Lecture 1 – AERM Macro

Florian Trouvain
University of Oxford

October 17, 2024

1
Introduction
• Goal of this class is to expose you to research methods in macro
• Hard to do in 3 weeks – macro nowadays combines methods from many fields
• Methods you will learn can be applied in many other contexts
◦ Dynamic optimization
− Perturbation methods (extremely flexible)
− Global solution methods (Bellman, discrete vs. continous time)

◦ General equilibrium
• Things I won’t be teaching
◦ Time-Series econometrics (Plagborg-Møller and Wolf, 2021; Sims, 1980)
◦ Applied microeconometrics (Nakamura and Steinsson, 2014, 2018)
2
What we will do

• Combination of theory, coding, and empirics

• Focus on particular context to apply methods

1. Barro and Sala-i Martin (1992): Growth across US regions

2. Aiyagari (1994), Achdou et al. (2022): Heterogenous agent general equilibrium
model
3. Micro data and macro model (Aiyagari + HH panel data)

3
1. Barro and Sala-i Martin (1992)

• Paper studies economic growth ( YL ) across US regions

◦ Revisit neoclassical model

◦ Perturbation method
◦ Related to large literature on cross-country income differences

• A classic – perhaps a precursor to more recent stuff in style (Nakamura and

Steinsson, 2014)

4
2. Aiyagari + 3. Household Level Data
1. Aiyagari (1994): heterogenous agent general equilibrium model
◦ Combine neoclassical model with uninsurable household income risk
◦ Global solution methods & GE
◦ Use micro moments (could be from an “identified” regression but not necessary)

2. Micro data and macro model (Aiyagari + HH panel data)

◦ Discipline income process
◦ Effect of a cash transfer
− Along the income distribution
− PE vs. GE
− How does macro matter for micro, and micro matter for macro
5
Methods focus – you will learn how to use Julia
• All coding in Julia Language
◦ Why Julia?
− Easy to learn
− Fast (very fast if used well)
◦ Alternatives (but not for this class)
− Matlab (common in macro, loops are slow, well maintained but expensive)
− Python (many many applications, needs numpy to be as fast as Julia)
− Fortran (fast, harder to learn)
◦ HW: Two common ways to use Julia, make sure you can use either on your
computer
− Jupiter note book (Quant Econ: https://julia.quantecon.org/intro.html)
− VS code (Cameron Pfiffer’s intro “Julia for Economists 2022: Julia Basics”, youtube)6
Problem Sets & Exam
• Weekly problem sets
◦ Mix of theory, coding, and empirical work
◦ Submit a pdf using latex/lyx + code
◦ Fail or pass grade – I want to see effort, that’s it
◦ Discussion among colleagues is okay, copy-pasting is considered cheating
◦ I will post the problemset on Thursday after class, and the submission deadline is
Wednesday night at 23:55
• Final exam
◦ Take home
◦ I will post a mock exam in due time, but similar to problem sets
7
Communication

• If you have questions about the material, just ask

• If you have organizational questions that matter for everyone, ask in class

• I will offer office hours

◦ Office hours great when you come prepared

◦ Office hours terrible when you i) miss class, and ii) do not come prepared

• Explaining complicated concepts via email is not effective

8
How to get most out of this class

• Fundamental understanding beats memorizing

• You have to do the problem set to learn how to code

• Can be a bit painstaking – code that doesn’t work can be frustrating

• Learning vs. your own research

◦ In this class, we write code from scratch

◦ In your own work, I recommend building on other researcher’s code
− Someone with some standing in the profession

9
Neoclassical Model

10
The Neoclassical Model... Again!
• The neoclassical model is an extremely useful benchmark
◦ Normative point of view
− Think flexible vs. sticky prices in business cycle
− Think knowledge spillover and monopoly distortion in long-run growth
◦ Computational point of view
− Neoclassical model is extremely well behaved: unique, smooth, closed form solution
− Most models nest the neoclassical model, which helps you i) understand how your
model works and ii) find bugs in your code

◦ Example: Suppose I take your Aiyagari economy with income risk, and I make the
income risk very small. It better be that the interest rate is close to the discount
rate r ≈ ρ
11
Neoclassical Production

• Competitive goods and factor markets

• Constant returns to scale

• Final good is the numeraire and can be consumed or invested in capital, which
depreciates at rate δ

Y =AK α L1−α (1)

K̇ =Y − C − δK (2)

12
Neoclassical Household

• Representative agent maximizes utility (think of dynasty if you like)

• Still, a price taker!

• Hold assets B which pay safe return r , and earn labor income wL

• Labor supplied inelastically but growing at constant rate gL

Z ∞
Cs
max V = e −(ρ−gL )s log (cs ) ds, cs := (3)
{Cs } 0 Ls
s.t. Ḃ =rB + wL − C (4)

13
Equilibrium
• Production
◦ Marginal product equals factor price
◦ Rent capital from competitive banking sector that mediates between household and
firms with rental rate R
∂Y
=⇒ ∂K = R, ∂Y
∂L = w

• Household
◦ Optimal consumption obeys Euler equation

ċ
=r −ρ (5)
c

• Banking sector: r = R − δ, B = K 14
Steady State

• Given population growth, steady state only relative to some normalization, k := K

L

 1
 1−α
αA
k∗ =
ρ+δ

• c∗ , y∗ , r∗ , w∗ follow immediately

15
Pertubation Method

16
Perturbation Method

• We want to assess whether the dynamics implied by the neoclassical growth model
explain actual growth patterns
◦ Across regions: Barro/Sala-i-Martin (1992)
◦ Across countries: Mankiw/Romer/Weil (1993)

• Clearly, have to solve for these transition dynamics first

• Note how the model guides our thinking here, also in terms of what data is needed

17
Linearized System

• Perturbantion Method ≈ Linearize system of equations around something steady

• Standard application linearizes around deterministic steady state

• Works in discrete and continous time

◦ Slight differences but conceptually the same

• Extremely powerful because because no “curse of dimensionaliy”

◦ How many state variables? Doesn’t matter...

◦ Theorems when the system is well behaved so linearization works
− Local uniqueness and convergence results

18
Linearized System: A Warning

• Approximation could be extremely bad for large shocks

• Personally: MIT shocks and DSGE always felt like a misnomer to me

◦ Nothing stochastic other than that one shock that no one ever saw coming...
◦ Risk doesn’t matter to first order (certainty equivalence)
◦ Frontier research in business cycle nowadays combines linearization methods with
global solution methods
− Auclert et al. (2021)
− Sequence-space Jacobian beyond the scope of the class but frontier in HANK literature

19
Linearized System: We want a linear dynamic system
• In continous time, consider a vector x ∈ R k and matrix A ∈ R k×k

• Consider the following dynamic system

x˙t = Λt xt (6)

◦ Obs 1: The system in (6) is general in the sense that any dynamic problem can be
written this way (up to adding a constant term) because Λ itself possibly non-linear
function of x, i.e., Λ = Λ (x)
◦ Obs 2: We know what Λt is, we want to find xt ∀t (multi-valued function of time)
◦ Obs 3: x0 contains some variables that are predetermined, but some that are not

• Neoclassical model: xt = [kt ct ]0 (if you knew c0 and k0 , you’d be done!)

20
Linearized System: We want a linear dynamic system

• Example neoclassical model

 c 
k˙t = k α−1 − − (δ + gL ) kt
k
c˙t = (rt − ρ) ct

 
α−1 1
(k − (δ + gL )) − k 
• Now let ẋ = [k̇ ċ]0 , and Λt = 
0 rt − ρ
• If you knew c0 and k0 , you’d be done!

• k0 given, c0 the hard part

21
Linearized System: What if Λt = Λ

• Linearization all about going from Λt → Λ

• Let’s do ċ k̇
c
≈ together, make sure you can derive k
≈

 
ρ+δ
(ρ − gL ) (gL + δ − α 
)
Λ=
−(1 − α)(ρ + δ) 0

• Where ẋ = [ d log
dt
k d log c
dt
] and x = [log( kk∗ ) log( cc∗ )]

• Why is this so useful? Because ẋ = Λx has an analytical solution

22
Linearized System: Solution

• Solution comes from Jordan form and matrix multiplication

• Easy to see in scalar case ẏ = κyt =⇒ yt = Ce κt

• More general:
xt = V Diag e λ1 e λ2 ... e λk b

• Where

◦ λ are eigenvalues of Λ
◦ V the stacked eigenvectors
◦ b is a column vector pinned down by initial condition and terminal condition

23
Linearized System: Convergence Dynamics

• Particularly nice to linearize around well-defined steady state

◦ Uniqueness
◦ Convergence

• λ different signs: “saddle path stable”: need neg. value for state variable

• λ all positive: “unstable”

• λ all negative: “stable”

• Ossiciliation possible for complex roots

• Discrete time: yt+1 − yt = ρyt , here eigenvalues about |λ| < 1

24
Alternative: Finite-difference with Shooting Algorithm
• Recall
 c 
k˙t = k α−1 − − (δ + gL ) kt
k
c˙t = (rt − ρ) ct

• Approximate slope at discrete points in time k̇t+l ≈ kt+l+∆∆−kt+l

 
α−1 ct
kt+∆ = ∆ ∗ kt − − (δ + gL ) kt + kt
kt
ct+∆ = ∆ ∗ (rt − ρ) ct + ct

• Find c0 s.t. cT = csteady state for large T

25
Alternative: Finite-difference with Shooting Algorithm

• Global solution method – no small shock assumption

• Tractable in simple model with one jump variable

• A lot harder with more control variables

26
Application: Mankiw,Romer, Weil

• First time we make touch with the data

• Peadagogical: think of how one can try to go from model to data

• Big Question: Why are some countries richer than others

• Classic answer by Mankiw et al. (1993): factor accumulation

◦ A bit outdated, but 50% of researchers find this paper essential!

◦ Note: Lucas (1990), Caselli (2005) on TFP as key driver of cross-country inequality
◦ Measurement challenges; role of technology adoption Parente and Prescott (1994),
Trouvain (2024)

27
Mankiw, Romer, Weil

• Neoclassical model with two types of capital, “physical” and “human”

◦ Y = K α H β L1−α−β
◦ K̇ = sK Y − δK , Ḣ = sH Y − δH

• Use steady state to derive estimating equation

α β
log Y /L = stufft + log sK + log sH
1−α−β 1−α−β

• Theory tells you what data you need: Y /L, sK , sH , gL , δ, ...

28
What about Growth Patterns and Speed of Convergence
• Maybe we are not in steady state – what sort of growth patterns do we expect?
• Use first-order log-linear approximation to show

yt = y0 e λt + yss 1 − e λt
 
(7)

• Homework: prove that λ = − (1 − α − β) (gF + gL + δ)

◦ Kind of suprising because you have two moving pieces but only one root...
◦ Compare to version with only physical capital (β = 0) w.r.t. convergence speed
◦ Take this to the data
yt − y0
≈ α + λy0
t
29
Mankiw, Romer, Weil: Pros and Cons for your own research
• Pros
◦ Structure to interpret the data
◦ Nice benchmark: neoclassical model with human capital
◦ At the time: new data, pushing out the frontier
◦ Warning: without structure, very hard to learn anything from the data in macro
• Cons
◦ Nowadays, model “low-tech”, too many exogenous pieces
− Tip: start out that way, and keep endogenizing more things
◦ Today, people would expect “cleaner” variation (Nakamura-Steinsson later)
− Opinion: Too bad... very little good work on cross country growth
− Opinion: I am very skeptical long-run growth is well approximated by spatial growth
30
Extra slides

• MRW & Growth Accounting Details

• Work in progress

31
MRW & Growth Accounting

• MRW and Lucas’ “why doesn’t capital flow to poor countries” paper give rise to
growth accounting literature

• Obvious endogeneity concerns in MRW

• Try something slightly less ambitious and ask:

“How much cross-country inequality can we account for with factors alone”
◦ Meaningful even if no causal interpretation? Yes
− Forces that inhibit factor accumulation (easier)
− Models where factor accumulation undistorted but somehow residual TFP low

Back

32
MRW & Growth Accounting

• Read beautiful handbook chapter by Caselli (05)

• Suppose output takes capital and labor efficiency units hL and residual TFP A

◦ h usually reflects quality adjustment for human capital, like years of schooling
Y
= Ak α h1−α
L
• One version: how much of the variance in log YL = y due to each piece?

Back

33
MRW & Growth Accounting
• Answer not as easy as it seems
◦ Measurement of K tricky, but h way worse due to quality differences in schooling
1
◦ A question of how to do this exactly because k, h ∝ A 1−α in benchmark model
◦ Some people (myself included) prefer
  α
k 1−α 1
y =Z· · h, Z := A 1−α
y

◦ No unique variance decomposition because no way k, h, A orthogonal

◦ h usually reflects quality adjustment for human capital, say years of schooling
• Bottom line: 50% or more because of TFP
• Suprising fact: agricultural TFP differences most important!
Back
34
References

35
Achdou, Y., J. Han, J.-M. Lasry, P.-L. Lions, and B. Moll (2022). Income and wealth
distribution in macroeconomics: A continuous-time approach. The review of economic
studies 89 (1), 45–86.

Aiyagari, S. R. (1994). Uninsured idiosyncratic risk and aggregate saving. The Quarterly
Journal of Economics 109 (3), 659–684.

Auclert, A., B. Bardóczy, M. Rognlie, and L. Straub (2021). Using the sequence-space
jacobian to solve and estimate heterogeneous-agent models. Econometrica 89 (5),
2375–2408.

Barro, R. J. and X. Sala-i Martin (1992). Convergence. Journal of political Economy 100 (2),
223–251.

35
Caselli, F. (2005). Accounting for cross-country income differences. Handbook of economic
growth 1, 679–741.

Lucas, R. E. (1990). Why doesn’t capital flow from rich to poor countries? The American
Economic Review 80 (2), 92–96.

Mankiw, N. G., D. Romer, and D. N. Weil (1993). A contribution to the empirics of economic
growth. The Quarterly Journal of Economics.

Nakamura, E. and J. Steinsson (2014). Fiscal stimulus in a monetary union: Evidence from us
regions. American Economic Review 104 (3), 753–792.

Nakamura, E. and J. Steinsson (2018). Identification in macroeconomics. Journal of

Economic Perspectives 32 (3), 59–86.

35
Parente, S. L. and E. C. Prescott (1994). Barriers to technology adoption and development.
Journal of political Economy 102 (2), 298–321.

Plagborg-Møller, M. and C. K. Wolf (2021). Local projections and vars estimate the same
impulse responses. Econometrica 89 (2), 955–980.

Sims, C. A. (1980). Macroeconomics and reality. Econometrica: journal of the Econometric

Society , 1–48.

Trouvain, F. (2024). Technology adoption, innovation, and inequality in a global world.

35