ECNS 561

Topics in Multiple Regression Analysis

Scaling Data
• For the simple regression case, we already discussed the effects of
changing the units of measurement
– Nothing different here
– Coefficients, SEs, CIs, t and F stats change in ways that preserve all
measured effects and testing outcomes
– Consider the textbook example of regressing child birth weight in ounces vs.
pounds on the number of cigarettes smoked by the mother while pregnant

Insert more
general proof
 Functional Form
• Logarithmic function form
– Recall the following

Independent
Model Dependent Variable Variable Interpretation of β
Level-level y x Δy = βΔx
Level-log y Log(x) Δy=(β/100)%Δx

Log-level Log(y) x %Δy = (100β) Δx

Log-log Log(y) Log(x) %Δy = β% Δx

• Why again do we use logs in applied work?
– Linear interpretation may not be reasonable
• Suppose 𝛽 1 = -5.00 in a county-level regression on the relationship between crime and
education (with an interpretation as follows “an increase in the average level of education in the
county by one year, leads to five fewer crimes per 1,000 population annually”…here, we have
defined the dependent variable as a rate)
• This means that the marginal impact of an added year of education is the same if we go from 9
to 10 years or from, say, 13 to 14 years
• Perhaps a better characterization of how crime changes with education levels is that each year
of education decreases crime by a constant percentage.
– We can be ignorant about the units of measurement of variables appearing in log
form because the slope coefficients are invariant to rescalings
– When y > 0, models using log(y) as the dependent variable often satisfy the
classical linear model assumptions more closely than models using the level of y
– Taking the log of a variable often will narrow its range.
• Ex. Annual sales, baseball players’ salaries, population variables.
• Narrowing the range of the dependent and independent variables can make OLS estimates less
sensitive to outliers
• Not always the case. If y is a ratio b/w 0 and 1, then log(y) can be very large in magnitude,
whereas the original variable, y, is bounded b/w 0 and 1.
• Models with Quadratics
– Suppose we are interested in econometrically modeling the age-crime profile

– Or, the crime time of day profile

• In both cases, crime depends on a factor in a non-linear
fashion…looks more like a quadratic relationship
• We would write the crime-age regression as
Crimei = β0 + β1Agei + β2Agei2 + εi
• The marginal effect of Age on Crime depends on Age itself
𝜕𝐶𝑟𝑖𝑚𝑒
= 𝛽1 + 2𝛽2 𝐴𝑔𝑒
𝜕𝐴𝑔𝑒
• For the relationship between age and crime, what do we expect the
signs on β1 and β2 to be?
• We can easily solve for the “turning point”…i.e., where the crime-
age profile is at its peak…how would we do this?
𝛽1
𝐴𝑔𝑒 ∗ = | |
2𝛽2
• The relationship need not be parabolic…e.g., we could have a U-
shaped relationship
• In this case, 𝛽1 < 0
and 𝛽2 > 0
• If 𝛽1 and 𝛽2 share the
same sign, there is no
turning point for values
of the independent
variable > 0.
• For example, if both estimates are positive, the smallest expected
value of y is at x = 0, and increases in x always have a positive and
increasing effect on y.
• Interaction Terms
– Often, we want to allow for the partial effect of the dependent variable with
respect to an explanatory variable to depend on the magnitude of yet another
explanatory variable
– For example, perhaps we want to allow the education and crime relationship
to differ by gender
Crimei = β0 + β1Educi + β2Malei + β3Educi*Malei + εi
– Here, the marginal effect of education on crime would be
𝜕𝐶𝑟𝑖𝑚𝑒
= 𝛽1 + 𝛽3 𝑀𝑎𝑙𝑒
𝜕𝐸𝑑𝑢𝑐
– If 𝛽1 < 0 (as we would probably expect) and 𝛽3 > 0, then this implies that one
more year of education has less of a crime-reducing effect for males than it
does for females…i.e., there is an interaction effect between gender and
education.
[show example of triple interaction in MDA and crime paper]
 Crimei = β0 + β1Educi + β2Malei + β3Educi*Malei + εi

– Interpreting the other variables in the model can get tricky when interaction
terms have been included.
– Q. What is the interpretation of β2 ?
– Ans. It is the relationship between gender and crime for individuals with zero
years of education.
• Obviously, this is not a very interesting interpretation
– So, when evaluating the relationship between gender and crime, we must be
sure to do so for interesting values of Educ.
• Q. What might be some “interesting” values of Educ?
Some practice problems (from Ch. 6)
1.) The following equation was estimated for firm salaries as a
function of sales and return on equity
log(𝑠𝑎𝑙𝑎𝑟𝑦) = 4.322 + .276log(sales) + .0215roe - .00008roe2
(.324) (.033) (.0129) (.00026)
N = 209, R2 = .282, standard deviation of roe ≈ 25
This equation allows roe to have a diminishing effect on log(salary).
Is this generally necessary? Explain why or why not.
Ans. No, this is generally not necessary.
– The t state on roe2 is only about -.30…i.e., this variable is nowhere near
statistically significant
– In addition, the inclusion of this squared term has only a very minor effect on
the slope even when the values of roe are large
• Approximate slope is .0215 - .00016roe. So, even a one standard deviation increase in roe
doesn’t matter much for the slope.
3.) The following equation was obtained via OLS
𝑟𝑑𝑖𝑛𝑡𝑒𝑛𝑠 = 2.613 + .00030sales - .0000000070sales2
(.429) (.00014) (.0000000037)
N = 32
R2 = .1484.
Note: sales are in millions of dollars
(i) At what point does the marginal effect of sales on rdintens become
negative?
(ii) Would you keep the quadratic term in the model?
(iii) Define salesbil = sales/1000. Rewrite the estimated equation.
(iv) For the purpose of reporting the results, which equation do you
prefer?
4.) The following model allows the return to education to depend
upon the total amount of both parents’ education, called pareduc:

log(wage) = β0 + β1educ + β2educ*pareduc + β3exper + β4tenure + ε

(i) Solve for the marginal effect of education

𝜕log(wage)/𝜕educ = β1 + β2pareduc
What sign do you expect for β2? Why?
(ii) Suppose we estimate the following
log(𝑤𝑎𝑔𝑒) = 5.65 + .047educ + .00078educ*pareduc + .019exper
(.13) (.010) (.00021) (.004)
+ .010tenure
(.003)
N = 722, R2 = .169
-Interpret the coefficient on the interaction term.
-It might help to choose two specific values for pareduc and to
compare the estimated return to educ…what might be reasonable
values to choose for evaluation?
(iii) When pareduc is added as a separate variable to the equation, we
get:
log(𝑤𝑎𝑔𝑒) = 4.94 + .097educ + .033pareduc - .0016educ*pareduc
(.38) (.027) (.017) (.0012)
+ .020exper + .010tenure
(.004) (.003)
N = 722, R2 = .174
Does the estimated return to education now depend positively on
parent education? Test the null that the return to education does not
depend on parent education.
8.) Suppose we want to estimate the effects of alcohol consumption
(alcohol) on college grade point average (colGPA). In addition to
collecting information on grade point averages and alcohol usage, we
also obtain attendance information (say, percentage of lectures
attended, called attend). A standardized test score (say, SAT) and high
school GPA (hsGPA) are also available.

(i) Should we include attend along with alcohol as explanatory

variables in a multiple regression model?

(ii) Should SAT and hsGPA be included as explanatory variables?

(iii) Extra question I’ve added…how might we actually get credible

causal estimates on this relationship of interest?
• The sharp increase in alcohol consumption at the minimum legal
drinking age has been exploited by researchers who are interested in
the determinants of academic performance as measured by grades.

• For instance, using data from the United States Air Force Academy,
Carrell et al. (2011) found that gaining legal access to alcohol at age
21 (the minimum legal drinking age) leads to a reduction in test
scores of one-tenth of a standard deviation.
Source: Carrell , Scott E., Mark Hoekstra , and James E. West. 2011. “Does Drinking
Impair College Performance? Evidence from a Regression Discontinuity Approach”
Journal of Public Economics, 95: 54-62.