Dummy Variables in Regression


A dummy variable or binary variable is a variable that takes on a value of 0 or 1 as an indicator that the observation has some kind of characteristic. Common examples:

It is common to use dummy variables as explanatory variables in regression models, if binary categorical variables are likely to influence the outcome variable.

1 Example: Factors Affecting Monthly Earnings

Let us examine a data set that explores the relationship between total monthly earnings (MonthlyEarnings) and a number of variables on an interval scale (i.e. numeric quantities) that may influence monthly earnings including including each person’s IQ (IQ), a measure of knowledge of their job (Knowledge), years of education (YearsEdu), and years experience (YearsExperience), years at current job (Tenure).

The data set also includes dummy variables that may explain monthly earnings, including whether or not the person is black / African American (Black), whether or not the person lives in a Southern U.S. state (South), and whether or not the person lives in an urban area (Urban).

The code below downloads data on the above variables from 1980 for 663 individuals and assigns it to a dataframe called df.

The following call to lm() estimates a multiple regression predicting monthly earnings based on the eight explanatory variables given above, which includes three dummy variables. The next call to summary() displays some summary statistics for the estimated regression.

## 
## Call:
## lm(formula = MonthlyEarnings ~ IQ + Knowledge + YearsEdu + YearsExperience + 
##     Tenure + Black + South + Urban, data = df)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -899.61 -228.64  -38.36  188.89 2138.04 
## 
## Coefficients:
##                 Estimate Std. Error t value Pr(>|t|)    
## (Intercept)     -475.254    143.374  -3.315 0.000968 ***
## IQ                 3.169      1.223   2.590 0.009816 ** 
## Knowledge          6.412      2.226   2.881 0.004097 ** 
## YearsEdu          45.563      8.664   5.259 1.97e-07 ***
## YearsExperience   13.356      3.969   3.365 0.000810 ***
## Tenure             3.711      2.954   1.256 0.209459    
## Black           -107.905     55.539  -1.943 0.052460 .  
## South            -37.840     31.072  -1.218 0.223732    
## Urban            174.627     31.904   5.474 6.29e-08 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 360.6 on 654 degrees of freedom
## Multiple R-squared:  0.2226, Adjusted R-squared:  0.2131 
## F-statistic: 23.41 on 8 and 654 DF,  p-value: < 2.2e-16

The p-values in the right-most column reveal that all of the coefficients are statistically significantly different from zero at the 5% significance level. We have statistical evidence that all of these variables influence monthly earnings.

The coefficient on Black is equal to -107.91. This means that even after accounting for the effects of all the other explanatory variables in the model (includes educational attainment, experience, location, knowledge, and IQ), black / African American people earn on average $107.91 less per month than non-black people.

The coefficient on South is -37.84. Accounting for the impact of all the variables in the model, people that live in Southern United States earn on average $37.84 less per month than others.

The coefficient on Urban is 174.63. Accounting for the impact of all the variables in the model, people that live in urban areas earn $174.63 more per month, which probably reflects a higher cost of living.

We can compute confidence intervals for these effects with the following call to confint()

##            2.5 %     97.5 %
## Black -216.96154   1.151043
## South  -98.85201  23.172580
## Urban  111.98018 237.273680

2 Dummy Interactions with Numeric Explanatory Variables

We found that black people have lower monthly earnings on average than non-black people. In our regression equation, this implies that the intercept is lower for black people than non-black people. We can also test whether a dummy variable affects the slope multiplying other variables.

For example, are there differences in the returns to education for black versus non-black people? To answer this, we include an interaction effect between Black and YearsEdu:

## 
## Call:
## lm(formula = MonthlyEarnings ~ IQ + Knowledge + YearsEdu + YearsExperience + 
##     Tenure + Black + South + Urban + Black * YearsEdu, data = df)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -899.84 -225.63  -43.73  185.46 2144.61 
## 
## Coefficients:
##                 Estimate Std. Error t value Pr(>|t|)    
## (Intercept)     -506.754    144.023  -3.519 0.000464 ***
## IQ                 3.118      1.221   2.553 0.010897 *  
## Knowledge          6.755      2.228   3.032 0.002530 ** 
## YearsEdu          47.836      8.727   5.481 6.04e-08 ***
## YearsExperience   12.803      3.971   3.224 0.001328 ** 
## Tenure             3.600      2.948   1.221 0.222488    
## Black            562.808    354.447   1.588 0.112805    
## South            -36.621     31.015  -1.181 0.238128    
## Urban            174.006     31.841   5.465 6.60e-08 ***
## YearsEdu:Black   -52.080     27.184  -1.916 0.055821 .  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 359.9 on 653 degrees of freedom
## Multiple R-squared:  0.227,  Adjusted R-squared:  0.2163 
## F-statistic:  21.3 on 9 and 653 DF,  p-value: < 2.2e-16

We see here that when accounting for an interaction effect between race and education, the coefficient on the Black dummy variable becomes insignificant, but the coefficient on the interaction term is negative and significant at the 10% level. The coefficient on the interaction term equal to -52.08 means the slope on education is 52.08 less when \(Black=1\).

The coefficient on the interaction term is interpreted as the additional marginal effect of the numeric variable for the group associated with the dummy variable equal to 1. For this example:

  • The marginal effect on monthly earnings for non-black people for an additional year of education is equal to $47.84 (i.e. when \(Black=0\)).

  • The marginal effect on monthly earnings for black people for an additional year of education is equal to $47.84 - $52.08 = -$4.24 (i.e. when \(Black=1\)), which implies a near zero and possibly negative return to education on income for the black population (you would need to test the hypothesis for the linear combination to answer this).

3 Interacting Dummy Variables with Each Other

Let us interact two of the dummy variables to understand this interpretation and motivation. In the call to lm() below, we use our baseline model and interact South and Urban:

## 
## Call:
## lm(formula = MonthlyEarnings ~ IQ + Knowledge + YearsEdu + YearsExperience + 
##     Tenure + Black + South + Urban + South * Urban, data = df)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -905.50 -222.90  -37.24  190.52 2128.95 
## 
## Coefficients:
##                 Estimate Std. Error t value Pr(>|t|)    
## (Intercept)     -538.952    148.641  -3.626 0.000310 ***
## IQ                 3.289      1.224   2.686 0.007405 ** 
## Knowledge          6.374      2.223   2.867 0.004278 ** 
## YearsEdu          46.735      8.685   5.381 1.03e-07 ***
## YearsExperience   14.011      3.985   3.516 0.000469 ***
## Tenure             3.675      2.950   1.246 0.213343    
## Black           -105.899     55.487  -1.909 0.056762 .  
## South             34.930     55.085   0.634 0.526234    
## Urban            213.078     39.922   5.337 1.30e-07 ***
## South:Urban     -104.974     65.652  -1.599 0.110315    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 360.2 on 653 degrees of freedom
## Multiple R-squared:  0.2256, Adjusted R-squared:  0.215 
## F-statistic: 21.14 on 9 and 653 DF,  p-value: < 2.2e-16

To interpret the meaning of the coefficient on South, Urban, and South*Urban, we will ignore (hold constant) all the terms in the regression equation that do not include one of these variables.

3.1 Difference between Urban and Rural Workers in the North/East/West

Workers in the North / East / and West U.S. have \(South=0\). Here \(South=0\), \((South~x~Urban)=0\), so neither the coefficient on the interaction nor the coefficient on South come into play.

The coefficient for \(b_{Urban}\) implies that in the Non-Southern U.S., urban workers earn on average $213.08 more in monthly earnings than rural workers.

3.2 Difference between Urban and Rural Workers in the South

When focusing on workers in the South, \(South=1\) and the interaction term comes into play.

  • Impact for urban workers in the south = \(b_{South}(1) + b_{Urban}(1) + b_{Urban*South}(1)\)
  • Impact for rural workers in the south = \(b_{South}(1) + b_{Urban}(0) + b_{Urban*South}(0)\)
  • Difference = \(b_{Urban} + b_{Urban*South} = 213.08 - 104.97 =\) $108.11

In the Southern U.S. states, urban workers on average earn $108.11 more in monthly earnings than rural workers.

3.3 Difference between Southern and North/East/West Monthly Earnings for Urban Workers

  • Impact for Southern urban workers = \(b_{South}(1) + b_{Urban}(1) + b_{Urban*South}(1)\)
  • Impact for Non-Southern urban workers = \(b_{South}(0) + b_{Urban}(1) + b_{Urban*South}(0)\)
  • Difference = \(b_{South} + b_{Urban*South} = 34.93 - 104.97 =\) -$70.04

For urban workers, workers in the South earn $70.04 less in monthly earnings than workers in the North/East/West.

3.4 Difference between Southern and North/East/West Monthly Earnings for Rural Workers

Rural workers have \(Urban=0\) and so the interaction term \(Urban~x~South=0\), so we can ignore both of those coefficients. The coefficient for \(b_{South}\) implies that Southern rural workers earn on average $34.93$ more per month than Non-Southern rural workers.

3.5 Three-Way Interactions and Higher!

What?! Things aren’t complicated enough for you?! Do at your own peril!

I have seen people include higher order interaction effects like South * Urban * Black * YearsEdu in their regressions. It has never been obvious to me that they understood what their results meant.

James M. Murray, Ph.D.
University of Wisconsin - La Crosse
R Tutorials for Applied Statistics