Scribd
Upload
3 views6 pages

Week 7

The document discusses the reasons why Cov(X, ε) is not equal to zero, highlighting issues like reverse causality, omitted variable bias, and inaccurate proxy variables. It introduces the Generalized Method of Moments (GMM) estimator as a solution to these problems, emphasizing the importance of having valid instruments that meet relevance and exogeneity conditions. Additionally, it reviews the Maximum Likelihood Estimation (MLE) approach and the necessity of conducting trinity tests to validate estimations.

Uploaded by

Mostafa Allam
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
3 views6 pages

Week 7

The document discusses the reasons why Cov(X, ε) is not equal to zero, highlighting issues like reverse causality, omitted variable bias, and inaccurate proxy variables. It introduces the Generalized Method of Moments (GMM) estimator as a solution to these problems, emphasizing the importance of having valid instruments that meet relevance and exogeneity conditions. Additionally, it reviews the Maximum Likelihood Estimation (MLE) approach and the necessity of conducting trinity tests to validate estimations.

Uploaded by

Mostafa Allam
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

Week 7

1. Why theoretically 𝐶𝑜𝑣 (𝑋, 𝜀 ) ≠ 0?


1. Reverse Causality (Endogeneity): X values are pre-determined
by the Y values.
2. Omitted Variable Bias. (Ramsey Test): If we did not include an
essential variable in our regression that is already correlated with
the X, it will be captured in the error term. This inflates the error
term and create correlation with this error term (that includes this
variable) and the X.
3. Inaccurate Proxy Variables.

𝑃𝑎𝑡𝑒𝑛𝑡𝑠 (𝑁𝑢𝑚𝑏𝑒𝑟)
𝑅𝐷 (𝐵𝑢𝑑𝑔𝑒𝑡 𝐸𝐺𝑃)
𝐼𝑛𝑛𝑜𝑣𝑎𝑡𝑖𝑜𝑛 (ℎ𝑎𝑠 𝑛𝑜 𝑢𝑛𝑖𝑡𝑠) = {
𝐶𝑜𝑝𝑦𝑟𝑖𝑔ℎ𝑡𝑠
𝑇𝑟𝑎𝑑𝑒𝑚𝑎𝑟𝑘𝑠

𝐼𝑛𝑐𝑜𝑚𝑒 𝑖𝑠 𝑖𝑛 𝐸𝐺𝑃 𝑜𝑟 𝑈𝑆𝐷

Many people in innovation literature use a proxy Trademarks.


I might have an innovation that is not recorded as Trademark.
This suggests that the proxy of trademark for innovation might
be associated with some errors.

𝑿 → 𝑰𝒏𝒏𝒐𝒗𝒂𝒕𝒊𝒐𝒏
𝑿∗ → 𝑻𝒓𝒂𝒅𝒆𝒎𝒂𝒓𝒌

𝑿 = 𝑿∗ + 𝒆 → 𝑻𝒓𝒖𝒕𝒉

𝑿∗ = 𝑿 − 𝒆

Regression:
𝒀 = 𝜶 + 𝜷𝑿 + 𝜺

𝑷𝒓𝒐𝒇𝒊𝒕𝒔 = 𝜶 + 𝜷 𝑰𝒏𝒏𝒐𝒗𝒂𝒕𝒊𝒐𝒏 + 𝜺

In practice what you did the below as you use data on


trademarks instead of innovation!

𝒀 = 𝜶 + 𝜷(𝑿∗ + 𝒆) + 𝜺

4. Non-Random Sampling techniques.

The Estimator GMM deals with the above issues! It assumes an availability of the
instrumental variable (Z) to be able to apply the estimator!

A good instrument should fulfill two conditions:

(1) Relevance condition: Strongly Correlated with the X.


(2) Exogeneity (Exclusion) Condition: Unrelated to Y (𝜀).

Two types of instruments:


(1) External instrument: Fulfills the two conditions.
(2) Internal instrument: Lagged information of the X. This might provide
weaker findings as exclusion condition might be violated!

GMM estimator have different forms.


Session 2
Summary of GMM (Conceptual):

Note 1:
𝑅𝑒𝑣𝑒𝑟𝑠𝑒 𝐶𝑎𝑢𝑠𝑎𝑙𝑖𝑡𝑦 (𝐸𝑛𝑑𝑜𝑔𝑒𝑖𝑛𝑡𝑦)
𝑂𝑚𝑖𝑡𝑡𝑒𝑑 𝑉𝑎𝑟𝑖𝑎𝑏𝑙𝑒 𝐵𝑖𝑎𝑠
𝑪𝒐𝒗(𝑿, 𝜺) ≠ 𝟎 {
𝐼𝑛𝑎𝑐𝑢𝑟𝑟𝑎𝑡𝑒 𝑃𝑟𝑜𝑥𝑦 𝑚𝑒𝑎𝑠𝑢𝑟𝑒
𝑁𝑜𝑛 − 𝑅𝑎𝑛𝑑𝑜𝑚 𝑆𝑎𝑚𝑝𝑙𝑖𝑛𝑔

Note 2:
𝑪𝒐𝒗(𝑿, 𝜺) ≠ 𝟎 cannot be detected by simple calculations. X might be
dependent on the error term 𝜺 in higher order moments (more complex
forms).

How do we capture the 𝐶𝑜𝑣(𝑋, 𝜀) ≠ 0?


(1) Assuming by logic/ previous literature.
(2) Test Durbin-Hausman-Wu formal test.
Note 3
What are the consequences that 𝑪𝒐𝒗(𝑿, 𝜺) ≠ 𝟎?
(1) Biased 𝛽̂ parameter either for OLS or MLE.
(2) Cannot evaluation MLE by Information matrix.

Note 4
What is the solution?
Solution is provided by the Generalized Method of Moments Estimator.
To apply this estimator, we need to have instruments (Z). We have two types of
instruments:
(1) Internal Instruments: Using Lag information of the Endogenous variable X.
(2) External Instrument:
a. Relevance: 𝐶𝑜𝑣(𝑋, 𝑍) ≠ 0.
b. Exclusion: 𝐶𝑜𝑣(𝑍, 𝜀) = 0.

Note 5
How do we get a lag in Econometrics?
Time GDP(t) GDP(t-1) GDP(t-2)
2015 2
2016 3 2
2017 -1 3 2
2018 4 -1 3
2019 1 4 -1
2020 0 1 4
2021 0 0 1
2022 1 0 0
2023 2 1 0
2024 3 2 1

𝑿 = 𝑮𝑫𝑷(𝒕)
𝒁 = 𝑰𝒏𝒕𝒆𝒓𝒏𝒂𝒍 𝑰𝒏𝒔𝒕𝒓𝒖𝒎𝒆𝒏𝒕 = 𝑮𝑫𝑷(𝒕 − 𝟐)
Review on MLE

(1) MLE is flexible Estimator compared to OLS. It assumes also that


𝐶𝑜𝑣(𝑋, 𝜀) = 0.
(2) The Advantage of this estimator is that it can deal with different structure
(distributions) of the Y.
a. Categorical (Education): 0,1,2,3,4.
b. Binary (Decisions): 0,1 (Dividends).
c. Censored (Minimum Wage).
d. Categorical (no order): A Company might decide to expand its
activities by Greenfield (new side), M&A, Extension of the current
site.
(3) We assume that we have information about PDF in Statistics

𝑁𝑜𝑟𝑚𝑎𝑙 → 𝑃𝑁
𝐵𝑖𝑛𝑜𝑚𝑖𝑎𝑙 → 𝑃𝐵
Θ = 𝑃𝑜𝑖𝑠𝑠𝑜𝑛 → 𝑃𝑃
..
( 𝑊𝑒𝑖𝑏𝑢𝑙𝑙 → 𝑃𝑤 )

(𝑃𝜃: 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝐷𝑒𝑛𝑠𝑖𝑡𝑦 𝐹𝑢𝑛𝑐𝑡𝑖𝑜𝑛).

When we apply MLE, we assume a given distribution for the 𝑌.

(4) To Find the Maximum Likelihood


a. Assume a Given PDF.
1 1 2
exp(− 𝜀 )
𝜎√2𝜋 2𝜎 2
b. Introduce the Joint Distribution

1 1 2
Π exp(− 𝜀 )
𝜎√2𝜋 2𝜎 2

c. We take log and Maximize


1
𝑙 = −𝑁 𝑙𝑛𝜎 − 𝑁 ln √2𝜋 − ∑𝜀 2
2𝜎 2

d. Every log Likelihood can be solved either by first order condition if


we have closed solution, or iterations if we have open formula.

(5) We cannot trust the estimations above unless we apply the trinity tests
(Wald, Likelihood Ratio test).

Those tests rely on three things:


(1) Gradient Vector ∇𝐺
(2) Hessian Matrix 𝐻
(3) Fisher information Matrix 𝐼(𝜃)

𝑊 = (𝜃𝑖 − 𝜃 ∗ )′ 𝐼(𝜃)(𝜃𝑖 − 𝜃 ∗ )~𝜒 2

𝐿𝑅 = 𝜆 = −2 × 𝐷𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝑜𝑓 𝐿𝑜𝑔 𝐿𝑖𝑘𝑒𝑙𝑖ℎ𝑜𝑜𝑑~𝜒 2

You might also like