# Tariffs and Growth: Heterogeneous Eﬀects by Economic Structure

*Mateo Hoyos*
Link

## Abstract

This article presents evidence that the impact of tariffs on GDP per capita is mediated by economic structure. I use a panel of 161 countries from 1960 to 2019 to study the impact of changes in average tariff rates on GDP per capita. Using a local projections difference-in-differences (LP-DiD) approach allows me to flexibly control for the surge in GDP that precedes tariff reductions, which if ignored could bias the estimates, and to estimate medium-term dynamic effects. The results, consistent with a specific strand of the trade theory literature, establish that the tariff-growth nexus is contingent on economic structure: tariff reductions led to lower GDP per capita for nonmanufacturer countries, but higher GDP per capita for manufacturers. Additionally, the effects are persistent even twenty years after tariff reductions. The validity of the baseline estimates is confirmed by several robustness checks, especially the control for relevant confounders in the tariffs-growth nexus and a clean controls analysis aimed to address biases from heterogeneity as highlighted by recent differencein-differences literature. The results seem to be driven by heterogeneous effects in productivity and capital accumulation, in turn related to changes in the manufacturing share of GDP.

## Results

Main finding of this paper is that the effects of globalization (proxied for by local tariffs) aren’t not uniformly positive for developing countries. The benefits depend importantly on the manufacturing export intensity of the exporter. Countries with high manufacturing export share see strong growth in GDP per capita post lowering tariffs. However, countries with low manufacturing intensity see declines in GDP per capita after dropping tariffs.

Good example of leveraging insights and methods from @dubeLocalProjectionsApproach. Perhaps useful methodology for my The Universal Law of SaaS Growth efforts, since the setup is quite similar (panel data, per unit impulses)

Makes the important point that local projections can suffer from pre-trends issues reminiscent of difference-in-differences if the outcome variable exhibits a trend relative to the timing of the impulse variable. It’s a form of selection bias. For example, countries that reduce tariffs in general were seeing rising GDP per capita ahead of tariff reduction, implying that the positive trend afterward may have already been baked in the cake. This kind of chart can easily be generated by simply running the same LP estimator on negative horizons (in levels, lagged outcome variable minus level at t-1). This can be avoided by including lags of the outcome variable:

countries reducing their tariffs are on different pretrends from those not changing them. In particular, the former countries display a relative surge in GDP before tariff reductions as compared to the latter. In other words, tariff changes are endogenous to the evolution of GDP, such that countries that decide to decrease tariffs do so after GDP has been on a relative increase. Failure to control for this surge constitutes a clear violation of the parallel trends assumption and may lead to biases in the treatment effect estimates.

They also test for non-linearity in the interaction with manufacturing by using six quantiles of manufacturing intensity.

## Interesting tidbits

Notably, they only include time fixed effects in their panel local projections because the data is already differenced. This is a good point that I hadn’t previously considered, but it’s aligned with the standard time series econometrics pedagogy:

I include only time fixed effects, as the equation is already in differences

I do worry about applying this too broadly – in my areas of research growth rates of companies tend to be highly variable, much more so than the growth rates of countries. Seems like even after differencing if you know that there’s large differences in growth rates across units then unit fixed effects might still be a good idea.

Simple way to eyeball how many lags you need in your local projections to avoid pre-trend: graph the negative horizon impulse response with various lag lengths. Keep adding lags until pre-trends are no more:

Good example of running local projections with interactions to explore heterogeneity: $y_{c,t+h}−y_{c,t−1}=β_{h}ΔTA_{c,t}+θ_{h}int_{c,t}+ϕ_{h}m_{c,t}+∑_{j=1}σ_{h}g_{c,t−j}+α_{t}+ϵ_{c,t}$

The initial share of manufacturing exports, mc,t, is calculated as the average of this variable in the five years before tariff reductions, to avoid contemporaneous endogeneity that may run from GDP to manufacturing exports.

With this specification, the impact of tariff changes on growth varies with the initial level of the manufacturing share of exports. For example, if I want to calculate the cumulative change in GDP per capita at time t + h in relation to a one standard-deviation tariff reduction for a country with an initial manufacturing share of exports of 29 percent, I estimate it by calculating $(−1)∗SD(ΔTA)∗(β_{h}+29∗θ_{h})$.

Interesting definition of a clean control – country which hasn’t experienced a >1 standard deviation tariff change in the last 10 years, with the term determined by observing that local projection effects seem to stabilize after 10 years.