Econometrics – II

BA (H) Econ Core– Spring, 2024 (February to May)

Chapter 6:Multiple Regression Analysis: Further Issues

Wooldridge, J.M. (2015) Introductory Econometrics, 6e

Tirtha Chatterjee
Section 6.3: More on Goodness of Fit and selection of regressors

Tirtha Chatterjee
 R-square- goodness of fit

• 𝑅! is an estimate of how much variation in y is explained by 𝑥", 𝑥!, … 𝑥# in the

population.
• It is the ratio of the explained variation(SSE) compared to the total variation (SST)
• Fraction of the sample variation in y that is explained by x
$$%
R2 = $$&
• Relative change in 𝑅! when variables are added to an equation is very useful
• The F statistic for testing the joint significance crucially depends on the difference in 𝑅! s between
the unrestricted and restricted models.
• Small 𝑅! does not mean biased estimators
• Unbiased estimators of the ceteris paribus effects of the independent variables are determined by
the zero conditional mean assumption MLR.4

Tirtha Chatterjee
 Consequences and implication of small 𝑅 "

• Makes prediction difficult- Because most of the variation in y is explained by

unobserved factors.
• Implies that we have not accounted for several factors that affect y
• Implies that the error variance is large relative to the variance of y
• Makes it difficult to get precise estimators of 𝛽'
• Large data set can be helpful here- we may be able to precisely estimate the partial effects even
though we have not controlled for many unobserved factors.

Tirtha Chatterjee
 Adjusted R-square
• Population 𝑅! is defined as 𝜌! = 1 − σ!(/ σ!)
• Measures proportion of the variation in y in population explained by the independent variables.
• Numerator is population variance of error term, u & denominator is population variance of y
• 𝑅! is supposed to be estimating population 𝑅! or 𝜌!
%%&⁄'
• estimates σ!" by 𝑆𝑆𝑅⁄𝑛 and σ!# by 𝑆𝑆𝑇/𝑛 . Hence, 𝑅! = 1 −
%%(⁄'

• But both numerator and denominator of 𝑅! are biased estimators of σ!( and σ!) .
• Replacing them with unbiased estimators we get adjusted R- square
[$$,⁄ -.#." ] 3!
4
𝑅, ! = 1 − [$$&/(-.")] = 1-[[$$&/(-.")]]
• But the ratio of two unbiased estimators is not an unbiased estimator
• Therefore, it is incorrect to assume that adjusted R-square is a corrected/better
estimator of population R-square.

Tirtha Chatterjee
 Adjusted R-square

• 𝑅, ! imposes a penalty for adding additional independent variables to a model.

• 𝑅! can never fall if we add a new independent variable as SSR falls but 𝑅, ! depends
on k or the number of independent variables
• As we increase k, SSR falls but degrees of freedom also falls
• SSR/(n-k-1) can go up or fall when new independent variable is added to a regression
• 𝑅, ! will increase as k increases only if
• t statistic on the new variable > 1 in absolute value or
• for a group of newly added independent variable, the F statistic for joint significance of the new
variables is greater than unity

Tirtha Chatterjee
 Adjusted R-square

(".,! )(-.")
• Adjusted R-square can be written in terms of 𝑅! as: 𝑅, ! =1−
(-.#.")
• For small n and large k, 𝑅, ! can be substantially below 𝑅!
• For example- if 𝑅!=0.30, n=51 and k=10, then
(".6.86)(9".") 6.:6 96 89
, !
𝑅 =1- (9"."6.") = 1 − = 1 − ;6 = 0.125
;6
• 𝑅, ! can also be negative for a small usual R-squared and small n − k − 1
• Indicating a very poor fit
• Often both 𝑅, ! and 𝑅! are reported in regression results
• Note that F statistics is computed using 𝑅! and not 𝑅, !

Tirtha Chatterjee
 How to Choose between Non-nested Models

• Non-nested models are those where none of the models are restricted versions of the
other.
• F-statistics are used to compare restricted model & unrestricted models-nested models
• Consider two models-
log 𝑠𝑎𝑙𝑎𝑟𝑦 = 𝛽) + 𝛽* 𝑦𝑒𝑎𝑟𝑠 + 𝛽! 𝑔𝑎𝑚𝑒𝑠𝑦𝑟 + 𝛽+ 𝑏𝑎𝑣𝑔 + 𝛽, ℎ𝑟𝑢𝑛𝑠𝑦𝑟 + 𝑢
log 𝑠𝑎𝑙𝑎𝑟𝑦 = 𝛽) + 𝛽* 𝑦𝑒𝑎𝑟𝑠 + 𝛽! 𝑔𝑎𝑚𝑒𝑠𝑦𝑟 + 𝛽+ 𝑏𝑎𝑣𝑔 + 𝛽, 𝑟𝑏𝑖𝑠𝑦𝑟 + 𝑢
• One possibility is to create a composite model that contains all explanatory variables
from the original models and then to test each model against the general model using
the F test.
• What do we do if both models get rejected or neither get rejected
• We might not be able to compare the models in that case

Tirtha Chatterjee
 Using 𝑅" " to Choose between Non-nested Models

log 𝑠𝑎𝑙𝑎𝑟𝑦 = 𝛽6 + 𝛽"𝑦𝑒𝑎𝑟𝑠 + 𝛽!𝑔𝑎𝑚𝑒𝑠𝑦𝑟 + 𝛽8𝑏𝑎𝑣𝑔 + 𝛽;ℎ𝑟𝑢𝑛𝑠𝑦𝑟 + 𝑢

log 𝑠𝑎𝑙𝑎𝑟𝑦 = 𝛽6 + 𝛽"𝑦𝑒𝑎𝑟𝑠 + 𝛽!𝑔𝑎𝑚𝑒𝑠𝑦𝑟 + 𝛽8𝑏𝑎𝑣𝑔 + 𝛽;𝑟𝑏𝑖𝑠𝑦𝑟 + 𝑢

• We know that F-stat cannot be used to compare these two models

• R-square and adjusted R-square can be used here for non-nested models
• 𝑅> ! for the regression containing hrunsyr is .6211 & 𝑅> ! for the regression containing rbisyr is
.6226.
• Thus, based on the 𝑅> ! , there is a very slight preference for the model with rbisyr.
• But the difference is very small
• Yet we can still prefer one model over the other

Tirtha Chatterjee
 Using 𝑅! ! to Choose between Non-nested Models- different functional form

• Consider two models with different functional forms

𝑟𝑑𝑖𝑛𝑡𝑒𝑛𝑠 = 𝛽6 + 𝛽" log 𝑠𝑎𝑙𝑒𝑠 + 𝑢

𝑟𝑑𝑖𝑛𝑡𝑒𝑛𝑠 = 𝛽6 + 𝛽"𝑠𝑎𝑙𝑒𝑠 + 𝛽!𝑠𝑎𝑙𝑒𝑠 ! + 𝑢

• 𝑅! is 0.061 and 0.148 respectively for the first and second model.
• It appears that quadratic model is better.
• But we know that 𝑅! does not penalize more complicated models and number of
parameters are different in the two models
• So we need 𝑅, !: 𝑅, ! is 0.030 and 0.090 respectively for the two models.
• Even this comparison gives the same results

Tirtha Chatterjee
 Using 𝑅! ! -Non-nested Models- different functional form for dependent variable

• 𝑅, ! cannot be used to compare non-nested models with different functional forms for
the dependent variable
• For example- y and log(y)
• 𝑅!s measure the explained proportion of the total variation in whatever dependent
variable we are using in the regression
• And different nonlinear functions of the dependent variable will have different
amounts of variation to explain.
• For example, the total variations in y and log(y) are not the same and are often very different.
• Comparing 𝑅, ! from regressions with these different forms of the dependent variables
does not tell us anything about which model fits better
• they are fitting two separate dependent variables.

Tirtha Chatterjee
 Deciding which explanatory variables to use

• This is critical for estimating the model

• We have to control for all relevant variables
• But we should not over control – i.e. add too many variables
• This makes the regression meaningless
• Also, we have to be careful that the variables are not correlated with each other
• Further while looking at policy impact, we have to carefully select control variables
• Unfortunately, the issue of whether or not to control for certain factors is not always
clear-cut.

Tirtha Chatterjee
 Controlling for too many factors in regression analysis

• We often over control for factors in multiple regression.

• This is because we are concerned about Omitted Variable Bias
• But it is important to remember the ceteris paribus nature of multiple regression.
• In some cases, it makes no sense to hold some factors fixed precisely because they
should be allowed to change when a policy variable changes
• Example- suppose we are doing a study to assess the impact of state beer taxes on
traffic fatalities.
• The idea is that a higher tax on beer will reduce alcohol consumption, and likewise drunk driving,
resulting in fewer traffic fatalities

Tirtha Chatterjee
 Controlling for too many factors in regression analysis

• Notice we don’t control for a variable measuring per capita beer consumption.
• If we control for beer consumption in this equation
• 𝛽" measures the difference in fatalities due to a one percentage point increase in tax,
holding beercons fixed.
• This has no meaning
• We should not be controlling for differences in beercons across states, unless we want
to test for some sort of indirect effect of beer taxes.
• Other factors, such as gender and age distribution, should be controlled for.

Tirtha Chatterjee
 Controlling for too many factors in regression analysis- 2nd example

• Effects of pesticide usage among farmers on family health expenditures.

• In addition to pesticide usage amounts, should we include the number of doctor visits
as an explanatory variable?
• The answer is No.
• Health expenditures include doctor visits, and we would like to pick up all effects of pesticide use
on health expenditures.
• If we include the number of doctor visits as an explanatory variable, then we are only measuring
the effects of pesticide use on health expenditures other than doctor visits.
• It makes more sense to use number of doctor visits as a dependent variable in a
separate regression on pesticide amounts.

Tirtha Chatterjee
 Adding regressors to reduce error variance

• What happens when we add an explanatory variable

• Reduces error variance
• But can also increase multicollinearity problem
• We include independent variables that affect y & are uncorrelated with other
independent variables of interest
• The issue is not unbiasedness here: Our aim is getting an estimator with a smaller
sampling variance.

Tirtha Chatterjee
 Adding regressors to reduce error variance

• Example- consider estimating the individual demand for beer as a function of the
average county beer price.
• It is important to control for individual characteristics like age, education
• affect demand
• They will be uncorrelated with county-level prices &
• Will give more precise estimates of price elasticity of beer demand.
• Might significantly reduce the error variance and standard error of the price
coefficient will be smaller

Tirtha Chatterjee
 Adding regressors when we are studying policy impact
• Another issue is endogeneity of our explanatory variables
• We should not worry that some of our explanatory variables are endogenous –
provided they are not themselves affected by the policy
• For example, in studying the effect of hours in a job training program on labor
earnings
• we can include the amount of education reported prior to the job training program.
• We need not worry that schooling might be correlated with omitted factors, such as “ability,”
• Because we are not trying to estimate the return to schooling.
• We are trying to estimate the effect of the job training program,
• And we can include any controls that are not themselves affected by job training without biasing
the job training effect.
• We must avoid including a variable such as the amount of education after the job
training program,
• Some people may decide to get more education because of how many hours they were assigned to
the job training program.

Tirtha Chatterjee