Recent Advances in Shift-Share IV

Peter Hull
U Chicago and NBER

November 2020
Introduction
Many canonical instrumental variables (IVs) leverage the quasi-
random assignment of some z` across observations `
Angrist (1990): randomly assigned draft lottery number z` as an
instrument for individual `’s service in the Vietnam War

But some z` are more complicated, combining variation across both

observations and some other common dimension n
Bartik (1991): predicted employment growth z` = ∑n s`n gn as an IV for
region `’s employment growth, where gn is the national growth of
industry n and the s`n ∈ [0, 1] are lagged employment shares
Similar z` : Blanchard & Katz (1992), Card (2009), Autor et al. (2013)

A recent methodological literature studies when/how such

“shift-share” IVs (SSIVs) can be used for causal inference
Formalizes two paths to identification: via “shocks” gn or “shares” s`n
Raises new practical considerations for SSIV estimation and inference

1
Autor, Dorn, & Hanson (ADH; 2013): “The China Shock”
ADH study the effects of rising Chinese import competition on US
local labor markets, 1990-2007
Share of US spending on Chinese goods: 0.6%→4.6%
Share of working-age pop employed in manufacturing: 12.6%→8.4%
Reverse causality concern: weak markets more likely to import
To address endogeneity challenge, they use a SSIV z` = ∑n s`n gn
gn : industry n’s growth of Chinese imports in eight non-U.S. economies
s`n : lagged share of mfg. industry n in total employment of location `
Treatment x` : local growth of Chinese imports ($1,000/worker)
Main outcome y` : local change in manufacturing employment share
ADH derive this instrument from a simple trade model:
“Our IV strategy will identify the Chinese productivity and trade-shock
component of United States import growth if the common within-
industry component of rising Chinese imports to the United States and
other high-income countries stems from China’s rising comparative
advantage and (or) falling trade costs in these sectors.” (p. 2129)
2
ADH First Stage and Reduced Form

Main IV estimate (state-clustered SE), RF/FS: -0.596 (0.099)

If causal (and setting aside general equilibrium effects, etc.), would
explain 33% of the fall in manufacturing employment
How might one assess causality with this IV?
Hard to think of changes in predicted import exposure like a lottery #...
3
Card (2009) Immigration “Enclave” IV

Card studies the effect of local immigration on local wages

Outcome y`j : log wage gap between immigrant and native men in skill
group j and region `
Treatment x`j : log ratio of immigrant to native hours in (`, j)
Seek to estimate immigrant-native inverse elasticity of substitution

He constructs a SSIV z`j by combining lagged shares s`n of immigrants

from countries n in region ` & national immigration rates gjn
Intended to address endogeneity from local labor demand shocks
“To the extent that initial immigrant shares are correlated with other
unobserved features that affect relative wage differentials in a city, an
enclave-based identification strategy may be less attractive...” (p. 15)
Again, hard to think of predicted immigration inflows like a lottery #

4
Card Reduced Form

5
Outline

1. The SSIV Setting

2. Identification from Shares: Goldsmith-Pinkham et al. (2020)

3. Identification from Shocks: Borusyak et al. (2020)

4. New Inference Challenges: Adão et al. (2019)

5. Further Settings: Borusyak and Hull (2020)

6
The SSIV Setting
Suppose we are interested in estimating some parameter β of a linear
causal or structural model y` = β x` + e`
Straightforward to generalize to heterogeneous treatment effects

Residualize e` on a vector of observed controls w` to get second stage:

y` = β x` + w`0 γ + ε` ,

where w` and ε` are orthogonal by construction: E [∑` w` ε` ] = 0

We instrument x` with z` = ∑n s`n gn , where ∑n s`n = 1 (for now)
Call s`n the “exposure shares” and gn the “shocks”
Share vary across observations, shocks do not
1 
IV is valid if E L ∑` z` ε` = 0; identification follows from a first stage
1 
Note no iid assumption, E L ∑ ` z` ε ` =6 E [z` ε` ]; will be important later

7
The SSIV Estimator
SSIV divides the regression of y` on z` , controlling for w` , (“reduced
form”) by the regression of x` on z` , controlling for w` (“first stage”)

By the Frisch-Waugh-Lovell theorem, this estimator can be written

∑ z` y ⊥ ∑ ∑ s`n gn y`⊥
βˆ = ` `⊥ = ` n ,
∑` z` x` ∑` ∑n s`n gn x`⊥

where v`⊥ denotes sample residuals from regressing v` on w`

Plugging in the model y` = β x` + w`0 γ + ε` gives

∑ ∑ s`n gn ε`⊥
βˆ = β + ` n ,
∑` ∑n s`n gn x`⊥

p p p
Consistency: βˆ −
→ β if 1
L ∑` ∑n s`n gn ε`⊥ −
→ 0, 1
L ∑` ∑n s`n gn x`⊥ −
→ π 6= 0
Asymptotic inference: find a σL such that (βˆ − β )/σL ⇒ N(0, 1)
8
Outline

1. The SSIV Setting

2. Identification from Shares: Goldsmith-Pinkham et al. (2020)

3. Identification from Shocks: Borusyak et al. (2020)

4. New Inference Challenges: Adão et al. (2019)

5. Further Settings: Borusyak and Hull (2020)

9
Goldsmith-Pinkham, Sorkin, and Swift (GPSS; 2020)
GPSS are interested in understanding when/how SSIV can be seen as
leveraging quasi-experimental variation across observations
Viewing the gn as fixed, z` = ∑n s`n gn is a linear combination of shares
It follows that z` is a valid instrument when the shares are exogenous

Formally, GPSS establish a numerical equivalence:

βˆ can be obtained from an overidentified IV procedure that uses N
share instruments s`n and a weight matrix based on the shocks gn

Sufficient condition for identification: quasi-experimental shock

exposure across observations
 
1 1
E [ε` | s`n ] = 0, ∀n =⇒ E ∑ `
z` ε` = ∑ ∑ gn E [s`n E [ε` | s`n ]] = 0
L L ` n
Diff-in-diff logic: when ε` are unobserved outcome trends (as in ADH)
E [ε` | s`n ] = 0 is akin to a “parallel trends” assumption
Consistency/inference follow from standard conditions (e.g. iid data)
10
Rotemberg Weights
The GPSS view of SSIV is one of many share instruments
In Card, N = 38. In ADH, N = 397 (!)

They propose “opening the black box” of overidentified IV by deriving

the weights SSIV implicitly puts on each share instrument
Builds on Rotemberg (1983), so they call these “Rotemberg weights”

∑ s`n y`⊥ gn ∑` s`n x`⊥

βˆ = ∑ α̂n βˆn , where βˆn = ` and α̂n =
n ∑` s`n x`⊥ ∑n0 gn0 ∑` s`n0 x`⊥
| {z } | {z }
n-specific IV estimate Rotemberg weight

Intuitively, more weight is given to share instruments with higher

shocks gn and larger first stages ∑` s`n x`⊥
Weights can be negative (potential problem with heterogeneous effects)
Under constant effects, show a formal link to Conley et al. (2012)’s
and Andrews et al. (2017)’s measures of sensitivity-to-misspecification
11
Rotemberg Weights in ADH (via GPSS)

Negative weights, large heterogeneity in individual βˆn estimates

12
Rotemberg Weights in Card (via GPSS)

No negative weights, low heterogeneity in individual βˆn estimates

13
Is Share Exogeneity a Plausible Identifying Assumption?
Several ways to probe the plausibility of exogenous s`n ex post:
Balance/pre-trend tests, overidentification tests (w/constant effects)
Straightforward to implement; no different than any other IV
GPSS find these tests broadly pass for Card, but fail badly for ADH

In some settings, share exogeneity is ex ante implausible

Suppose unobserved shocks νn affect ε` via the same/correlated shares
E.g. in ADH, unobserved technology shocks across industries n can
affect labor markets via employment shares, along with observed gn
Then share exogeneity cannot hold: the shares drive outcomes through
both observed and unobserved channels
Formally, if ε` = ∑n s`n νn + ε̃` , then s`n and ε` cannot be uncorrelated
in large samples even if they are randomly assigned to observations

Likely share endogeneity calls for a new approach to SSIV...

14
Outline

1. The SSIV Setting

2. Identification from Shares: Goldsmith-Pinkham et al. (2020)

3. Identification from Shocks: Borusyak et al. (2020)

4. New Inference Challenges: Adão et al. (2019)

5. Further Settings: Borusyak and Hull (2020)

15
Borusyak, Hull, and Jaravel (BHJ; 2020)
BHJ are interested in understanding when/how SSIV can be seen as
leveraging quasi-random variation in the shocks
Like GPSS, they establish a numerical equivalence:
βˆ can be obtained from a just-identified shock-level IV procedure that
uses gn to instrument for a shock-level “aggregate” of the treatment:

∑ z` y ⊥ ∑ ∑ s`n gn y`⊥ ∑n gn ∑` s`n y`⊥ ∑n sn gn ȳn⊥

βˆ = ` `⊥ = ` n = = ,
∑` z` x` ∑` ∑n s`n gn x`⊥ ∑n gn ∑` s`n x`⊥ ∑n sn gn x̄n⊥

where sn = L1 ∑` s`n are weights capturing the average importance of

shock n, and v̄n = ∑∑` s`n
s
v`
is an exposure-weighted average of v`
` `n

It follows that βˆ is consistent iff this shock-level IV procedure is

They then derive new conditions for SSIV identification + consistency

Want to view gn as random shocks, so can’t assume z` = ∑n s`n gn is iid

16
BHJ Baseline Assumptions
A1 (Quasi-random shock assignment): E [gn | ε̄, s] = µ, ∀n
Each shock has the same expected value, conditional on the shock-level
unobservables ε̄n and average exposure sn
Implies SSIV validity: E L1 ∑` z` ε` = E [∑n sn gn ε̄n ] = 0
 

A2 (Many uncorrelated shocks): E [∑n sn2 ] → 0 and ∀(n, n0 ) with

n0 6= n, Cov (gn , gn0 | ε̄, s) = 0
First part: expected Herfindahl index of average shock exposure
converges to zero (implies N → ∞)
Second part: shocks are mutually uncorrelated given the unobservables
1 p
Imply a shock-level law of large numbers: L ∑` z` ε` = ∑n sn gn ε̄n −
→0

Both assumptions, while novel for SSIV, would be standard for a

shock-level IV regression with weights sn and instrument gn
p
Identification of β follows given a first stage: 1
L ∑` z` x`⊥ −
→ π 6= 0
Sufficient condition: most observations are mostly exposed to a small
number of shocks affecting treatment
17
BHJ Extensions
Conditional Quasi-Random Assignment: E [gn | ε̄, q, s] = qn0 µ for
some observed shock-level variables qn
Consistency follows when w` = ∑n s`n qn is controlled for in the IV

Weakly Mutually Correlated Shocks: gn | (ε̄, q, s) are clustered or

otherwise mutually dependent
Consistency follows when mutual correlation is not too strong

Panel Data: Have (y`t , x`t , s`nt , gnt ) across ` = 1, . . . , L, t = 1, . . . , T

Consistency can follow from either N → ∞ or T → ∞
Unit fixed effects “de-mean” the shocks, if s`nt are time-invariant
Also see Jaeger et al. (2019) for dynamic biases in panel SSIVs

Estimated Shocks: gn = ∑` w`n g`n proxies for an infeasible gn∗

Consistency may require a “leave-out” adjustment: z` = ∑` w`n g̃`n for
g̃`n = ∑`0 6=` w`0 n g`0 n (akin to JIVE solution to many-IV bias)

Multiple shocks: Propose new overidentified SSIV procedures

18
The “Incomplete Shares” Issue

So far, we have assumed a constant sum-of-shares: S` ≡ ∑n s`n = 1,

but in some settings S` varies across observations
E.g. in ADH, S` is region `’s share of non-manufacturing employment
since s`n is the share of manufacturing industry n in total employment

BHJ show that A1/A2 are not enough for identification in this case
The IV implicitly uses variation across S` , which may be endogenous

Controlling for the sum-of-shares S` isolates clean shock variation

Can be seen as a special case of conditional quasi-random assignment:
“dummying out” the non-manufacturing sector, in ADH
Further controls needed when A1 holds conditional on qn ; e.g. isolating
within-period variation in panels requires interacting S` with period FE

19
A Taxonomy of SSIV Settings
BHJ distinguish between three cases of SSIVs in the literature
Case 1: the IV is based on a set of shocks which can itself be thought
of as an instrument (i.e. many, plausibly quasi-randomly assigned)
E.g. Acemoglu et al. (2016) use the ADH shocks to conduct an
industry-level IV analysis
BHJ shows how this identifying variation can be mapped to estimate
effects at a different “level” (i.e. industries → local labor markets)

Case 2: the researcher does not directly observe many quasi-random

shocks, but can estimate them in-sample
Canonical setting of Bartik (1991), where gn are average industry
growth rates (thought to proxy for latent demand shocks)
See also Card (2009), where national immiration rates are estimated

Case 3: the gn cannot be naturally viewed as an instrument

Either too few (small N) or implausibly exogenous, even given some qn
Identification may (or may not) instead follow from share exogeneity
20
Ex Ante vs. Ex Post Validity

BHJ emphasize that the decision to pursue a “shocks” vs. “shares”

identification strategy should be made ex ante
Undesirable to base identifying assumptions on ex post tests, though
balance/pre-trend tests can be used to falsify assumptions
The two identification strategies may have different economic content

They suggest thinking about whether shares are “tailored” to the

economic question / treatment, or are “generic”
Generic shares (e.g. ADH): unobserved νn are likely to enter ε` via the
same or similar shares, violating share exogeneity
Tailored shares (e.g. Mohnen 2019) have a DD feel; don’t even need
the shocks, except to possibly improve power / avoid many-IV bias

21
ADH Revisited
BHJ show how ADH can be seen as leveraging quasi-random shocks
Ex ante plausible (unlike exogenous shares): imagine random industry
productivity shocks in China affecting imports in U.S. & elsewhere
Many shocks (industries), plausibly weakly mutually correlated

Evaluate A1 by regional and industry-level balance tests

Industry shocks are uncorrelated with five observables considered by
Acemoglu et al. (2016) (e.g. lagged capital to value-added ratios)

Evaluate A2 by studying variation across industries

Effective sample size (1/HHI of sn weights): 58-192
Shocks appear mutually uncorrelated across sectors (SIC3)

Check sensitivity to adjusting for potential industry-level confounders

Control for w` = ∑n s`n qn , where qn include the Acemoglu et al. (2016)
observables, sector FE, industry pre-trends ...

22
BHJ do ADH

Robust coefficient of ≈ −0.3, after accounting for incomplete shares

23
Outline

1. The SSIV Setting

2. Identification from Shares: Goldsmith-Pinkham et al. (2020)

3. Identification from Shocks: Borusyak et al. (2020)

4. New Inference Challenges: Adão et al. (2019)

5. Further Settings: Borusyak and Hull (2020)

24
Adão, Kolesar, and Morales (AKM, 2019)
AKM study a novel inference challenge when SSIV identification
leverages quasi-random shocks
Observations with similar shares s`1, , . . . , s`N are likely to have correlated
z` , even when not “clustered” in conventional ways (e.g. by distance)
When ε` is similarly clustered (e.g. ε` = ∑n s`n νn + ε̃` ), the large-sample
distribution of βˆ may not be well-approximated by standard CLTs

They show by simulation that this can lead to large size distortions
Tests with nominal 5% rejection rates can reject true nulls in 55% of
placebo shock realizations (ADH-based Monte Carlo)
Reminiscent of Bertrand et al. (2004) study of robust SEs in diff-in-diff

They then derive a new CLT + SEs to address “exposure clustering”

“Design-based:” leverage iidness of shocks, not observations, building
on BHJ identification framework
Also develop null-imposed (AKM0) CIs, which help in finite samples
25
ADH Monte Carlos (Robust/Clustered)

26
ADH Monte Carlos (AKM/AKM0)

27
Exposure-Robust SEs
BHJ show a convenient solution to exposure clustering, via their
equivalent shock-level IV regression
Usual robust/clustered SEs can be valid when βˆ is given by estimating

ȳn⊥ = α + β x̄n⊥ + qn0 τ + ε̄n⊥ ,

instrumenting x̄n⊥ by gn and weighting by sn

Numerically identical IV estimate, when controls include ∑n s`n qn
Null-imposed CIs similarly straightforward at the shock level

Clustering logic: can get valid SEs by estimating the IV at the level of
identifying variation (here, shocks)
Same logic applies to performing valid balance/pre-trend tests and
evaluating first-stage strength of the instrument
New Stata package ssaggregate helps translate data to the shock level,
after which researchers can proceed with familiar estimation commands
28
Install with ssc install ssaggregate; please send us comments to improve!
29
Outline

1. The SSIV Setting

2. Identification from Shares: Goldsmith-Pinkham et al. (2020)

3. Identification from Shocks: Borusyak et al. (2020)

4. New Inference Challenges: Adão et al. (2019)

5. Further Settings: Borusyak and Hull (2020)

30
Borusyak and Hull (BH; 2020)
Many instruments may be seen as being SSIV-like, combining a set of
exogenous shocks and measures of non-random shock exposure
Nonlinear SSIVs: z` = f (g1 , . . . , gN , s`1 , . . . , s`N ) for nonlinear f (·)
Network treatments/instruments: z` combines shocks to other
nodes with observed network linkages
Transportation instruments: z` combines transportation
infrastructure upgrades with geography and nearby market sizes
Simulated eligibility instruments: z` combines variation in state
policies with individual demographics / income / etc.
BH develop a general framework for such settings, building on BHJ
Identification generally requires an adjustment for non-random
exposure, akin to the adjustment for “incomplete shares” in linear SSIV
Inference leverages “design” of the shocks to account for non-random
“exposure clustering” (randomization inference)

BH illustrate this framework by addressing bias in “market access”

regressions & boosting power in a simulated instrument setting
31
Summary

We’ve learned a lot about shift-share IV over the past few years
Identification can come from exogenous shock exposure (akin to a DD)
But as-good-as-random shocks may be a more plausible identifying
assumption; then consistency/inference is non-standard
Many new tools to solve practical issues in either case

General advice for researchers hoping to use a SSIV:

Decide in advance whether exogenous shares or exogenous shocks is a
plausible assumption (BHJ taxonomy may be a helpful guide)
Apply appropriate tests to probe your a priori claims (i.e. GPSS / BHJ)
If exogenous shocks, address exposure clustering and be careful with
the “incomplete shares” issue, especially in panels

32