp
Dartmouth (Tuck)                134         84
Pennsylvania (Wharton)          129         75
Chicago                         121         71
Columbia                        120         74
Yale                            119         84
Stanford                        115         70
                       YearsToPayback SalaPreMBA
Dartmouth (Tuck)                  2.9         54
Pennsylvania (Wharton)            3.2         64
Chicago                           3.3         60
Columbia                          3.1         57
Yale                              2.8         46
Stanford                          3.2         58
                       SalaPostMBA Tuition GMAT
Dartmouth (Tuck)               165      75  710
Pennsylvania (Wharton)         177      75  716
Chicago                        164      75  710
Columbia                       160      75  707
Yale                           134      72  700
Stanford                       160      78  710
                             Location State
Dartmouth (Tuck)              Hanover    NH
Pennsylvania (Wharton)   Philadelphia    PA
Chicago                       Chicago    IL
Columbia                     New York    NY
Yale                        New Haven    CT
Stanford                     Stanford    CA

可用summary()函数对定量数据做一个描述性统计。

summary(x)

 >FiveYearGain      FiveYearGp     YearsToPayback 
 Min.   :  2.00   Min.   :  2.00   Min.   :2.200  
 1st Qu.: 62.00   1st Qu.: 57.25   1st Qu.:2.900  
 Median : 79.00   Median : 67.00   Median :3.100  
 Mean   : 78.83   Mean   : 65.93   Mean   :3.190  
 3rd Qu.: 99.25   3rd Qu.: 76.00   3rd Qu.:3.325  
 Max.   :134.00   Max.   :126.00   Max.   :4.900  
                                                  
   SalaPreMBA     SalaPostMBA       Tuition     
 Min.   :24.00   Min.   : 70.0   Min.   :22.00  
 1st Qu.:38.00   1st Qu.: 94.0   1st Qu.:47.75  
 Median :42.50   Median :105.5   Median :61.50  
 Mean   :43.75   Mean   :112.4   Mean   :58.07  
 3rd Qu.:48.75   3rd Qu.:128.5   3rd Qu.:69.25  
 Max.   :64.00   Max.   :180.0   Max.   :78.00  
                                                
      GMAT                Location      State   
 Min.   :560.0     Los Angeles: 3    CA    : 7  
 1st Qu.:640.0     Atlanta    : 2    NY    : 5  
 Median :660.0     Boston     : 2    MA    : 4  
 Mean   :659.9     New York   : 2    TX    : 4  
 3rd Qu.:690.0     Ann Arbor  : 1    FL    : 3  
 Max.   :716.0     Athens     : 1    GA    : 3  
                 (Other)      :49   (Other):34 

3. 相关性度量

这里选取SalaPostMBA作为被解释变量，则需要寻找和其具有较高相关关系的变量构建回归方程。可用cor()函数，$为索引号。例如度量SalaPreMBA与SalaPostMBA的皮尔逊相关系数：

cor(x$SalaPostMBA,x$SalaPreMBA,method = "pearson")

>0.9242337

可见该变量与被解释变量的相关性很高，可以作为解释变量进行回归。因此有

cor(x[,1:7])   #最后两列变量是字符，删去

  FiveYearGain  FiveYearGp YearsToPayback SalaPreMBA SalaPostMBA    Tuition
FiveYearGain      1.0000000  0.69466576    -0.70818368  0.5310755  0.69842102  0.4540424
FiveYearGp        0.6946658  1.00000000    -0.95323083 -0.1491163  0.03342518 -0.2422810
YearsToPayback   -0.7081837 -0.95323083     1.00000000  0.1222730 -0.01691694  0.1623029
SalaPreMBA        0.5310755 -0.14911629     0.12227301  1.0000000  0.92423368  0.7836609
SalaPostMBA       0.6984210  0.03342518    -0.01691694  0.9242337  1.00000000  0.7811176
Tuition           0.4540424 -0.24228099     0.16230293  0.7836609  0.78111759  1.0000000
GMAT              0.6672689  0.13096436    -0.20274377  0.8250422  0.77723442  0.6621838
                     GMAT
FiveYearGain    0.6672689
FiveYearGp      0.1309644
YearsToPayback -0.2027438
SalaPreMBA      0.8250422
SalaPostMBA     0.7772344
Tuition         0.6621838
GMAT            1.0000000

同时，可以引入car包中的scatterplotMatrix()可迅速得到可视化的相关关系如下图。

install.package("car")
library(car)
scatterplotMatrix(x,spread=F,lty.smooth=2)

4.回归方程构建

第5行代表被解释变量SalaPostMBA与所有变量之间的相关关系，可见第1、4、6、7个变量具有较高的相关关系。我们选择这四个变量，利用lm()最小二乘法进行回归方程构建。

lm.sol <- lm(SalaPostMBA ~ FiveYearGain + SalaPreMBA + Tuition + GMAT,x)
# 方程具体摘要
summary(lm.sol)

> Call:
lm(formula = SalaPostMBA ~ FiveYearGain + SalaPreMBA + Tuition + 
    GMAT, data = x)

Residuals:
     Min       1Q   Median       3Q      Max 
-12.3369  -6.0117   0.7479   4.4076  15.2953 

Coefficients:
             Estimate Std. Error t value Pr(>|t|)    
(Intercept)  65.72676   26.70138   2.462  0.01700 *  
FiveYearGain  0.32540    0.04442   7.325 1.11e-09 ***
SalaPreMBA    2.50162    0.23279  10.746 3.98e-15 ***
Tuition       0.21477    0.10131   2.120  0.03853 *  
GMAT         -0.15296    0.05185  -2.950  0.00466 ** 
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 7.15 on 55 degrees of freedom
Multiple R-squared:  0.9307,	Adjusted R-squared:  0.9256 
F-statistic: 184.6 on 4 and 55 DF,  p-value: < 2.2e-16

得到方程形式在Coefficients组件中，可用summary(lm.sol)$coef查看。

5.系数检验

回归方程构建完了需要对方程及其系数做显著性检验。
对方程来说，最后一行中

F-statistic: 184.6 on 4 and 55 DF, p-value: < 2.2e-16

方程的P值小于0.05，表明回归具有意义。对各系数检验，从

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 65.72676 26.70138 2.462 0.01700 *
FiveYearGain 0.32540 0.04442 7.325 1.11e-09 ***
SalaPreMBA 2.50162 0.23279 10.746 3.98e-15 ***
Tuition 0.21477 0.10131 2.120 0.03853 *
GMAT -0.15296 0.05185 -2.950 0.00466 **

可见各系数的P值均小于0.05，因此不显著，变量都具有意义。
因此方程可以投入使用，判定系数$R^2$= 0.9256,解释效果很好。

全部代码如下：

x <- read.table("E:/R/bschool0.txt",header = T,sep=",")
dim(x)
head(x)
cor( x$SalaPostMBA, x$SalaPreMBA, method = "pearson")
library(car)
scatterplotMatrix(x, spread = F, lty.smooth = 2)
cor(x[, 1:7]) 
lm.sol <- lm(SalaPostMBA ~ FiveYearGain + SalaPreMBA + Tuition + GMAT,x)
# 系数检验
summary(lm.sol)
# 一元画图
lm.sol2 <- lm(x$SalaPostMBA ~  x$Tuition , data = x)
plot(x$Tuition, x$SalaPostMBA)
abline(lm.sol2)