linear regression slides

7/29/2019 Linear Regression Slides

1/129

Univariate RegressionMultivariate Regression

Specification IssuesInference

Linear Regression

Michael R. Roberts

Department of FinanceThe Wharton School

University of Pennsylvania

October 5, 2009

Michael R. Roberts Linear Regression 1/129


2/129



BasicsOrdinary Least Squares (OLS) EstimatesUnits of Measurement and Functional FormOLS Estimator Properties

Motivation

Linear regression is arguably the most popular modeling approachacross every field in the social sciences.

1 Very robust technique2 Linear regression also provides a basis for more advanced empirical

methods.3 Transparent and relatively easy to understand technique4 Useful for both descriptive and structural analysis

Were going to learn linear regression inside and out from an appliedperspective

focusing on the appropriateness of different assumptions, modelbuilding, and interpretation

This lecture draws heavily from Wooldridges undergraduate andgraduate texts, as well as Greenes graduate text.



3/129




Terminology

The simple linear regression model (a.k.a. - bivariate linearregression model, 2-variable linear regression model)

y = + x + u (1)

y = dependent variable, outcome variable, response variable,explained variable, predicted variable, regressand

x = independent variable, explanatory variable, controlvariable, predictor variable, regressor, covariate

u = error term, disturbance

= intercept parameter

= slope parameter



4/129




Details

Recall model is

y = + x + u

(y, x, u) are random variables(y, x) are observable (we can sample observations on them)

u is unobservable = no stat tests involving u(, ) are unobserved but estimable under certain conds

Model implies that u captures everything that determines y exceptfor x

In natural sciences, this often includes frictions, air resistance, etc.

In social sciences, this often includes a lot of stuff!!!


U i i R i B i


5/129




Assumptions

1 E(u) = 0As long as we have an intercept, this assumption is innocuousImagine E(u) = k = 0. We can rewrite u = k + w =

yi = ( + k) + E(xi) + w

where E() = 0. Any non-zero mean is absorbed by the intercept.

2 E(u|x) = E(u)Assuming q u (= orthogonal) is not enough! Correlation onlymeasures linear dependence

Conditional mean independenceImplied by full independence q u (= independent)Implies uncorrelatedIntuition: avg of u does not depend on the value of qCan combine with zero mean assumption to get zero conditionalmean assumption E(u

|q) = E(u) = 0


U i i t R i B i


6/129




Conditional Mean Independence (CMI)

This is the key assumption in most applications

Can we test it?

Run regression.

Take residuals u = y y & see if avg u at each value of x = 0?Or, see if residuals are uncorrelated with xDoes these exercise make sense?

Can we think about it?

The assumption says that no matter whether x is low, medium, or

high, the unexplained portion of y is, on average, the same (0).But, what if agents (firms, etc.) with different values of x aredifferent along other dimensions that matter for y?


Uni ariate Regression Basics


7/129




CMI Example 1: Capital Structure

Consider the regression

Leveragei = + Profitabilityi + ui

CMI = that average u for each level of Profitability is the sameBut, unprofitable firms tend to have higher bankruptcy risk and

should have lower leverage than more profitable firms according totradeoff theoryOr, unprofitable firms have accumulated fewer profits and may beforced to debt financing, implying higher leverage according to thepecking orderThese e.g.s show that the average u is likely to vary with the levelof profitability

1st e.g., low profitable firms will be less levered implies lower avg ufor less profitable firms2nd e.g., low profitable firms will be more levered implies higher avg

u for less profitable firmsMichael R. Roberts Linear Regression 7/129

Univariate Regression Basics


8/129




CMI Example 2: Investment

Consider the regression

Investmenti = + qi + ui

CMI =

that average u for each level of q is the same

But, firms with low q may be in distress and invest less

Or, firms with high q may have difficultly raising sufficient capital tofinance their investment

These e.g.s show that the average u is likely to vary with the levelof q

1st e.g., low q firms will invest less implies higher avg u for low qfirms2nd e.g., high q firms will invest less implies higher avg u for low qfirms




9/129




Population Regression Function (PRF)

PRF is E(y|x). It is fixed but unknown. For simple linear regression:PRF = E(y|x) = + x (2)

Intuition: for any value of x, distribution of y is centered about

E(y|x)




10/129




OLS Regression Line

We dont observe PRF, but we can estimate via OLS

yi = + xi + ui (3)

for each sample point iWhat is ui? It contains all of the factors affecting yi other than xi.

= ui contains a lot of stuff! Consider complexity ofy is individual food expendituresy is corporate leverage ratiosy is interest rate spread on a bond

Estimated Regression Line (a.k.a. Sample Regression Function

(SRF))

y = + x (4)

Plug in an x and out comes an estimate of y, y

Note: Different sample =

different (, )




11/129

Uni ariate RegressionMultivariate Regression



OLS Estimates

Estimators:

Slope = = Ni=1(xi x)(yi y)

Ni=1(xi x)2Intercept = = y x

Population analogues

Slope = Cov(x, y)Var(x)

= Corr(x, y) SD(y)SD(x)

Intercept = E(y) E(x)




12/129

gMultivariate Regression


Ordinary Least Squares (OLS) EstimatesUnits of Measurement and Functional FormOLS Estimator Properties

The Picture




13/129

gMultivariate Regression



Example: CEO Compensation

Model

salary = + ROE + y

Sample 209 CEOs in 1990. Salaries in $000s and ROE in % points.

SRF

salary = 963.191 + 18.501ROE

What do the coefficients tell us?Is the key CMI assumption likely to be satisfied?

Is ROE the only thing that determines salary?Is the relationship linear? = estimated change is constant acrosssalary and ROE

dy/dx = indep of salary & ROE

Is the relationship constant across CEOs?




14/129

Multivariate RegressionSpecification Issues

Inference


PRF vs. SRF


Univariate RegressionM l i i R i

BasicsO di L S (OLS) E i


15/129


Inference


Goodness-of-Fit (R2)

R-squared defined as

R2 = SSE/SST = 1 SSR/SSTwhere

SSE = Sum of Squares Explained =Ni=1

(yi y)2

SST = Sum of Squares Total =N

i=1(yi y)2

SSR = Sum of Squares Residual =

Ni=1

(ui u)2 =Ni=1

ui2

R2 = [Corr(y, y)]2


Univariate RegressionM lti i t R i

BasicsO di L t S (OLS) E ti t


16/129


Inference


Example: CEO Compensation

Model

salary = + ROE + y

R2 = 0.0132

What does this mean?



BasicsOrdinary Least Squares (OLS) Estimates


17/129


Inference


Scaling the Dependent Variable

Consider CEO SRF

salary = 963.191 + 18.501ROE

Change measurement of salary from $000s to $s. What happens?

salary = 963, 191 + 18, 501ROE

More generally, multiplying dependent variable by constant c =OLS intercept and slope are also multiplied by c

y = + x + u

cy = (c) + (c)x + cu(Note: variance of error affected as well.)

Scaling = multiplying every observation by same #No effect on R2 - invariant to changes in units





18/129


Inference


Scaling the Independent Variable

Consider CEO SRF

salary = 963.191 + 18.501ROE

Change measurement of ROE from percentage to decimal (i.e.,multiply every observations ROE by 1/100)

salary = 963.191 + 1, 850.1ROE

More generally, multiplying independent variable by constantc = OLS intercept is unchanged but slope is divided by c

y = + x + u

y = + (/c)cx + cuScaling = multiplying every observation by same #No effect on R2 - invariant to changes in units





19/129


Inference


Changing Units of Both y and x

Model:

y = + x + u

What happens to intercept and slope when we scale y by c and x byk?

cy = c + cx + cu

cy = (c) + (c/k)kx + cu

intercept scaled by c, slope scaled by c/k





20/129


Inference


Shifting Both y and x

Model:

y = + x + u

What happens to intercept and slope when we add c and k to y andx?

c + y = c + + x + u

c + y = c + + (x + k)

k + u

c + y = (c + k) + (x + k) + u

Intercept shifted by k, slope unaffected





21/129

gSpecification Issues

Inference

y q ( )Units of Measurement and Functional FormOLS Estimator Properties

Incorporating Nonlinearities

Consider a traditional wage-education regression

wage = + education + u

This formulation assumes change in wages is constant for all

educational levelsE.g., increasing education from 5 to 6 years leads to the same $increase in wages as increasing education from 11 to 12, or 15 to16, etc.

Better assumption is that each year of education leads to a constantproportionate (i.e., percentage) increase in wages

Approximation of this intuition captured by

log(wage) = + education + u





22/129

gSpecification Issues

Inference

y q ( )Units of Measurement and Functional FormOLS Estimator Properties

Log Dependent Variables

Percentage change in wage given one unit increase in education is

%wage (100)educ

Percent change in wage is constant for each additional year ofeducation

= Change in wage for an extra year of education increases aseducation increases.

I.e., increasing return to education (assuming > 0)

Log wage is linear in education. Wage is nonlinear

log(wage) = + education + u

= wage = exp ( + education + u)





23/129


Units of Measurement and Functional FormOLS Estimator Properties

Log Wage Example

Sample of 526 individuals in 1976. Wages measured in $/hour.Education = years of education.

SRF:

log(wage) = 0.584 + 0.083education, R2 = 0.186

Interpretation:

Each additional year of education leads to an 8.3% increase in wages(NOT log(wages)!!!).For someone with no education, their wage is exp(0.584)...this ismeaningless because no one in sample has education=0.

Ignores other nonlinearities. E.g., diploma effects at 12 and 16.





24/129



Constant Elasticity Model

Alter CEO salary model

log(salary) = + log(sales) + u

is the elasticity of salary w.r.t. salesSRF

log(salary) = 4.822 + 0.257log(sales), R20.211

Interpretation: For each 1% increase in sales, salary increase by0.257%

Intercept meaningless...no firm has 0 sales.



S ifi i I

BasicsOrdinary Least Squares (OLS) EstimatesU i f M d F i l F


25/129



Changing Units in Log-Level Model

What happens to intercept and slope if we units of dependentvariable when its in log form?

log(y) = + x + u

log(c) + log(y) = log(c) + + x + u log(cy) = (log(c) + ) + x + u

Intercept shifted by log(c), slope unaffected because slope measuresproportionate change in log-log model



S ifi ti I

BasicsOrdinary Least Squares (OLS) EstimatesU it f M t d F ti l F


26/129



Changing Units in Level-Log Model

What happens to intercept and slope if we units of independentvariable when its in log form?

y = + log(x) + u

log(c) + y = + log(x) + log(c) + u y = ( log(c)) + log(cx) + u

Slope measures proportionate change



Specification Issues

BasicsOrdinary Least Squares (OLS) EstimatesUnits of Measurement and Functional Form


27/129



Changing Units in Log-Log Model

What happens to intercept and slope if we units of dependentvariable?

log(y) = + log(x) + u

log(c) + log(y) = log(c) + + log(x) + u log(cy) = ( + log(c)) + log(x) + u

What happens to intercept and slope if we units of independentvariable?

log(y) = + log(x) + u

log(c) + log(y) = + log(x) + log(c) + u log(y) = ( log(c)) + log(cx) + u






28/129



Log Functional Forms

Dependent Independent InterpretationModel Variable Variable of

Level-level y x dy = dx

Level-log y log(x) dy = (/100)%dxLog-level log(y) x %dy = (100)dxLog-log log(y) log(x) %dy = %dx

E.g., In Log-level model, 100

= % change in y for a 1 unit

increase in x (100 = semi-elasticity)

E.g., In Log-log model, = % change in y for a 1% change in x (= elasticity)






29/129



Unbiasedness

When is OLS unbiased (i.e., E() = )?1 Model is linear in parameters2 We have a random sample (e.g., self selection)3 Sample outcomes on x vary (i.e., no collinearity with intercept)4 Zero conditional mean of errors (i.e., E(u|x) = 0)

Unbiasedness is a feature of sampling distributions of and .

For a given sample, we hope and are close to true values.






30/129



Variance of OLS Estimators

Homoskedasticity = Var(u|x) = 2Heterokedasticity = Var(u|x) = f(x) R+






31/129

pInference OLS Estimator Properties

Standard Errors

Remember, larger error variance = larger Var() = biggerSEs

Intuition: More variation in unobservables affecting y makes it hard

to precisely estimate

Relatively more variation in x is our friend!!!

More variation in x means lower SEs for

Likewise, larger samples tend to increase variation in x which also

means lower SEs for

I.e., we like big samples for identifying !




Mechanics and InterpretationOLS Estimator PropertiesFurther Issues


32/129

pInference Binary Independent Variables

Basics

Multiple Linear Regression Model

y = 0 + 1x1 + 2x2 + ... + kxk + u

Same notation and terminology as before.Similar key identifying assumptions1 No perfect collinearity among covariates2 E(u|x1,...xk) = 0 = at a minimum no correlation and we have

correctly accounted for the functional relationships between y and

(x1, ..., xk)SRF

y = 0 + 1x1 + 2x2 + ... + kxk






33/129

Inference Binary Independent Variables

Interpretation

Estimated intercept beta0 is predicted value of y when all x = 0.Sometimes this makes sense, sometimes it doesnt.

Estimated slopes (1,...k) have partial effect interpretations

y =

1x1 + ... +

kxk

I.e., given changes in x1 through xk, (x1,..., xk), we obtain thepredicted change in y.

When all but one covariate, e.g., x1, is held fixed so

(x2,..., xk) = (0,..., 0) then

y = 1x1

I.e., 1 is the coefficient holding all else fixed (ceteris paribus)



Specification Issuesf



34/129


Example: College GPA

SRF of college GPA and high school GPA (4-point scales) and ACTscore for N = 141 university students

colGPA = 1.29 + 0.453hsGPA + 0.0094ACT

What do intercept and slopes tell us?Consider two students, Fred and Bob, with identical ACT score buthsGPA of Fred is 1 point higher than that of Bob. Best prediction ofFreds colGPA is 0.453 points higher than that of Bob.

SRF without hsGPA

colGPA = 1.29 + 0.0271ACT

Whats different and why? Can we use it to compare 2 people withsame hsGPA?


Univariate RegressionMultivariate RegressionSpecification Issues

I f

Mechanics and InterpretationOLS Estimator PropertiesFurther IssuesBi I d d t V i bl


35/129


All Else Equal

Consider prev example. Holding ACT fixed, another point on highschool GPA is predicted to inc college GPA by 0.452 points.

If we could collect a sample of individuals with the same high schoolACT, we could run a simple regression of college GPA on highschool GPA. This holds all else, ACT, fixed.

Multiple regression mimics this scenario without restricting the

values of any independent variables.



Inference

Mechanics and InterpretationOLS Estimator PropertiesFurther IssuesBinar Inde endent Variables


36/129


Changing Multiple Independent Variables Simultaneously

Each corresponds to the partial effect of its covariate

What if we want to change more than one variable at the sametime?

E.g., What is the effect of increasing the high school GPA by 1

point and the ACT score by 1 points?

colGPA = 0.453hsGPA + 0.0094ACT = 0.4624

E.g., What is the effect of increasing the high school GPA by 2

point and the ACT score by 10 points?

colGPA = 0.453hsGPA + 0.0094ACT

= 0.453 2 + 0.0094 10 = 1



Inference

Mechanics and InterpretationOLS Estimator PropertiesFurther IssuesBinary Independent Variables


37/129


Fitted Values and Residuals

Residual = ui = yi

yi

Properties of residuals and fitted values:1 sample avg of residuals = 0 = y = y2 sample cov between each indep variable and residuals = 03 Point of means (y, x1, ..., xk) lies on regression line.



Inference



38/129


Partial Regression

Consider 2 independent variable model

y = 0 + 1x1 + 2x2 + u

Whats the formula for just 1?

1 = (r

1r1)1r1y

where r1 are the residuals from a regression of x1 on x2.

In other words,1 regress x1 on x2 and save residuals2 regress y on residuals3 coefficient on residuals will be identical to 1 in multivariate

regression



Inference



39/129


Frisch-Waugh-Lovell I

More generally, consider general linear setup

y = XB + u = X1B1 + X2B2 + u

One can show that

B2 = (X2M1X2)1(X2M1y) (5)

where

M1 = (I P1) = I X1(X1X1)1X1)P1 is the projection matrix that takes a vector (y) and projects itonto the space spanned by columns of X1

M1 is the orthogonal compliment, projecting a vector onto the spaceorthogonal to that spanned by X1



Inference



40/129


Frisch-Waugh-Lovell II

What does equation (5) mean?

Since M1 is idempotent

B2 = (X2M1M1X2)1(X2M1M1y)

= (X2X2)1(X2y)

So B2 can be obtained by a simple multivariate regression of y on X2

But y and X2 are just the residuals obtained from regressing y andeach component of X2 on the X1 matrix



Inference



41/129

y p

Omitted Variables Bias

Assume correct model is:y = XB + u = X1B1 + X2B2 + u

Assume we incorrectly regress y on just X1. Then

B1 = (X

1X1)1X1y

= (X1X1)1X1(X1B1 + X2B2 + u)

= B1 + (X

1X1)1X1X2B2 + (X

1X1)1X1u

Take expectations and we get

B1 = B1 + (X

1X1)

1

X

1X2B2Note (X1X1)

1X1X2 is the column of slopes in the OLS regressionof each column of X2 on the columns of X1OLS is biased because of omitted variables and direction is unclear depending on multiple partial effects



Inference



42/129

Bivariate Model

With two variable setup, inference is easier

y = 0 + 1x1 + 2x2 + u

Assume we incorrectly regress y on just x1. Then

1 = 1 + (x

1x1)1x1x22

= 1 + 2

Bias term consists of 2 terms:1 = slope from regression of x2 on x12 2 = slope on x2 from multiple regression of y on (x1, x2)

Direction of bias determined by signs of and 2.

Magnitude of bias determined by magnitudes of and 2.



Inference



43/129

Omitted Variable Bias General Thoughts

Deriving sign of omitted variable bias with multiple regressors inestimated model is hard. Recall general formula

B1 = B1 + (X

1X1)1X1X2B2

(X

1X1)

1X

1X2 is vector of coefficients.Consider a simpler model

y = 0 + 1x1 + 2x2 + 3x3 + u

where we omit x3

Note that both 1 and 2 will be biased because of omission unlessboth x1 and x2 are uncorrelated with x3.

The omission will infect every coefficient through correlations



Inference



44/129

Example: Labor

Consider

log(wage) = 0 + 1education + 2ability + u

If we cant measure ability, its in the error term and we estimate

log(wage) = 0 + 1education + w

What is the likely bias in ? recall

1 = 1 + 2

where is the slope from regression of ability on education.

Ability and education are likely positively correlated = > 0Ability and wages are likely positively correlated = 2 > 0So, bias is likely positive =

1 is too big!



Inference



45/129

Goodness of Fit

R2 still equal to squared correlation between y and y

Low R2 doesnt mean model is wrong

Can have a low R2 yet OLS estimate may be reliable estimates of

ceteris paribus effects of each independent variableAdjust R2

R2a = 1 (1 R2)n 1

n

k

1

where k = # of regressors excluding intercept

Adjust R2 corrects for df and it can be < 0



Inference



46/129

Unbiasedness

When is OLS unbiased (i.e., E() = )?1 Model is linear in parameters

2 We have a random sample (e.g., self selection)3 No perfect collinearity4 Zero conditional mean of errors (i.e., E(u|x) = 0)

Unbiasedness is a feature of sampling distributions of and .

For a given sample, we hope and are close to true values.



Inference



47/129

Irrelevant Regressors

What happens when we include a regressor that shouldnt be in the

model? (overspecified)No affect on unbiasedness

Can affect the variances of the OLS estimator



Inference



48/129

Variance of OLS Estimators

Sampling variance of OLS slope

Var(j) =2

Ni=1(xij xj)2(1 R2j )

for j = 1,..., k, where R2j is the R2 from regressing xj on all other

independent variables including the intercept and 2 is the varianceof the regression error term.

Note

Bigger error variance (2) = bigger SEs (Add more variables tomodel, change functional form, improve fit!)More sampling variation in xj = smaller SEs (Get a larger sample)Higher collinearity (R2j ) = bigger SEs (Get a larger sample)



Inference



49/129

Multicollinearity

Problem of small sample size.No implication for bias or consistency, but can inflate SEs

Consider

y = 0 + 1x1 + 2x2 + 3x3 + u

where x2 and x3 are highly correlated.Var(2) and Var(3) may be large.

But correlation between x2 and x3 has no direct effect on Var(1)

If x1 is uncorrelated with x2 and x3, then R21 = 0 and Var(1) is

unaffected by correlation between x2 and x3Make sure included variables are not too highly correlated with thevariable of interest

Variance Inflation Factor (VIF) = 1/(1 R2j ) above 10 issometimes cause for concern but this is arbitrary and of limited use






50/129

Data Scaling

No one wants to see a coefficient reported as 0.000000456, or1,234,534,903,875.

Scale the variables for cosmetic purposes:1 Will effect coefficients & SEs2

Wont affect t-stats or inferenceSometimes useful to convert coefficients into comparable units, e.g.,SDs.

1 Can standardize y and xs (i.e., subtract sample avg. & divide bysample SD) before running regression.

2 Estimated coefficients = 1 SD in y given 1 SD in x.Can estimate model on original data, then multiply each coef bycorresponding SD. This marginal effect = in y units for a 1SD in x






51/129

Log Functional Forms

Consider

log(price) = 0 + 1log(pollution) + 2rooms + u

Interpretation1

1 is the elasticity of price w.r.t. pollution. I.e., a 1% change inpollution generates an 1001% change in price.2 2 is the semi-elasticity of price w.r.t. rooms. I.e., a 1 unit change in

rooms generates an 1002% change in price.

E.g.,

log(price) = 9.23 0.718log(pollution) + 0.306rooms + u

= 1% inc. in pollution = 0.72% dec. in price= 1 unit inc. in rooms = 30.6% inc. in price






52/129

Log Approximation

Note: percentage change interpretation is only approximate!Approximation error occurs because as log(y) becomes larger,approximation %y 100log(y) becomes more inaccurate. E.g.,

log(y) = 0 + 1log(x1) + 2x2

Fixing x1 (i.e., x1 = 0) = log(y) = 2x2Exact % change is

log(y) = log(y)

logy(y) = 2x2 = 2(x

2

x2)

log(y/y) = 2(x2 x2)y/y = exp(2(x

2 x2))

(y y)/y

% = 100

exp(2(x

2 x2)) 1





Fi f L A i i


53/129

Figure of Log Approximation

Approximate % change y : log(y) = 2x2

Exact % change y : (y/y)% = 100

exp(2x2)





U f l f L


54/129

Usefulness of Logs

Logs lead to coefficients with appealing interpretationsLogs allow us to be ignorant about the units of measurement ofvariables appearing in logs since theyre proportionate changes

If y > 0, log can mitigate (eliminate) skew and heteroskedasticity

Logs of y or x can mitigate the influence of outliers by narrowingrange.

Rules of thumb of when to take logs:

positive currency amounts,variable with large integral values (e.g., population, enrollment, etc.)

and when not to take logsvariables measured in years (months),proportions

If y [0, ), can take log(1+y)Michael R. Roberts Linear Regression 54/129




P P P i Ch


55/129

Percentage vs. Percentage Point Change

Proportionate (or Relative) Change(x1 x0)/x0 = x/x0

Percentage Change

%x = 100(x/x0)

Percentage Point Change is raw change in percentages.

E.g., let x = unemployment rate in %If unemployment goes from 10% to 9%, then

1% percentage point change,(9-10)/10 = 0.1 proportionate change,100(9-10)/10 = 10% percentage change,

If you use log of a % on LHS, take care to distinguish betweenpercentage change and percentage point change.





M d l ith Q d ti


56/129

Models with Quadratics

Consider

y = 0 + 1x + 2x2 + u

Partial effect of x

y = (1 + 22x)x = dy/dx = 1 + 22x

= must pick value of x to evaluate (e.g., x)1 > 0, 2 < 0 =

parabolic relation

Turning point = Maximum = 1/(22)Know where the turning point is!. It may lie outside the range of x!Odd values may imply misspecification or be irrelevant (above)

Extension to higher order straightforward





M d l ith I t ti


57/129

Models with Interactions

Considery = 0 + 1x1 + 2x2 + 3x1x2 + u

Partial effect of x1

y = (1

+ 3

x2

)x1

=

dy/dx1

= 1

+ 3

x2

Partial effect of x1 = 1 x2 = 0. Have to ask if this makessense.

If not, plug in sensible value for x2 (e.g., x2)

Or, reparameterize the model:

y = 0 + 1x1 + 2x2 + 3(x1 1)(x2 2) + uwhere (1, 2) is the population mean of (x1, x2)

2(1) is partial effect of x2(x1) on y at mean value of x1(x2).







58/129


Reparameterized model

y = 0 + 1x1 + 2x2 + 3(x1x2 + 12 x12 x21) + u= (0 + 312)

0+ (1 + 32)

1x1

+ (2 + 31) 2

x2 + 3x1x2 + u

For estimation purposes, can use sample mean in place of unknown

population meanEstimating reparameterized model has two benefits:

Provides estimates at average value (1, 2)Provides corresponding standard errors





Predicted Values and SEs I


59/129

Predicted Values and SEs I

Predicted value:y = 0 + 1x1 + ... + kxk

But this is just an estimate with a standard error. I.e.,

= 0 + 1c1 + ... + kck

where (c1, ..., ck) is a point of evaluation

But is just a linear combination of OLS parameters

We know how to get the SE of this. E.g., k = 1

Var() = Var(0 + 1c1)

= Var(0) + c21 Var(1) + 2c1Cov(0, 1)

Take square root and voila! (Software will do this for you)


Univariate Regression


Inference

Mechanics and Interpretation

OLS Estimator PropertiesFurther IssuesBinary Independent Variables

Predicted Values and SEs II


60/129

Predicted Values and SEs II

Alternatively, reparameterize the regression. Note = 0 + 1c1 + ... + kck = 0 = 1c1 ... kck

Plug this into the regression

y = 0 + 1x1 + ... + kxk + uto get

y = 0 + 1(x1 c1) + ... + k(xk ck) + uI.e., subtract the value c

jfrom each observation on x

jand then run

regression on transformed data.

Look at SE on intercept and thats the SE of the predicated value ofy at the point (c1,..., ck)

You can form confidence intervals with this too.




Inference



Predicting y with log(y) I


61/129

Predicting y with log(y) I

SRF:

log(y) = 0 + 1x1 + ... + kxk

Predicted value of y is not exp(log(y))

Recall Jensens inequality for convex function, g:

g

fd

g fd g(E(f)) E(g(f))

In our setting, f = log(y), g=exp(). Jensen =exp{E[log(y)]} E[exp{log(y)}]

We will underestimate y.




Inference



Predicting y with log(y) II


62/129

Predicting y with log(y) II

How can we get a consistent (no unbiased) estimate of y?If u X

E(y|X) = 0exp(0 + 1x1 + ... + kxk)where 0 = E(exp(u))

With an estimate of , we can predict y asy = 0exp(log(y))

which requires exponentiating the predicted value from the logmodel and multiplying by 0

Can estimate 0 with MOM estimator (consistent but biasedbecause of Jensen)

0 = n1

ni=1

exp(ui)




Inference



Basics


63/129

Basics

Qualitative information. Examples,1 Sex of individual (Male, Female)2 Ownership of an item (Own, dont own)3 Employment status (Employed, Unemployed

Code this information using binary or dummy variables. E.g.,

Malei = 1 if person i is Male

0 otherwise

Owni = 1 if person i owns item0 otherwise

Empi = 1 if person i is employed

0 otherwise

Choice of 0 or 1 is relevant only for interpretation.




Inference



Single Dummy Variable


64/129

Single Dummy Variable

Consider

wage = 0 + 0female + 1educ + u

0 measures difference in wage between male and female given samelevel of education (and error term u)

E(wage|female = 0, educ) = 0 + 1educE(wage|female = 1, educ) = 0 + + 1educ

= = E(wage|female = 1, educ) E(wage|female = 0, educ)Intercept for males = 0, females = 0 + 0




Inference



Intercept Shift


65/129

Intercept Shift

Intercept shifts, slope is same.




Inference



Wage Example


66/129

Wage Example

SRF with n = 526, R2 = 0.364

wage = 1.57 1.81female + 0.571educ + 0.025exper + 0.141tenureNegative intercept is intercept for men...meaningless because other

variables are never all = 0Females earn $1.81/hour less than men with the same education,experience, and tenure.

All else equal is important! Consider SRF with n = 526, R2 = 0.116

wage = 7.10 2.51female

Female coefficient is picking up differences due to omitted variables.




Inference



Log Dependent Variables


67/129

og epe de t a ab es

Nothing really new, coefficient has % interpretation.

E.g., house price model with N = 88, R2 = 0.649

price = 1.35 + 0.168log(lotsize) + 0.707log(sqrft)+ 0.027bdrms + 0.054colonial

Negative intercept is intercept for non-colonial homes...meaninglessbecause other variables are never all = 0A colonial style home costs approximately 5.4% more than otherwise

similar homesRemember this is just an approximation. If the percentage change islarge, may want to compare with exact formulation



68/129



Inference



Ordinal Variables


69/129

Consider credit ratings: CR (AAA, AA,..., C, D)If we want to explain bond interest rates with ratings, we couldconvert CR to a numeric scale, e.g., AAA = 1, AA = 2,... and run

IRi = 0 + 1CRi + ui

This assumes a constant linear relation between interest rates andevery rating category.

Moving from AAA to AA produces the same change in interest rates

as moving from BBB to BB.Could take log interest rate, but is same proportionate change muchbetter?




Inference



Converting Ordinal Variables to Binary


70/129

g y

Or we could create an indicator for each rating category, e.g.,CRAAA = 1 if CR = AAA, 0 otherwise; CRAA = 1 if CR = AA, 0otherwise, etc.

Run this regression:

IRi = 0 + 1CRAAA + 2CRAA + ... + m1CRC + ui

remembering to exclude one ratings category (e.g., D)

This allows the IR change from each rating category to have a

different magnitudeEach coefficient is the different in IRs between a bond with a certaincredit rating (e.g., AAA, BBB, etc.) and a bond with a ratingof D (the omitted category).




Inference



Interactions Involving Binary Variables I


71/129

g y

Recall the regression with four categories based on (1) marriagestatus and (2) sex.

log(wage) = 0.321 + 0.213marriedMale 0.198marriedFemale+

0.110singleFemale+ 0.079education

We can capture the same logic using interactions

log(wage) = 0.321 0.110female + 0.213married

+ 0.301femaile married + ...Note excluded category can be found by setting all dummies = 0

= excluded category = single (married = 0), male (female = 0)




Inference



Interactions Involving Binary Variables II


72/129

Note that the intercepts are all identical to the original regression.Intercept for married male

log(wage) = 0.321 0.110(0) + 0.213(1)

0.301(0) (1) = 0.534Intercept for single female

log(wage) = 0.321 0.110(1) + 0.213(0) 0.301(1) (0) = 0.211

And so on.

Note that the slopes will be identical as well.




Inference



Example: Wages and Computers


73/129

Krueger (1993),N

= 13, 379 from 1989 CPSlog(wage) = beta0 + 0.177compwork + 0.070comphome

+ 0.017compwork comphome + ...(Intercept not reported)

Base category = people with no computer at work or home

Using a computer at work is associated with a 17.7% higher wage.(Exact value is 100(exp(0.177) - 1) = 19.4%)

Using a computer at home but not at work is associated with a

7.0% higher wage.Using a computer at home and work is associated with a100(0.177+0.070+0.017) = 26.4% (Exact value is100(exp(0.177+0.070+0.017) - 1) = 30.2%)




Inference



Different Slopes


74/129

Dummies only shift intercepts for different groups.

What about slopes? We can interact continuous variables withdummies to get different slopes for different groups. E.g,

log(

wage) = 0 + 0

female+ 1

educ+ 1

educ

female+

u

log(wage) = (0 + 0female) + (1 + 1female)educ + u

Males: Intercept = 0, slope = 1

Females: Intercept = 0 + 0, slope = 1 + 1

= 0 measures difference in intercepts between males and females= 1 measures difference in slopes (return to education) between

males and females




Inference



Figure: Different Slopes I


75/129





Inference



Figure: Different Slopes I


76/129




Multivariate RegressionSpecification IssuesInference



Interpretation of Figures


77/129

1st figure: intercept and slope for women are less than those for men

= women earn less than men at all educational levels2nd figure: intercept for women is less than that for men, but slopeis larger

=

women earn less than men at low educational levels but the gap

narrows as education increases.= at some point, woman earn more than men. But, does this point

occur within the range of data?

Point of equality: Set Women eqn = Men eqn

Women: log(wage) = (0 + 0) + (1 + 1)educ + u

Men: log(wage) = (0) + 1educ + u

= e = 0/1Michael R. Roberts Linear Regression 77/129





Example 1


78/129

Consider N = 526, R2 = 0.441

log(wage) = 0.389 0.227female + 0.082educ 0.006female educ + 0.29exper 0.0006exper2 + ...

Return to education for men = 8.2%, women = 7.6%.

Women earn 22.7% less than men. But statistically insignif...why?Problem is multicollinearity with interaction term.

Intuition: coefficient on female measure wage differential betweenmen and women when educ = 0.Few people have very low levels of educ so unsurprising that we cant

estimate this coefficient precisely.More interesting to estimate gender differential at educ, for example.Just replace female educ with female (educ educ) and rerunregression. This will only change coefficient on female and itsstandard error.






Example 2


79/129

Consider baseball players salaries N = 330, R2 = 0.638

log(salary) = 10.34 + 0.0673years + 0.009gamesyr + ...

0.198black 0.190hispan+ 0.0125black percBlack+ 0.0201hispan percHisp

Black players in cities with no blacks (percBlack = 0) earn 19.8%less than otherwise identical whites.

As percBlack inc ( = percWhite dec since perchisp is fixed),black salaries increase relative to that for whites. E.g., ifpercBalck = 10% =

blacks earn -0.198 + 0.0125(10) = -0.073,

7.3% less than whites in such a city.

When percBlack = 20% = blacks earn 5.2% more than whites.Does this = discrimination against whites in cities with largeblack pop? Maybe best black players choose to live in such cities.




Functional Form Misspecification

Using Proxies for Unobserved VariablesRandom Coefficient ModelsMeasurement Error

Single Parameter Tests


80/129

Any misspecification in the functional form relating dependentvariable to the independent variables will lead to bias.

E.g., assume true model is

y = 0 + 1x1 + 2x2 + 3x22 + u

but we omit squared term, x22 .

Amount of bias in (0, 1, 2) depends on size of 3 and correlationamong (x1, x2, x

22 )

Incorrect functional form on the LHS will bias results as well (e.g.,log(y) vs. y)

This is a minor problem in one sense: we have all the sufficient data,so we can try/test as many different functional forms as we like.

This is different from a situation where we dont have data for arelevant variable.






RESET


81/129

Regression Error Sepecification Test (RESET)Estimate

y = 0 + 1x1 + ... + kxk + u

Compute predicted values yEstimate

y = 0 + 1x1 + ... + kxk + 1y2 + 2y

3 + u

(choice of polynomial is arbitrary.)

H0 : 1 = 2 = 0

Use F-test with F F2,nk3

Michael R. Roberts Linear Regression 81/129 Univariate Regression




Tests Against Nonnested Alternatives


82/129

What if we wanted to test 2 nonnested models? I.e., we cantsimply restrict parameters in one model to obtain the other.

E.g.,

y = 0 + 1x1 + 2x2 + u

vs.

y = 0 + 1log(x1) + 2log(x2) + u

E.g.,

y = 0 + 1x1 + 2x2 + u

vs.

y = 0 + 1x1 + 2z + u





Davidson-MacKinnon Test


83/129

Test

Model 1: y = 0 + 1x1 + 2x2 + u

Model 2: y = 0 + 1log(x1) + 2log(x2) + u

If 1st model is correct, then fitted values from 2nd model, (y),

should be insignificant in 1st modelLook at t-stat on 1 in

y = 0 + 1x1 + 2x2 + 1y + u

Significant 1 = rejection of 1st model.Then do reverse, look at t-stat on 1 in

y = 0 + 1log(x1) + 2log(x2) + 1y + u

where y are predicted values from 1st model.Significant 1 = rejection of 2nd model.





Davidson-MacKinnon Test: Comments


84/129

Clear winner need not emerge. Both models could be rejected orneither could be rejected.

In latter case, could use R

2

to choose.Practically speaking, if the effects of key independent variables on yare not very different, the it doesnt really matter which model isused.

Rejecting one model does not imply that the other model is correct.





Omitted Variables


85/129

Consider

log(wage) = 0 + 1educ + 2exper + 3ability + u

We dont observe or cant measure ability.

= coefficients are unbiased.What can we do?

Find a proxy variable, which is correlated with the unobserved

variable. E.g., IQ.

Michael R Roberts Linear Regression 85/129 Univariate Regression




Proxy Variables


86/129

Consider

y = 0 + 1x1 + 2x2 + 3x

3 + u

x3 is unobserved but we have proxy, x3

x3 should be related to x3 :

x3 = 0 + 1x3 + v3

where v3 is error associated with the proxys imperfect

representation of x3Intercept is just there to account for different scales (e.g., abilitymay have a different average value than IQ)





Plug-In Solution to Omitted Variables I


87/129

Can we just substitute x3 for x3 ? (and run

y = 0 + 1x1 + 2x2 + 3x3 + u

Depends on the assumptions on u and v3.

1 E(u|x1, x2, x3 ) = 0 (Common assumption). In addition,E(u|x3) = 0 = x3 is irrelevant once we control for (x1, x2, x3 )(Need this but not controversial given 1st assumption and status ofx3 as a proxy

2 E(v3|x1, x2, x3) = 0. This requires x3 to be a good proxy for x3E(x3 |x1, x2, x3) = E(x3 |x3) = 0 + 1x3

Once we control for x3, x

3 doesnt depend on x1 or x2





Plug-In Solution to Omitted Variables II


88/129

Recall true model

y = 0 + 1x1 + 2x2 + 3x

3 + u

Substitute for x3 in terms of proxy

y = (0 + 30) 0 + 1x1 + 2x2 + 33x3 + u+ 3v3 eAssumptions 1 & 2 on prev slide = E(e|x1, x2, x3) = 0 = wecan est.

y = 0 + 1x1 + 2x2 + 3x3 + eNote: we get unbiased (or at least consistent) estimators of(0, 1, 2, 3).

(0, 3) not identified.





Example 1: Plug-In Solution


89/129

In wage example where IQ is a proxy for ability, the 2nd assumptionis

E(ability|educ, exper, IQ) = E(ability|IQ) = 0 + 3IQThis means that the average level of ability only changes with IQ,not with education or experience.

Is this true? Cant test but must think about it.





Example 1: Cont.


90/129

If proxy variable doesnt satisfy the assumptions 1 & 2, well get

biased estimatesSuppose

x3 = 0 + 1x1 + 2x2 + 3x3 + v3

where E(v3|x1, x2, x3) = 0.Substitute into structural eqn

y = (0 + 30) + (1 + 31)x1 + (2 + 32)x2 + 33x3 + u+ 3v3

So when we estimate the regression:

y = 0 + 1x1 + 2x2 + 3x3 + ewe get consistent estimates of (0 + 30), (1 + 31),(2 + 32), and 33 assuming E(u+ 3v3|x1, x2, x3) = 0.Original parameters are not identified.





Example 2: Plug-In Solution


91/129

Consider q-theory of investment

Inv = 0 + 1q+ u

Cant measure q so use proxy, market-to-book (MB),

q = 0 + 1MB + v

Think about identifying assumptions1 E(u|q) = 0 theory say q is sufficient statistic for inv2 E(q|MB) = 0 + 1MB = avg level of q changes only with MB

Even if assumption 2 true, were not estimating 1 in

Inv = 0 + 1MB + e

Were estimating (0, 1) where

Inv = (0 + 10)

0

+ 11

1

MB + e



Inference

Functional Form MisspecificationUsing Proxies for Unobserved VariablesRandom Coefficient ModelsMeasurement Error

Using Lagged Dependent Variables as Proxies


92/129

Lets say we have no idea how to proxy for an omitted variable.

One way to address is to use the lagged dependent variable, whichcaptures inertial effects of all factors that affect y.

This is unlikely to solve the problem, especially if we only have onecross-section.

But, we can conduct the experiment of comparing to observationswith the same value for the outcome variable last period.

This is imperfect, but it can help when we dont have panel data.



Inference


Model I


93/129

Consider an extension to the basic model

yi = i + ixi

where i is an unobserved intercept and the return to educationdiffers for each person.

This model is unidentified: more parameters (2n) than observations(n)

But we can hope to identify avg intercept, E(i) = , and avgslope, E(i) = (a.k.a., Average Partial Effect (APE).

i = +ci, i = +

di

where ci and di are the individual specific deviation from averageeffects.

= E(ci) = E(di) = 0Michael R Roberts Linear Regression 93/129




Model II


94/129

Substitute coefficient specification into model

yi = + xi + ci + dixi + xi + uiWhat we need for unbiasedness is E(ui|xi) = 0

E(ui|xi) = E(ci + dixi|xi)This amounts to requiring

1 E(ci|xi) = E(ci) = 0 = E(i|xi) = E(i)2 E(di|xi) = E(di) = 0 = E(i|xi) = E(i)

Understand these assumptions!!!! In order for OLS to consistentlyestimate the mean slope and intercept, the slopes and interceptsmust be mean independent (at least uncorrelated) of theexplanatory variable.



Inference


What is Measurement Error (ME)?


95/129

When we use an imprecise measure of an economic variable in aregression, our model contains measurement error (ME)

The market-to-book ratio is a noisy measure of qAltmans Z-score is a noisy measure of the probability of defaultAverage tax rate is a noisy measure of marginal tax rate

Reported income is noisy measure of actual incomeSimilar statistical structure to omitted variable-proxy variablesolution but conceptually different

Proxy variable case we need variable that is associated withunobserved variable (e.g., IQ proxy for ability)

Measurement error case the variable we dont observe has awell-defined, quantitative meaning but our recorded measure containserror



Inference


Measurement Error in Dependent Variable


96/129

Let y be observed measure of y

y = 0 + 1x1 + ... + kxk + u

Measurement error defined as e0 = y yEstimable model is:

y = 0 + 1x1 + ... + kxk + u+ e0

If mean of ME = 0, intercept is biased so assume mean = 0If ME independent of X, then OLS is unbiased and consistent and

usual inference valid.If e0 and u uncorrelated than Var(u+ e0) > Var(u) =measurement error in dependent variable results in larger errorvariance and larger coef SEs



Inference


Measurement Error in Log Dependent Variable


97/129

When log(y) is dependent variable, we assume

log(y) = log(y) + e0

This follows from multiplicative ME

y = ya0

where

a0 > 0e0 = log(a0)



Inference


Measurement Error in Independent Variable


98/129

Model

y = 0 + 1x

1 + u

ME defined as e1 = x1

x1

AssumeMean ME = 0u x1 , x1, or E(y|x1 , x1) = E(y|x1 ) (i.e., x1 doesnt affect y aftercontrolling for x1 )

What are implications of ME for OLS properties?

Depends crucially on assumptions on e1

Econometrics has focused on 2 assumptions

Mi h l R R b t Li R i 98/129 Univariate Regression


Inference


Assumption 1: e1 x1


99/129

1s

t assumption is ME uncorrelated with observed measureSince e1 = x1 x1 , this implies e1 x1Substitute into regression

y = 0 + 1x1 + (u

1e1)

We assumed u and e1 have mean 0 and are with x1= (u 1e1) is uncorrelatd with x1.= OLS with x1 produces consistent estimator of coefs= OLS error variance is 2u + 21 2e1

ME increases error variance but doesnt affect any OLS properties(except coef SEs are bigger)

Mi h l R R b t Li R i 99/129 Univariate Regression


Inference


Assumption 2: e1 x1This is the Classical Errors in Variables (CEV) assumption and


100/129

This is the Classical Errors-in-Variables (CEV) assumption and

comes from representation:

x1 = x

1 + e1

(Still maintain 0 correlation between u and e1)

Note e1 x

1 =Cov(x1, e1) = E(x1e1) = E(x

1 e1) + E(e21 ) =

2e1

This covariance causes problems when we use x1 in place of x

1 since

y = 0

+ 1

x1

+ (u

1

e1

) and

Cov(x1, u 1e1) = 12e1I.e., indep var is correlatd with error = bias and inconsistent OLSestimates




Functional Form MisspecificationUsing Proxies for Unobserved Variables

Random Coefficient ModelsMeasurement Error

Assumption 2: e1 x1 (Cont.)


101/129

Amount of inconsistency in OLS

plim(1) = 1 +Cov(x1, u 1e1)

Var(x1)

= 1 +

12e1

2x1+ 2e1

= 1

1

2e1

2x1+ 2e1

= 1 2x12x1

+ 2e1






CEV asymptotic bias

From previous slide:


102/129

From previous slide:

plim(1) = 1 2x1

2x1+ 2e1

Scale factor is always < 1 = asymptotic bias attenuatesestimated effect (attenuation bias)If variance of error (2e1 ) is small relative to variance of unobservedfactor, then bias is small.

More than 1 explanatory variable and bias is less clear

Correlation between e1 and x1 creates problem. If x1 correlated with

other variables, bias infects everything.Generally, measurement error in a single variable casuesinconsistency in all estimators. Sizes and even directions of thebiases are not obvious or easily derived.






Counterexample to CEV Assumption


103/129

Consider

colGPA = 0 + 1smoked + 2hsGPA + u

smoked = smoked + e1

where smoked

is actual # of times student smoked marijuana andsmoked is reported

For smoked = 0 report is likely to be 0 = e1 = 0For smoked > 0 report is likely to be off = e1 = 0

= e1 and smoked

are correlated estimated effect (attenuation bias)I.e., CEV Assumption does not hold

Tough to figure out implications in this scenario




Regression Errors

HeteroskedasticityHypothesis Tests

Statistical Properties


104/129

At a basic level, regression is just math (linear algebra andprojection methods)

We dont need statistics to run a regression (i.e., compute

coefficients, standard errors, sums-of-squares, R2, etc.)What we need statistics for is the interpretation of these quantities(i.e., for statistical inference).

From the regression equation, the statistical properties of y come

from those of X and u




Regression Errors


What is heteroskedasticity (HSK)?


105/129

Non-constant variance, thats it.

HSK has no effect on bias or consistency properties of OLS

estimatorsHSK means OLS estimates are no longer BLUE

HSK means OLS estimates of standard errors are incorrect

We need an HSK-robust estimator of the variance of the coefficients.



106/129



Regression Errors


HSK-Robust LM-Statistics


107/129

The recipe:1 Get residuals from restricted model u2 Regress each independent variable excluded under null on all of the

included independent variables; q excluded variables =

(r1, ..., rq)

3 Compute the products between each vector rj and u4 Regression of 1 (a constant 1 for each observation) on all of the

products rju without an intercept5 HSK-robust LM statistic, LM, is N SSR1, where SSR1 is the sum

of squared residuals from this last regression.

6 LM is asymptotically distributed 2q




Regression Errors


Testing for HSK

The model


108/129

y = 0 + 1x1 + ... + kxk + u

Test H0 : Var(y|x1,..., xk) = 2E(u|x1, ..., xk) = 0 = this hypothesis is equivalent toH0 : E(u

2

|x1,..., xk) =

2 (I.e., is u2 related to any explanatoryvariables?)

u2 = 0 + 1x1 + ... + kxk + u

Test null H0 : 1 = ... = k = 0

F-test : F =R2u2

(1 R2u2

/(n k 1)LM-test : LM = N R2u2 (BP-test sort of)




Regression Errors


Weighted Least Squares (WLS)

Pre HSK-robust statistics, we did WLS - more efficient than OLS if


109/129

Pre HSK robust statistics, we did WLS more efficient than OLS if

correctly specified variance form

Var(u|X) = 2h(X), h(X) > 0X

E.g., h(X) = x21 or h(x) = exp(x)

WLS just normalizes all of the variables by the square root of thevariance fxn (

h(X)) and runs OLS on transformed data.

yi/

h(Xi) = 0/

h(Xi) + 1/(xi1/

h(Xi)) + ...

+ k/(xik/h(Xi)) + ui/h(Xi)yi = 0x

0 + 1x

1 + ... + kx

k + u

where x0 = 1/

h(Xi)



110/129



Regression Errors


Feasible Generalized Least Squares (FGLS) Recipe


111/129

Consider variance form:

Var(u|X) = 2exp(0 + 1x1 + ... + kxk)FGLS to correct for HSK:

1

Regress y on X and get residuals u2 Regress log(u2) on X and get fitted values g3 Estimate by WLS

y = 0 + 1x1 + ... + kxk + u

with weights 1/exp(g), or transform each variable (includingintercept) by multiplying by 1/exp(g) and estimate via OLS

FGLS estimate is biased but consistent and more efficient than OLS.




Regression ErrorsHeteroskedasticityHypothesis Tests

OLS + Robust SEs vs. WLS


112/129

If coefficient estimates are very different across OLS and WLS, itslikely E(y|x) is misspecified.If we get variance form wrong in WLS then

1 WLS estimates are still unbiased and consistent2 WLS standard errors and test statistics are invalid even in large

samples3 WLS may not be more efficient than OLS





Single Parameter Tests

Model


113/129

y = 0 + 1x1 + 2x2 + ... + kxk + u

Under certain assumptions

t(j) =j jse(j)

tnk1

Under other assumptions, asymptotically ta N(0, 1)

Intuition: t(j) tells us how far in standard deviations our

estimate j is from the hypothesized value (j)

E.g., H0 : j = 0 = t = j/se(j)E.g., H0 : j = 4 = t = (j 4)/se(j)





Statistical vs. Economic Significance


114/129

These are not the same thing

We can have a statistically insignificant coefficient but it may beeconomically large.

Maybe we just have a power problem due to a small sample size, or

little variation in the covariateWe can have a statistically significant coefficient but it may beeconomically irrelevant.

Maybe we have a very large sample size, or we have a lot of variationin the covariate (outliers)

You need to think about both statistical and economic significancewhen discussing your results.





Testing Linear Combinations of Parameters I

Model


115/129

y = 0 + 1x1 + 2x2 + ... + kxk + u

Are two parameters the same? I.e.,H0 : 1 = 2 (1 2) = 0The usual statistic can be slightly modified

t =1 2

se(1 2) tnk1

Careful: when computing the SE of difference not to forgetcovariance term

se(1 2) =

se(1)2 + se(2)

2 2Cov(1, 2)1/2





Testing Linear Combinations of Parameters II


116/129

Instead of dealing with computing the SE of difference, canreparameterize the regression and just check a t-stat

E.g., define = 1 2 = 1 = + 2 and

y = 0 + ( + 2)x1 + 2x2 + ... + kxk + u

= 0 + x1 + 2(x1 + x2) + ... + kxk + u

Just run a t-test of new null, H0 : = 0 same as previous slide

This strategy always works.





Testing Multiple Linear Restrictions

Consider H0 : 1 = 0, 2 = 0, 3 = 0 (a.k.a., exclusion


117/129

restrictions), H1 : H0nottrueTo test this, we need a joint hypothesis testOne such test is as follows:

1 Estimate the Unrestricted Model

y = 0 + 1x1 + 2x2 + ... + kxk + u

2 Estimate the Restricted Model

y = 0 + 4x4 + 5x5 + ... + kxk + u

3 Compute F-statistic

F = SSRR SSRU)/qSSRU/(n k 1) Fq,nk1

where q = degrees of freedom (df) in numerator = dfR dfU,n k 1 = df in denominator = dfU,





Relationship Between F and t Statistics

t2

has an F1,nk1 distribution.


118/129

n k 1,

All coefficients being individually statistically significant (significantt-stats) does not imply that they are jointly significant

All coefficients being individually statistically insignificant(insignificant t-stats) does not imply that they are jointly

insignificantR2 form of the F-stat:

F =R2U R2R)/q

(1

R2U)/(n

k

1)

(Equivalent to previous formula.)

Regression F-Stat tests H0 : 1 = 2 = ... = k = 0





Testing General Linear Restrictions I

Can write any set of linear restrictions as follows


119/129

H0 : R q = 0H1 : R q = 0

dim(R) = # of restrictions # of parameters. E.g.,H0 : j = 0 = R = [0, 0,..., 1, 0, ..., 0], q = 0H0 : j = k = R = [0, 0, 1,..., 1, 0, ..., 0], q = 0H0 : 1 + 2 + 3 = 1 = R = [1, 1, 1, 0,..., 0], q = 1H0 : 1 = 0, 2 = 0, 3 = 0 =

R =

1 0 0 0 ... 00 1 0 0 ... 00 0 1 0 ... 0

, q = 000





Testing General Linear Restrictions II

Note that under the null hypothesis


120/129

E(R q|X) = R0 q = 0Var(R q|X) = RVar(|X)R = 2R(XX)1R

Wald criterion:

W = (R q)[2R(XX)1R]1(R q) 2Jwhere J is the degrees of freedom under the null (i.e., the # ofrestrictions, the # of rows in R)

Must estimate 2, this changes distribution

F = (R q)[2R(XX)1R]1(R q) FJ,nk1where the n k 1 are df of the denominator (2)





Differences in Regression Function Across Groups I

Consider

7/29/2

linear regression slides

Documents