A Practical Guide to Shift-Share Instruments (and What I Learned Replicating the China Shock)

A Practical Guide to Shift-Share Instruments (and What I Learned Replicating the China Shock)

Imagine two U.S. towns in the early 1990s. One town is heavily dependent on furniture and textiles. The other is dominated by healthcare and tech services. Over the next two decades, Chinese imports surge in furniture and textiles, but not in healthcare and tech. Unsurprisingly, the manufacturing town is hit much harder than the service town.

We want to know: What is the causal effect of increased import competition on local labor market outcomes? But regions that specialize in vulnerable industries might already differ from other regions in ways we cannot fully observe. Simply comparing their employment changes to everyone else will not give us a clean causal effect. This is the kind of setting where shift-share instruments shine. 

Shift-share designs offer a structured way to build an instrument that captures how strongly each region is exposed to a common set of shocks. In this post, I’ll explain how this works, why it’s trickier than it first appears, and what I learned when I used the Borusyak–Hull–Jaravel (BHJ) framework to replicate the classic Autor, Dorn, and Hanson (2013) “China shock” paper.

What Problem Are Shift-Share Instruments Solving?

A typical empirical question might look like this:

How does exposure to a particular economic shock affect outcomes across regions?

The outcome could be changes in employment, wages, voting, migration, or something else. The main regressor of interest is often a measure of exposure to the shock—for example, how much import competition a region faces.

The problem is that this exposure is rarely exogenous. Regions that are more exposed to the shock might have different initial industry structures, political histories, or demographic trends. Those unobserved differences can be correlated with both exposure and outcomes, which biases a simple OLS regression.

Instrumental variables (IV) are one way to address this. IVs provide a way to isolate variation in the treatment that is as good as random, allowing us to estimate causal effects even when the main explanatory variable is endogenous.

 Shift-share IVs build an instrument by combining:

• information on how initially exposed each unit is (the shares), and

• information on how big the shock was in each category (the shifts).

If constructed and justified carefully, this instrument can mimic random variation in exposure across regions.

What Is a Shift-Share Instrument, in Plain Language?

The core idea of a shift-share instrument is very simple:

1. Start with baseline shares that describe the composition of each unit.

For regions, these might be industry employment shares or historical immigrant shares by origin country.

2. Combine those shares with category-level shocks that occur later.

For industries, this could be national import growth by industry. For immigration, it could be national inflows by country of origin.

Mathematically, for each region i, we construct:

Shift Share Instrument:

z_i = ∑_k s_ik * g_k

Where

s_ik = share of region i in category k (e.g., industry k employment share),

g_k = shock to category k (e.g., Chinese import growth in industry k).

In the Autor, Dorn, and Hanson (2013) paper:

• Units i are U.S. commuting zones.

• Categories k are manufacturing industries.

• Shares s_ik are lagged industry employment shares in each commuting zone.

• Shifts g_k are growth in Chinese imports to other high-income countries by industry.

Intuitively, this gives us a measure of how much each commuting zone “should” be affected by the China shock if the only thing that mattered were its initial industry structure.

Where Things Get Tricky: Incomplete Shares, Dominant Categories, and Anticipation

On paper, the shift-share IV formula is deceptively simple. In practice, several issues can quietly undermine identification.

One subtle but important issue is incomplete shares. In many applications, the shares don’t cover the entire economy. For example, in the China-shock setting, the shares often only cover manufacturing industries. Non-manufacturing employment is implicitly treated as a “missing category” with zero shock.

Borusyak, Hull, and Jaravel show that when shares do not sum to one, variation in the total share—such as the overall manufacturing share in a region—can contaminate the instrument unless you explicitly control for it. In other words, part of what you think of as “shock-driven variation” might actually be simple cross-sectional differences in how manufacturing-intensive each region is.

Another issue arises when one or two categories dominate the instrument. If almost all of the action comes from a single industry or origin country, the “many shocks” logic breaks down. In immigration settings, for example, a huge Mexican share can dominate the instrument, making the design behave like a single-shock experiment rather than a many-shock one. That raises questions about whether the law of large numbers really helps here.

A third complication is anticipation. Shares are supposed to reflect pre-shock exposure. If the baseline period is chosen too late—after people have started re-sorting or adjusting in anticipation of the shock—then the shares themselves are no longer exogenous.

Exogenous Shifts Approach Checklists: Replicating Autor, Dorn, and Hanson (2013)

Step 1: Idealized Experimental Design

In the ADH setting, the “ideal” experiment would randomly assign different levels of import growth across industries, so that exposure across commuting zones would be as-if random.

Step 2: Bridge the gap between the observed and ideal shifts

Apparently, real import growth is not randomized. Our goal is to approximate this ideal with controls and diagnostics. Therefore, we need to:

1. specify some control variables, and

2. describe how observed shifts proxy for the ideal ones.

You want to include controls for confounders at both the shock level (e.g., industry skill intensity) and the unit level (e.g., regional characteristics).

Step 3: Control for Incomplete Shares

Now return to the question: do the shares sum to one?

If they do not—for example, if you only have manufacturing industries—you should construct the total exposure share:

S_i = ∑_k s_ik

This step is crucial for separating variation driven by shocks from variation driven by simple differences in how specialized each unit is.

Borusyak et al. (2022) recommend using a more lagged S_ℓ,t−1 to better capture long-term manufacturing trends and eliminate bias. “In the labour supply example, quasi-experimental variation in manufacturing shocks is isolated provided one controls for a region’s lagged manufacturing share S_ℓ.

See below my replication. Column 7 controls for S_ℓ,t−1. 

Step 4: Lag shares to the beginning of the natural experiment

It is recommended to lag shares. However, lagging shares beyond what is necessary would typically make the instrument weaker.

Step 5: Report descriptive statistics for shifts in addition to observations. (My replication is based on the methods provided in Borusyak et al. (2022). I get very similar but not exactly the same results.)

For example, in my replication, a few industries experienced disproportionately large import shocks, which underscores the importance of inspecting the distribution of g_k before relying on the “many shocks” assumption.

Step 6: Implement balance tests for shifts in addition to the instrument

We need to implement the balance tests for the instrument at the level of units, and also directly for g_k at the level of shifts. Using covariates from Acemoglu et al. (2016), I tested whether industry-level shocks were correlated with pre-trends or baseline industry characteristics that might threaten shift exogeneity. The balance tests below show whether the shocks behave as if they were as good as random conditional on controls.

Step 7: Estimate the Model With Correct Standard Errors and Sensitivity Checks

When estimating the main IV results, it’s important to use shift-consistent standard errors, such as the Adão, Kolesár, and Morales (2019) variance estimator or the shift-level 2SLS approach proposed by Borusyak, Hull, and Jaravel (2022). After obtaining the baseline estimates, check robustness by trying alternative control sets and comparing results with and without unit-level weights (e.g., population weights in regional analyses). If the estimates remain stable across these variations, it provides reassurance that the design is capturing true quasi-experimental variation rather than modeling artifacts.

A Note on the Exogenous Shares Approach

If you take the exogenous shares perspective, the key task is to ensure that the baseline shares you use are truly predetermined and not simply capturing other confounding shocks. This involves choosing appropriate unit-level controls, checking which shares drive most of the identifying variation (often using Rotemberg weights), and running balance or pre-trend tests on both the shares and the instrument itself. It’s also good practice to try different combinations of share instruments to confirm that the results are not overly dependent on one particular share or category.

Conclusion

Shift-share instruments sit at an interesting intersection between theory and practice. On one hand, their basic structure is intuitive: exposure times shocks. On the other hand, getting the assumptions right requires careful thinking about incomplete shares, the level at which randomness enters the design, and how inference should be carried out.

References

1. Autor, D., Dorn, D., & Hanson, G. (2013). The China Syndrome: Local Labor Market Effects of Import Competition in the United States.
2. Borusyak, K., Hull, P., & Jaravel, X. (2022). Quasi-Experimental Shift-Share Research Designs.
3. Borusyak, K., Hull, P., & Jaravel, X. (2025). A Practical Guide to Shift-Share Instruments.
   
4. Adão, R., Kolesár, M., & Morales, E. (2019). Shift-Share Designs: Theory and Inference.