Spillovers Travel Down, Not Across

The knowledge-transfer argument for foreign investment usually gets made horizontally. A multinational arrives, local competitors observe its methods, copy what they can, hire away its trained staff, and the whole industry improves. It is a plausible story and it is the one most investment strategies implicitly tell. It is also the version the evidence has consistently failed to find.

01Lithuania, Firm by Firm

The study that reoriented this literature used firm-level data from an annual enterprise survey covering firms accounting for about 85 percent of output in each sector. Rather than looking within industries, it looked across them, testing whether domestic firms supplying inputs to sectors with a large foreign presence performed differently from those that were not.

They did. A one-standard-deviation increase in foreign presence in the sourcing sectors, which the study quantifies as an increase of 4 percentage points in its backward-linkage variable, was associated with a 15 percent rise in output of each domestic firm in the supplying industry. The author describes the productivity of Lithuanian firms as positively correlated with the extent of potential contacts with multinational customers.

The same analysis found no such effect in the same industry as the foreign entrant. The coefficient on the horizontal variable was not statistically significant, consistent, as the study notes, with an existing literature that had repeatedly failed to find positive intraindustry spillovers. Nor was there evidence of gains flowing downward from multinational suppliers of intermediate inputs.

The mechanism is unglamorous and entirely plausible. A multinational buying locally has a direct commercial interest in its suppliers becoming better: it specifies quality standards, sometimes provides technical assistance, and pays for reliability. A multinational competing locally has the opposite interest. Knowledge moves along a commercial relationship, and the supply chain is where the relationship is.

02Three Thousand Six Hundred Estimates

A single country study invites the obvious objection. That objection was answered in 2011 by a meta-analysis that collected 3,626 estimates of spillovers from 57 studies, with a median of 45 estimates per study, coding 55 variables of study design for each. More than 100 researchers had by then examined productivity spillovers from foreign affiliates to local firms in upstream or downstream sectors, with results that varied broadly across methods and countries.

The pooled conclusion supports the direction found in Lithuania. Once biases are taken into account, the average spillover to suppliers is economically significant, while the spillover to buyers is statistically significant but small. For same-industry spillovers, the corrected effect is not significantly different from zero.

03Exaggerated Fivefold

The second finding is the one that should change how any reader treats this entire literature, including the parts this series has relied on. The arithmetic average of all published estimates of backward spillovers is 0.88. The same effect, corrected for publication bias, is 0.178. In the authors' words, because of publication bias the average estimate of spillovers reported in peer-reviewed journals is exaggerated fivefold.

The mechanism is ordinary academic incentive. Journals select relatively large estimates for publication, and authors, editors and referees all prefer results that are statistically significant or consistent with theory. Estimates that are small or wrongly signed disproportionately never appear.

Two details sharpen it. First, no such selection was found for working papers, only for published journal articles, which means the filter operates at publication rather than at research. Second, the selection is more prominent among the results deemed most important. Because backward spillovers determine the main message of these studies, the authors observe, they are more likely to be polished, while forward and horizontal results, which function as bonus findings, are less distorted. The more a number matters to a paper's argument, the more it has been inflated.

The more central a finding is to a paper's argument, the more inflated it turns out to be. Bias concentrates on the number the author most needs.

04Where This Literature Disagrees With Itself

An earlier piece in this series described a financial-development condition on the benefit of foreign investment, on the basis that domestic firms need credit to act on the opportunity a multinational creates. That reading has support, and the meta-analysis names the researchers who make it. It also finds the opposite.

Its result is that greater spillovers are received by countries with underdeveloped financial systems. The proposed explanation is that foreign investment can itself relieve the credit constraint, by bringing scarce capital and by making supplier status a form of collateral. Survey evidence from the Czech Republic supports it: a quarter of suppliers to foreign affiliates said that supplier status helped them obtain more financing. Where credit is tight, becoming a multinational's supplier is worth more, not less.

Both mechanisms are real and they push in opposite directions. Weak credit markets make it harder for a domestic firm to gear up to serve an entrant, and they also make the entrant's business more valuable when it arrives. Which dominates is an empirical question, the studies disagree, and the honest position is that the financial condition on FDI benefit is unsettled rather than established. We should not have implied otherwise, and the amendment matters more than the tidiness of the earlier claim.

05Ownership Structure Decides Too

One further finding has direct consequences for how a deal is negotiated. In the Lithuanian data, the productivity effect originated from investments with joint foreign and domestic ownership and not from fully owned foreign affiliates. The meta-analysis reaches the same place: the degree of foreign ownership matters, and the dummy for wholly foreign-owned investments carries a negative sign.

The explanation offered is straightforward, which is that jointly owned projects source more locally. A partner with domestic ownership brings existing supplier relationships and reasons to use them; a wholly owned subsidiary is freer to import from its established global network. Since local sourcing is the channel through which the spillover travels, an ownership structure that reduces local sourcing reduces the spillover.

The meta-analysis also reports that protection of intellectual property rights is insignificant for the magnitude of spillovers, and that firms in countries open to international trade benefit more, being accustomed to trading with foreign firms and more likely to produce the intermediate goods foreign affiliates need.

06The Builder's Reading

1. Target investors who will buy locally. The measured benefit runs to suppliers. An entrant that imports every input generates the jobs and little else the strategy promised.
2. Measure local procurement, not headline investment value. Procurement share is the variable connected to the effect. Announced capital is not.
3. Prefer joint ownership where the choice exists. The productivity gain came from partially owned projects. Wholly owned affiliates source less locally and transmit less.
4. Build supplier capability, not competitor capability. Programs aimed at helping local firms meet multinational quality and reliability standards sit directly on the channel that works.
5. Discount published effect sizes. A fivefold correction on the headline number of this literature is not a rounding error. Treat any single published estimate as an upper bound.

That last point deserves to close the piece, because it applies to the series that contains it. Ten articles have argued that institutions should know what their instruments achieved and should report the result honestly. This literature did something harder: it measured its own distortion, published the correction, and reported that its most important findings were the most inflated. The number to take away is not 0.88 or 0.178. It is that somebody checked, and that the checking is what makes the remaining estimate worth using at all.

The Lithuanian firm-level results are drawn from an annual enterprise survey covering firms accounting for approximately 85 percent of output in each sector. The 15 percent figure corresponds to a one-standard-deviation increase in foreign presence in downstream sectors, equal to 4 percentage points in the study's backward-linkage variable, and is an association rather than an established causal effect. The meta-analysis pools 3,626 semi-elasticity estimates from 57 studies, with the oldest published in 2002 and the median in 2008. Corrected effects are those adjusted for publication selection bias. The financial-development result is reported as a finding of the meta-analysis and is in tension with other work in the same literature, as discussed above. No IEPA engine outputs are used in this piece.