QUANTITATIVE RESEARCH

METHODS II
VERA E. TROEGER
DEPARTMENT OF ECONOMICS

E-MAIL: [email protected]
OFFICE HOURS: BY APPOINTMENT
COURSE OUTLINE
1. Introduction to the Course: the OLS model, Gauss-Markov
Assumptions and Violations
2. Heteroskedasticity, cross-sectional correlation,
multicollinearity, omitted variable bias: tests and common
solutions.
3. Specification issues in Linear Models: Non-Linearities and
Interaction Effects
4. Dynamics, serial correlation and dependence over time
5. Pooling Data 1: Heterogeneity: How to choose the right
model – how good are the available tests.
COURSE OUTLINE CONT.
6. Pooling Data 2: Slope and Parameter Heterogeneity:
Seemingly Unrelated Regression and Random Coefficient
Models, modelling parameter instability
7. Specification Issues: Endogeneity and spatial
econometrics: instrumental variable approaches and
simultaneous equation models
8. Limited Dependent Variable Models: Binary and Ordered
Choice Models
9. Limited Dependent Variable Models with Pooled and Panel
Data
10. Specification Issues in Limited Dependent Variable
Models: Dynamics and Heterogeneity
11. Wrap Up, Q&A
THE SIMPLE LINEAR REGRESSION
MODEL
Regression models help investigating bivariate and multivariate
relationships between variables, where we can hypothesize that 1 variable
depends on another variable or a combination of other variables.

Relationships between variables in political science and economics are not

exact – unless true by definition, but relationships include most often a
non-structural or random component, due to the probabilistic nature of
theories and hypotheses in PolSci, measurement errors etc.

Regression analysis enables to find average relationships that may not be

obvious by just „eye-balling“ the data – explicit formulation of structural and
random components of a hypothesized relationship between variables.

Example: positive relationship between unemployment and government

spending
SIMPLE LINEAR REGRESSION
ANALYSIS
Linear relationship between x (explanatory variable) and y
(dependent variable)
Epsilon describes the random component of the linear relationship
between x and y
yi      x i  i
15
10
5
y
0
-5
-10

-2 0 2 4 6
x
 yi      x i  i

Y is the value of the dependent variable

(spending) in observation i (e.g. in the UK)
Y is determined by 2 components:
1. the non-random/ structural component alpha+beta*xi
– where x is the independent/ explanatory variable
(unemployment) in observation i (UK) and alpha and
beta are fixed quantities, the parameters of the
model; alpha is called constant or intercept and
measures the value where the regression line
crosses the y-axis; beta is called coefficient/ slope,
and measures the steepness of the regression line.
2. the random component called disturbance or error
term epsilon in observation i
A simple example:
x has 10 observations: 0,1,2,3,4,5,6,7,8,9
The true relationship between y and x is: y=5+1*x, thus, the true
y takes on the values: 5,6,7,8,9,10,11,12,13,14
There is some disturbance e.g. a measurement error, which is
standard normally distributed: thus the y we can measure
takes on the values: 6.95,5.22,6.36,7.03,9.71,9.67,10.69,13.85,
13.21,14.82 – which are close to the true values, but for any
given observation the observed values are a little larger or
smaller than the true values.
the relationship between x and y should hold on average true
but is not exact
When we do our analysis, we don‘t know the true relationship
and the true y, we just have the observed x and y.
We know that the relationship between x and y should have the
following form: y=alpha+beta*x+epsilon (we hypothesize a
linear relationship)
The regression analysis „estimates“ the parameters alpha and
beta by using the given observations for x and y.
The simplest form of estimating alpha and beta is called
ordinary least squares (OLS) regression
 OLS-Regression:
Draw a line through the scatter plot in a way to minimize the deviations of
the single observations from the line:

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

y7
epsilon7
yhat7

y1
y

alpha
yhat1

0 1 2 3 4 5 6 7 8 9 10
0 2 4 6 8 10 x
x
y Fitted values

ŷi  ˆ  ˆ  x i  ˆ i  ˆ i  yi  ˆ  ˆ  x i
Minimize the sum of all squared deviations from the line (squared
 
residuals)
This is done mathematically by the statistical program at hand
the values of the dependent variable (values on the line) are called
predicted values of the regression (yhat): 4.97,6.03,7.10,8.16,9.22,
10.28,11.34,12.41,13.47,14.53 – these are very close to the „true
values“; the estimated alpha = 4.97 and beta = 1.06
 OLS REGRESSION
Ordinary least squares regression: minimizes the squared
residuals
yi      x i  i  i  yi      x i
ŷi  ˆ  ˆ  x i  ˆ i
n n

 i i  i  min
(Y
i 1
Yˆ ) 2
(
ˆ ) 2

i 1

15
Components:

10
Component plus residual
DY: y; at least 1 IV: x
5
Constant or 0

intercept term: alpha

-5

Regression coefficient,
slope: beta
-10

-2 0 2 4 6
Error term, residuals: epsilon x
DERIVATION
Linear relationship between x (explanatory variable) and y (dependent
variable)
yi      x i  i
n n

 (Y Yˆ )   (ˆ )
i 1
i i
2

i 1
i
2
 min

 
n
 ˆ   yi  ˆ  x i  ˆ  y  ˆ  x
i 1

y 
n

i y
 ˆ  i 1
n

x
i 1
i  x

y 
n
 y  xi  x 
i
Cov  x, y 
 ˆ  i 1
  X ' X 1X ' y
n
V ar  x 
  xi  x 
2

i 1
Naturally we still have to verify whether ̂ and ̂
really minimize the sum of squared residuals and
satisfy the second order conditions of the
minimizing problem. Thus we need the second
derivatives of the two functions with respect to
alpha and beta which are given by the so called
Hessian matrix (matrix of second derivatives). (I
spare the mathematical derivation)

The Hessian matrix has to be positive definite (the

determinant must be larger than 0) so that ̂ and ̂
globally minimize the sum of squared residuals.
Only in this case alpha and beta are optimal
estimates for the relationship between the
dependent variable y and the independent variable
x.
Regression coefficient:

 (x i  x)(yi  y)
ˆ yx  i 1
n

 (x
i 1
i  x) 2

Beta equals the covariance between y and x divided by the

variance of x.
Interpretation: example
alpha=4.97, beta=1.06
Education and earnings: no education gives you a
minimal hourly wage of around 5 pounds. Each
additional year of education increases the hourly
wage by app. 1 pound:

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
beta=1.06

alpha

0 1 2 3 4 5 6 7 8 9 10
x

y Fitted values
Properties of the OLS estimator:

Since alpha and beta are estimates of the unknown parameters,

ŷi  ˆ  ˆ  x i estimates the mean function or the systematic part of
the regression equation. Since a random variable can be predicted best
by the mean function (under the mean squared error criterion), yhat can
be interpreted as the best prediction of y. the difference between the
dependent variable y and its least squares prediction is the least
squares residual: e=y-yhat =y-(alpha+beta*x).
A large residual e can either be due to a poor estimation of the
parameters of the model or to a large unsystematic part of the
regression equation
For the OLS model to be the best estimator of the relationship between x
and y several conditions (full ideal conditions, Gauss-Markov conditions)
have to be met.
If the „full ideal conditions“ are met one can argue that the OLS-
estimator imitates the properties of the unknown model of the
population. This means e.g. that the explanatory variables and the error
term are uncorrelated.
GAUSS-MARKOV ASSUMPTIONS,
FULL IDEAL CONDITIONS OF OLS
The full ideal conditions consist of a collection of assumptions about the true
regression model and the data generating process and can be thought of as a
description of an ideal data set. Ideal conditions have to be met in order for
OLS to be a good estimate (BLUE, unbiased and efficient)
Most real data do not satisfy these conditions, since they are not generated by an
ideal experiment. However, the linear regression model under full ideal
conditions can be thought of as being the benchmark case with which other
models assuming a more realistic DGP should be compared.
One has to be aware of ideal conditions and their violation to be able to control for
deviations from these conditions and render results unbiased or at least
consistent:
1. Linearity in parameters alpha and beta: the DV is a linear function of a set of IV
and a random error component
→ Problems: non-linearity, wrong determinants, wrong estimates; a relationship that
is actually there can not be detected with a linear model
2. The expected value of the error term is zero for all observations
E  i   0
→ Problem: intercept is biased
3. Homoskedasticity: The conditional variance of the error term is
constant in all x and over time: the error variance is a measure of
model uncertainty. Homoskedasticity implies that the model
uncertainty is identical across observations.

V  i   E  i ²   2  cons tan t

→ Problem: heteroskedasticity – variance of error term is different

across observations – model uncertainty varies from observation
to observation – often a problem in cross-sectional data, omitted
variables bias
4. Error term is independently distributed and not correlated, no
correlation between observations of the DV.

Cov  i ,  j   E  i  j   0, i  j

→ Problem: spatial correlation (panel and cross-sectional data), serial

correlation/ autocorrelation (panel and time-series data)
5. Xi is deterministic: x is uncorrelated with the error term since xi is
deterministic: Cov  X i , i   E  X i i   E  X i  * E  i 
 X i E  i   X i E  i   sin ce X i is det
0

→ Problems: omitted variable bias, endogeneity and simultaneity

6. Other problems: measurement errors, multicolinearity

If all Gauss-Markov assumptions are met than the OLS estimators alpha
and beta are BLUE – best linear unbiased estimators:
best: variance of the OLS estimator is minimal, smaller than the
variance of any other estimator
linear: if the relationship is not linear – OLS is not applicable.
unbiased: the expected values of the estimated beta and alpha
equal the true values describing the relationship between x and y.
INFERENCE
Is it possible to generalize the regression results for the sample under
observation to the universe of cases (the population)?
Can you draw conclusions for individuals, countries, time-points beyond
those observations in your data-set?
Significance tests are designed to answer exactly these questions.
If a coefficient is significant (p-value<0.10, 0.05, 0.01) then you can
draw conclusions for observations beyond the sample under
observation.
But…
Only in case the samples matches the characteristics of the population
This is normally the case if all (Gauss-Markov) assumptions of OLS
regressions are met by the data under observation.
If this is not the case the standard errors of the coefficients might be
biased and therefore the result of the significance test might be wrong
as well leading to false conclusions.
SIGNIFICANCE TEST: THE T-TEST

 2
 
2
,
N *Var  X 

The t-test:
T-test for significance: testing the H0 (Null-Hypothesis) that beta
equals zero: H0: beta=0; HA: beta≠0
The test statistic follows a student t distribution under the Null
ˆ  r ˆ  r
  t n  2
 
SE  ˆ SSR
N *Var  X 
ˆ ˆ
  t n  2
 
SE ˆ SSR
N *Var  X 

t is the critical value of a t – distribution for a specific number of

observations and a specific level of significance: convention in
statistics is a significance level of 5% (2.5% on each side of the t-
distribution for a 2-sided test) – this is also called the p-value.
Significance test – rule of thumb:
If the regression-coefficient (beta) is at least twice as large as the
corresponding standard error of beta the result is statistically significant
at the 5% level.
POWER OF A TEST
For a given test statistic and a critical region of a given significance
level we define the probability of rejecting the null hypothesis as the
power of a test
The power would be optimal if the probability of rejecting the null would
be 0 if there is a relationship and 1 otherwise.
This is, however, not the case in reality. There is always a positive
probability to draw the wrong conclusions from the results:

One can reject the null-hypothesis even though it is true (type 1 error,
alpha error):
alpha=Pr[Type I Error]
=Pr[rejecting the H0 | H0 is true]

Or not reject the null-hypothesis even though it is wrong (type 2 error, beta
error)
beta=Pr[Type II Error]
=Pr[accepting the H0 | Ha is true]
CONFIDENCE INTERVALS
Significance tests assume that hypotheses come before
the test: beta≠0. however, the significance test leaves us
with some vacuum since we know that beta is different from
zero but since we have a probabilistic theory we are not
sure what the exact value should be.

Confidence intervals give us a range of numbers that are

plausible and are compatible with the hypothesis.

As for significance test the researcher has to choose the

level of confidence (95% is convention)

Using the same example again: estimated beta is 1 and the

SE(beta) is 0.4 ; the critical value of the two-sided t-
distribution are 1.96 and -1.96
Calculation of the confidence interval:
The question is how far can a hypothetical value differ from the
estimated result before they become incompatible with the estimated
value?
The regression coefficient b and the hypothetical value beta are
incompatible if either
b b
 tcrit or  tcrit
SE  b  SE  b 

That is if beta satisfies the double inequality:

b  SE  b  * tcrit    b  SE  b  * tcrit
Any hypothetical value of beta that satisfies this inequality will therefore
automatically be compatible with the estimate b, that is will not be
rejected. The set of all such values, given by the interval between the
lower and upper limits of the inequality, is known as the confidence
interval for b. The centre of the confidence interval is the estimated b.
If the 5% significance level is adopted the corresponding confidence
interval is known as the 95% confidence interval (1% - 99%).
DEFINITIONS
Total Sum of Squares (SST):
n
SST    yi  y 
2

i 1

Explained
n
(Estimation) Sum of Squares (SSE):
SSE    yˆi  y 
2

i 1

Residualn Sum ofn

Squares or Sum of Squares Residuals (SSR):
SSR   ˆi2    yi     xi 
2

i 1 i 1
GOODNESS OF FIT
How well does the explanatory variable explain the dependent
variable?
How well does the regression line fit the data?

The R-squared (coefficient of determination) measures how much

variation of the dependent variable can be explained by the explanatory
variables.

The R² is the ratio of the explained variation compared to the total

variation: it is interpreted as the fraction of the sample variation in y that
is explained by x.

Explained variation of y / total variation of y:

n

 (Yˆ  Yˆ ) 2
SSE SSR
R 
2 i 1
n
  1

SST SST
(Y  Y ) 2

i 1
Properties of R²:

0 ≤ R² ≤ 1, often the R² is multiplied by 100 to get the percentage of the

sample variation in y that is explained by x
If the data points all lie on the same line, OLS provides a perfect fit to the
data. In this case the R² equals 1 or 100%.
A value of R² that is nearly equal to zero indicates a poor fit of the OLS
line: very little of the variation in the y is captured by the variation in the
y_hat (which all lie on the regression line)
R²=(corr(y,yhat))²
The R² follows a complex distribution which depends on the explanatory
variable
Adding further explanatory variables leads to an increase the R²
The R² can have a reasonable size in spurious regressions if the
regressors are non-stationary
Linear transformations of the regression model change the value of the R²
coefficient
The R² is not bounded between 0 and 1 in models without intercept
PROPERTIES OF AN ESTIMATOR
1. FINITE SAMPLE PROPERTIES
There are often more than 1 possible estimators to estimate a
relationship between x and y (e.g. OLS or Maximum Likelihood)
How do we choose between two estimators: the 2 mostly used
selection criteria are bias and efficiency.
Bias and efficiency are finite sample properties, because they describe
how an estimator behaves when we only have a finite sample (even
though the sample might be large)
In comparison so called “asymptotic properties” of an estimator have
to do with the behaviour of estimators as the sample size grows
without bound
Since we always deal with finite samples and it is hard to say whether
asymptotic properties translate to finite samples, examining the
behaviour of estimators in finite samples seems to be more important.
UNBIASEDNESS
UnBiasedness: the estimated coefficient is on average true:
That is: in repeated samples of size n the mean outcome of the estimate equals
the true – but unknown – value of the parameter to be estimated.


E   ˆ  0 
If an estimator is unbiased, then its probability distribution has an expected
value equal to the parameter it is supposed to be estimating. Unbiasedness
does not mean that the estimate we get with any particular sample is equal to
the true parameter or even close. Rather the mean of all estimates from
infinitely drawn random samples equals the true parameter.
3.5
3
2.52
Density
1.5 1
.5
0

.5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1
b1
SAMPLING VARIANCE OF AN
ESTIMATOR
Efficiency: is a relative measure between two estimators – measures
the sampling variance of an estimator: V(beta)
Let ̂ and  be two unbiased estimator of the true parameter  . With
variances V ˆ  and V   . Then ̂ is called to be relative more efficient
    than  .
than  if is smaller V  
V ˆ   
The property of relative efficiency only helps us to rank two unbiased
estimators.
.08

.4
.06

.3
Density

Density
.04

.2
.02

.1
0

-20 - 10 0 10 20 -4 -2 0 2 4
b1 b2
 TRADE-OFF BETWEEN BIAS AND
EFFICIENCY
With real world data and the related problems we sometimes have only
the choice between a biased but efficient and an unbiased but
inefficient estimator. Then another criterion can be used to choose
between the two estimators, the root mean squared error (RMSE). The
RMSE is a combination of bias and efficiency and gives us a measure
of overall performance of an estimator.
RMSE:


1 K
RMSE     
2.5

2
ˆ ˆ   true
K k 1
2


MSE  E    true 


2
xtfevd ˆ
 
1.5


MSE  Var   Bias ,  true 

 
Density

2
ˆ ˆ
 
1
.5

fixed effects
k measures the number
0

-.5 0 .5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5 of experiments, trials or
coefficient of z3 simulations
ASYMPTOTIC PROPERTIES OF
ESTIMATORS
We can rule out certain silly estimators by studying the asymptotic or large
sample properties of estimators.
We can say something about estimators that are biased and whose
variances are not easily found.
Asymptotic analysis involves approximating the features of the sampling
distribution of an estimator.
Consistency: how far is the estimator likely to be from the true parameter as
we let the sample size increase indefinitely.
If N→∞ the estimated beta equals the true beta:

lim Pr  ˆ n       0, p lim ˆ n  ,
n   
lim E ˆ n   
n 

Unlike unbiasedness, consistency involves that the variance of the

estimator collapses around the true value as N approaches infinity.
Thus unbiased estimators are not necessarily consistent, but those whose
variance shrink to zero as the sample size grows are consistent.
MULTIPLE REGRESSIONS
In most cases the dependent variable y is not just a function
of a single explanatory variable but a combination of several
explanatory variables.
THUS: drawback of binary regression: impossible to draw
ceteris paribus conclusions about how x affects y (omitted
variable bias).
Models with k-independent variables:

yi    1x i1  2 x i2  ...  k x ik  i

Control for omitted variable bias

But: increases inefficiency of the regression since
explanatory variables might be collinear.
OBTAINING OLS ESTIMATES IN MULTIPLE
REGRESSIONS
yi   0  1 xi1   2 xi 2   i

n n n n

 x  x2  x  x1  yi  y     xi1  x1  xi 2  x2    xi 2  x2  yi  y 
2
i2 i1
ˆ1  i 1 i 1 i 1 i 1
2
n n
 n 
 x  x1   x  x2      xi1  x1  xi 2  x2  
2 2
i1 i2
i 1 i 1  i 1 

ˆ  X ' X 1 X ' y

The intercept is the predicted value of y when all explanatory variables

equal zero.
The estimated betas have partial effect or ceteris paribus interpretations.
We can obtain the predicted change in y given the changes in each x.
when x_2 is held fixed then beta_1 gives the change in y if x_1 changes
by one unit.
“HOLDING OTHER FACTORS FIXED”

The power of multiple regression analysis is that it provides

a ceteris paribus interpretation even though the data have
not been collected in a ceteris paribus fashion.

Example: multiple regression coefficients tell us what effect

an additional year of education has on personal income if
we hold social background, intelligence, sex, number of
children, marital status and all other factors constant that
also influence personal income.
STANDARD ERROR AND SIGNIFICANCE
IN MULTIPLE REGRESSIONS

 
2 n
1
Var ˆ1 
SST1 1  R12 
 ˆ 2
 i
n   k  1 i 1

ˆ 2

n
SST1    xi1  x1 
2

i 1
n

  xˆi1  x1 
2

SSE
R12 for the regression of xi1 on xi 2 : R12   i 1
n

  xi1  x1 
SST 2

i 1


   
SD ˆ1  SE ˆ1 
SST1 1  R12 
F – TEST: TESTING MULTIPLE
LINEAR RESTRICTIONS
t-test (as significance test) is associated with any OLS coefficient.
We also want to test multiple hypotheses about the underlying parameters
beta_0…beta_k.
F

SSR  SSR / q r ur 
SSRur /  n  k  1
F  Fq , n  k 1
The F-test, tests multiple restriction: e.g. all coefficients jointly equal zero:
H0: beta_0=beta_1=…=beta_k=0
Ha: H0 is not true, thus at least one beta differs from zero
The F-statistic (or F-ratio) is defined as:
The F-statistic is F distributed under the
Null-Hypothesis.

F-test for overall significance of a regression: H0: all coefficients are jointly
zero – in this case we can also compute the F-statistic by using the R² of the
Regression:
R² / k
1  R ²  /  n  k  1
SSR_r: Sum of Squared Residuals of the restricted model (constant
only)
SSR_ur: Sum of Squared Residuals of the unrestricted model (all
regressors)
SSR_r can be never smaller than SSR_ur  F is always non-negative
k – number of explanatory variables (regressors), n – number of
observations, q – number of exclusion restrictions (q of the variables
have zero coefficients): q = df_r – df_ur (difference in degrees of
freedom between the restricted and unrestricted models; df_r > df_ur)
The F-test is a one sided test, since the F-statistic is always non-
negative
We reject the Null at a given significance level if F>F_critical for this
significance level.
If H0 is rejected than we say that all explanatory variables are jointly
statistically significant at the chosen significance level.
THUS: The F-test only allows to not reject H0 if all t-tests for all single
variables are insignificant too.
Goodness of Fit in multiple Regressions:

As with simple binary regressions we can define SST, SSE and SSR. And
we can calculate the R² in the same way.
BUT: R² never decreases but tends to increase with the number of
explanatory variables.
THUS, R² is a poor tool for deciding whether one variable of several
variables should be added to a model.
We want to know whether a variable has a nonzero partial effect on y in
the population.
Adjusted R²: takes the number of explanatory variables into account since
the R² increases with the number of regressors:

n 1
R 2
adj  1
nk
1  R 2

k is the number of explanatory variables and n the number of observations
COMPARING COEFFICIENTS
The size of the slope parameters depends on the scaling of the
variables (on which scale a variables is measured), e.g. population in
thousands or in millions etc.
To be able to compare the size effects of different explanatory variables
in a multiple regression we can use standardized coefficients:
ˆ x j ˆ j
bˆ j  for j  1,..., k
ˆ y
Standardized coefficients take the standard deviation of the dependent
and explanatory variables into account. So they describe how much y
changes if x changes by one standard deviation instead of one unit. If x
changes by 1 SD – y changes by b_hat SD. This makes the scale of
the regressors irrelevant and we can compare the magnitude of the
effects of different explanatory variables (the variables with the largest
standardized coefficient is most important in explaining changes in the
dependent variable).
SESSION 2
Testing for Violations of Gauss-Markov Assumptions
Solutions to deal with most common violations

Specification Issues:
Non-linear Relationships
Interaction Effects and Dummy Variables
PROBLEMS IN MULTIPLE REGRESSIONS:
1. MULTICOLINEARITY

Perfect multicolinearity leads to drop out of one of the variables: if x1

and x2 are perfectly correlated (correlation of 1) – the statistical
program at hand does the job.

The higher the correlation the larger the population variance of the
coefficients, the less efficient the estimation and the higher the
probability to get erratic point estimates. Multicolinearity can result in
numerically unstable estimates of the regression coefficients (small
changes in X can result in large changes to the estimated regression
coefficients).

Trade off between omitted variable bias and inefficiency due to

multicolinearity.
TESTING FOR MULTICOLINEARITY
Correlation between explanatory variables: Pairwise colinearity can be
determined from viewing a correlation matrix of the independent
variables. However, correlation matrices will not reveal higher order
colinearity.

Variance Inflation Factor (vif): measures the impact of collinearity among

the x in a regression model on the precision of estimation. vif detects
higher order multicolinearity: one or more x is/are close to a linear
combination of the other x.
Variance inflation factors are a scaled version of the multiple correlation
coefficient between variable j and the rest of the independent1 variables.
Specifically, VIF 
1 R
j 2
j

where Rj is the multiple correlation coefficient.

Variance inflation factors are often given as the reciprocal of the above
formula. In this case, they are referred to as the tolerances.
If Rj equals zero (i.e., no correlation between Xj and the remaining
independent variables), then VIFj equals 1. This is the minimum value.
Neter, Wasserman, and Kutner (1990) recommend looking at the largest
VIF value. A value greater than 10 is an indication of potential
multicolinearity problems.
POSSIBLE SOLUTIONS
Reduce the overall error: by including explanatory variables
not correlated with other variables but explaining the
dependent variable

Drop variables which are highly multi-collinear

Increase the variance by increasing the number of

observations

Increase the variance of the explanatory variables

If variables are conceptually similar – combine them into a

single index, e.g. by factor or principal component analysis
2. OUTLIERS
Problem:

The OLS principle implies the minimization of squared

residuals. From this follows that extreme cases can have a
strong impact on the regression line. Inclusion/exclusion of
extreme cases might change the results significantly.

The slope and intercept of the least squares line is very

sensitive to data points which lie far from the true
regression line. These points are called outliers, i.e.
extreme values of observed variables that can distort
estimates of regression coefficients.
TEST FOR OUTLIERS
symmetry (symplot) and normality (dotplot) of dependent variable gives
first indication for outlier cases
Residual-vs.-fitted plots (rvfplot) indicate which observations of the DV
are far away from the predicted values
lvr2plot is the leverage against residual squared plot. The upper left
corner of the plot will be points that are high in leverage and the lower
right corner will be points that are high in the absolute of residuals. The
upper right portion will be those points that are both high in leverage
and in the absolute of residuals.
DFBETA: how much would the coefficient of an explanatory variable
change if we omitted one observation?
DFBETA measures how much impact each observation has on a
particular coefficient is. The DFBETA for an explanatory variable and for
a particular observation is the difference between the regression
coefficient calculated for all of the data and the regression coefficient
calculated with the observation deleted, scaled by the standard error
calculated with the observation deleted. The cut-off value for DFBETAs
is 2/sqrt(n), where n is the number of observations.
 .08
Belgium

.06 Belgium

Belgium
Japan
Netherlands
Leverage

Japan
Belgium Belgium Switzerland
IrelandSweden
UKFinland Ireland Ireland
.04

Belgium
Ireland Ireland
Ireland
UK
UK Japan
Japan
Belgium Ireland
UKJapan Switzerland
Ireland
Austria Austria
Netherlands
UK
UK Austria
Belgium
UK
Italy Switzerland
Switzerland
Ireland Ireland
UKBelgium
Japan
Belgium
Italy Finland
UKCanadaAustralia
Netherlands
Netherlands
UK Australia
Ireland
Italy
ItalyItaly
Belgium
UK AustraliaFinlandUK Japan Australia Sweden
Belgium
US UK
Austria
Ireland
US Canada
Ireland US Japan Switzerland
Switzerland
Switzerland
France
Italy
UK
Belgium
Italy
Canada
Italy Italy
Italy
Germany
UK Australia
US
UKFinland
Ireland
UK
UK Japan
Italy
GermanyJapan
Canada
Japan Australia
Australia Switzerland
.02

Italy
Japan
CanadaCanada
Germany
US Canada
Germany
Ireland
Netherlands
Ireland Canada
Germany
Denmark Ireland
Netherlands Japan Australia
Denmark Switzerland
Ireland
Ireland
US Netherlands
Germany
Denmark
Netherlands
Germany
Ireland
Belgium
Germany
CanadaItalyUS
Ireland
Belgium
Finland
Netherlands
Italy
Ireland
Finland Ireland
Canada
Australia Switzerland
Switzerland
SwedenSweden
Japan
Germany
Germany
Ireland
Ireland
Italy
Canada
Finland
France
Japan
Italy
US
UK Sweden
Canada
UK Ireland
Netherlands
Finland
Germany
Canada
ItalyNorway
Italy France
Belgium
Italy
Italy Ireland
Ireland Australia
France
Switzerland
Australia Switzerland
Switzerland
France
Japan
US
Austria
Belgium US
Netherlands
Ireland
Canada
Netherlands
Canada
Italy Netherlands
Netherlands
Netherlands
BelgiumBelgium
Australia
UK Denmark AustraliaSwitzerland
Australia Switzerland
Australia Switzerland
Switzerland
Finland
UK
UK
US Ireland
Italy
UK Germany
Canada
Norway
UK
France
US
Norway
Austria
Finland
Austria
Canada
UK
Canada
DenmarkItaly
Canada
Finland
Austria
Canada Netherlands
Netherlands
Norway
Canada
Finland
Austria
Finland
SwedenNetherlands
Germany
Denmark
Germany
Japan
US Belgium
Italy
Italy
NorwayBelgium
Australia
Netherlands Australia
Australia Sweden Sweden
Sweden
Sweden
Denmark
Belgium
Japan
Belgium
US
Netherlands
Belgium
US
US Austria
Finland
Austria
France
Belgium
UK Canada
Finland
Norway Norway
Denmark
Italy
Netherlands
Italy
Norway
Norway
Belgium
Denmark
Netherlands Canada
Italy Finland
Belgium
Sweden
Netherlands
Sweden
Australia
Austria
Netherlands
US
Denmark
France
Austria
Canada France
Austria
Australia
N France
Australia
orway Norway Denmark
Switzerland
FranceNorway
Denmark Denmark Sweden
Japan
Netherlands
Austria
ItalyNorway
Canada Finland
Belgium
Austria Austria
Finland
Sweden
Norway
Netherlands
Austria
Austria
Sweden
USFrance Norway
Sweden
Netherlands
Norway
Finland
Norway
Japan
Sweden
US France
Sweden
Norway
Japan Australia
Norway
Norway
Canada
CanadaSweden
Austria
Germany
Japan Australia
Netherlands
Australia
Australia
Norway Denmark
DenmarkDenmark
Norway
DenmarkDenmark
Denmark
Denmark Denmark
Sweden
Denmark Sweden Sweden
US
Belgium
Austria
Denmark
Austria Austria
Finland
Finland
Japan
UK
US
Finland
US
Finland
Sweden
Sweden
Germany
Norway
Germany
Canada
Norway
US
Denmark
US
Austria
Austria France
Sweden
France
Australia Norway France
Denmark
Netherlands
Norway
France France
France
Germany Denmark Denmark DenmarkSweden Sweden Sweden Sweden
Netherlands
Austria
Germany
Germany
US DenmarkFinland
France
Germany
US
Austria
Canada
Australia
France
Finland Sweden
Finland
Denmark
France
France France
France
Norway
France
France
Australia Germany
Australia
Netherlands
Finland
Finland France
Germany
Australia
France
France Finland
France
Germany Sweden
Germany
Netherlands Norway
0

0 .005 .01 .015 .02 .025

Normalized residual squared
SOLUTIONS: OUTLIERS
Include or exclude obvious outlier cases and check their impact on the
regression coefficients.
Logarithmize the dependent variable and possibly the explanatory
variables as well – this reduces the impact of larger values.

jacknife, bootstrap:
Are both tests and solutions at the same time: they show whether single
observations have an impact on the results. If so, one can use the
jacknifed and bootstrapped coefficients and standard errors which are
more robust to outliers than normal OLS results.
Jacknife: takes the original dataset, runs the same regression N-1 times,
leaving one observation out at a time.
Example command in STATA: „jacknife _b _se, eclass: reg spend unem growthpc
depratio left cdem trade lowwage fdi skand “
Bootstrapping is a re-sampling technique: for the specified number of
repetitions, the same regression is run for a different sample randomly
drawn from the original dataset.
Example command: „bootstrap _b _se, reps(1000): reg spend unem growthpc
depratio left cdem trade lowwage fdi skand “
3. OMITTED VARIABLE BIAS
The effect of omitted variables that ought to be included:
Suppose the dependent variable y depends on two explanatory
variables:
yi   0  1 xi1   2 xi 2   i
But you are unaware of the importance of x2 and only include x
yi   0  1 xi1   i

If x2 is omitted from the regression equation, x1 will have a “double”

effect on y (a direct effect and one mimicking x2)
The mimicking effect depends on the ability of x1 to mimic x2 (the
correlation) and how much x2 would explain y
Beta1 in the second equation is biased upwards in case x1 and x2 are
positively correlated and downward biased otherwise
Beta1 is only unbiased if x1 and x2 are not related (corr(x1,x2)=0
However: including variable that is unnecessary, because it does not
explain an variation in y the regression becomes inefficient and the
reliability of point estimates decreases.
TESTING FOR OMITTED VARIABLES
Heteroskedasticity of the error term with respect to the observation of a
specific independent variable is a good indication for omitted variable
bias:
Plot the error term against all explanatory variables
Ramsey RESET F-test for omitted variables in the whole model: tests
for wrong functional form (if e.g. an interaction term is omitted):
• Regress Y on the X’s and keep the fitted value Y_hat ;
• Regress Y on the X’s, and Y_hat² and Y_hat³.
• Test the significance of the fitted value terms using an F test.
Szroeter test for monotonic variance of the error term in the explanatory
variables
Solutions:
Include variables that are theoretically important and have a high
probability of being correlated with one or more variables in the model
and explaining significant parts of the variance in the DV.
Fixed unit effects for unobserved unit heterogeneity (time invariant
unmeasurable characteristics of e.g. countries – culture, institutions)
4. HETEROSKEDASTICITY
The variance of the error term is not constant in each observation but
dependent on unobserved effects, not controlling for this problem violates
one of the basic assumptions of linear regressions and renders the
estimation results inefficient.
Possible causes:
Omitted variables, for example: spending might vary with the economic
size of a country, but size is not included in the model.
Test:
plot the error term against all independent variables
White test if the form of Heteroskedasticity is unknown
Breusch-Pagan Lagrange Multiplier test if the form is known

Solutions:
Robust Huber-White sandwich estimator (GLS)
White Heteroskedasticity consistent VC estimate: manipulates the
variance-covariance matrix of the error term.
More substantially: include omitted variables
Dummies for groups of individuals or countries that are assumed to behave
more similar than others
Tests for Heteroskedasticity:
a. Breusch-Pagan LM test for known form of Heteroskedasticity:
groupwise 2
n
T 2
 si 
LM    2
2 i 1  s
 1 


si2 =sum of group-specific squared residuals

s 2 = OLS residuals
H0: homoskedasticity ~ Chi² with n-1 degrees of freedom
LM-test assumes normality of residuals, not appropriate if
assumption not met.

b. Likelihood Ratio Statistic

Residuals are computed using MLE (e.g. iterated FGLS, OLS loss
of power)
 
2 ln      NT  ln  2    T ln  i2  ~  2  dF  n  1
c. White test if form of Heteroskedasticity is unknown:

V  i | xi    2
H0:
Ha: V  i | xi    i2

2
1. Estimate the model under H0ei
2. Compute squared residuals:
3. Use squared residuals as dependent variable of auxiliary
regression: RHS: all regressors, their quadratic forms and
interaction terms
ei2   0  1 xi 2  ...   k 1 xik   k 1 xi22   k 1 xi 2 xi 3  ...   q xik2  i

4. Compute White statistic from R² of auxiliary regression:

n * R 2 
a
  (2q )
5. Use one-sided test and check if n*R² is larger than 95%
quantile of Chi²-distribution
Robust White Heteroskedasticity Consistent Variance-
Covariance Estimator:

 2
Normal Variance of beta:  
2
,
N *Var  X 


ˆ  ˆ  1 ˆ
Robust White VC matrix: V    
n
  n  X ' X  X ' DX  X ' X 
1 1
ˆ ˆ

Dˆ  diag  ei2 

D is a n*n matrix with off-diagonals=0 and diagonal the squared residuals.

The normal variance covariance matrix is weighted by the non-constant
error variance.
Robust Standard errors therefore tend to be larger.
GENERALIZED LEAST SQUARES
APPROACHES
The structure of the variance covariance matrix Omega is used not just to adjust
the standard errors but also the estimated coefficient.
GLS can be an econometric solution to many violations of the G-M conditions
(Autocorrelation, Heteroskedasticity, Spatial Correlation…), since the Omega
Matrix can be flexibly specified
Since the Omega matrix is not known, it has to be estimated and GLS becomes
FGLS (Feasible Generalized Least Squares)
All FGLS approaches are problematic if number of observations is limited – very
inefficient, since the Omega matrix has to be estimated
1
Beta:  N
  N 
    X i '  X i    X i '  1 yi  
1

 i 1   i 1 
1
 N
  N

  X i '  X i    X i '  yi 
ˆ  ˆ 1 ˆ 1

 i 1   i 1 

X 
1
1
Estimated covariance matrix: ' X
Omega matrix with heteroscedastic error structure and
contemporaneously correlated errors, but in principle
FGLS can handle all different correlation structures…:

 12  21 31   n1 
 
 12  22 32   n 2 
   13  23 32   n3 
 
      
  2n 3n   n2 
 1n
HETEROSKEDASTICITY IN POOLED DATA
Is a bigger problem than in pure cross-sections since the error term
of one unit i can be correlated with the error term of unit j at the same
point in time and different points in time!

Test: Breusch-Pagan test, Cook-Weisberg test (estat hettest after

reg) – for unspecific heteroskedasticity, Ramsey test for omitted
variables (estat ovtest after reg), Szroeter’s test of heteroskedasticity
for specific right-hand side variables (estat szroeter).

Solutions:
• White Heteroskedasticity consistent VC estimate
• More substantially: include omitted variables
• Parks-Kmenta approach: FGLS
• Beck/ Katz: Panel Corrected Standard Errors
• More theoretically: Dummies for groups of individuals or countries
that are assumed to behave more similar than others
THE PARKS-KMENTA APPROACH
Basically a Feasible Generalized Least Squares Estimator which
takes the two-dimensional nature of the pooled data into account.

As all FGLS – problematic if number of observations is limited – very

inefficient, since the Omega matrix has to be estimated

The estimated covariance matrix is

Parks (1967) and Kmenta suggested an Omega matrix with panel
specific AR1 error structure and contemporaneously correlated errors,
but in principle FGLS can handle all different correlation structures…
The Parks method combines two sequential FGLS transformations, first
eliminating serial correlation of the errors then eliminating
contemporaneous correlation of the errors.

This is done by initially estimating an OLS equation.

The residuals from this estimation are used to estimate the unit-specific
serial correlation of the errors, which are then used to transform the
model into one with serially independent errors.

Residuals from this estimation are then used to estimate the

contemporaneous correlation of the errors, and the data is once again
transformed to allow for OLS estimation with now spherical errors.

Assumptions: , cross section specific variation

, for i ≠ j, cross-sectional correlation

, first-order serial correlation

PANEL CORRECTED STANDARD ERRORS
Beck and Katz 1995: APSR – stata command: xtpcse – as a critique of
the parks method (FGLS) that only can be used if T is as least as large
as N: MC experiments show that OLS PCSE is more efficient than
FGLS and at the same time avoids over-confidence in the estimation
results

Coefficient estimates of OLS are consistent but inefficient in pooled

data; degree of inefficiency depends on data and exact error process

B&K suggest to use OLS and correct only for the panel structure in the
standard errors:

where
For panel models with contemporaneously correlated and panel
heteroscedastic errors, Omega is an NT x NT block diagonal
matrix with an N x N matrix of contemporaneous covariances,
Sigma, along the diagonal. To estimate Omega, we need an
estimate of Sigma. Since the OLS estimates are consistent, one
can use the OLS residuals to provide a consistent estimate of
Sigma. Let e_it be the OLS residual for unit i at time t. A typical
element of Sigma is estimated by:

With the estimate Sigma_hat being comprised of all these

elements. This is used to form the estimator Omega_hat by
creating a block diagonal matrix with the Sigma matrices along
the diagonal.
As the number of time points increases, Sigma_hat becomes an
increasingly better estimator of Sigma.
PCSE
Treats cross-sectional correlation as nuisance – however
sometimes spatial effects are of theoretical interest

Manipulation of the standard errors of the OLS estimates: the size

of the SE (beta) is corrected by accounting for the panel structure
of the data

Serial correlation has to be treated beforehand

PCSE estimates are biased whenever OLS estimates are biased

5. AUTOCORRELATION
The observation of the residual in t1 is dependent on the
observation in t0: not controlling for autocorrelation violates on
of the basic assumptions of OLS and may bias the estimation
of the beta coefficients
i t  i t 1  it
Options:
lagged dependent variable
differencing the dependent variable
differencing all variables
Prais-Winston Transformation of the data
HAC constitent VC matrix

Tests:
Durbin-Watson, Durbin’s m, Breusch-Godfrey test
Regress e on lag(e)
Autocorrelation
The error term in t1 is dependent on the error term in t0: not controlling
for autocorrelation violates on of the basic assumptions of OLS and may
bias the estimation of the beta coefficients

i t  i t 1  it
The residual of a regression model picks up the influences of those
variables affecting the DV that have not been included in the regression
equation. Thus, persistence in excluded variables is the most frequent
cause of autocorrelation.
Autocorrelation does make no predictions about a trend, though a trend
in the DV is often a sign for serial correlation.
Positive autocorrelation: rho is positive: it is more likely that a positive
value of the error-term is followed by a one and a negative by a negative
one.
Negative autocorrelation: rho is negative: it is more likely that a positive
value of the error-term is followed by a negative one and vice versa.
DW test for first order AC: T

  et  et 1 
2

d t 2
T

e
t 1
2
t

Regression must have an intercept

Explanatory variables have to be deterministic
Inclusion of LDV biases statistic towards 2
Efficiency problem of serial correlation can be fixed by Newey-West
HAC consistent VC matrix for Heteroskedasticity of unknown form and
AC of order p: Problem: VC matrix consistent but coefficient can still be
biased! (HAC is possible with “ivreg2” in stata)
VˆNW  ˆ   T  X ' X  S *  X ' X 
1 1

S   et xt xt    l et et 1  xt xt' l  xt l xt' 
* 1 T 2 ' 1 p T
T t 1 T t 1 t l 1
1
l  1 
p 1
OR a simpler Test:

Estimate the model by OLS

compute the residuals

Regress the residuals on all independent variables

(including the LDV if present) and the lagged residuals

If the coefficient on the lagged residual is significant (with

the usual t-test), we can reject the null of independent
errors.
LAGGED DEPENDENT VARIABLE
yit    0 yit 1  k x it  i t
The interpretation of the LDV as measure of time-persistency is
missleading
LDV captures average dynamic effect, this can be shown by Cochrane-
Orcutt distributive lag models. Thus LDV assumes that all x-variables
have an one period lagged effect on y
 make sure interpretation is correct – calculating the real effect of x -
variables
Is an insignificant coefficient really insignificant if coefficient of lagged y
is highly significant? y    y   x 
it 0 i ,t 1 1 it it

yit  (   0 yi ,t 1   it )
1 
xit
FIRST DIFFERENCE MODELS

Differencing only the dependent variable – only if theory

predicts effects of levels on changes
FD estimator assumes that the coefficient of the LDV is
exactly 1 – this is often not true
Theory predicts effects of changes on changes
Suggested remedy if time series is non-stationary (has a
single unit root), asymptotic analysis for T→ ∞.
Consistent
 
K
yi t  yi t 1  k  x k i t  x k i t 1 i t  i t 1
k 1
K
 yi t  k  x k i t  i t
k 1
PRAIS-WINSTEN TRANSFORMATION
Models the serial correlation in the error term – regression
results for X variables are more straight forwardly interpretable:
yi t  x i t   i t with i t  i t 1  it
The it are iid – with N  0,  2 
 1  2  T 1 
The VC matrix of the error term is  
 1   T  2 
1  2
   1  T 3 
1  2  
      
T 1 T  2 T  3  1 


The matrix is stacked for N units. Diagonals are 1.

Prais-Winston is estimated by GLS. It is derived from the AR(1)
model for the error term. The first observation is preserved
1. Estimation of a standard linear regression:

y i t  x i t   i t
2. An estimate of the correlation in the residuals is then obtained by the
following auxiliary regression:
i t  i t 1  it
3. A Cochrane-Orcutt transformation is applied for observations
t=2,…,n
 
yi t  yi t 1   x i t  x i t 1  i t
4. And the transformation for t=1 is as follows:

1  2 y1    
1  2 x1  1  2 1

5. With Iterating to convergence, the whole process is repeated until

the change in the estimate of rho is within a specified tolerance, the
new estimates are used to produce fitted values for y and rho is re-
estimated, by:
 
yi t  yˆ i t   yi t 1  yˆ i t 1  i t
DISTRIBUTED LAG MODELS
Simplest form is Cochrane-Orcutt – dynamic structure of all
independent variables is captured by 1 parameter, either in
the error term or as LDV
If dynamics are that easy – LDV or Prais-Winston is fine –
saves Degrees of Freedom
Problem: if theory predicts different lags for different right
hand side variables – than a miss-specified model leads
necessarily to bias
Test down – start with relatively large number of lags for
potential candidates:
yi t  x i t 1  x i t 12  x i t  23    x i t  n n 1  i t
n  1, , t  1
SPECIFICATION ISSUES IN MULTIPLE
REGRESSIONS: 1. NON-LINEARITY
One or more explanatory variables have a non-linear effect on the
dependent variable: estimating a linear model would lead to wrong
or/and insignificant results. Thus, even though in the population there
exist a relationship between an explanatory variable and the dependent
variable, but this relationship cannot be detected due to the strict
linearity assumption of OLS

Test:
Ramsay RESET F-test gives a first indication for the whole model
In general, we can use acprplot to verify the linearity assumption
against an explanatory variable – though this is just “eye-balling”
Theoretical expectations should guide the inclusion of squared terms.
 Augmented component plus residual
-10 -5 0 5 10

.2 .4 0 .6
institutional openness to trade standardized
1 .8

Augmented component plus residual

-15 -10 -5 0 5 10
4 6 0
2
level of democracy
8
10
 SOLUTIONS
Handy solutions without leaving the linear regression framework:
Logarithmize the IV and DV: gives you the elasticity, higher values are weighted
less (engel curve – income elasticity of demand). This model is called a log-log
model or a log-linear model
log yi  log     log x i  log i

• Different functional forms give parameter estimates that have different substantial
interpretations. The parameters of the linear model have an interpretation as marginal
effects. The elasticities will vary depending on the data. In contrast the parameters of
the log-log model have an interpretation as elasticities. So the log-log model assumes
a constant elasticity over all values of the data set. Therefore the coefficients of a log-
linear model can be interpreted as percentage changes – if the explanatory variable
changes by one percent the dependent variable changes by beta percent.
• The log transformation is only applicable when all the observations in the data set are
positive. This can be guaranteed by using a transformation like log(X+k) where k is a
positive scalar chosen to ensure positive values. However, careful thought has to be
given to the interpretation of the parameter estimates.
• For a given data set there may be no particular reason to assume that one functional
form is better than the other. A model selection approachyi  is
 to
1 estimate
x i  2  x i2 competing
i
models by OLS and choose the model with the highest R-squared.

include an additional squared term of the IV to test for U-shape and inverse U-
shape relationships. Careful with the interpretation! The size of the two
coefficients (linear and squared) determines whether there is indeed a u-shaped
or inverse u-shaped relationship.
Hausken, Martin, Plümper 2004: Government Spending and Taxation in
Democracies and Autocracies, Constitutional Political Economy 15, 239-
59.
polity_sqr polity govcon
0 0 20.049292
0.25 0.5 18.987946
1 1 18.024796
2.25 1.5 17.159842
4 2 16.393084
6.25 2.5 15.724521
9 3 15.154153
12.25 3.5 14.681982
16 4 14.308005
20.25 4.5 14.032225
25 5 13.85464
30.25 5.5 13.775251
36 6 13.794057
42.25 6.5 13.911059
49 7 14.126257
56.25 7.5 14.43965
64 8 14.851239
72.25 8.5 15.361024
81 9 15.969004
90.25 9.5 16.67518
100 10 17.479551
The „u“ shaped relationship between democracy and government
spending:

21
government consumption in % of GDP
20

19

18

17

16

15

14

13
0 2 4 6 8 10
degree of democracy
2. INTERACTION EFFECTS
Two explanatory variables do not only have a
direct effect on the dependent variable but also a
combined effect
yi    1  x 1i  2  x 2i  3  x1i  x 2i  i

Interpretation: combined effect:

b1*SD(x1)+b2*SD(x2)+b3*SD(x1*x2)
Example: monetary policy of currency union has
a direct effect on monetary policy in outsider
countries but this effect is increased by import
shares.
Example:
government
spending
Government spending in % of GDP 50

45
Low cristian democratic
portfolio
High cristian democratic
portfolio
40

35
Low unemployment High unemployment
 70
6050
spend
40 30
20

0 5 10 15 20
unem

government spending trade+1sd

trade mean trade-1sd
 80
60
spend
40
20

0 50 100 150
trade

government spending unem+1sd

unem mean unem-1sd
INTERACTION EFFECTS OF
CONTINUOUS VARIABLES

Marginal Effect of Unemployment on Spending as Trade Openness changes

Dependent Variable: Government Spending
2
Marginal Effect of Unemployment
1.5
1
.5
0
-.5

0 20 40 60 80 100 120 140 160

Trade Openness

Marginal Effect of Unemployment on Spending

95% Confidence Interval
 2

.015
Marginal Effect of unemployment

Kernel Density Estimate of trade

1.5
on government spending

.01
.5 1

.005
0

Mean of international trade exposure

-.5

0
0 50 100 150
international trade exposure
Thick dashed lines give 95% confidence interval.
Thin dashed line is a kernel density estimate of trade.
3. DUMMY VARIABLES
An explanatory variable that takes on only the
values 0 and 1
Example: DV: spending, IV: whether a country
is a democracy (1) or not (0).
yi      D  i

Alpha then is the effect for non-democracies

and alpha+beta is the effect for democracies.