Methodological Research Seminar

Panel Data Analysis and Instrumental Variables in Action in Finance

Galina Besstremyannaya

Professor, Faculty of Economic Sciences, Senior Researcher at the International Laboratory for Macroeconomic
Analysis, National Research University Higher School of Economics, Russian Federation, Moscow

Apr 1, 2024

Besstremyannaya Apr 1, 2024 1 / 34

Overview

Apr 1, 2024
Panel data analysis (review and work at home)
Instrumental variables (in class)
Online resources on advanced econometrics and work with Harvard Dataverse (in class,
sections)
Apr 15, 2024
Heterogeneity: latent class models and quantile regressions (in class)
Online resources on advanced econometrics
Data archives and repositories: data and codes for replications
An example of a replication paper on instrumental variable quantile regression
Work with MIT data archives on returns to education papers (in class, sections)

Besstremyannaya Apr 1, 2024 2 / 34

Homework (in groups) - due Apr 25, 2024
Use: Harvard Dataverse to search for papers by Bloom and/or van Reenen OR MIT data
archives (web pages of any MIt faculty member) OR replicate the paper by Hall and
Jones (1999) (only 1 group of of 14) OR Smith, B. (2019). Dutch disease and the oil
boom and bust. Canadian Journal of Economics/Revue canadienne d’économique, 52(2),
584-623 (only 1 group out of 14). NO OTHER online resource or papers for replications
are allowed!
In MIT data archives find a paper on panel data analysis, endogeneity or heterogeneity
(quantile regression or latent class models) with data and codes for replications. In
Harvard Data verse find any paper coauthored by Bloom or van Reenen on panel data
analysis, endogeneity or heterogeneity (quantile regression or latent class models). NO
OTHER topics are allowed.

Besstremyannaya Apr 1, 2024 3 / 34

Homework

File 1: Write a short essay (in Word/LaTex) with 14 pt size for all text AND codes AND tables. Essays with
codes and tables of smaller font are not graded! The essay should include
Your names and groups within HSE!! Nonamed files are not graded!
(Grade 6) Your search strategy and Contents of data archive: what files are available and how they can be
used as supporting documentation for the paper and/or for replication
(Grade 7) Short CRITICAL outline of the paper (topic, research question, methods, results) with details
on what results are supported by codes and/or data files (e.g. with tabulation). Do not copy-paste from
the paper, write in your own words. Include equations and describe notations and indices.
(Grade 8) Replicate the key regression table in the paper. Include commands and their output IN the text
of your essays. Provide comments to the command. If they are present in the paper, include the parts of
analysis based on the necessary steps in each method as studied in class. E.g. Tests on the quality of
instruments and verbal arguments; verbal arguments in favor or RE or FE models; robustness checks
including various corrections of the variance-covariance matrix (e.g. for clustered data),
inclusion/exclusion of time effects, interaction terms; tests of equality of coefficients in latent class models
and condition al quantile regression.

Besstremyannaya Apr 1, 2024 4 / 34

Homework (cont.)

(Grade 9) If they are absent in the paper, conduct the parts of analysis based on the necessary steps in
each method as studied in class. E.g. Tests on the quality of instruments and verbal arguments; verbal
arguments in favor or RE or FE models; robustness checks including various corrections of the
variance-covariance matrix (e.g. for clustered data), inclusion/exclusion of time effects, interaction terms;
tests of equality of coefficients in latent class models and condition al quantile regression. Include
commands and their output IN the text of your essays. Provide comments to the command.
(Grade 10) Do any reasonable addition to the analysis in the paper. A Show what extra results you
obtained and why they are interesting/important.
File 2: Your code with command (with comments, where you distinguish YOUR comments from the pre-exiting
author comments). File 3: Your log file/output file with commands and results of their execution.
Email files 1,2 and 3 to gbesstremyannaya@hse.ru by 10.00 April 25, 2024. LATE WORK IS NOT ACCEPTED!

Besstremyannaya Apr 1, 2024 5 / 34

Panel Data Analysis

Brooks, C. (2019) Introductory Econometrics for Finance. 4th Edition, Cambridge University
Press.
Chapter 11, slides 32–43.

Besstremyannaya Apr 1, 2024 6 / 34

Instrumental variables

Theoretical overview: relevance and exogeneity of instruments

Example from productivity analysis: Hall and Jones (1999)
Example from corporate finance: Bloom et al. (2007)
Example from stock market: Mironov and Petronevich (2015), HSE

Besstremyannaya Apr 1, 2024 7 / 34

Part 1: Theoretical overview

Setup: y = X β + u, y is n × 1, X is n × k.
Some of the variables in X are endogenous.
One reason for endogeneity is REVERSE CAUSALITY.
Let Z be a matrix of instrumental variables.
Identification conditions:
1) Instruments are not weak: X is correlated with Z .
2) Instruments are exogenous: Z is uncorrelated with u.

Besstremyannaya Apr 1, 2024 8 / 34

Identification tests
Setup
y = Y β1 + X1 β2 + u = X β + u (a)
Y = X1 Π1 + X2 Π2 + V = Z Π + V (b),
where n is sample size,
Y is n × p matrix of endogenous regressors,
X1 is n × k1 matrix of exogenous regressors, k1 + p = k,
X2 is n × k2 matrix of excluded exogenous instruments, k2 ≥ p,
Z = [X1 X2 ] matrix of instruments, X = [Y X1 ] matrix of regressors, k1 + k2 ≥ k1 + p.
1. Test for weak instruments: the goodness-of-fit statistic in the first stage regression
Y = X1 Π1 + X2 Π2 + V = Z Π + V
2. Test for exogeneity of instruments: does not exist!
Test of overidentifying restrictions: Simultaneously tests whether “the instruments are
uncorrelated with the error term” [and] “one or more of the excluded exogenous variables
should in fact be included in the structural equation [(a)]” (Stata Base Reference Manual.
Release 17. p.1161).
Besstremyannaya Apr 1, 2024 9 / 34
Note on the first stage statistic
The term “weak instruments” appeared in Stager and Stock (1994)
The tests and tabulation of the first-stage statistic are given in Stock, Wright and Yogo
(2002) and Stock and Yogo (2005) in case of homoskedasticity. See ivregress
postestimation, ffirst in Stata
F statistic for one-dimensional Y , Stock and Yogo (2005) tests for multi-dimensional Y
(NOT F-statistic in each first-stage regression for each component of Y ).
The coefficient for a particular instrument may be insignificant, but the Stock and Yogo
(2005) test may prove the validity of instruments!
Another test in case of heteroskedasticity, serial correlation and clustered standard errors
was developed in Montiel Olea and Pflueger (2013) and implemented by Pflueger and
Wang (2015) weakivtest
Montiel Olea, J. L., and C. E. Pflueger. 2013. A robust test for weak instruments. Journal of Business and Economic Statistics 31: 358–369.
Staiger, D.O. and Stock, J.H., 1994. Instrumental variables regression with weak instruments. NBER WP0151.
Stock, J. H., J. H. Wright, and M. Yogo. 2002. A survey of weak instruments and weak identification in generalized method of moments. Journal of Business and
Economic Statistics 20: 518–529.
Stock, J. H., and M. Yogo. 2005. Testing for weak instruments in linear IV regression. In Identification and Inference for Econometric Models: Essays in Honor of
Thomas Rothenberg, ed. D. W. K. Andrews and J. H. Stock, 80–108. New York: Cambridge University Press.
Besstremyannaya Apr 1, 2024 10 / 34
Note on the overidentification test

Suppose there are k1 = 4 exogenous regressors and p = 2 endogenous regressors.

Z includes k1 exogenous regressors and k2 excluded regressors. The requirement here is
k2 ≥ p.
The overidentification test applies to the case when the number k2 of the excluded
regressors is greater than the number of endogenous regressors: k2 > p.
The overidentification test essentially demands that p of the excluded (aka additional)
regressors (instruments) are exogenous and tests whether the remaining k2 − p
instruments are exogenous.
Accordingly, the result of overidentification test is INSUFFICIENT for full justification of
the exogeneity of instruments. So the exogeneity assumption MUST be based on
economic grounds.

Besstremyannaya Apr 1, 2024 11 / 34

Part 2: Hall and Jones (1999)

Hall, R. E., Jones, C. I. (1999). Why do some countries produce so much more output per
worker than others? The Quarterly Journal of Economics, 114(1), 83-116. NBER WP 1998.
Empirical estimation of causes of variation in output per worker in different countries.
11,780 citations in Google scholar as of Oct 2021!
Uses version 5.6 of Penn World Tables, Data for 127 countries in 1950–1992, Cross-sectional
data for 1988, IV regression to explain variation in output per worker.

Besstremyannaya Apr 1, 2024 12 / 34

Overview of the paper

Field: Macroeconomics, Growth, Growth decomposition, Solow model of exogenous growth

and estimates of the total factor productivity
Novelty: Decompose growth for a large sample of countries. Evaluate the impact of
technological change on growth. Estimate the causes of variation in the amount of per capita
output.
Methodology: Statistical analysis of growth decomposition. Instrumental variable regression of
per capita output, social infrastructure is an endogenous variable and is instrumented.
Note: Dataset and Stata codes are available for REPLICATIONS. See www.empiricalde.com
Dataset for Ch.2 of Soderbom et al. (2015) Empirical Development Economics

Besstremyannaya Apr 1, 2024 13 / 34

Two parts in the paper: Statistical part

Yi = Kiα (Ai Hi )1−α , (1)

where Yi is output in country i, Li is labor, Hi is human capital (a function of years of
schooling), Ai is the labor-augmenting measure of productivity.

Yi  K α/(1−α) H
i i
= Ai (2)
Li Yi Li

Hit = e ϕ(Eit ) Lit , (3)

where ϕ is a function of the number of years of schooling Eit .
Piece-wise linear ϕ(Eit ), returns to primary education are based on estimates by
Psacharopoulus (1994) for sub-Saharan Africa, OECD average estimates for returns to
secondary education.
Besstremyannaya Apr 1, 2024 14 / 34
Two parts in the paper: Econometric part

Results from the statistical analysis: output per worker differs across countries
Explanation: countries use technology inefficiently and the reason is differences in the social
infrastructure.
Social infrastructure (P.14 of the WP): “the institutions and government policies that provide
the incentives for individuals and firms in an economy”. May either stimulate productive
activities (innovation, investment in human capital) or cause predatory actions (crime,
corruption).
However, social infrastructure is endogenous to production (REVERSE CAUSALITY), so the
use of instrumental variable approach is required for estimation.

Besstremyannaya Apr 1, 2024 15 / 34

Social infrastructure: defined as an average of the two measures
1) government anti-diversion policies (country-level data from Political Risk Services, reports
24 variables (categories) on risk to international investors in the country),
the Knack and Keefer (1995) approach of using an average of 5 categories: law and order;
business bureaucratic quality; government protection from corruption; from risk of
expropriation; government repudation of contracts. An aggregate index on a [0,1] scale.
2) openness to international trade (Sachs and Warner (1995) index). Fraction of years during
the period 1950 to 1994 when economy was open; on a [0,1] scale.

Besstremyannaya Apr 1, 2024 16 / 34

Econometric part of Hall and Jones (1999), cont’d
Instruments for social infrastructure: exposure to Western culture/expansion of Western
culture.
Approximated by: distance from equator, predicted trade share of national economy (Frankel
and Romer, 1996) and the extent for which 5 European languages (English, French, German,
Portuguese, Spanish) are spoken as first languages.
More specifically,
1) fraction of population speaking English,
2) fraction of population speaking a European language (any of the above 5 languages,
including English)
3) distance from equator,
4) predicated trade share based on a gravity model of international trade.
Frankel and Romer (1996)
https://www.nber.org/system/files/working papers/w5476/w5476.pdf
Trade share is ”country’s total trade (exports plus imports) divided by its GDP”.
Besstremyannaya Apr 1, 2024 17 / 34
Non weak instruments

Economic grounds (P.21 of the WP): “Our instruments are positively correlated with social
infrastructure. Western Europe discovered the ideas of Adam Smith, teh importance of
property rights, and the system of checks and balances in government, and the countries that
were strongly influenced by Western Europe were, other things equal, more likely to adopt
favorable infrastructure”.
1) Languages of the Western Europe spoken as a native language - “naturally” (P.21 of the
WP) correlated with the exposure to Western culture.
2) Distance to equator: Western Europeans were likely to migrate to regions FAR from
equator (sparsely populated and similar in climate).
Econometric analysis: results of the first-stage regression (Tables 2-3).
Homework (not to be graded): find the value of F statistic in the paper.

Besstremyannaya Apr 1, 2024 18 / 34

Exogeneity of instruments

“To satisfy this criterion, we must ask if European influence was somehow more intensively
targeted toward regions of the world that are more likely to have high output per worker
today.” (P.22 of the WP).
The answer by Hall and Jones (1999) is negative, since Europeans tended to settle in sparsely
populated regions where land was NOT productive. (P.22 of the WP).

Besstremyannaya Apr 1, 2024 19 / 34

Part 3: Bloom and van Reenen (2007)

Bloom N., van Reenen JV. (2006) Measuring and explaining management practices across
firms and countries. NBER Working Paper 12216. QJE journal paper 2007. 3,500 citations in
google scholar as of Oct 2021.
Differences in productivity of firms may be attributed to the effect of the quality of
management.
Novelty of the paper: a survey to measure managerial practices across 732 medium-sized
manufacturing firms in the US, UK, France in Germany.
Conducted in 2004.
The purpose of the paper
1) To investigate the effect of management on production.
2) To account for endogeneity of managerial practices by using instruments:
CEO choice in family firms and product market competition.

Besstremyannaya Apr 1, 2024 20 / 34

Quantification of management

18 variables (practices) on management, expertly evaluated on a rank scale from 1 (the worst)
to 5 (the best).
4 areas:
operations (5 measures);
monitoring (5 measures);
targets (5 measures);
incentives (5 measures).
Each practice (measure) is translated into z-score (standardization to have zero mean and
standard deviation equal to one):
z = (x − µ)/σ.
An average of all measures is taken as the z-score for management at the firm.
Results - pics for clustering of firms in each country

Besstremyannaya Apr 1, 2024 21 / 34

Management as an explanatory variable in productivity regression

Specification at the country level. For a given country

logYit = αl logLit + αk logKit + αn logNit + βMi + γZit + uit , (4)

where i is firm, t is year, Y is deflated sales of the firm, K is capital, L is labor, N is materials,
M is management score (time invariant!), z are control variables (denoted ”general controls”:
log(average hours worked), log(firm age), dummy for listed firm, dummy for consolidated firm,
share of workforce with degrees, share of workforce with MBAs, dummies for 3 digit code
industries) and u is error term.
Data are then pooled across firms and countries (and years), with introduction of four country
dummies interacted with the full set of annual dummies. Coefficients in the regression
equation are assumed not to vary across countries. See Table 2 on P.59 of the NBER WP.

Besstremyannaya Apr 1, 2024 22 / 34

Reverse causality

See P.43-44 of the NBER WP

“Our estimated effects of the “true effect” of management on productivity could be biased
upwards or downwards due to reverse causality.
For example, positive feedback could occur if higher productivity enabled cash-constrained to
invest more resources in improving managerial practices. This would bias our coefficient on
management upwards.
Negative feedback could occur if higher productivity allows managers to reduce their input of
effort. This would bias the coefficient on management downwards.”
Homework (not to be graded): Is upward or downward bias discovered by the authors?

Besstremyannaya Apr 1, 2024 23 / 34

Instruments for management
Note: This section is in the NBER WP ONLY, not in the QJE article See WP: “V.E
Instrumental variable estimates of management practices in the production function”
1) Family firms (assumption of worse management at these firms)
Family largest shareholder
Family largest shareholder and family CEO
Family largest shareholder and family CEO and primo geniture (the eldest son is
approinted as CEO at the family firm)
2) Product market competition (lower competition implies worse management)
Measures computed as average values across years (see P.38):
Degree of import penetration in a country (share of total imports in domestic production);
Lerner index of competition = 1- profits/sales;
Survey question on the number of competitors (a rank variable, with the following
categories: 0 for no competitors, 1 for less than 5, 2 for 5 or more competitors).
Besstremyannaya Apr 1, 2024 24 / 34
Non weak instruments
Economicjustification: PP.1355–1360 of the QJE paper
1) Negative association between family ownership and management.
“the impact of changing CEOs and CFOs” (P.1355),
“the knowledge that family members will receive management positions in the future may
generate a “Carnegie effect” of reducing their investment in human capital earlier in life.
These selection and Carnegie effects are likely to be much more negative for primogeniture
family firms, in which the eldest son is destined to control the firm from birth” (P.1359).
2) Positive association between product market competition and management.
“the productivity-enhancing effects of competition work through improving average
management practices” (P.1355).
“Higher competition can increase managerial effort, as the fear of bankruptcy is higher
(Schmidt 1997). In addition, the sensitivity of market share to marginal cost differences is
greater under higher competition, so this increases themarginal return to managerial effort.”
(P.1368)
The results of the first stage regression: Tables 4-6 on PP.61–63 of the WP
Besstremyannaya Apr 1, 2024 25 / 34
Exogenenity of instruments

“We need to .. assume that ... the mechanism by which competition (and ownership) impacts
productivity is solely through improving managerial practices.
Based on these admittedly very strong identification assumptions the instrumental variable
strategy identifies the causal effect of management on performance.”

Besstremyannaya Apr 1, 2024 26 / 34

Part 4: Dutch disease in Russia

References
Sonin, K. (2012) Economics of “The Natural Resource Curse”
https://cpb-us-w2.wpmucdn.com/voices.uchicago.edu/dist/1/2748/files/2019/04/SoninResourceCurseJuly2013.pdf

Kalcheva, K., Oomes, N. (2007). Diagnosing Dutch disease: does Russia have the
symptoms? No. 7-207. BOFIT Discussion Papers. Bank of Finland, BOFIT Institute for
Economies in Transition
Mironov, V. V., Petronevich, A. V. (2015). Discovering the signs of Dutch disease in
Russia. Resources Policy, 46, 97-112.
Dobrynskaya, V., Turkisch, E. (2010). Economic diversification and Dutch disease in
Russia. Post-Communist Economies, 22(3), 283-302.
Su, C. W., Qin, M., Tao, R., Umar, M. (2020). Does oil price really matter for the wage
arrears in Russia? Energy, 208, 118350.

Besstremyannaya Apr 1, 2024 27 / 34

Dutch disease: history

“The Dutch disease”. The Economist. Saturday, 26 November 1977. Volume: 265, Issue:
7004. Pages 82–83.
The discovery of the large quantity of natural gas in Groningen (the Netherlands) in 1959 led
to subsequent decrease in the industrial production.
Other maladies:
1) “too strong a currency”(P.82)
2) “real incomes are higher... 10% than they would have been” (PP.82-83)
3) “use of gas revenues” (P.83). The lion’s share goes to government (80%).

Besstremyannaya Apr 1, 2024 28 / 34

Dutch disease: overview

Increase in revenues from export causes deceleration of growth in the manufacturing

sector.
So a country rich in rents from natural resources may experience lower growth in
manufacturing and lower long-run growth.
Rich natural resources also cause political instability, conflicts etc. adverse effects on
economic institutions.
Indeed, countries rich in natural resources on average grow slower than other countries.

Besstremyannaya Apr 1, 2024 29 / 34

Symptoms

1 Appreciation of the real exchange rate (e.g. due to surplus in balance of payments owing
to rise in oil price)
2 Decrease in the growth of manufacturing sector (manufacturers in the tradable sectors of
home economy lose their competitiveness due to real exchange rate appreciation)
3 Increase in wages (the increase in wages is higher than growth in productivity)
4 Growth in the service sector

Besstremyannaya Apr 1, 2024 30 / 34

Russia as an energy “superpower”
U.S. Energy Information Administration. ”What countries are the top producers and
consumers of oil?” https://www.eia.gov/tools/faqs/faq.php?id=709&t=6 [Data as of
2019]
Russia is currently the third largest oil producer, over 11% of crude oil production across
the whole world.
Largest oil producers: US (became number one in 2013, number two in 2012, 19% of
production), Saudi Arabia (12%), Russia (11%).
Report of energy minister A.Novak at Oil and Gas Forum (2013)
https://minenergo.gov.ru/node/3400
About 1/3 of GDP and 2/3 of export comes from oil and gas.
Ministry of Finance data on execution of budgets since 2006
https://minfin.gov.ru/ru/statistics/fedbud/execute/?id 65=80041-

yezhegodnaya informatsiya ob ispolnenii federalnogo byudzhetadannye s 1 yanvarya 2006 g.

As of 2019: 41% of revenues of the Federal budget come from oil and gas.
Besstremyannaya Apr 1, 2024 31 / 34
Dutch disease in Russia: empirical strategy

Time series analysis, e.g. dependent variable is real exchange rate and the list of
explanatory variables includes oil price (See Mironov and Petronevich, 2015); or
dependent variable is output in manufacturing and the list of explanatory variables
includes oil price (See Kalcheva and Oomes, 2007).
Test dependent variable and explanatory variables for stationarity.
In case of non-stationarity, apply cointegration test to variable in the model. Estimate the
model in the first differences in absence of cointegration.
In presence of cointegration, search for cointegration vector(s) for the dependent and
explanatory variables. Then estimate (vector) error correction model: the dependent
variable and explanatory variables in the first differences, error correction term as an
additional regressor (See Mironov and Petronevich, 2015, Kalcheva and Oomes, 2007).

Besstremyannaya Apr 1, 2024 32 / 34

Mironov and Petronevich (2015)

Evaluates the major symptom of the Dutch disease in Russia: appreciation of the real
exchange rate due to rise in oil price.
The dependent variable is the real effective exchange rate.
Endogenous variables:
oil price
the volume of exported oil
differences in labor productivity in Russia verses its trade partners

One reason for endogenity here: simultaneity in time-series models.

Control variables: government expenditure as share of GDP and net international reserves.

Besstremyannaya Apr 1, 2024 33 / 34

Vector error correction model

Vector model: multi-dimentional dependent variable (e.g. exchange rate and oil price are
taken as a vector).
VAR model: autoregressive lags are included.
Estimation strategy:
Cointegration equation. (Each component of) the dependent variable is a function of
endogenous variables and controls. Obtain the residuals. (Table 1 on P.102, with
exception of the last line)
Error correction model: first difference in (each component) of the dependent variable is a
function of the first difference in the explanatory variables and of the error correction
term. (Table 1 on P.102, the last line - a part of the results of the VECM for real
exchange rate as dependent variable).

Besstremyannaya Apr 1, 2024 34 / 34