Lecture 1:

Data and Models

Prof. Sharyn OHalloran
Sustainable Development U9611
Econometrics II
Why Statistics Matters: Harpers Index
Example 1: Logical Inconsistencies!?
Number of U.S. terrorism trials brought before a jury
since September 11, 2001 : 1
Number of terrorism convictions resulting : 2
Number of them dismissed in June due to a "pattern
of mistakes" by the prosecution : 2
Example 2: Comparing Different Units
Average number of clothing items an adult American
acquired in 2002 : 52
Estimated average amount of textiles thrown out by
each U.S. household in 2001, in pounds : 66
Lesson: Numbers dont speak for themselves.
General Approach: Data Visualization
Standard econometric approach
emphasizes:
Calculating the variance-covariance matrix
Applying fixes to get the right answer
Ease of statistical programs and increased
computing power emphasis has shifted to
a multifaceted view of data analysis
Graphical presentation
Question driven estimation
Interpretation and inference
General Approach: Question Driven
Modeling complex social science phenomena
become a series of choices:
What is the process by which the data will be
generated?
Random
Experimental
Observational
What is the appropriate estimation techniques?
Linear
Probabilistic
What is the scope of inference?
How general are the findings?
Has a lot to do with the research design.
Game Plan
Examine variables individually
Transform variables as needed
Examine key relations
Identify appropriate estimation techniques
OLS, Probit, Logit, etc
Define model specification
Which variables to include in the analysis
Derived both from inspection of the data and theory
Then run analysis
Perform post-regression diagnostics
Tests for significance, graphical analysis, simulations
Repeat!
Example: Growth and Democracy
Question: Does economic growth promote
transitions to democracy?
Traditional answer had been Yes.
Democracy is like a luxury good
This is one of the classic findings in political economy
Recent rejoinder (PACL) says money does not predict
transitions to democracy, but can help you stay there
once youre rich.
The focus here quickly turns to political institutions such as
property rights and the rule of law. (Rodrik, Shleifer et. al.)
The importance of getting this right is more than
academic.
A fundamental policy question is whether to promote
economic or political reform first.
Example: Growth and Democracy
How to address this debate?
Collect data yearly, across all countries
Measure of economic growth
GDP per capita
Political institution types:
Democracy
Autocracy
Other covariates
Education
Total Population

This is an example of the type of analysis that

we will be doing in the course.
And for your final paper!
Example: Growth and Democracy

Excel sheet
of data

Observational Data
Example: Growth and Democracy

Comma separated text,

as viewed in Notepad

Insheet data file into Stata

Insheet using fileName
The Importance of Being Normal
Before running any analysis, check the
distributions of all key variables.
Easiest to work with if they are normal:
Comparing normal distributions involves only
comparing means and standard deviations
Some statistical procedures assume variables
are normally distributed
Other procedures work better with normality

Other authors used straight GDP/capita

Example: Growth and Democracy
GDP Growth = (GDP-GDPt-1)/GDPt-1
Variable Obs Mean Std. Dev. Min Max
gdpgrowth 4269 0.0188657 0.0655765 -0.4189689 0.7769101

Unimodal Distribution

Kdensity to look at
distribution of
single variable.
 Example: Growth and Democracy
Polity Score = Ordinal ranking of how democratic a
country is, on a -10 to 10 scale
Variable Obs Mean Std. Dev. Min Max
polxnew 5671 -0.4456004 7.579176 -10 10

Distribution of Polity
.1

scores is bimodal.
Autocratic
.08

This means that we

might want to
Density
.06

separate the data into

Democratic two distinct
.04

categories.

Plus, number of
.02

partial democracies
-10 -5 0 5 10
polity score w/ -66=., -7 7=0, -8 8=next non-88 value is growing
 0 .2 .4 .6 .8 1

1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974

A u to c ra cy
1975
1976
1977

Fu ll D e m o cra c y
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
P a rtia l D e m o cra c y

1995
Over the last 50 years, move to democracy.

1996
1997
1998
Example: Growth and Democracy

1999
2000
The Importance of Being Normal
Per Capita GDP = Total GDP/ Population

Variable Obs Mean Std. Dev. Min Max

pwtrgdpc 4417 5713.802 5838.832 281.2581 27060.44
.0002
.00015

Per capita GDP is

Density

a unimodal but
.0001

skewed distribution
.00005
0

0 10000 20000 30000

real gdp per capita (chain in de x, constant 1996 international $)
re al gdp per capita (chain index, co nstant 1996 i nternational $)
0 10,000 20,000 30,000 Box Plot

Median
Outliers

1.5 X IQR
The Importance of Being Normal

Interquartile Range
 The Importance of Being Normal
Symmetry Plot Quantile-Normal Plot

re al gdp per capita (chain index, co nstant 1996 i nternational $)

30000
real gdp per capita (chain index, constant 1996 international $)
25000

20000
10000 15000 20000
Distance above median

-20000 -10000 0 10000

5000 0

0 1000 2000 3000 -20000 -10000 0 10000 20000 30000

Distance below median Inverse Normal

Shows that the values at the Relative to a normal

high end of the distribution distribution, there is more
are farther from the median weight at the left tail and less
than those at the low end. in the middle.
The Importance of Being Normal
Log of Per capita GDP = Log (Total GDP/ Population)
.4

Distribution is more
.3

symmetric now, even

if the tails are a little
Density
.2

thin.
.1
0

5 6 7 8 9 10
loggdp
The Importance of Being Normal
Note: Much more
Box Plot symmetric, with no
outliers.
10

1.5 X IQR
9

Median
Interquartile Range
8
7
6
 The Importance of Being Normal
Symmetry Plot Quantile-Normal Plot

12
loggdp
2.5

10
2
Distance above median
1.5

l og gdp
8
1

6
.5
0

4
0 .5 1 1.5 2 2.5 4 6 8 10 12
Distance below median Inverse Normal

Now this follows the 45 Much more similar to a

degree line almost exactly. normal distribution, except
for the thin tails.
 Inspecting Key Relationships
Growth and Democracy
10 5
Level of Democracy

Hard to tell
whats going
0

on here.
-5 -10

6 7 8 9 10
Log of GDP per capita
 Inspecting Key Relationships
Growth and Democracy
Add a lowess,
10

or local regression
line.
Level of Democracy
5

A data summary
technique.
0

Shows a clear
-5

positive relation
between the
-10

6 7 8 9 10 variables.
Log of GDP per capita
bandwidth = .8
 Inspecting Key Relationships
.15

But we should
check to see if
Autocratic
.1

democratic and
Probability

autocratic transitions
act differently.
.05

Democratic Both show a clear

impact of GDP on
transition
0

6 7 8 9 10 probabilities.
Log of GDP Per Capita, Lagged One Year

Transitions to Democracy Transitions to Autocracy

Estimation Techniques
Say we decide to look at transitions:
Democracy
Autocracy
Democracy Autocracy

Then the dependent variable has only two

values: Transition or No Transition
This type of qualitative dependent variable occurs
often in social science:
Voting for a Republican or Democrat
Supreme Court Decisions overrule or uphold
Yea and Nay votes when passing legislation, etc
Appropriate estimation technique is Probit.
Estimates nonlinear probabilities
 Comparing Model Specifications
PACL regress transitions on:
GDP per capita
GDP per capita squared

GDP growth

Adjusting for previous regime type

------------------------------------------------------------------------------
anydem | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Lpdem | 7.993817 7.57117 1.06 0.291 -6.845404 22.83304
Lgdp | -1.284701 1.053302 -1.22 0.223 -3.349136 .7797335
Lpdemgdp | -1.575778 1.951307 -0.81 0.419 -5.400269 2.248714
Lgdp2 | .0909221 .0676323 1.34 0.179 -.0416348 .2234789
Lpdemgdp2 | .1274521 .1248902 1.02 0.307 -.1173281 .3722323
Lgrowth | -.4754997 .6829151 -0.70 0.486 -1.813989 .8629892
Lpdemgrowth | -1.48283 1.18237 -1.25 0.210 -3.800234 .8345726
_cons | 2.670554 4.066588 0.66 0.511 -5.299812 10.64092
------------------------------------------------------------------------------
 Comparing Model Specifications
PACL regress transitions on:
GDP per capita
GDP per capita squared

GDP growth

Adjusting for previous regime type

------------------------------------------------------------------------------
anydem | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Lpdem | 7.993817 7.57117 1.06 0.291 -6.845404 22.83304
Lgdp | -1.284701 1.053302 -1.22 0.223 -3.349136 .7797335
Lpdemgdp | -1.575778 1.951307 -0.81 0.419 -5.400269 2.248714
Lgdp2 | .0909221 .0676323 1.34 0.179 -.0416348 .2234789
Lpdemgdp2 | .1274521 .1248902 1.02 0.307 -.1173281 .3722323
Lgrowth | -.4754997 .6829151 -0.70 0.486 -1.813989 .8629892
Lpdemgrowth | -1.48283 1.18237 -1.25 0.210 -3.800234 .8345726
_cons | 2.670554 4.066588 0.66 0.511 -5.299812 10.64092
------------------------------------------------------------------------------
Comparing Model Specifications
Why include GDP per capita and its
square as independent variables?
You would do this to check if a variable has a
curvilinear effect.
For example, higher levels of incomes have a
negative impact on transitions.
But if the impact is not significant and there is
no good theoretical reason to include it, it
should be dropped from the regression.
 Alternative Model Specification
So let us try the same analysis without the
square term.
------------------------------------------------------------------------------
anydem | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Lpdem | -.0334039 .7746361 -0.04 0.966 -1.551663 1.484855
Lgdp | .1283458 .0646331 1.99 0.047 .0016672 .2550243
Lpdemgdp | .4573869 .0991983 4.61 0.000 .2629618 .651812
Lgrowth | -.4904828 .6819738 -0.72 0.472 -1.827127 .8461612
Lpdemgrowth | -1.515618 1.176133 -1.29 0.198 -3.820797 .789561
_cons | -2.756896 .4980252 -5.54 0.000 -3.733007 -1.780784
------------------------------------------------------------------------------
 Alternative Model Specification
So let us try the same analysis without the
square term.
------------------------------------------------------------------------------
anydem | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
Lpdem | -.0334039 .7746361 -0.04 0.966 -1.551663 1.484855
Lgdp | .1283458 .0646331 1.99 0.047 .0016672 .2550243
Lpdemgdp | .4573869 .0991983 4.61 0.000 .2629618 .651812
Lgrowth | -.4904828 .6819738 -0.72 0.472 -1.827127 .8461612
Lpdemgrowth | -1.515618 1.176133 -1.29 0.198 -3.820797 .789561
_cons | -2.756896 .4980252 -5.54 0.000 -3.733007 -1.780784
------------------------------------------------------------------------------

It works!
Significant results with other covariates
added, as well (population, education, etc.)
Regression Diagnostics
So GDP predicts democracy
But does this really have a substantive
impact?
Raising GDP from its
Impact of GDP on Transitions to Democracy minimum to maximum
.07

value increases the

probability of a
.06

Fitted democratic transition

Transition to Democracy

Regression from 2% to 6.5%.

.05

Line
This decreases the
.04

expected life of an
.03

autocratic regime from

34 years to just 10
.02

6 7 8 9 10 years.
Log of GDP Per Capita, Lagged One Year
Regression Diagnostics
So GDP predicts Democracy,
But does this really have a substantive
impact?
Impact of GDP on Transitions to Autocracy
.25

Raising GDP from its

minimum to maximum
.2

value decreases the

Fitted
probability of an
.15

Regression
autocratic transition
paut

Line
from 23% to
.1

(essentially) 0%.
.05
0

6 7 8 9 10
Log of GDP Per Capita, Lagged One Year
The End (Or the Beginning)
What else could you do with this analysis?
Add more covariates
Education
Population

Resource Curse

Treat data differently

Use entire democracy-autocracy scale, rather than
dividing it into discrete categories
Treat this as a survival problem

Others?
Class Organization
Text: Statistical Sleuth
Website: CourseWorks
Grades:
Homework: 35%
Midterm: 25%
Final Paper and Presentation: 35%
Participation: 5%