EC212: Introduction to Econometrics

Introduction
(Wooldridge, Ch. 1)

Tatiana Komarova

London School of Economics

Summer 2021

1
1. What is econometrics?
(Wooldridge, Ch. 1.1)

2
 Econometrics & Economic data

• Econometrics is set of tools to analyze data by economists or

other social scientists

• We can use econometrics to estimate economic relationships,

test economic theories, evaluate policy, and forecasting

3
 Data

• In contrast to natural sciences, data from controlled

experiments (called experimental data) are harder to obtain
in social sciences

• We usually have access to nonexperimental data (also called

observational data) which are collected in passive manner
after we observe outcomes on individuals, firms, schools, etc

4
2. Steps in empirical analysis
(Wooldridge, Ch. 1.2)

5
 Steps for empirical analysis

• Step 1: Carefully pose a question

• Step 2: Specify economic or conceptual model
• Step 3: Turn economic model into econometric model
• Step 4: Collect data on variables and use statistical methods
to estimate parameters, construct confidence intervals and
test hypotheses

6
 Example: Effect of job training

• To study effect of job training on worker productivity

(measured by hourly wage), we can start with equation like

wage = f (educ, exper , training )

where educ is measure of schooling, exper is measure of

workforce experience and training is measure of time spent in
job training (variable of most interest)

7
 Issues

• To turn this economic model into econometric model, we

resolve certain difficulties and ambiguities concerning
economic model

• 1. How should we measure these economic variables?

• 2. What is exact functional relationship among variables?

• 3. How do we account for unobserved factors that make

relationships among variables inexact?

8
 Example: Econometric model

• For job training example, we can specify an econometric

model as

wage = β0 + β1 educ + β2 exper + β3 training + u

• Constants β0 , β1 , β2 and β3 are parameters of the model

and we hope to estimate them (especially β3 in this example)

• Ideally we can collect data on wage, educ, exper and training

from large group of working people

9
• Last term u is called error term. It plays very important role.
It represents all other factors that can affects on wage
(e.g. intelligence, motivation, measurement error, etc)

• We use statistical methods to estimate and test hypotheses

about parameters by data. For example, hypothesis that job
training has no effect on wage is β3 = 0. Hypothesis that one
year of experience is worth one year of education is β1 = β2

10
3. Structure of economic data
(Wooldridge, Ch. 1.3)

11
 Cross-sectional data

• Economic data come in several different forms. First half of

this course focuses on cross-sectional data. Second half
discusses on time series and panel data

• Cross-sectional data are collected on individuals, families,

firms, schools, or some other units at given point in time

• In this course, we assume that cross-sectional data set

represents a random sample. That is, each unit in
population has same chance of appearing in sample and draws
are statistically independent of one another

12
 Random sampling

• Random sampling generates observations that are independent

and identically distributed

• Random sample is representative of population of interest and

gives us the best chance of learning about population

• Quick intuition: Population is everyone in this room. Sample

is people in some row randomly chosen
• Formally, if Y1 , Y2 , . . . , Yn are independent random
variables with a common probability distribution, then
{Y1 , ..., Yn } is said to be a random sample from this
distribution

13
When random sampling assumption is not suitable:
Example 1

• Sample selection

• Imagine we study the accumulation of wealth

• Some families might refuse to report their wealth

• If, for example, wealthier families are less likely to disclose their
wealth, then the resulting sample on wealth is not a random
sample

14
When random sampling assumption is not suitable:
Example 2

• When we sample from units that are large relative to the

population

• For example, imagine we want to explain new business activity

across different East Coast states in the US as a function of
wage rates, energy prices, tax rates, and other state
characteristics

• It is unlikely that business activities in states near one another

are independent

15
4. Causality and notion of ceteris
paribus
(Wooldridge, Ch. 1.4)

16
 Causality and ceteris paribus
• Concept of causality is key in econometrics
• How can we know that more spending causes better student
performance (on average)? How can we know another year of
education causes an increase in wages (on average)?
• The definition of the causal effect of x on y :

“How does variable y change if variable x is changed

but all other relevant factors are held constant”?

• Finding correlations in data might be suggestive but is rarely

conclusive
• To establish causality, notion of ceteris paribus (all relevant
factors equal) is crucial
• If we succeed in “holding fixed” other relevant factors, then
sometimes we establish that changes in one variable (say,
education) “cause” changes in another variable (wage) 17
Example: Does new fertilizer increase soybean yield?

• One way to determine causal effect of fertilizer is to conduct

experiment. Choose several one-acre plots of land, apply
different amounts of fertilizer to plots and record soybean
yields

• This experiment does not hold other relevant factors fixed

because, for example, land quality differs across plots

• It is impossible to truly hold all other factors fixed. But here is

important point: If fertilizer amounts are assigned
independently of other land factors that affect yield, then
simple regression analysis gives us “good” estimate of causal
effect

18
Example: What is value of another year of education on
earnings?

• Imagine experimental situation. At birth, each child is

randomly given highest grade that he/she must complete.
This is like randomly assigning fertilizer amounts (years of
education) to plots of land (individuals). Then we record their
earnings

• Such experiment is not feasible. We must usually rely on

nonexperimental data

• Problem for inferring causality from simple correlation analysis

(between years of education and earnings) is: individuals and
their parents largely determine amount of education

19
• Probably on average people who are smarter choose to
become better educated. But smarter people would earn more
on average

• Positive correlation between years of education and earnings

does not necessarily imply that it is due to education. Other
confounding factors (e.g. intelligence and past experience)
could explain most of difference in earnings

• We will learn to use multiple regression analysis to

convincingly (we hope) estimate causal effects

20
So, WHY do Correlation and Causation Differ?
Drowning, Ice Cream and Temperature

21
A “causal” relationship between two variables

22
So, WHY do Correlation and Causation Differ?
A third variable causes both ice cream consumption and drowning

23
 Reasons why A and B are correlated

• A causes B

• B causes A (reverse causality)

• A third variable C causes both A and B

• Dumb luck

Statistical techniques usually establish association, causality

requires interpretation

24
So, WHY do Correlation and Causation Differ?
Reverse causality...

25
So, WHY do Correlation and Causation Differ?
Reverse causality...

• Sometimes timing can help disentangle the causality but these

are only in the most simple of cases.
• In many situations, we require econometric techniques to
decipher the causal effect. Think about hospitals and health
status for example. A naive approach would conclude that
hospitals make people ill.
• Much of econometrics is designed to disentangle correlation
and causation.

26
 Spurious correlations (1)

Correlation=0.9979

27
 Spurious correlations (2)

Correlation=0.8701

28