This example shows how to perform a regression with categorical covariates using categorical arrays andfitlm
.
loadcarsmall
The variableMPG
contains measurements on the miles per gallon of 100 sample cars. The model year of each car is in the variableModel_Year
, andWeight
contains the weight of each car.
Draw a scatter plot ofMPG
againstWeight
, grouped by model year.
figure() gscatter(Weight,MPG,Model_Year,'bgr','x.o') title('MPG vs. Weight, Grouped by Model Year')
The grouping variable,Model_Year
, has three unique values,70
,76
, and82
, corresponding to model years 1970, 1976, and 1982.
Create a table that contains the variablesMPG
,Weight
, andModel_Year
. Convert the variableModel_Year
to a categorical array.
cars = table(MPG,Weight,Model_Year); cars.Model_Year = categorical(cars.Model_Year);
Fit a regression model usingfitlm
withMPG
as the dependent variable, andWeight
andModel_Year
as the independent variables. BecauseModel_Year
is a categorical covariate with three levels, it should enter the model as two indicator variables.
The scatter plot suggests that the slope ofMPG
againstWeight
might differ for each model year. To assess this, include weight-year interaction terms.
The proposed model is
whereI[1976] andI[1982]是虚拟变量表示模型ars 1976 and 1982, respectively.I[1976] takes the value 1 if model year is 1976 and takes the value 0 if it is not.I[1982] takes the value 1 if model year is 1982 and takes the value 0 if it is not. In this model, 1970 is the reference year.
fit = fitlm(cars,'MPG~Weight*Model_Year')
fit = Linear regression model: MPG ~ 1 + Weight*Model_Year Estimated Coefficients: Estimate SE ___________ __________ (Intercept) 37.399 2.1466 Weight -0.0058437 0.00061765 Model_Year_76 4.6903 2.8538 Model_Year_82 21.051 4.157 Weight:Model_Year_76 -0.00082009 0.00085468 Weight:Model_Year_82 -0.0050551 0.0015636 tStat pValue ________ __________ (Intercept) 17.423 2.8607e-30 Weight -9.4612 4.6077e-15 Model_Year_76 1.6435 0.10384 Model_Year_82 5.0641 2.2364e-06 Weight:Model_Year_76 -0.95953 0.33992 Weight:Model_Year_82 -3.2329 0.0017256 Number of observations: 94, Error degrees of freedom: 88 Root Mean Squared Error: 2.79 R-squared: 0.886, Adjusted R-Squared: 0.88 F-statistic vs. constant model: 137, p-value = 5.79e-40
The regression output shows:
fitlm
recognizesModel_Year
as a categorical variable, and constructs the required indicator (dummy) variables. By default, the first level,70
, is the reference group (usereordercats
to change the reference group).
The model specification,MPG~Weight*Model_Year
, specifies the first-order terms forWeight
andModel_Year
, and all interactions.
The modelR2= 0.886, meaning the variation in miles per gallon is reduced by 88.6% when you consider weight, model year, and their interactions.
The fitted model is
Thus, the estimated regression equations for the model years are as follows.
Model Year | Predicted MPG Against Weight |
---|---|
1970 |
|
1976 |
|
1982 |
|
The relationship betweenMPG
andWeight
has an increasingly negative slope as the model year increases.
Plot the data and fitted regression lines.
w = linspace(min(Weight),max(Weight)); figure() gscatter(Weight,MPG,Model_Year,'bgr','x.o') line(w,feval(fit,w,'70'),'Color','b','LineWidth',2) line(w,feval(fit,w,'76'),'Color','g','LineWidth',2) line(w,feval(fit,w,'82'),'Color','r','LineWidth',2) title('Fitted Regression Lines by Model Year')
Test for significant differences between the slopes. This is equivalent to testing the hypothesis
anova(fit)
ans = SumSq DF MeanSq F pValue Weight 2050.2 1 2050.2 263.87 3.2055e-28 Model_Year 807.69 2 403.84 51.976 1.2494e-15 Weight:Model_Year 81.219 2 40.609 5.2266 0.0071637 Error 683.74 88 7.7698
0.0072
(from the interaction row,Weight:Model_Year
), so the null hypothesis is rejected at the 0.05 significance level. The value of the test statistic is5.2266
. The numerator degrees of freedom for the test is2
, which is the number of coefficients in the null hypothesis.
There is sufficient evidence that the slopes are not equal for all three model years.
fitlm
|categorical
|reordercats
|anova