lOMoARcPSD|15344507

Topic 3 econometrics: interactions between regressors

Econometrics I (Universitat Pompeu Fabra)

Studocu is not sponsored or endorsed by any college or university

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)
 lOMoARcPSD|15344507

Topic 3. Nonlinear Regression Functions

General comments
The regression functions so far have assumed a linear relationship between the X’s
and Y.
However, the linear approximation is not always a good prior.
The multiple regression model can handle regression functions that are nonlinear in
one or more of the dependent variables.
Nonlinear regression functions on parameters can be estimated with nonlinear least
squares.
A function f(x) is linear if the slope of f(x) is the same for all values of x.
Concretely, if the regression model assumes a linear relationship between X and Y, the
corresponding ß is the marginal effect of X on Y which is constant for all possible
values of X.

The general nonlinear population regression function

The change in Y associated with a change in X

• The change in Y associated with a change in X1, ∆𝑋1 holding constant X2, …, Xk is the
difference between the expected value of Y when the independent variables take the
value X1 + ∆𝑋1, X2, …, Xk.
• ∆𝑌, the expected change in Y, is what happens on average with Y when X1 change in
∆𝑋1 units, while the other variables remain constant.
• In a nonlinear regression model, this effect will be

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

How to model nonlinear relationships using multiple regression

1. Identify a possible nonlinear relationship
2. Specify a nonlinear function and estimate its parameter by OLS
3. Determine if the nonlinear relationship improves the linear model
4. Graph the estimated nonlinear regression function
5. Estimate the effect on Y of a change in X

Identify a possible nonlinear relationship

- The economic theory should serve as a guide to suggest a possible nonlinear
relationship
- Even before looking at the data, ask ourselves if the slope of the relationship between
X and Y should depend on the value of X (or another independent variable).
- For example, thinking about class dynamics with 11-year-olds suggests that cutting
the size of the class between 11 and 12 should be different than between 30 and 31.

How to determine if the nonlinear relationship improves the linear model

o Believing that the regression function is nonlinear doesn’t meant that it’s
o We must empirically determine if the nonlinear model is appropriate
o Normally with the t-statistic or the F-statistic we can test the null hypothesis that the
relationship is linear.
Is the relationship between price and size of the dwellings linear?
o Dwellings are usually larger than 30 m2
o In case of very large apartment, we might expect a different effect on price than
on smaller apartments.

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

If a relation between Y and X is nonlinear:

- The marginal effect of X on Y is not constant. This means that the effect on Y of a
change in X depends on the value of X.
- If we estimate an equation that specifies a constant relationship between X and Y, then
the regression is mis-specified because the functional form is wrong.
- The estimator of the effect on Y of X is biased: in general, it isn’t even right on average.
- The solution is to estimate a regression function that is nonlinear in X
- Let’s explore some possible alternatives

Polynomial models
Quadratic regression model

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Application to the case of the price and size of the dwelling

- What is the predicted `rice of the dwelling if the size increases by 1 squared meter?
- Based on the regression results, let’s look at two cases
1. In the first case, size changes from 80 to 81
2. In the second case, size change from 150 to 151

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Standard error of the estimated coefficients

The estimator of the effect of Y of changing X1 depends on the estimated value of the
population regression function 𝑓̂, which, like any estimator, changes form sample to
sample.
One way to quantify the uncertainty associated with the estimated effect is to
calculate the confidence interval of the true population parameter.
To do this, we must compute the standard error of ∆𝑌̂
In order to do it in a nonlinear regression, we mut use the same tools as when testing
a constraint with multiple coefficients.

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

General case

• One way to specify a nonlinear regression is through a polynomial in X

• In general, r denotes the degree of the highest polynomial in X included in the
regression.
• When r = 2 we have quadratic model and if r = 3 it’s a cubic equation.
• The coefficients are difficult to interpret one by one, but the regression function itself
is interpretable.
• Since the effect on Y depends on the value of X, it’s important not to extrapolate
beyond what the data allows.

How to decide the degree of polynomial to use?

1. Choose the maximum value of r and estimate a polynomial for that r
2. Use the t-statistic to test if the coefficient of Xr (ßr) is zero. If we reject the null
hypothesis, then we can assume that Xr belongs in the regression.
3. If we cannot reject the null hypothesis ßr = 0, we remove Xr from the regression and
estimate a polynomial of degree r-1
4. Iterate until the null hypothesis is rejected
5. Normally the r is not greater than 4

Logarithmic models
Logarithms
• Another way to specify nonlinear regression function is with natural logarithms
• When increments are small, logarithms transform changes in variables to percentage
changes.
• Thus, for example, if we want to study questions as the “gender gap” in wages or
elasticity of demand, the specification in logarithms can be very useful.

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Summary --> logarithmic transformations

Three cases differ in which variables (Y or X) is transformed by taking logarithms.
The regression is linear in the new variables ln (Y) and/or ln(X) and the coefficients
can be estimated by OLS
Hypothesis tests and confidence intervals are now implemented in the same way than
before.
The interpretation of ß1 differs from case to case
The choice of specification (functional form) should be guided by judgement (which
interpretation makes the most sense in your application), but also with tests and
plotting predicted values.
An interesting property that the estimated models have when the explanatory
variable is in logarithms, is that if we change the unit of measurement the estimated
coefficient doesn’t change, as it does when the variables is not expressed in logarithm.
The intuitive reason for this is that the interpretation of the estimated coefficient is
that an increase of 1% in the explanatory variables is associated with change of 0.01ß1
in the case of the model linear-log and ß1% in the case of the log-log model.

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Dummy variables

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Interpretation
we can work the numbers for Alternative 3 to check that the three make the exact
same prediction; that’s because they are the same model just expressed differently
(the R2 is differently in Alternative 3 because we are not including a constant).

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Interactions between regressors

So far, we have assumed that the effect of a larger area is independent f the value of
the rest of the variables.
That is, it could be that the effect on price of an extra meter squared is different for
dwelling with heating than for dwelling without it?
How can we model these types of interactions?
We will see three cases: interactions between two categorical variables, interactions
between a continuous and a dummy variable and interaction between two continuous
variables.

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)

 lOMoARcPSD|15344507

Example
- Let’s estimate the price of a dwelling as a
function of the area, the number of dorms
and an interaction term
- As we can see, now the interaction term is
negative but not significant
- That is, we cannot reject the null hypothesis
that the effect of area on rice doesn’t depend
on the number of rooms.

Downloaded by Etsegenet Tafese (etsegenettafesse12@gmail.com)