The Promise and the Baseline: What the Agreement Was Supposed to Do

The North American Free Trade Agreement entered political argument through a promise before the agreement entered force through law. Sponsors in the administration of President Bill Clinton presented the trilateral pact as a route to expanded exports, to job creation, and to rising living standards in the United States. The claim joined three distinct expectations about NAFTA employment impact into a single case for approval. Expanded exports would open larger markets for American producers in Mexico and Canada. Job creation would follow from that expansion as firms serving those markets hired additional workers. Rising living standards would complete the sequence as trade raised productivity, widened consumer choice, and lowered prices. Judged against that stated aim, the record assembled in the research literature supports only a qualified verdict. The pact did expand trade within North America, and the aggregate gains for the United States were positive in the leading general equilibrium estimates, but the gains were small relative to the size of the American economy. The promise of a broad national transformation, stated in the language of a campaign for ratification, was not met at the scale at which the promise had been made.

The qualified verdict requires the second half of the sentence as well. A small aggregate effect describes an average across an entire economy, and an average can conceal concentrated losses in particular industries, occupations, and places. The evidence on the North American agreement contains both patterns at once. National welfare effects were modest, and employment effects at the national level were modest, while wage effects for less educated workers in the industries and localities most exposed to tariff reduction were severe. The thread that holds the findings together, and the claim that organizes the assessment that follows, is small in aggregate, severe in place. Neither half of that claim cancels the other half. The first half answers the sponsors on their own chosen ground, the performance of the national economy as a whole. The second half answers a different question, the distribution of gains and losses across workers who faced very different exposure to the same change in tariff protection. An honest measurement has to keep both questions in view at the same time, because the political history of the pact turned on the gap between them.

The warnings offered against ratification in 1992 and 1993 described the opposite geography of effects. Ross Perot, campaigning for the presidency in 1992, warned of a giant sucking sound as American production moved south toward lower Mexican wages. The phrase survived the campaign because the image compressed a specific fear into a memorable form. Production that had located in the United States, on that account, would relocate across the border once tariff protection and other barriers fell, and American workers would lose employment not through gradual competition but through a direct transfer of plants and orders to Mexico. Organized labor advanced a related displacement fear through the ratification debate. Union opposition rested on the proposition that firms would use access to lower wage labor, combined with duty free return of finished goods to the American market, to pressure American wages downward and to move work out of unionized plants. Stated as attributed positions, the warnings and the promises defined the terms on which the agreement would later be judged, and the research record assembled over the following two decades gave each side a portion of the evidence and denied each side the whole of it.

The formal identity of the measure that carried those promises and warnings into law is stated once here, because every later measurement traces back to the same legal instrument. The NAFTA Implementation Act, Public Law 103-182, enacted as H.R. 3450 in the 103rd Congress, was signed by President Bill Clinton on December 8, 1993, and the statute is codified principally at 19 U.S.C. 3301 and following. The implementing act approved the North American Free Trade Agreement, which had been signed on December 17, 1992, and the agreement entered into force on January 1, 1994. The House of Representatives passed the implementing bill on November 17, 1993, by 234 to 200, and the Senate passed the bill on November 20, 1993, by 61 to 38. Vice President Al Gore debated Ross Perot over the agreement on November 9, 1993, in the most closely watched public confrontation of the ratification campaign. Tariff phase outs under the agreement ran as long as fifteen years for the most sensitive agricultural goods, a schedule that matters for measurement because the legal change did not arrive as a single alteration completed on the first day of 1994.

That legal identity also fixes the object of evaluation. The subject under assessment is not trade with Mexico or Canada in general, and the subject is not globalization stated as an abstraction. The subject is a defined set of tariff reductions, rules, and institutional commitments approved through a defined statute, operating from a defined start date, across a defined phase in period. Precision on that point protects the assessment from a common error in public argument, the attribution of every change in North American production after 1993 to a single agreement. The provisions evaluated here are the provisions approved through the implementing act, and the effects credited to the trilateral agreement have to be effects that those provisions could plausibly have produced, measured against a baseline that describes North American commerce as the agreement found that commerce at the end of 1993.

The baseline begins with a fact that cuts against both the largest promises and the largest warnings. Trade with the two partners was already substantial before 1994. Canada and Mexico were not distant or marginal markets that the pact connected to the United States for the first time, and supply relationships already crossed both borders in volume. A trade agreement between partners with extensive prior exchange operates differently from an agreement that creates a channel where almost no channel existed. The agreement changes the terms on which an existing flow moves, but the starting flow itself places a bound on the size of the national effect that tariff change alone can generate. The survey by Mary E. Burfisher, Sherman Robinson, and Karen Thierfelder, published in the Journal of Economic Perspectives in 2001, recorded the consensus that the agreement had little effect on the United States labor market at the national level, a result connected to the scale of the preexisting trade relationship and to the modest size of the tariff change that remained to be made.

American tariffs on Mexican goods supply the clearest illustration of the baseline problem. By 1993, American tariffs on imports from Mexico were already low, and much Mexican output entered the United States duty free under preference programs and under the maquiladora regime. A tariff that stands at a low rate before liberalization can only fall by a small amount, and output that already enters without duty experiences no tariff change at all when a free trade agreement takes effect. The margin of change, rather than the headline label of free trade, determines the economic force of an agreement, and for a substantial share of Mexican goods that margin was narrow on the American side of the border. Mexican producers gained more secure and more permanent access than preference programs had offered, but the immediate price change at the border was limited. Any measurement that treats January 1, 1994 as the moment when Mexican goods passed from heavy taxation to free entry misstates the starting point.

The Mexican side of the baseline had also moved before the trilateral agreement took effect. Mexico had liberalized trade unilaterally from 1985 onward, reducing protection through decisions taken in Mexico City rather than through negotiation with Washington or Ottawa. By the time negotiators completed the North American text, Mexican firms and Mexican consumers had already lived through several years of tariff reduction, import licensing reform, and opening to foreign competition, and accounts by the Office of the United States Trade Representative and the United States International Trade Commission describe a Mexican tariff structure that fell sharply across that unilateral phase. The sequence matters for attribution, because factories that reorganized in response to the opening that began in 1985 and farmers who faced import competition before 1994 belong to the baseline rather than to the effects column of the 1993 statute. The agreement locked in, extended, and made reciprocal a liberalization that Mexico had already begun on a unilateral basis.

Canada presents the baseline problem in the purest form of the three countries. Trade between Canada and the United States already operated under the bilateral free trade agreement of 1988, so the principal tariff liberalization between the two largest partners in the trilateral arrangement had taken place six years before the North American agreement entered into force. For Canadian American commerce, the marginal change delivered in 1994 was the smallest of the three bilateral relationships, and the trilateral text mattered to Canada chiefly through access to Mexico and through the investment and dispute provisions of a three country framework. Measurement that ignores the 1988 agreement will assign to the 1993 statute changes that the earlier bilateral agreement had already produced, since liberalization that has already happened cannot happen a second time.

Production along the Mexican border supplies the fourth element of the baseline, and the element most often mishandled in popular accounts. The maquiladora regime dated from 1965, almost three decades before the trilateral agreement entered into force. Under that regime, plants in Mexico assembled or processed imported components for export, and the border region around those plants developed a specialized industrial geography long before negotiators drafted the North American text. Cross border production sharing, the practice of dividing stages of manufacture between plants in the United States and plants in Mexico, was an established feature of North American industry by 1993 rather than an invention of the agreement. The agreement altered the incentives surrounding the regime, changed the tariff treatment available to border plants, and, in specific subsectors, stimulated additional activity, but the agreement entered a landscape that cross border assembly had already shaped. Qualitative treatment belongs at the baseline stage, because the employment figures for the border sector belong to the detailed findings that later sections carry. The baseline point stands without those figures. In 1993, the border assembly system already existed at scale, and any account that dates Mexican border industrialization from 1994 erases the institution that made rapid expansion possible.

The phase in schedule deepens the measurement difficulty created by the baseline. Tariff reductions proceeded in stages, and a staged liberalization produces staged effects, with some industries facing full change in 1994 and others facing gradual reduction across a decade or more. Research designs that compare outcomes before and after a single date capture the early stages and miss the later stages, while designs that use a long window admit recessions, exchange rate movements, and trade shocks originating outside North America into the comparison. Part of the disagreement among published estimates follows from that timing alone, because a study of the first five years measured a different legal reality from a study of the first fifteen years.

The peso crisis of December 1994 illustrates the attribution hazard in the sharpest available form. Within the first year of the agreement, Mexico devalued the peso and then allowed the currency to float, and the Mexican economy contracted through a severe crisis that a United States led support package followed. Exchange rate movement of that scale changes the relative price of Mexican and American goods by far more than the tariff reductions scheduled for the first year of the pact, and a collapse in Mexican incomes changes Mexican demand for American exports through a channel that has nothing to do with treaty text. Mexico recorded a trade surplus with the United States in 1995 because the crisis crushed Mexican import demand, an artifact of the crisis rather than an effect of the agreement. Effects measured across 1994 through 1996 therefore carry a confound at the center of the observation window, and honest reporting states that confound wherever effects from those years enter the discussion.

A later confound, larger in scale for American manufacturing localities, arrived through China. Permanent normal trade relations with China in 2000 and Chinese accession to the World Trade Organization in December 2001 produced a shock to import competition that the research literature treats as larger than the North American shock for the decade after 2000. Localities that lost manufacturing employment after 2000 faced overlapping exposure to both openings, and a simple count of plant closures in such places cannot separate the two causes. The strongest local labor market research on the North American agreement addressed that problem through design, using data from the 1990s and controlling for import competition from China, and the finding of a distinct effect for the earlier agreement rests on that separation.

Five findings carry the answer that the evidence supports, and each finding answers a narrower question than the public debate usually asked. The first finding concerns aggregate estimates, the economy wide effects on welfare and trade that general equilibrium models and survey reviews produced for the three countries. The second finding concerns distributional wage effects, the relative experience of workers sorted by education, industry, and locality according to the protection that each position had enjoyed before the agreement. The third finding concerns divergence among published job loss figures and the methods that produce that divergence, from general equilibrium modelling through local labor market econometrics through deficit accounting through administrative certification counts. The fourth finding concerns the certification count itself, the number of workers whom the Department of Labor certified as displaced under the agreement specific adjustment assistance program that the implementing act created, a count drawn from adjudicated petitions rather than from a model. The fifth finding concerns the partner countries and the border, the experience of Mexican agriculture and Mexican border assembly, the limited marginal change for Canada, and the peso confound that complicates every reading of the first years. Named in that order, the five findings move from the national average to the exposed locality, from modelled estimates to an administrative record, and from the American side of each border to the partner side.

The order of the findings enacts the central methodological distinction of an impact assessment, the distinction between aggregate and distributional effects. Aggregate research asks whether the national economy gained or lost on net, and the models built for that question reallocate workers between industries by construction, so the models measure net gains and cannot show displacement as experienced in a particular town. Local labor market research asks whether workers in places and industries with high prior protection fell behind comparable workers elsewhere. Deficit accounting converts a bilateral trade balance into an employment equivalent through fixed coefficients, and administrative certification counts adjudicated petitions. Confusion enters when an answer to one question is presented as an answer to another.

Political salience follows from that same distinction. Gains from trade liberalization, in the account that sponsors relied upon, arrive in diffuse form, as lower prices reach millions of consumers in amounts too small for any single household to attribute to a treaty and as export gains reach firms scattered across many industries. Losses arrive in concentrated and visible form, since a plant closure names a town, a date, and a group of workers. Concentrated visible losses outweigh diffuse invisible gains in political weight, even when the aggregate arithmetic favors the diffuse side, and the ratification debate previewed that asymmetry in full.

Both political uses of the agreement therefore fail against the full record. The claim that the trilateral agreement devastated American manufacturing cannot be supported by the aggregate literature, which found national effects that were small relative to the economy, nor by the baseline, which shows that the remaining tariff margin in 1994 was too narrow to carry devastation on a national scale. The claim that the agreement was costless cannot be supported by the localized wage research, which found relative losses for the least educated workers in the most exposed places, nor by the certification counts, which recorded adjudicated displacement across the operating period of the agreement specific program. The statute deserves assessment against the aims stated for the statute in 1993, in the language of exports, employment, and living standards for the country as a whole. On that ground the verdict is qualified success at modest scale, with the distributional record adding the qualification that gives the assessment its lasting interest, because a policy can clear a national test and still impose severe costs on the citizens least positioned to absorb those costs.

The baseline converts that verdict from an assertion into an expectation that the evidence could have falsified. Substantial preexisting trade and low preexisting American tariffs each predicted a bounded aggregate effect. Mexican liberalization that began in 1985 predicted that part of any observed opening belonged to an earlier policy, the Canadian agreement of 1988 predicted the smallest marginal change on the northern border, and the maquiladora regime that dated from 1965 predicted that border production would expand from an established base rather than begin from nothing. The convergence of those elements explains why economists who disagreed on much else converged on small aggregate estimates. The same baseline leaves distributional effects entirely open, because a small national margin can still fall with full force on the narrow set of industries that enjoyed the remaining protection, and the workers in those places faced the largest change in competitive conditions that the 1993 agreement delivered. Severity in place is the baseline prediction for distribution, just as smallness in aggregate is the baseline prediction for the nation.

Finding One: The Aggregate Estimates and Why They Are Small

The aggregate literature on the North American agreement returned a result that sat uneasily beside the scale of the political argument that preceded adoption. The pact had been debated as if it would remake the American economy, either by opening a large new field for exports or by draining industrial employment toward lower Mexican wages. The economy-wide studies that followed measured something far more modest. For the United States, the estimated welfare effect was positive and small, the estimated effect on the labor market as a whole was limited, and the estimated expansion of trade inside the bloc was large in proportional terms without producing a correspondingly large change in national welfare. Trade could grow rapidly from an established base, prices and production could adjust across many sectors, and the net gain for a very large economy could remain a small fraction of one percent, all at the same time and without contradiction.

What did the general equilibrium models actually measure?

The models measured economy-wide welfare and trade flows after tariff removal, comparing a simulated economy with the agreement against one without it. Lorenzo Caliendo and Fernando Parro estimated welfare and intra-bloc trade changes, not counts of displaced workers, because their model reallocated labor across sectors by construction.

A computable general equilibrium model began from a representation of an entire economy rather than from a single industry, firm, or bilateral balance. Its structure linked producers, consumers, intermediate inputs, and trade flows across sectors, so that a tariff reduction in one place altered relative prices elsewhere and drew resources toward some activities and away from others. The comparison at the center of such an exercise was counterfactual. One version of the modeled economy retained the tariff structure that the trilateral agreement removed, and another version incorporated the removal, and the difference between the two versions produced the estimate. What emerged was therefore not a count of events observed in administrative records and not a tally of workers who had lost particular jobs. It was a modeled difference in welfare and in the volume of trade, expressed for each country as a whole, under the assumptions built into the model.

That method defined both the strength and the boundary of the result. Because the model represented the economy as an interconnected system, it could capture effects that a study of one tariff line would miss, including changes in the cost of intermediate goods that one industry bought from another and the reallocation of labor and capital toward sectors where each country held a comparative position. Because labor was reallocated across sectors by construction, however, the same model could not display displacement as an outcome. A worker in the model moved within the system as sectors expanded and contracted, and the adjustment was part of the mechanism rather than a measured cost with a duration, a wage history, or a locality attached to it. The model could estimate welfare and trade effects, but it could not show which workers bore the cost of moving or whether a particular town recovered the activity it lost. Those limits belonged to the method itself.

The fullest expression of that approach for the 1993 agreement came from Lorenzo Caliendo and Fernando Parro, in “Estimates of the Trade and Welfare Effects of NAFTA,” published in the Review of Economic Studies, volume 82, number 1, in 2015, at pages 1 to 44, and first circulated as National Bureau of Economic Research Working Paper 18508 in 2012. The study applied a computable general equilibrium model to the three signatory economies and estimated the welfare and trade consequences attributable to the tariff reductions the agreement carried out. Rather than asking what happened to one product category crossing one border, Caliendo and Parro asked what the removal of the tariff structure did to welfare in Mexico, the United States, and Canada when the interactions among sectors and countries were allowed to work through the model together.

The welfare results placed the three countries in sharply different positions, and the ordering carried much of the meaning. Caliendo and Parro estimated that Mexico gained 1.31 percent in welfare, that the United States gained 0.08 percent, and that Canada lost 0.06 percent. For Mexico, a gain above one percent represented a measurable improvement relative to the size of the Mexican economy, and it reflected the scale of the change the agreement represented for the smallest and most protected of the three economies at the point of adoption. For the United States, a gain of eight hundredths of one percent represented a real but very small improvement, positive in direction and minor in scale against the whole of American output and consumption. For Canada, a loss of six hundredths of one percent sat so close to zero that the modeled figure indicated no substantial welfare movement in either direction. Read together, the three figures described an agreement whose aggregate consequences were concentrated where preexisting barriers and the relative size of the trade relationship made them largest, and diluted where the economy was largest and the barriers lowest.

The modeled trade results told a different kind of story, and the contrast between the trade figures and the welfare figures explained much of the confusion surrounding popular readings of the research. Caliendo and Parro modeled increases in trade within the bloc of 118 percent for Mexico, 41 percent for the United States, and 11 percent for Canada. A modeled increase of 41 percent in intra-bloc trade for the United States sounded dramatic when quoted alone, and a modeled increase of 118 percent for Mexico sounded even more dramatic. Yet a large proportional increase in trade did not translate mechanically into a large proportional change in national welfare, because trade with the two partners represented only one portion of a much larger economy, because many goods moving across the borders were intermediate inputs counted within production chains rather than as final gains, and because welfare depended on the price and efficiency changes the trade produced rather than on the gross value of goods that crossed a border. The model therefore returned both rapid trade expansion and a small net welfare change for the United States, and the two results reinforced rather than undermined each other.

Canada functioned within the Caliendo and Parro results as the limiting case that made the underlying logic visible. A welfare change of minus 0.06 percent and a modeled intra-bloc trade increase of 11 percent, the smallest figures for any of the three countries, followed from the position Canada occupied before the trilateral agreement entered into force. Liberalization between Canada and the United States had already been carried largely to completion under the bilateral free trade agreement of 1988, so the marginal tariff change between the two largest partners in the bloc was modest by 1994. Canada had a stake in the trilateral text, but that stake ran through access to Mexico and through the investment and dispute chapters rather than through tariff change with the United States. Where preceding liberalization had been most complete, the additional measured effect was smallest.

A second and earlier line of evidence reached a compatible conclusion by a different route. Mary E. Burfisher, Sherman Robinson, and Karen Thierfelder published “The Impact of NAFTA on the United States” in the Journal of Economic Perspectives, volume 15, number 1, in 2001, at pages 125 to 144. Their article surveyed the studies produced after the agreement entered into force and assessed what the accumulated research had established during the first years of operation. A survey of that kind carried particular weight because it did not depend on the assumptions of one model or the data choices of one research team. Where a survey and a later economy-wide model pointed in the same general direction, despite different methods and publication dates, the convergence strengthened the aggregate finding beyond what either source would have supported on its own.

On benefits, the survey by Burfisher, Robinson, and Thierfelder concluded that both the United States and Mexico benefited from the agreement, with much larger relative benefits for Mexico. The asymmetry matched the Caliendo and Parro ordering published fourteen years later, and it rested on a foundation the survey made explicit through the body of work it reviewed. Mexico entered the agreement with higher barriers, a smaller economy, and a trade relationship with the United States that represented a far larger share of Mexican economic activity than the same relationship represented for the United States. Removing barriers between two economies of very different size therefore produced relative gains of very different scale, even where the absolute value of trade moving in each direction was large. The American benefit was genuine in the estimation of the post-agreement studies, but it was spread across an economy of continental dimensions, and a benefit spread on that scale registered as a small percentage however it was aggregated.

On the labor market, the verdict of the survey was direct. Burfisher, Robinson, and Thierfelder concluded that the agreement had little effect on the United States labor market. That finding addressed the labor market in aggregate, as a national system of employment and wages, and within that frame the authors found no large movement attributable to the pact. The conclusion carried the same methodological boundary that governed the general equilibrium literature as a whole. An aggregate labor market effect could be small while particular industries faced strong pressure and particular localities absorbed concentrated losses, because the national measure netted expanding sectors against contracting ones and averaged workers in exposed places together with workers in places the tariff changes barely touched. The survey finding therefore settled one question and left another open. It established that the post-agreement research had not found a large national labor market effect, and it did not, by its own construction, measure the distribution of adjustment costs among workers and communities where exposure had been greatest.

On the character of the new trade, Burfisher, Robinson, and Thierfelder concluded that trade creation greatly exceeded trade diversion, and that the pattern held especially in intermediate goods. Trade creation described trade that arose because a lower tariff allowed a more efficient producer to supply a good, or allowed a producer to obtain an input at a lower cost, in a way that raised real income. Trade diversion described trade that merely shifted from a more efficient supplier outside the bloc to a less efficient supplier inside the bloc because the preference, rather than the underlying cost, had changed. A finding that creation greatly exceeded diversion indicated that the dominant effect of the tariff reductions, in the studies the survey reviewed, ran through lower costs and expanded production rather than through the rerouting of purchases toward higher cost sources sheltered by a preference. The emphasis on intermediate goods reinforced that reading, since much of the trade at issue consisted of components and materials bought by firms for use in further production, so a tariff reduction on those goods lowered production costs inside the United States as well as abroad.

Burfisher, Robinson, and Thierfelder also recorded an observation about the relationship between the evidence and the debate that had preceded the agreement. The results, the survey authors wrote, confirmed the consensus opinion of economists at the time of the debate. Before adoption, the professional expectation among economists who supported the agreement had generally been that the aggregate American effects would be positive and small, precisely because American tariffs had already stood low and because the American economy had dwarfed the trade flows at stake. The post-agreement studies, produced after several years of operating data had become available, matched that expectation in direction and in scale. The literature had therefore not reversed the professional judgment rendered during the legislative fight. It had tested that judgment against experience and found the broad outline intact, even though the same years also produced evidence, outside the scope of the aggregate finding, about concentrated costs that aggregate models were structurally unable to display.

The third aggregate assessment in the record approached the question through consumer gains rather than through a modeled welfare percentage. Gary Clyde Hufbauer and Jeffrey J. Schott, in NAFTA Revisited: Achievements and Challenges, published by the Peterson Institute for International Economics in 2005, estimated consumer gains of about 5.5 billion dollars per year across the first nine years of the agreement, amounting to about 49.5 billion dollars in total. The estimate had to be carried at exactly the strength the authors and the compiled record allowed. The qualifier “about” belonged to the figure, the period covered the first nine years of operation, and the gain described a consumer benefit as Hufbauer and Schott assessed it, not a comprehensive accounting of every gain and loss across producers, workers, and governments. Stated within those limits, the estimate gave the aggregate finding a dollar form that complemented the percentage form of the general equilibrium results, and its modest scale per household aligned with the small welfare percentage Caliendo and Parro later estimated by a different technique.

Three structural reasons, taken together, explained why the aggregate American effect had to be small, and each reason rested on conditions that predated the agreement rather than on any judgment about how well the agreement operated. The first reason was relative size. The United States economy was so large in comparison with the trade involved that even a substantial proportional change in commerce with Mexico and Canada moved only a small share of total American production, consumption, and employment. A tariff removal could transform the economics of a particular supply chain, redirect sourcing for a particular product, or expand a particular export line by a wide margin, and yet register as a minor change when divided across the whole of the American economy. The denominator governed the headline percentage. Studies that expressed their results as shares of national welfare or of the national labor market were dividing by the full scale of the largest economy in the bloc, and any plausible numerator produced by the tariff changes at issue yielded a small quotient.

The second reason was the state of American tariffs and entry channels before 1994. American tariffs on Mexican goods were already low when the agreement was negotiated, and much Mexican output already entered the United States duty free under preference programs and under the maquiladora regime, the border assembly system that had operated since 1965. An agreement that removes a tariff extracts an efficiency gain roughly in proportion to the barrier it eliminates, and where the barrier was already low, or where goods already crossed the border without a duty, the room for a further gain was correspondingly narrow. Large portions of the trade the agreement governed had therefore already been conducted on terms close to the terms the agreement formalized. The pact locked in and extended access that in many product lines already existed in practice, and the measured aggregate effect reflected the difference between the old terms and the new terms rather than the difference between open trade and no trade at all.

The third reason ran in the opposite direction across the southern border. Mexico had liberalized on its own initiative well before the trilateral agreement entered into force, beginning in 1985, and that unilateral opening had already carried out a large share of the adjustment that a trade agreement would otherwise have been credited with producing. Accounts from the Office of the United States Trade Representative and the United States International Trade Commission, as carried in the record, placed Mexican average tariffs at above 25 percent in 1985 and at a trade-weighted average of about 10 percent by the mid-1990s. By the time the North American agreement took effect, Mexican barriers had already fallen substantially, and the additional change operated on an economy that had been moving toward openness for nearly a decade. For the United States, the export opportunities and the import competition created in 1994 represented an increment on a process already underway, not a sudden passage from a closed Mexican market to an open one, and increments of that kind produced smaller measured effects than a single step across the full distance would have produced.

A related condition governed the northern relationship and completed the structural account. Trade between Canada and the United States was already substantial before 1994, and it already operated under the bilateral free trade agreement concluded in 1988. The trilateral agreement therefore layered a three-country framework onto a northern trade relationship whose central liberalization had already occurred, adding the Mexican dimension and a set of common rules while changing the underlying tariff position between the United States and Canada by comparatively little. Combined with the position on the southern border, the structural picture pointed in one direction from every side. The agreement formalized and extended an integration that was already far advanced on both borders before January 1, 1994, and the aggregate studies measured the effect of that formalization and extension, a quantity that preexisting conditions had already ensured would be small for the United States.

Measured trade growth, set against that background, showed how rapidly commerce could expand while the welfare effect remained minor. Burfisher, Robinson, and Thierfelder reported that Mexico’s share of United States imports rose from 6.9 percent in 1993 to 11.5 percent in 2000. The shift represented a rapid reorientation of sourcing toward Mexico, and it demonstrated that the agreement coincided with a genuine deepening of North American production integration. Yet even after that increase, imports from Mexico remained a modest share of American imports, and imports themselves represented only one component of an economy dominated by domestic production and services. Two-way trade across the bloc also grew several-fold in nominal terms across the first decade of operation. Neither figure contradicted the welfare estimates. Both measured flows, while the welfare studies measured the net gain generated by the changed terms on which those flows occurred, and a flow could multiply from a modest base while the net gain on each dollar of it remained small.

The aggregate finding therefore stood on converging evidence from three directions. Caliendo and Parro estimated a welfare gain of 0.08 percent for the United States, Burfisher, Robinson, and Thierfelder found little effect on the American labor market and trade creation greatly exceeding trade diversion, and Hufbauer and Schott put consumer gains at about 5.5 billion dollars per year across the first nine years. Each source had a defined method and a defined limit, and the three together described an agreement whose national effect on the United States was positive, diffuse, and small relative to the economy it touched, for structural reasons visible before the first tariff line changed.

Finding Two: The Workers the Averages Conceal

The aggregate estimates describe an economy so large that a trade agreement with two smaller partners moved national totals by fractions of a percent. That description answers one question, and it leaves another question untouched. A national average is built by adding gains and losses across every worker, every industry, and every place, and then dividing the result by the whole. Where a loss fell on a narrow group, the division itself hides the event. The second finding in the research record starts at the other end of that arithmetic. It asks not what the trilateral agreement did to the United States as a single unit, but what the pact did to the particular workers and the particular towns that stood directly in the path of the tariff reductions it enacted. The answer differs from the aggregate answer in scale and in character. Rather than modeling the whole economy and allowing labor to move between sectors until a new equilibrium forms, the study followed actual workers through Census records across the decade in which the agreement took effect, and it compared workers who had been sheltered by high tariffs before 1994 with workers who had not. The sheltered group lost ground in relative terms, year after year in cumulative effect, and the size of that relative loss was large enough to shape earnings trajectories for workers with the least formal education.

The study at the center of this finding is the work of Shushanik Hakobyan and John McLaren, published as “Looking for Local Labor Market Effects of NAFTA” in the Review of Economics and Statistics, volume 98, number 4, in 2016, at pages 728 to 741. The research had circulated earlier as National Bureau of Economic Research Working Paper 16535, issued in 2010. Hakobyan and McLaren built the analysis on United States Census data for 1990 to 2000, a choice of endpoints that carries much of the argument. The 1990 Census recorded the American labor force before the North American agreement was signed in December 1992 and before it entered into force on January 1, 1994, while the 2000 Census recorded the same labor force after six years of operation under the pact and after a substantial portion of its tariff phase out schedule had run. A comparison across those two points therefore measures something different from a reaction in a single year. It measures wage growth across a full decade for workers classified by the exposure of the industry that employed them and the exposure of the locality where they lived, and wage growth across a decade is the form in which a labor market shock becomes a life outcome, because slower growth compounds into a permanently lower earnings path even where no single year contained a dramatic fall.

The exposure design deserves close attention, because it determines what the findings can and cannot establish. Hakobyan and McLaren measured exposure by the tariff protection an industry or a locality enjoyed before the agreement. An industry that had entered the 1990s behind a high American tariff on Mexican goods faced a large change when the implementing act dismantled that protection on a set schedule, since Mexican producers gained access to the American market on terms that the tariff had previously denied to them. An industry that had entered the 1990s with little or no tariff protection faced a far smaller change, because the agreement removed a barrier that had barely existed. The same logic applied to places. A locality whose employment rested heavily on protected industries carried a high measure of exposure even though the tariff itself was a national instrument, while a locality whose employment rested on unprotected industries or on activities untouched by trade with Mexico carried a low measure. Exposure, in this design, is therefore not a guess about which places felt harmed and not a count of imports after the fact. It is a measurement taken from the structure of protection as that structure stood before the policy changed, applied forward to the workers and towns whose earnings the Census then recorded. Because the measure of vulnerability was fixed by the pre agreement tariff schedule, the differing exposure came from a policy decision enacted in the statute, rather than from choices those workers or towns made during the decade under study.

The industry result states the distributional finding at its sharpest. For workers who had not completed high school, wage growth in the most protected industries ran up to 17 percentage points lower across 1990 to 2000 than wage growth for comparable workers in initially unprotected industries. Every element of that sentence carries weight. The figure is expressed in percentage points of wage growth, not as a percentage decline in the level of wages, and the distinction separates the finding from cruder claims made in public argument. Hakobyan and McLaren did not find that the wages of high school dropouts in protected industries fell by 17 percent. The study found that their wages grew more slowly, by as much as 17 percentage points over the decade, than the wages of similar workers whose industries had never enjoyed protection. A worker can receive raises across a decade and yet sustain a loss of exactly this kind, because the loss is measured against the path that comparable workers followed. Comparable workers in initially unprotected industries faced the same national business cycle, the same monetary policy, and the same technological changes of the 1990s, so the gap between the two groups isolates the part of the protected workers experience that the unprotected workers did not share. The education qualifier is equally specific. The large figure belongs to high school dropouts, the workers with the fewest credentials to carry into another industry and the least bargaining position within their own, and the study reported the effect at that strength for that group rather than for the workforce as a whole.

The locality result extends the finding from the industry a worker served to the place a worker inhabited. In the most vulnerable localities, wage growth for a high school dropout ran 8 percentage points slower across the same decade. The number is smaller than the industry figure, and the relationship between the two numbers is itself instructive. An industry classification follows a worker into whichever town contains the plant, while a locality classification gathers every worker in a place, including workers whose own industries carried little protection but whose earnings depended on the spending of neighbors who worked in exposed plants. A town level average therefore mixes the directly exposed with the indirectly affected, and a mixed average produces a smaller measured gap even where the underlying harm to the most exposed households matched the industry result. The 8 percentage point figure captures the experience of place as such. It records that residence in a vulnerable locality predicted slower wage growth for a worker without a high school diploma, over and above the national experience of workers with the same education. Place operated as a labor market condition in its own right during the decade after the 1993 agreement entered into force, because local labor markets clear locally for workers whose skills, family obligations, housing, and finances bind them to a commuting area rather than to a continental market. A worker in an exposed town competed within a local pool of employers, and when the anchor employers of that pool faced new competition from Mexican production, the terms available across the pool deteriorated together.

The study also traced the effect beyond the factory gate, into the local service industries of exposed places. Workers in local service industries in those places were also affected, through local demand. A protected plant that loses orders or closes a line reduces the earnings of its own workforce first, and those reduced earnings then circulate as reduced spending at the retailers, restaurants, repair shops, medical offices, and personal service businesses that depend on local customers rather than on national markets. Service workers in such a town sold their labor to the households around them, and when the households around them earned less, or grew more slowly, the demand supporting service wages contracted with it. This spillover breaks any simple division between trade exposed workers and everyone else. A hairdresser, a cook, or a clerk in an exposed locality held no job that Mexican imports could take directly, and no tariff had ever protected that job, yet the wage growth available in that occupation depended on the incomes generated by the protected industries nearby. The distributional harm of the pact did not stop at the boundary of the traded goods sector. It propagated through the ordinary commercial life of the towns where traded goods production had concentrated, reaching workers whose connection to international trade ran entirely through the spending of their neighbors.

Education divided the results more cleanly than any other characteristic in the study. For college educated workers, the effects were small and statistically insignificant. Statistically insignificant carries a precise meaning in this context: the estimated effects for the college educated were close enough to zero that the data could not distinguish them from no effect at all, given the variation present in the Census records. The contrast between that result and the dropout results gives the finding its distributional shape. The same tariff reductions, applied through the same statute, across the same decade, left one educational group with wage growth close to the experience of its unprotected counterparts while imposing a large relative penalty on another. College educated workers held credentials that employers in expanding sectors recognized, and the study found no measurable wage penalty attached to their exposure. Workers without a high school diploma held the least transferable position in the labor market, and the study found the penalty concentrated there. The exposed workforce of the protected industries also stood within a longer institutional history, since the terms on which less educated industrial workers held their jobs had been built across decades of collective bargaining, seniority rules, and statutory protection described in the wider history of American labor legislation since 1950, and a tariff was only one layer of the structure that had supported those wages. When the tariff layer was removed, the workers with the thinnest alternative supports absorbed the measured loss, while the workers whose earnings rested on portable credentials did not.

A later study by the same authors added a gender dimension to the distributional record, and it receives one passage here because the fact sheet of the research assigns it a supporting rather than a central place. In “NAFTA and the Gender Wage Gap,” issued as Upjohn Institute Working Paper 17-270 in 2017, Hakobyan and McLaren found that married blue collar women in exposed places were among the hardest hit. The result refines the education finding rather than replacing it, since it identifies a group within the blue collar workforce of vulnerable localities whose wage outcomes deteriorated most sharply, and it is best read as an extension of the 2016 design by its own authors, pointing toward a finer division of the harm inside the exposed places, rather than as a separate estimate competing with the Census results already described.

The mechanism behind these results follows from the geography of protection as it stood before 1994. Tariff protection in the United States had never been spread evenly across the economy. Protection had concentrated in specific industries, and those industries had seated themselves in specific towns and regions, where a plant or a cluster of plants anchored local employment across generations. That concentration was the product of industrial history rather than of the agreement, and it meant that the pre agreement tariff schedule amounted to a map of identifiable places. Removing the protection therefore landed on identifiable people in identifiable places. A tariff cut written as a line in a schedule became, on the ground, a change in the competitive position of named employers in named towns, and through them a change in the earnings of workers whose names appeared on local payrolls and whose children attended local schools. The national framing of trade policy, in which a tariff is a small percentage applied to a category of goods, obscures that translation, and the Census based design restored it by following the consequences from the location of protection to the location of workers. Where protection had been thin, the agreement changed little and the study found little. Where protection had been thick and had supported a local industrial base, the agreement withdrew a support on which wages in that base had partly rested, and the measured wage growth of the least educated workers in those places fell behind the national comparison for a decade.

The duration of the adjustment forms the second half of the mechanism. The wage effects in the study run across 1990 to 2000, the full span between the two Censuses, and the finding describes an adjustment that ran for a decade rather than a season. Tariff phase outs under the North American agreement themselves ran on schedules, in some categories extending across as much as fifteen years, so the competitive change arrived in stages rather than in a single break. Local adjustment took longer. A worker displaced from a protected industry, or held in it at slower growing wages, faced retraining that took time, alternative local employers that paid less, and a housing and family position that made relocation costly, while a town whose anchor industry contracted faced a shrinking tax base and declining demand for every local service at once. None of those frictions resolves within a single business cycle, and the decade long window of the study is what allows the record to show persistence rather than a brief disturbance. A design that had compared only 1993 with 1995 would have captured dislocation without trajectory. The 1990 to 2000 design captured trajectory, and the trajectory it captured for high school dropouts in exposed industries and exposed places was one of cumulative divergence, in which each year of slower growth widened the distance between those workers and their counterparts in unprotected employment, so that by 2000 the gap represented not a temporary dip but a decade of forgone earnings growth.

Reconciling this finding with the aggregate finding requires no strain, because the two findings describe the same events at different scales and the arithmetic of scale explains both. National averages divide a concentrated loss across a continental labor force. The American labor force of the 1990s spread across services, government, construction, finance, health care, education, and every other sector that the trilateral agreement touched lightly or not at all. A severe relative wage penalty borne by high school dropouts in a defined set of protected industries and vulnerable localities enters that average as a small quantity, diluted by the earnings growth of the vast majority of workers whose industries carried no protection to lose. General equilibrium estimates of the kind summarized in the first finding register the net result of that dilution, together with the consumer gains and the efficiency gains that freer trade within the bloc generated, and they return a figure close to zero for the United States because the net result at national scale was close to zero. Both statements hold at once. The agreement produced small aggregate effects for the American economy, and it produced large relative wage losses for an identifiable minority of American workers. Each method answers the question it asked, and confusion enters public argument only when a figure built to answer the national question is used to deny the local answer, or when a local figure built on relative comparisons is presented as a measure of national loss. Hakobyan and McLaren measured relative effects, the difference between exposed workers and comparable unexposed workers, and relative effects of that design characterize who bore the cost within the economy rather than the net cost to the economy as a whole.

The separation of this effect from import competition with China demanded particular care, and Hakobyan and McLaren built that separation into the study rather than asserting it afterward. The authors controlled for import competition from China and found the NAFTA effect separate from it. Permanent normal trade relations with China arrived in 2000, and China entered the World Trade Organization in December 2001, and the shock that followed across American manufacturing localities during the decade after 2000 was the larger of the two. A study of wage outcomes that ended in 2000 rests largely on the decade before that shock took hold, and the Hakobyan and McLaren design worked on 1990s data for that reason, while also carrying an explicit control for Chinese import competition within its estimates. The finding reported here is therefore a finding about the 1993 agreement and its tariff schedule, not a repackaging of the China shock under a Mexican label. The care runs in both directions. Nothing in the distributional result licenses attributing the whole post 2000 decline of American manufacturing employment to the North American agreement, since the design that produced the result separated the two shocks and assigned the larger later shock to its own cause. What the design supports is narrower and, for that reason, harder to dismiss: within the 1990s, among workers sorted by the protection their industries and towns had enjoyed before 1994, the removal of that protection tracked a measurable, education graded, place bound penalty in wage growth, and that penalty survived the inclusion of the China control.

Read together, the elements of the second finding compose a single picture of concentrated adjustment. A statute signed in December 1993 and in force from January 1994 withdrew tariff protection that had been concentrated, by decades of industrial history, in particular industries seated in particular places. The workers most exposed to that withdrawal were those with the least formal education, and their wage growth fell behind comparable workers by as much as 17 percentage points in the most protected industries and by 8 percentage points in the most vulnerable localities across the following decade. The penalty reached past the exposed plants into local service employment through the contraction of local demand, and it left college educated workers measurably untouched. A later working paper by the same authors located some of the sharpest harm among married blue collar women in the exposed places. The aggregate literature and the local literature describe one economy from two distances, and the distance determines the visibility of the harm. From the altitude of the national accounts, the pact registered as a small net gain shared thinly across a continent. From the altitude of the Census tract and the payroll, it registered as a decade of lost ground for workers who could least afford the loss, in towns built around the very protection the statute removed.

Finding Three: Four Methods, Four Different Agreements

The published estimates of what the North American agreement did to American employment appear, at first reading, to describe different agreements. One body of research reported effects so small that they registered as rounding differences beside the size of the American economy. Another body of research reported relative wage losses concentrated in particular industries and places that persisted across a decade. A third body of research reported job displacement in the hundreds of thousands. A fourth record counted workers one petition at a time and produced a total that was neither a model output nor an economy wide estimate. The divergence did not arise primarily from disagreement over the underlying trade flows, which were drawn from the same customs records and the same national accounts. The divergence arose from method. Each method asked a different question, counted a different population, applied a different counterfactual, and permitted a different set of adjustments to occur inside the calculation. A figure produced by one method therefore cannot be placed beside a figure produced by another method as though both measured the same object on the same scale.

The distinction between net, relative, gross, and adjudicated measurement explains most of the distance between the smallest and the largest numbers in the literature. A net estimate asks what happened in the economy as a whole after gains in some sectors offset losses in others. A relative estimate asks what happened to a defined group compared with a similar group that faced less exposure, and the answer can be large for the group even when the national total moved very little. A gross estimate asks how much domestic production a given trade flow might have supported under stated assumptions, without tracing where displaced workers went afterward. An adjudicated count records only the cases that entered and survived an administrative process that required evidence, investigation, and a formal determination. Confusion enters the public argument when a gross figure is quoted as though it were a net figure, or when a relative effect on an exposed locality is quoted as though it described the national labor market. The studies themselves, read with their methods attached, conflict far less than the quotations drawn from them.

General equilibrium modelling belongs to the first family, and its design determines its characteristic result. A general equilibrium model builds a representation of the whole economy, or of several linked economies, in which industries buy inputs from one another, households supply labor and consume output, prices respond to changes in costs, and trade flows adjust when tariffs change. The researcher then removes or reduces the tariffs specified in the agreement inside the model and compares the resulting equilibrium with a baseline in which the tariffs remained in place. The comparison yields estimates of changes in trade volumes, output by sector, and welfare, expressed as a change in real income. Lorenzo Caliendo and Fernando Parro, in Estimates of the Trade and Welfare Effects of NAFTA, published in the Review of Economic Studies in 2015, applied such a model to the trilateral agreement and estimated a welfare gain of 0.08 percent for the United States, a welfare gain of 1.31 percent for Mexico, and a welfare loss of 0.06 percent for Canada. The same model estimated increases in trade within the bloc of 41 percent for the United States, 118 percent for Mexico, and 11 percent for Canada.

The method imposes a particular view of the labor market before any data on workers enters the calculation. In the standard form of such a model, labor moves between sectors as relative prices change, and the economy returns to an equilibrium in which the available labor force is employed, although possibly in different industries and at different wages than in the baseline. Displacement, in the sense of a worker losing a job in an import competing plant and experiencing a spell without work or a permanent reduction in earnings, does not appear as a separate outcome, because the model solves for the allocation after adjustment rather than tracing the path of adjustment worker by worker. The strength of the method lies in consistency across the whole system, so that export expansion, import competing contraction, input price changes, and household purchasing power all appear inside a single framework that cannot count the same dollar twice. Mary E. Burfisher, Sherman Robinson, and Karen Thierfelder, in The Impact of NAFTA on the United States, published in the Journal of Economic Perspectives in 2001, surveyed the post agreement studies and reported that both the United States and Mexico benefited, with much larger relative benefits for Mexico, that the agreement had little effect on the United States labor market, and that trade creation greatly exceeded trade diversion, especially in intermediate goods. Neither that survey conclusion nor the Caliendo and Parro estimates, on their own terms, denies that adjustment fell heavily on particular places.

The smallness of the aggregate estimates also reflected conditions that preceded the pact. Trade with Canada and Mexico was already substantial before 1994, American tariffs on Mexican goods were already low, and much Mexican output entered the United States duty free under preference programs and under the maquiladora regime. Mexico had liberalized unilaterally from 1985, reducing average tariffs from above 25 percent in 1985 to a trade weighted average of about 10 percent by the mid 1990s, according to accounts by the Office of the United States Trade Representative and the United States International Trade Commission. Canada and United States trade already operated under the bilateral free trade agreement of 1988. A model that removes a tariff that was already low, on trade that was already large, produces a small percentage change in national welfare even when the absolute value of the trade affected is considerable.

Local labor market econometrics belongs to the second family, and its design produces a different kind of answer from the same agreement. Rather than modelling the entire economy, an econometric study of local labor markets identifies places and industries that were exposed to the agreement in measurably different degrees, and then compares outcomes for workers in more exposed settings with outcomes for comparable workers in less exposed settings across a defined period. Exposure in the leading study came from the tariff protection that an industry or a locality had enjoyed before the agreement took effect. A place whose employment depended heavily on industries that had been sheltered by high tariffs faced a larger reduction in protection when the agreement phased those tariffs down, and a place whose industries had faced little protection faced a smaller change. The comparison isolates the effect of the change in protection from national trends that affected all places, provided that the comparison groups were otherwise similar and that other shocks can be controlled for in the design.

Shushanik Hakobyan and John McLaren, in Looking for Local Labor Market Effects of NAFTA, published in the Review of Economics and Statistics in 2016, applied that design to United States Census data for 1990 to 2000. The authors found that wage growth for high school dropouts in the most protected industries was up to 17 percentage points lower across 1990 to 2000 than wage growth for comparable workers in initially unprotected industries, and that in the most vulnerable localities wage growth for a high school dropout ran 8 percentage points slower. Workers in local service industries in exposed places were also affected, through local demand, because a reduction in earnings at the traded goods employers of a town reduced spending at businesses that sold only to that town. Effects for college educated workers were small and statistically insignificant in the same study. The authors controlled for import competition from China and found the effect of the trilateral agreement separate from that competition. None of those findings supplies a national count of jobs gained or lost. The findings measure relative wage growth for defined groups against a counterfactual drawn from less exposed groups, and the measurement is powerful precisely because the scope is narrow. A town can fall far behind a national trend that itself moved very little, and both statements can hold at once without contradiction.

The second family also carries limits that follow from its design. A relative estimate depends on the choice of comparison group, on the accuracy of the exposure measure, and on the assumption that no other shock struck the exposed places in the same pattern during the same years. The method can estimate the gap between exposed and unexposed workers, but it cannot add gains and losses across every group into a national employment total, because the national total was never the object of the comparison. It also measures wages more readily than job counts, since census records follow the earnings of workers who remained observable in the data more completely than they follow every separation and every exit from the labor force.

Bilateral deficit accounting belongs to the third family, and it differs from the first two families in a fundamental respect. The method does not estimate a counterfactual from a model of the economy, and it does not compare exposed and unexposed groups in observed data. The method starts from an observed trade balance between two countries and converts that balance into an employment equivalent through arithmetic. Because the starting point is an observed deficit rather than an estimated change caused by the agreement, and because the conversion applies to the whole balance rather than to a measured tariff effect, the resulting figures stand apart in scale from everything produced by the other families. Understanding the figures requires following the conversion step by step, since the conclusion is contained in the steps and cannot be separated from them.

Why does the deficit method produce the largest numbers?

Deficit accounting produces the largest numbers because it converts an entire bilateral trade balance into jobs with fixed labor content coefficients and treats the full deficit as displaced domestic production, without allowing wages, prices, exchange rates, or reallocation to offset any part of the balance in the calculation itself over time.

The first step records the bilateral balance. Trade statisticians total the value of goods that the United States imported from a partner during a period and subtract the value of goods that the United States exported to that partner during the same period, and a negative result for the United States is reported as a bilateral deficit with that partner. The second step assigns a labor content to each dollar of trade. Analysts using the method take the amount of American labor associated with a dollar of domestic production in the relevant industries, expressed as a number of jobs per million dollars of output or an equivalent coefficient, and hold that relationship fixed for the calculation. The third step applies the coefficient to the balance. Imports are treated as production that could have occurred in the United States and therefore as employment that the imports displaced, while exports are treated as production that did occur in the United States and therefore as employment that the exports supported. The fourth step nets the two sides against each other, subtracting the export supported employment from the import displaced employment, and reports the remainder as jobs displaced by the deficit. A fifth step, present in multi year presentations of the method, repeats the calculation across a span of years or applies it to the balance of a given year and presents the total as the employment consequence of trade with that partner over the period stated by the author.

Robert E. Scott, in Economic Policy Institute Briefing Paper 308, published on May 3, 2011, applied the method to trade with Mexico and reported that trade deficits with Mexico totaling 97.2 billion dollars in 2010 had displaced 682,900 United States jobs across 1994 to 2010, net of the jobs that the method credits to exports. An earlier Economic Policy Institute briefing paper, published in 2003, applied the method to deficits with Canada and Mexico combined and put displacement through 2002 at 879,280 jobs. The two figures illustrate the mechanics of the method as clearly as they illustrate the scale of the results. The earlier figure covered both partners and a shorter period yet exceeded the later figure, because the arithmetic follows the size of the bilateral balances fed into it and the coefficients applied to those balances, rather than following a measured change in tariffs or a count of observed separations. A year with a larger deficit yields a larger displacement figure under the method even if tariff rates did not change during that year, and a partner with a larger bilateral balance yields a larger figure than a partner whose trade affected a similar number of workers through other channels. The output moves with the balance, by construction.

The arithmetic must produce the largest numbers in the literature for reasons that follow directly from the steps. The method begins with the entire bilateral balance, measured in tens of billions of dollars, rather than with the portion of trade that changed because the agreement reduced tariffs. Any cause that widened the deficit, including faster American growth that raised import demand, an exchange rate movement that changed relative prices, or a shift in saving and investment between the two economies, enters the starting figure in full and then passes through the labor content coefficient into the job total. The coefficient itself is held fixed, so the calculation assumes that each million dollars of imports corresponds to the same amount of American labor in every year and at every scale, without allowance for productivity growth or for differences in the composition of imports beyond the categories the analyst employed. No stage reduces the total for workers who moved into export production, services, or other manufacturing after a displacement, because the method converts a balance rather than tracking workers. No stage reduces the total for the effect of lower import prices on household purchasing power or on the costs of producers that used imported inputs, because price effects have no place in a fixed coefficient conversion. The result is a gross measure in the strict sense, stated before any of the adjustments that the other method families build into their designs.

Administrative certification counts belong to the fourth family, and the family is introduced here only as a method, since the count itself and the program record behind it belong to a separate part of the evidence. Under the adjustment assistance program created for the agreement, groups of workers petitioned the Department of Labor, the department investigated each petition, and certification followed only where the investigation found that the statutory tests were met, including a connection to increased imports from either partner or to a shift of production to either partner. The method therefore counts adjudicated cases at petitioning firms. Every worker in the resulting total passed through a file, an investigation, and a determination, and no worker entered the total through a coefficient, a model solution, or an extrapolation from a sample. The same feature that gives the count its documentary solidity also defines its boundary. Workers who never petitioned do not appear, workers whose petitions were denied do not appear, and workers affected indirectly through the decline of a local economy do not appear unless a petition covered them and the investigation sustained it. The certification method yields a documented floor, a minimum that rests on case records.

Placed side by side, the four families form a coherent map of the evidence rather than a contest with a single winner. General equilibrium modelling answers the net national question, local labor market econometrics answers the distributional question, deficit accounting answers a gross conversion question, and certification counting answers an administrative question grounded in determinations made case by case. A national debate that mixes the four answers without naming the four questions manufactures disagreement where the underlying research describes different dimensions of the same episode.

The criticism of deficit accounting was direct, published, and specific. Economists at the Cato Institute published a rebuttal of the Economic Policy Institute model in 2011, the same year as the Scott briefing paper, and the objections recorded in the survey literature proceeded along three lines. The first objection concerned the starting point. In the position stated by the critics in the survey literature, a bilateral balance is determined by macroeconomic saving and investment flows and by exchange rates rather than by a trade agreement, so a deficit with Mexico could widen or narrow for reasons that had no connection to the tariff schedule of the North American agreement, and a method that converted the whole balance into displacement attributed every one of those movements to trade in goods. The second objection concerned the coefficient. The method holds labor content per dollar fixed, in the critics stated position, so the conversion cannot register productivity change, cannot register shifts in the composition of imports between goods that embody different amounts of labor, and cannot distinguish a dollar of imports that replaced domestic production from a dollar of imports that supplied an input to domestic production. The third objection concerned adjustment. The method does not model wage or price adjustment, in the critics stated position, so a fall in import prices produces no offsetting gain to consumers or to producers using imported inputs inside the calculation, and a change in wages or exchange rates produces no reallocation of production or employment. Taken together, the objections stated that the method answered a narrower question than its headline figures suggested, and that a gross conversion presented as a net employment loss overstated what the arithmetic could support.

Fair statement of the dispute also requires the position that defenders of the method would advance for the quantity being measured. In that position, deficit accounting captures gross displacement pressure in import competing production, the volume of domestic output and employment placed under direct competition by the flow of imports recorded in the bilateral balance, before offsets elsewhere in the economy are considered. A gross measure, in that position, has a legitimate use, because the pressure on a specific factory or industry is real even if the national economy later absorbs the affected workers elsewhere, and because net measures can render that pressure invisible by construction. The dispute therefore turns less on arithmetic, which any reader of the stated steps can reproduce, than on naming. A figure described as gross displacement pressure under fixed coefficients states what the method did. A figure described simply as jobs lost by the agreement claims a causal and net result that the method, as the Cato Institute rebuttal and the survey literature objections stated, was not built to establish.

The rule for reading every figure in the employment literature on the 1993 agreement follows from the four families and from the dispute over the third family in particular. The method behind a figure matters more than the figure. An estimate quoted without its method is not evidence, because the same trade records yield a near zero net welfare change under general equilibrium assumptions, a double digit wage growth gap for the most exposed dropouts under a local comparison design, hundreds of thousands of displaced job equivalents under a fixed coefficient conversion of a bilateral balance, and a documented floor under an adjudicated petition process. Each statement requires its author, its period, its population, and its method to carry any meaning at all. Stripped of those attachments, a number retains only its size, and size alone persuades without informing. Attached to its method, a number states a defined claim that evidence can support or contest on stated grounds.

Finding Four: The Certification Count, the One Number Nobody Modelled

Every other number in the debate over the North American agreement came out of a model, a conversion, or an extrapolation. The certification count came out of case files. When the implementing act created NAFTA Transitional Adjustment Assistance, it built a small administrative machine whose only task was to decide, group by group and plant by plant, whether particular American workers had lost their jobs for reasons the statute recognized as connected to trade with Mexico or Canada. Each decision left a record. Added together across the years in which the agreement-specific program operated, those records produced a total that no economist estimated and no advocate chose, because the total was simply the sum of adjudicated cases that the Department of Labor had certified between 1994 and 2002. That quality, derivation from decisions rather than from assumptions, gives the figure its standing among modelled welfare gains, measured wage gaps, and deficit conversions resting on contested arithmetic.

The program sat inside a much older structure. Trade Adjustment Assistance predated the trilateral agreement by decades, and the authority under which the general program operated derived from the Trade Act of 1974 guide, which had established the basic idea that workers displaced by import competition could receive a defined package of federal assistance. What the 1993 implementing act added was a component built for one agreement alone. NAFTA Transitional Adjustment Assistance, known by the shortened form NAFTA-TAA, was an agreement-specific component of Trade Adjustment Assistance, created for workers whose job loss was attributable to trade with Mexico or Canada under the new pact. Eligibility turned on the source of the trade pressure. A worker group displaced by imports from outside North America had no claim under the agreement-specific component, however severe the displacement, while a group displaced by imports from a partner country, or by an employer’s decision to move production across a partner border, fell squarely within the design.

The statutory tests were stated in causal terms, and the causal language did real work. A group qualified where increased imports from Mexico or Canada contributed to the separations, or where production had shifted from an American workplace to Mexico or Canada. The first captured the factory that lost orders because a competing good crossed the border under the new tariff schedules. The second captured the factory whose own output crossed the border in the opposite direction, as product lines and jobs relocated to a partner country while finished goods returned to American buyers. Requiring one of those channels to be shown in each case kept the program from becoming a general remedy for industrial decline. A plant closing because of domestic competition, obsolete equipment, changing consumer taste, or a recession fell outside the tests unless the investigation connected the loss to one of the two trade channels the statute named. The tests therefore filtered the caseload toward losses with a demonstrable partner-country link, and the resulting count inherited that filter in full.

The process began with a petition. Worker groups petitioned the Department of Labor, and the department investigated before certifying eligible groups. No worker entered the count through a survey response, a statistical sample, or an economist’s coefficient. Entry required a filing, an official inquiry into the circumstances of the separation, and a determination that the statutory tests had been met on the facts of that workplace. Investigators had to satisfy themselves about what had happened at a particular firm: whether imports from a partner country had increased, whether production had shifted to a partner country, and whether the separations followed from that trade development. Certification was the outcome of that determination, and it unlocked the assistance the program existed to deliver. Certified workers gained access to income support, training, and relocation allowances, the three elements that the adjustment assistance idea had long combined: money to bridge the loss of wages, preparation for different employment, and help with the cost of moving where work existed.

Coverage reached beyond the plant gate in one defined direction. Workers at supplier firms secondarily affected were covered under a commitment in the Statement of Administrative Action, the document through which the administration set out its understanding of how the implementing legislation would operate. The inclusion recognized a basic feature of industrial production: a certified plant rarely stood alone, and firms that supplied components, materials, or services to a plant undone by partner-country imports could lose orders and shed workers as a consequence of a displacement that had formally occurred elsewhere. Secondary coverage extended the program’s reach along that supply relationship, although the extension had limits that the interpretation of the count must keep in view, because coverage of defined supplier workers under an administrative commitment was never the same as coverage of every downstream consequence in a local economy.

The program had barely begun operating before official Washington started measuring it. The General Accounting Office published “Dislocated Workers: An Early Look at the NAFTA Transitional Adjustment Assistance Program” in November 1994, less than a year into the agreement’s life. An evaluation issued at that stage could examine only the opening months of a program whose caseload depended on petitions filed, investigated, and decided after separations had occurred, so the count available then was necessarily a fragment. The fragment grew quickly all the same. Contemporary program reporting recorded that more than 64,709 workers were certified from January 1994 through April 1996, a little over two years into the program’s operation. That early subtotal already exceeded the scale of a pilot effort, and it accumulated during a period complicated by forces outside the agreement, since Mexico’s peso crisis of December 1994, when the incoming administration of President Ernesto Zedillo devalued the currency and then allowed it to float, convulsed trade flows across the border for reasons that belonged to Mexican macroeconomic management rather than to the tariff schedules.

What does certification prove, and what does it not prove?

Certification proves that a Department of Labor investigation found a defined worker group eligible under statutory tests tying its job loss to imports from, or production shifts to, Mexico or Canada. It does not prove total displacement, net employment loss, or causation for workers outside adjudicated petitions.

The distinction drawn in that answer organizes everything else the record can support. Against the background of the early subtotal, the full operating-period record accumulated year after year as petitions moved through investigation and determination. Hufbauer and Schott, in NAFTA Revisited: Achievements and Challenges, published in 2005 by the Peterson Institute for International Economics, compiled the Department of Labor certification records for 1994 through 2002 and reported about 525,000 workers certified across that period, with the compiled record totaling 525,094, an average of about 58,000 per year. The precision of the compiled total and the roundness of the reported figure describe the same body of records at different levels of presentation. Stated with that attribution, Department of Labor records for 1994 through 2002, compiled in Hufbauer and Schott, NAFTA Revisited, 2005, the number rests on administrative records rather than on anyone’s simulation of an economy that never existed, and a reader who doubts the figure can in principle return to the underlying certification decisions from which the compilation was built.

A second account of the same record, delivered in a congressional setting, corroborated the scale while adding composition. Sandra Polaski, a former Department of Labor official, testified before a Senate subcommittee on September 11, 2006 that about 500,000 workers were certified across 1994 through 2002. Polaski’s testimony carried the record further than a bare total could go. About half of the certified job losses, she testified, involved production shifts to Mexico, which means that the relocation channel, the mechanism that had dominated the ratification argument in 1993, accounted for roughly one in every two certified cases across the operating period, with the remainder arising through the import channel and shifts connected with the partner countries under the statutory tests. The testimony also distributed the certified workers across industries. Apparel accounted for about 28 percent of certified workers, electronics for 13 percent, automobiles and parts for 7 percent, and fabricated metals for 6 percent. Those four shares together cover a little over half of the certified total, leaving the rest spread across the wider manufacturing economy and beyond.

The industry composition matches, in broad outline, the structure of exposure that the rest of the research literature identified through entirely different methods. Apparel’s position at the head of the list fits an industry built on labor-intensive production, thin margins, and direct competition with lower wage producers, where tariff reductions could move orders across a border quickly. Electronics reflects an industry already organized around cross-border supply chains in which assembly stages migrated readily. Automobiles and parts and fabricated metals represent industries where continental integration had been proceeding since long before 1994, where capital intensity moderated the speed of relocation, and where displacement nonetheless registered in thousands of certified cases. Certification clustered where the statutory channels bit hardest, and the clustering is consistent with the finding elsewhere in the evidence that the agreement’s effects were modest in aggregate and severe in place.

The production-shift finding performs similar disciplinary work on the political memory of the agreement. During the ratification debate, relocation of American production to Mexico had been the most vivid prediction in circulation, crystallized in the warning of a giant sucking sound that Ross Perot had issued during the 1992 campaign. The aggregate modelling literature later found net effects for the United States that were small by any standard, but the certification record speaks to a more concrete question: in cases that officials actually adjudicated, how often did the file show production moving to Mexico? Polaski’s answer, about half of the certified job losses, confirms that relocation was not a marginal curiosity within the caseload but one of its two great channels, roughly coequal with displacement through increased imports. That confirmation neither vindicates the largest claims made in 1993 nor supports the claim that relocation fears were invented. Within the documented population of certified workers, the mechanism at the heart of the controversy operated at scale, year after year, in cases where investigators found the statutory test satisfied on the facts of the workplace.

The count’s period has a clean legal boundary, and the boundary matters for any use of the figure. The separate, agreement-specific program did not run indefinitely and did not fade away through administrative neglect. The Trade Act of 2002, Public Law 107-210, merged NAFTA-TAA into the general Trade Adjustment Assistance program and repealed the NAFTA-specific subchapter, as the Congressional Research Service recorded in report R44153. From that merger forward, workers displaced by trade with the partner countries sought assistance through the general program, and no separate agreement-specific certification series continued to accumulate. The operating period of the agreement-specific count is therefore 1994 through 2002, nine years that begin with the pact’s entry into force and end with the statutory consolidation of 2002. A total drawn from 1994 through 2002 cannot silently absorb later displacement, cannot be extended across the whole life of the agreement, and cannot be compared against estimates built for other periods without adjustment.

Interpretation begins where the arithmetic ends, and the count demands more interpretive care than its concreteness might suggest. The first discipline concerns what the machinery could not see. Certification is an administrative count of adjudicated cases at petitioning firms, and each word in that description marks an exclusion. Workers who never petitioned do not appear, whatever happened at their workplaces. Workers whose petitions were denied likewise stand outside the total, and denial could reflect the strictness of the statutory tests as readily as the absence of trade pressure, since a file that failed to establish increased imports from a partner country or a shift of production to a partner country failed regardless of how trade-exposed the industry felt to the workers inside it. Most downstream effects escape the count altogether. Supplier coverage under the Statement of Administrative Action reached defined secondarily affected workers, but the diner that lost customers when a plant closed, the municipal tax base that thinned as wages left a town, and the second round of layoffs in local services as spending contracted never passed through a petition, and none of that stands in the 525,094.

Those exclusions make the figure a floor rather than a total, and the floor character of the count should govern every sentence that deploys it. A floor states that documented displacement reached at least this level under a demanding evidentiary process, and it makes no claim about how much higher the full figure might run. Honest use of the number therefore moves in one direction only. The count can refute a claim that documented displacement under the pact was trivial, because half a million certified workers across nine years, each attached to a decided case, cannot be described as trivial without emptying the word of meaning. The count cannot refute a claim that total displacement exceeded the certified figure, since the program’s own design guarantees undercounting. Arguments that treat 525,094 as the whole of the agreement’s displacement mistake an administrative intake for a census, while arguments that dismiss the figure because larger totals exist elsewhere discard the one total that was built from examined cases.

A second discipline concerns the difference between gross and net. The certification count is not a net employment estimate and must never be described as one. Certification recorded workers displaced in cases meeting the statutory tests; it recorded nothing about employment created elsewhere in the economy, nothing about jobs supported by exports to the partner countries, nothing about positions that arose in logistics, distribution, or services as continental trade expanded, and nothing about the reemployment of the certified workers themselves after training or relocation. Net employment effects belong to the general equilibrium literature, whose models reallocate labor across the economy by construction and which found, in the work surveyed elsewhere in the evidence, aggregate effects for the United States that were small. The modelled net figure describes an economy adjusting as a whole, with gains and losses offsetting inside the arithmetic, while the certification count describes a documented population of losers whose losses were real regardless of gains accruing to other workers in other places. An economy can register a small net effect and still contain 525,094 certified displacements, and the evidence assembled across the five findings indicates that the American economy under the trilateral agreement did exactly that.

Set against the other method families in the literature, the certification count occupies a position that none of the others can take from it. General equilibrium models estimate economy-wide net effects and cannot show displacement, because their structure reallocates workers rather than following them. Local labor market econometrics measures relative effects on exposed groups and places with great sensitivity, expressing harm as slower wage growth across a decade rather than as a roster of the displaced. Deficit accounting converts a bilateral balance into jobs through fixed labor-content coefficients and produces the largest figures in the debate, figures whose method drew the standard objection recorded in the survey literature, pressed in the Cato Institute’s published 2011 rebuttal of the institute model, namely that a bilateral balance reflects macroeconomic saving and investment flows and exchange rates rather than the terms of a trade agreement. Officials read files, applied a statute, and counted the workers in the groups that qualified. That is why the figure is the most defensible single number in the article: every worker in the total corresponds to a determination made under law on facts investigated by the department charged with the program, a chain of custody from workplace to number that no modelled estimate can replicate.

Defensibility is not completeness, and the closing discipline is to hold both properties together. The certification record establishes a documented minimum of displacement attributable to the agreement under statutory tests, distributed across industries in proportions that testimony preserved, driven about half by production shifts to Mexico in the cases adjudicated, sustained at an average pace of about 58,000 workers a year across nine years, and terminated as a separate series by a statutory merger that had nothing to do with the underlying trade flows. It does not establish the agreement’s net effect on American employment, the full count of workers harmed by partner-country trade, or the experience of communities where displacement propagated beyond petitioning firms. Used within its limits, the figure anchors the evidence, because every challenge to it must take the form of a challenge to thousands of individual determinations rather than to a set of assumptions. Used beyond its limits, it becomes merely another weapon in the argument it was meant to discipline. The record itself, more than 64,709 workers certified from January 1994 through April 1996 and about 525,000 across the operating period, supports a restrained conclusion: displacement under the pact was documented, substantial, concentrated, and adjudicated, and it coexisted with aggregate effects that the modelling literature found small. Both halves of that sentence rest on evidence, and the certification count is the half that nobody modelled.

Finding Five: The Mexican Side of the Ledger

An account of what the North American agreement did to employment in the United States remains incomplete if it stops at the border. The pact bound three labor markets together, and the pressures that moved American workers had counterparts, often larger and often harsher, on the other side of each frontier. Mexico entered the agreement as the partner with the most at stake. Its economy was the smallest of the three, its tariff reductions were the deepest, and its exposure to American competition was the most direct. The Mexican record therefore tests the agreement in its strongest form. Where the effects on the United States were small in aggregate and concentrated in particular places, the effects on Mexico touched whole sectors of the countryside and whole districts along the northern border, and they arrived entangled with a financial crisis that struck within the first year and that no honest reading of the period can separate from the agreement by simple inspection. The evidence for Mexico divides into three large subjects, the crisis that overwhelmed the opening years, the transformation of agriculture under liberalized grain trade, and the expansion and partial contraction of border assembly, with a shorter record for Canada and a contested one for migration across the border itself.

Why do the agreement’s first years resist clean measurement?

A currency crisis struck Mexico in December 1994, less than a year after the agreement entered into force. The peso devaluation, a 6.2 percent contraction in 1995, and inflation of about 52 percent overwhelmed any early tariff effects and distorted trade flows, so effects dated 1994 to 1996 cannot be assigned cleanly.

The sequence of the crisis deserves statement in full, because the order of events determines what any figure from those years can mean. In December 1994 the incoming administration of President Ernesto Zedillo devalued the peso and then allowed it to float. A managed exchange rate gave way, within the first twelve months of the trilateral agreement, to a currency whose value moved freely and fell hard. The contraction that followed was severe by any standard applied to a national economy. Mexican gross domestic product fell 6.2 percent in 1995, inflation ran about 52 percent in 1995, and unemployment peaked at 7.6 percent in August 1995, figures reported by the International Monetary Fund in Finance and Development in June 1998. Each of those numbers describes a macroeconomic shock, not a trade shock. A fall in output of that size reduces employment across construction, services, retail trade, and domestic manufacturing alike, in sectors the agreement touched lightly and in sectors it never touched at all. Inflation of about 52 percent rewrote every price in the economy, including the relative prices between imported and domestic goods that tariff reductions were supposed to adjust gradually. Any attempt to read a tariff effect out of the August 1995 unemployment peak would mistake the largest movement in the economy for one of the smallest.

The external response confirmed the scale of the emergency. A United States led support package of about 50 billion dollars followed the devaluation, an intervention sized for a financial crisis rather than for a trade adjustment. Recovery, when it came, was correspondingly broad. Growth resumed at 5.2 percent in 1996 and 7.0 percent in 1997, rates that reflected an economy rebounding from a deep contraction rather than an economy responding in measured steps to scheduled tariff reductions. A rebound of that character lifts sectors in proportion to their fall, on a timetable set by finance and demand rather than by the phase out schedules written into the agreement, which ran as long as fifteen years for the most sensitive agricultural goods.

The trade figures of the crisis period carry the same warning in a sharper form. Mexico ran a trade surplus with the United States in 1995, and a reader approaching that surplus without the crisis in view might treat it as evidence about the agreement, as though Mexican producers had gained an early advantage under the new rules or American producers had lost ground to liberalization. The surplus was a crisis artifact. The devaluation and the contraction crushed Mexican import demand, because households and firms with collapsing incomes and a depreciating currency buy less from abroad, and the bilateral balance moved for that reason rather than because the agreement had rearranged production. The direction of the balance therefore proves nothing about the pact. For that reason the discipline applied throughout the Mexican record is categorical. No effect dated 1994 to 1996 can be cleanly assigned to the agreement. Output, employment, prices, and trade flows in those years were governed by the devaluation, the contraction, the support package, and the recovery, and the tariff changes of the pact were a minor current running beneath a flood. Any study, any political claim, and any employment figure that treats the opening years as a clean test of the trilateral agreement has failed at the first step, whatever its method and whatever its conclusion.

The countryside was, on the evidence, where the agreement pressed hardest on Mexican livelihoods. The central commodity was maize, the staple grain of the Mexican diet and the crop around which small scale farming across much of the country was organized. Under the liberalization the agreement set in motion, United States corn exports to Mexico roughly quadrupled across the first decade, and imports rose from about 7 percent of Mexican corn consumption to about 34 percent. A change of that proportion alters the position of every domestic producer, because imported corn supplying about 7 percent of Mexican corn consumption exerts a different discipline from imported corn supplying about 34 percent. As the import share rose, imported supply disciplined the domestic price continuously, and the Mexican grower competed each season against the delivered cost of American corn. The quadrupling of exports therefore describes more than a trade flow. It describes a transfer of price setting power from the Mexican harvest to the international market, and with it a transfer of risk onto producers who had the least capacity to absorb a run of low prices.

The terms on which that competition took place were themselves contested, and the most developed research on the question found that the competition was not conducted at full cost. Timothy A. Wise of the Global Development and Environment Institute at Tufts University, in Agricultural Dumping Under NAFTA research published in 2010, found that American corn was exported to Mexico at an average dumping margin of 19 percent across 1997 to 2005, meaning that the corn was sold below the cost of production. That finding changes the character of the rural story. Displacement by a more efficient foreign producer is one kind of event, with one kind of policy response. Displacement by imports priced below their own production cost is another kind, because the Mexican producer in that case was not losing a fair contest on productivity and was not being asked to match an attainable standard. The producer was competing against a price that the exporting system itself did not cover. Wise estimated that some 3 million Mexican corn producers were affected, and that estimated losses to Mexican corn producers totaled 6.6 billion dollars over nine years. Nor was corn an isolated case. Wise estimated losses across eight products at 12.8 billion dollars for 1997 to 2005, which places the maize experience inside a broader pattern in which the basic food and feed crops of Mexican agriculture faced import competition priced on the same terms.

The Mexican state held an instrument that might have moderated that pressure and declined to use it. The agreement permitted a tariff rate quota on corn imports, a device that allows imports up to a set quantity on liberalized terms and applies a protective charge beyond that quantity, and Mexico never enforced the tariff rate quota the agreement permitted. The omission matters because it removes a possible defense of the outcome based on the text of the pact. The hardship of the corn producing regions cannot be explained entirely as the unavoidable consequence of commitments Mexico had no means to soften, because a softening mechanism existed within the commitments and went unused as import volumes climbed far beyond the levels of the early 1990s. In place of border protection, Mexican policy offered Procampo payments, the transition support provided to producers as the substitute for the shield that the quota would have supplied. A payment of that kind supports income without restoring the market price, and the distinction governed the lived experience of liberalization in the countryside. A farmer receiving a transition payment while the price of the crop fell toward the imported level was being compensated, in part, for a loss that the payment did not prevent, and the compensation arrived as a fiscal transfer dependent on the budget and the administration of the program rather than as earnings from the harvest itself.

Employment in Mexican agriculture contracted across the period on a scale that matched the severity of the price pressure, though the attribution of that contraction requires the full qualification its source attached to it. Mexican agricultural employment fell from about 8.1 million in the early 1990s to about 6 million in the first quarter of 2006, a decline of over 2 million workers. Sandra Polaski, a former Department of Labor official, stated the figure in testimony before a Senate subcommittee on September 11, 2006, and Polaski attached an explicit caveat that must travel with the number wherever the number goes. Not all of the decline can be attributed to the agreement, and the peso devaluation pushed agricultural trade in the opposite direction, making Mexican produce cheaper in foreign markets and imports dearer at home during the crisis years even as tariff reductions worked toward greater import penetration over the longer term. The caveat is not a formality. Agricultural employment in a developing economy declines for reasons that include mechanization, urbanization, the pull of higher wages in towns and in border industry, and the push of exhausted land and subdivided holdings, and those forces operated in Mexico through the 1990s and into the following decade alongside the trade changes. A decline of over 2 million therefore records the scale of the rural transformation without settling the share of that transformation for which the pact was responsible. Honest use of the figure keeps both halves of the statement together, the size of the fall and the refusal of its own source to assign the whole fall to the agreement.

Against the losses in grain and oilseed country stood one genuine offset, and its limits are as instructive as its existence. Exports of Mexican fruits and vegetables to the United States grew across the period, a result Polaski also recorded, as Mexican horticulture exploited the same open border that admitted American corn, selling produce into the American market in the seasons and varieties where Mexican growers held the advantage. It did not, however, keep pace with imports of American grains and oilseeds, in Polaski’s assessment, and the imbalance between the two flows captures the asymmetry of agricultural liberalization between the two countries. The expanding Mexican export sector was horticultural, concentrated in particular regions and seasons, and capable of employing only a fraction of the labor that grain farming had supported across the whole countryside, while the expanding American export sector was the staple crop sector itself, reaching into every corn buying market in Mexico. A partial offset that grows more slowly than the pressure it offsets leaves the sector smaller at the end of the process than at the beginning, and that is the shape of the agricultural ledger under the pact, substantial displacement in the staple economy, genuine growth in a narrower export economy, and a net rural contraction whose precise division among causes its most careful witnesses declined to fix.

The second great Mexican subject is the border assembly industry, the maquiladora sector, and its record under the agreement runs in the opposite direction from the countryside for most of the period, though the sector was neither created by the pact nor immune to forces larger than the pact. The border assembly program dates from 1965 and predates the agreement by almost three decades. The baseline is a matter of record. Employment in the sector stood at 546,600 in December 1993, according to INEGI, the Mexican statistics institute, on the eve of entry into force. From that substantial base, the sector expanded powerfully during the first seven years of the agreement. Polaski testified in 2006 that the sector added about 800,000 jobs between January 1994 and its peak in early 2001, and the sector reached about 1.3 million workers in 2000.

The agreement did contribute to that expansion, but the contribution was specific rather than general, and Polaski’s testimony identified the channel. The tariff cuts of the pact stimulated the apparel subsector of the maquiladoras specifically, because clothing production sat at the intersection of the rules the agreement rewrote, and preferential access to the American market for garments assembled under those rules gave border plants a reason to expand that they had not possessed in the same form before 1994. That specificity is the accurate measure of the pact in the border economy. The agreement did not invent assembly production in Mexico, which had operated since 1965 under its own program and its own logic of wage differentials and proximity to the American market. The agreement sharpened particular incentives, apparel above all, inside an established industry, and it did so during years when the devalued peso had also made Mexican labor cheaper in dollar terms for reasons that owed nothing to trade law. Attributing the full 800,000 job expansion to the pact would repeat, in a favorable direction, the same error that attributing the full agricultural decline to the pact commits in an unfavorable direction. In both cases the agreement was one force among several, strongest where its tariff changes bit directly, weakest where macroeconomic conditions and older programs were doing the work.

The contraction that followed the peak demonstrates the same point from the other side. After reaching about 1.3 million workers in 2000 and a peak in early 2001, the border plants shed about 125,000 jobs through January 2006, under pressure from the United States recession of 2001 and from competition with China. Neither cause was a provision of the North American agreement. A recession in the destination market reduces orders for assembled goods regardless of the tariff treatment those goods receive, and Chinese competition pressed the sector from outside the trilateral framework altogether, offering buyers an alternative assembly platform whose rivalry no clause of the 1993 text governed. Polaski assigned the shedding of those jobs to the recession and to Chinese competition rather than to the agreement, and the assignment carries a lesson that applies to the whole employment debate on both sides of the border. The maquiladora sector rose under the pact without the pact explaining the rise, and it fell after 2001 without the pact explaining the fall.

Canada presents the simplest case of the three, and its simplicity is itself evidence about how trade agreements produce employment effects. Liberalization between Canada and the United States was largely complete under the bilateral free trade agreement of 1988, which had already removed the principal barriers between the two economies six years before the trilateral agreement entered into force. The marginal change for Canada under the North American pact was therefore the smallest of the three partners, because the larger part of what a free trade agreement does to tariffs, investment rules, and market access between Canada and the United States had already been done. A second agreement layered over a completed liberalization has little left to move. The modelled result matches that logic exactly. Lorenzo Caliendo and Fernando Parro, in Estimates of the Trade and Welfare Effects of NAFTA, published in the Review of Economic Studies in 2015, modelled a welfare change for Canada of minus 0.06 percent, a figure that belongs in the category of rounding difference rather than in any account of national gain or loss. Canada’s stake in the trilateral text ran instead through access to Mexico and through the investment and dispute chapters, subjects that concerned the terms of participation in a continental framework rather than tariff change with the United States.

The last Mexican question is the one that connects the countryside to the border and to the United States itself, and it is the question on which the evidence permits the least certainty. If agricultural liberalization and the crisis years displaced rural workers on the scale the employment figures suggest, some share of those workers moved, to Mexican cities, to the border assembly districts, and across the northern frontier. Philip Martin of the University of California, Davis, writing in 2005, estimated that the unauthorized Mexican population in the United States rose from 2.5 million in 1995 to 4.5 million in 2000, an increase of 2 million people in five years. That increase is a demographic fact of the first order, and it coincided with the years in which corn imports were climbing toward a third of Mexican consumption and the rural economy was contracting. Coincidence, however, is not attribution, and the research record treats the connection as genuinely disputed. How much of the rural outflow the agreement caused, as against the peso crisis and longer trends, remains unsettled among researchers, and the dispute is substantive rather than rhetorical. The crisis of 1995 destroyed urban and rural livelihoods alike and gave millions of Mexicans a reason to move that had nothing to do with tariff schedules. Longer trends in Mexican demography, in the established networks linking particular Mexican communities to particular American destinations, and in American demand for immigrant labor all predated the agreement and continued through it, and a flow fed by established networks acquires momentum of its own as each settled migrant lowers the cost and risk of the journey for others. The careful conclusion is therefore a limited one. The pact plausibly contributed to the pressures that moved rural Mexicans, through the channel of import competition in the staple economy, at the same time as a currency crisis and older structural forces were moving the same population in the same direction, and no figure in the record separates those contributions with confidence. Martin’s estimates establish the size of the movement. They do not, and their presentation does not pretend to, establish the share of the movement for which the 1993 agreement was responsible.

Taken together, the Mexican, border, and Canadian records return the employment debate to the proportion that the aggregate research established for the United States. The agreement was a secondary force where economies were already open to each other and where macroeconomic shocks intervened. The agreement was a primary force where it exposed a large, vulnerable producing population to import competition on terms that included pricing below production cost, as in the Mexican maize economy, and where its tariff changes engaged a specific industry directly, as in border apparel. A ledger with entries of that size on both sides describes neither a triumph nor a catastrophe in the simple forms that political argument preferred. It describes a reallocation across a continent, heavily concentrated in its costs, uneven in its benefits, and confounded at its origin by a crisis that arrived within a year and that governs the reading of everything the agreement did first.

The NAFTA Employment Impact Evidence Table

The five-finding evidence table gathers the central results of the assessment in a single artifact, so that a reader can compare the findings without detaching any result from the research that produced the result. Each finding carries a settled or contested rating, and the rating describes the state of the research record for that finding rather than the size of the effect the finding reports. A settled rating marks a result that the cited research established within the limits of the method stated, while a contested rating marks a result whose scale or causal reading remained in dispute among the researchers who examined the same episode.

Finding	Direction	Period	Principal research	Method	Rating
Aggregate welfare and employment effects	Small welfare gain for the United States, negligible net employment effect	1994 to 2009	Caliendo and Parro (2015); Burfisher, Robinson and Thierfelder (2001)	General equilibrium modeling; literature survey	Settled
Wage effects on exposed workers	Wage growth losses concentrated on workers without high school qualifications in exposed industries and localities	1990 to 2000	Hakobyan and McLaren (2016)	Local labor market econometrics	Settled
Job displacement from the bilateral deficit	Large claimed displacement, reaching 682,900 jobs	1994 to 2010	Scott, Economic Policy Institute (2011)	Bilateral-deficit accounting	Contested
Workers certified as trade-displaced	525,094 workers certified	1994 to 2002	Department of Labor records, compiled in Hufbauer and Schott (2005)	Administrative count	Settled
Mexican agriculture and border assembly	Farm-sector losses alongside large assembly employment gains	Corn studies 1997 to 2005; assembly record 1993 to 2006	Wise (2010); Polaski testimony (2006); INEGI statistics	Sector studies and administrative data	Contested

Effects the Drafters Did Not Intend

A tariff schedule describes what a trade agreement changes at a border, and the North American agreement changed less at that border than the scale of the later argument suggests. What the schedule does not describe is the structure firms build after the percentages fall, because firms respond to durable access by reorganizing production itself. The deepest unintended effect lay in that reorganization. Burfisher, Robinson, and Thierfelder (2001) found that trade creation greatly exceeded trade diversion, especially in intermediate goods: components and materials that cross a border for further work rather than for final sale. Concentration in that category bound factories together across borders instead of simply opening three finished goods markets to one another.

That integration carried consequences a reading of tariff lines alone would miss. Once production was divided into cross-border stages, a disruption in one country reached employers and workers in the others through orders, inventories, and component supply rather than through consumer prices alone. Trade within the bloc deepened, and two-way trade grew several-fold in nominal terms across the first decade. The older picture of separate national industries competing at arm’s length described the economy less well as the phase-out proceeded, and later evaluation grew harder because the units measured were no longer cleanly national.

The second unintended effect ran toward a promise that reached fewer people than its design language suggested. The implementing act created an agreement-specific adjustment assistance component for workers whose job loss was attributable to trade with Mexico or Canada, through increased imports or a shift of production. Workers petitioned the Department of Labor, the department investigated, and certification unlocked income support, training, and relocation allowances. Assistance followed adjudication rather than displacement itself, and the burden of initiation rested on the workers and firms concerned. Those who never petitioned, whose petitions failed, or whose local businesses lost customers after a closure stood outside the certified rolls.

The compiled record gives that subset a scale and a boundary: about 525,000 workers were certified across 1994 to 2002, and the Trade Act of 2002 closed the agreement-specific count by merging the separate program into the general Trade Adjustment Assistance program.

Certification deserves its standing as the anchor number precisely because of its limits. It counts adjudicated cases at petitioning firms only, excluding workers who never petitioned, workers whose petitions were denied, and most downstream effects, and it is not a net employment estimate. Its strength is that nobody modelled it: each worker stood behind a petition, an investigation, and a determination under the statutory tests. The adjustment promise underdelivered through coverage, since the remedy reached a documented subset of the displaced rather than a census of all whose earnings the tariff reductions had disturbed.

The third unintended effect crossed the southern border and returned through migration. United States corn exports to Mexico roughly quadrupled across the first decade, and imports rose from about 7 percent of Mexican corn consumption to about 34 percent. Wise (2010) estimated American corn was exported at an average dumping margin of 19 percent across 1997 to 2005, meaning below the cost of production. Mexican agricultural employment fell sharply across the period, with the caveat Polaski attached: not all of that decline can be attributed to the agreement, and the peso devaluation pushed agricultural trade in the opposite direction.

The peso crisis supplies the confound that disciplines every claim about the first years. In December 1994 Mexico devalued the peso and then allowed it to float, and Mexican gross domestic product fell 6.2 percent in 1995. Mexico ran a trade surplus with the United States in 1995 because the crisis crushed Mexican import demand, an artifact of the crisis rather than an effect of the agreement. Against that background, rural displacement added to migration pressure. Martin (2005) estimated that the unauthorized Mexican population in the United States rose from 2.5 million in 1995 to 4.5 million in 2000. How much of that outflow the agreement caused, as against the peso crisis and longer trends, is genuinely disputed among researchers.

What the Statute Plainly Did Not Accomplish, and Why the Argument Never Ends

Measured against its sponsors’ largest claims, the 1993 agreement plainly did not transform aggregate American employment in either direction. Caliendo and Parro (2015) built a computable general equilibrium model and estimated small welfare effects across the bloc. Burfisher, Robinson, and Thierfelder (2001) likewise found much larger relative benefits for Mexico and little effect on the United States labor market. A welfare gain of eight hundredths of one percent is a real gain in a model and a negligible event in a national labor market of continental scale.

The smallness of the aggregate result was built into the starting conditions. Trade with the two partners was already substantial before 1994, American tariffs on Mexican goods were already low, and much Mexican output entered duty free under preference programs and the maquiladora regime. Mexico had liberalized unilaterally from 1985, and Canada-United States trade already operated under the bilateral agreement of 1988. An agreement that lowers barriers already lowered in large part, between economies already trading heavily, changes relative prices only at the margin, producing the small aggregate welfare numbers the models returned.

Convergence between Mexican and American living standards on the timetable its sponsors described was the second accomplishment the record does not show. The Mexican welfare gain in the Caliendo and Parro (2015) model was the largest in the bloc and far larger in percentage terms than the American gain, yet a gain of that size does not close a living standards gap between economies at very different income levels on that timetable. The border assembly sector illustrates both the real gains and their limits: the maquiladora program dated from 1965, predated the agreement, and reached about 1.3 million workers in 2000, and the tariff cuts stimulated the apparel subsector specifically. Plants then shed jobs through January 2006 under pressure from the United States recession of 2001 and from competition with China.

Compensation of the workers whom the wage research identifies was the third failure, following from a mismatch between where the losses fell and where the remedy stopped. Hakobyan and McLaren (2016) measured exposure by prior tariff protection. Wage growth for high school dropouts in the most protected industries was up to 17 percentage points lower across 1990 to 2000 than for comparable workers in initially unprotected industries. Local service workers in exposed places were also affected through local demand, effects for college-educated workers were small and statistically insignificant, and the authors found the NAFTA effect separate from import competition from China, which the design controlled for. The losses therefore fell on a group wider than the certified rolls, and a remedy keyed to certification could not reach the full group the research describes.

Chronology disciplines the rest of the argument, because the larger shock arrived after 2000 under a different statute. Permanent normal trade relations with China, enacted in 2000, and Chinese accession to the World Trade Organization in December 2001 opened the decade in which Chinese import competition pressed hardest on the same localities and industries that dominate manufacturing decline narratives. Hakobyan and McLaren’s design, built on 1990s data and controlling for import competition from China, separates the two shocks: an agreement signed in 1993, in force from January 1, 1994, with tariff phase-outs running as long as fifteen years for the most sensitive agricultural goods, cannot carry sole responsibility for a deterioration whose sharpest phase followed a separate grant of permanent normal trade relations and a separate accession, a sequence examined in the account of the larger concurrent trade shock from China. Assigning two decades of manufacturing decline to the 1993 act misreads the sequence and flattens two distinct policy decisions into one.

The mandatory counter-reading cuts against both political uses of the agreement, and an honest account disappoints both. The devastation claim fails against the aggregate literature. Its largest numbers come from deficit accounting, a gross conversion of a bilateral balance through fixed labor-content coefficients: Scott (2011) put displacement at 682,900 jobs across 1994 to 2010 on that method. Economists at the Cato Institute rebutted that model in 2011 on the grounds, recorded in the survey literature, that a bilateral balance reflects macroeconomic saving and investment flows and exchange rates rather than a trade agreement, that labor content per dollar is held fixed, and that wage and price adjustment goes unmodeled. That criticism stands as the critics’ position. The four method families each see a different object: general equilibrium models estimate net effects and reallocate labor by construction, so they cannot show displacement; local labor market econometrics estimate relative effects on exposed groups and places; deficit accounting produces the largest numbers; and administrative certification counts adjudicated cases only and supplies a documented floor.

The costless claim fails just as plainly against the other half of the record. A gain of 0.08 percent in American welfare in the Caliendo and Parro (2015) model says nothing about the distribution of the adjustment underneath it, and the distribution is where the costs lived. Wage growth penalties of up to 17 percentage points for high school dropouts in the most protected industries, measured across 1990 to 2000 by Hakobyan and McLaren (2016), are lasting losses for the workers who absorbed them. About 525,000 certified workers across 1994 to 2002, an administrative floor assembled case by case, cannot be reconciled with a description of the agreement as free of cost, and neither can a caseload in which production shifts to Mexico accounted for about half of the certified losses. Read across its methods, the record supports modest national effects and concentrated harm at the same time.

Why the politics ran so much larger than the measurements lies in the asymmetry between the two kinds of effect. Gains from the agreement arrived diffuse and largely invisible. Hufbauer and Schott (2005) estimated consumer gains of about 5.5 billion dollars per year across the agreement’s first nine years, and gains of that character spread across millions of purchases in slightly lower prices. Losses arrived concentrated and visible. A certified production shift has a plant, a gate, a final shift, a local tax base, and several hundred named workers in a single locality, and certification generated a public record of plant-level loss with no equivalent record of plant-level gain. Ross Perot’s warning of a “giant sucking sound,” dating from the 1992 campaign, fixed the employment question at the center of the 1993 ratification contest.

Concentrated and visible losses outweigh diffuse and invisible gains, because political organization follows identifiability: workers who lose a plant know what happened and whom to petition, while consumers who gain a few dollars have no event to organize around and no agreement to credit. Local losses also persist in place in a way that national averages erase, since a locality that loses its anchor employer loses the local service demand that Hakobyan and McLaren (2016) detected in exposed places, spreading the wage penalty to workers who never made an import-competing product. That persistence explains why the argument never ends even as the measurements settle. Each retelling can select a true figure, the modelled welfare gain, the certified half million, or the wage penalty, and each selection produces a different statute in the public memory while citing real research.

The succession belongs in the record as a dated fact and nothing more. The United States-Mexico-Canada Agreement was implemented by Public Law 116-113, signed January 29, 2020, and the successor agreement entered into force on July 1, 2020. No finding in the research surveyed here, covering an agreement in force from January 1, 1994 and studies whose periods end years before the succession, is extended to the successor agreement. The closing judgment therefore returns to the measurement the record sustains across its method families, from the general equilibrium models of Caliendo and Parro to the local wage estimates of Hakobyan and McLaren to the certification compilation of Hufbauer and Schott: the North American agreement was small in aggregate, severe in place.

Closing Assessment: Small in Aggregate, Severe in Place

The judgment that the evidence supports has a name, and the name can be stated plainly: the North American agreement was small in aggregate, severe in place. Small in aggregate describes the national accounts. Severe in place describes the payrolls, towns, and countrysides where the remaining tariff protection had stood before 1994 and where its removal landed. Neither half corrects the other half. The aggregate half prevents a local catastrophe from being read as a national one, and the distributional half prevents a small national average from being read as proof that the adjustment cost little to the people who absorbed the adjustment.

On the national question, the estimates converged. Caliendo and Parro (2015) modeled a welfare gain of 0.08 percent for the United States, against 1.31 percent for Mexico and a loss of 0.06 percent for Canada, and Burfisher, Robinson, and Thierfelder (2001) surveyed the post-agreement studies and found little effect on the American labor market. The structure of the starting position explains the convergence. Trade with both partners was already substantial, American tariffs on Mexican goods were already low, Mexican liberalization had run unilaterally from 1985, and Canadian American trade already operated under the bilateral agreement of 1988. A change at a narrow margin, applied to a continental economy, produced a narrow national result.

On the distributional question, the estimates also converged, toward severity. Hakobyan and McLaren (2016), following Census records from 1990 to 2000, found wage growth for high school dropouts running up to 17 percentage points lower in the most protected industries and 8 percentage points slower in the most vulnerable localities than for comparable workers, with spillovers into local services through local demand and with no measurable penalty for college educated workers. A decade of slower growth compounds into a lower earnings base for every later year, in places where an anchor employer, a tax base, and a service economy contracted together.

Between those two literatures stands the anchor of the assessment, the certification count. Department of Labor records for 1994 through 2002, compiled in Hufbauer and Schott (2005), totaled 525,094 workers certified under the agreement-specific adjustment assistance program, about 525,000 workers at an average of about 58,000 per year. The count anchors the argument because nobody modeled the count. Each worker in the total stood behind a petition, an investigation, and a determination under statutory tests. The count is a floor and not a net estimate: the count excludes workers who never petitioned, workers whose petitions were denied, and most downstream losses, and the count records nothing about gains elsewhere in the economy. A floor of that size, built case by case, refutes a claim of costlessness without pretending to measure the whole displacement.

The discipline that keeps the judgment honest is method discipline. A general equilibrium estimate answers a net national question and reallocates labor by construction. A local labor market estimate answers a relative question about exposed groups and places. A deficit conversion answers a gross accounting question and produces the largest figures on contested assumptions. A certification count answers an administrative question about adjudicated cases. A figure detached from the method that produced the figure can be made to prove opposite conclusions in successive paragraphs. For study of an episode of that kind, keep your statute notes, citations, and case chronologies together free on VaultBook, with each estimate filed beside its author, its period, and its method, because the filing habit enforces the reading habit.

No finding in the assessment extends to the successor agreement, which entered into force on July 1, 2020. The record judged here belongs to the agreement that entered into force on January 1, 1994, and to the studies that measured that agreement across the periods stated in each study. On that record, measured against the promises stated for the statute in 1993, the verdict stands in the form the evidence gave the verdict: small in aggregate, severe in place.

Frequently Asked Questions

Q: Did NAFTA cost American jobs?

The answer depends on the level at which cost is measured. Burfisher, Robinson, and Thierfelder (2001), surveying the post-agreement studies in the Journal of Economic Perspectives, concluded that the agreement had little effect on the United States labor market as a whole, and Caliendo and Parro (2015) modeled a welfare gain of 0.08 percent for the United States. Displacement nevertheless occurred and was documented case by case: Department of Labor records for 1994 through 2002, compiled in Hufbauer and Schott (2005), totaled 525,094 workers certified as displaced under the agreement-specific program. The agreement therefore produced a small net national effect alongside substantial documented displacement in exposed industries and places, and neither half of that record cancels the other half.

Q: How many workers were certified as displaced by NAFTA?

Department of Labor certification records for 1994 through 2002 totaled 525,094 workers, reported as about 525,000, with the compilation attributed to Hufbauer and Schott in NAFTA Revisited (2005). The average across the nine years of the agreement-specific program ran at about 58,000 workers per year. Sandra Polaski, a former Department of Labor official, testified before a Senate subcommittee on September 11, 2006 that about 500,000 workers were certified across the same period and that about half of the certified job losses involved production shifts to Mexico. Apparel accounted for about 28 percent of certified workers, electronics for 13 percent, automobiles and parts for 7 percent, and fabricated metals for 6 percent, according to that testimony.

Q: Did NAFTA lower wages for less educated workers?

Hakobyan and McLaren (2016), in the Review of Economics and Statistics, found relative losses in wage growth rather than a measured fall in wage levels. Using United States Census data for 1990 to 2000, the study found that wage growth for high school dropouts in the most protected industries was up to 17 percentage points lower than for comparable workers in initially unprotected industries, and that wage growth for a high school dropout in the most vulnerable localities ran 8 percentage points slower. Effects for college educated workers were small and statistically insignificant. The authors measured exposure by the tariff protection an industry or locality had enjoyed before the agreement, and the authors controlled for import competition from China and found the agreement effect separate from that competition.

Q: What did economists estimate about NAFTA welfare gains?

Caliendo and Parro (2015), in the Review of Economic Studies, applied a computable general equilibrium model and estimated welfare effects of a gain of 1.31 percent for Mexico, a gain of 0.08 percent for the United States, and a loss of 0.06 percent for Canada. The same model estimated increases in trade within the bloc of 118 percent for Mexico, 41 percent for the United States, and 11 percent for Canada. Burfisher, Robinson, and Thierfelder (2001) reached a compatible conclusion in a survey of the post-agreement studies, reporting that both the United States and Mexico benefited, with much larger relative benefits for Mexico. Hufbauer and Schott (2005) estimated consumer gains of about 5.5 billion dollars per year across the first nine years.

Q: Did NAFTA hurt Mexican farmers?

The record for Mexican grain producers was severe, with a qualification attached to the largest employment figure. Timothy A. Wise (2010), in Agricultural Dumping Under NAFTA research, found that American corn was exported to Mexico at an average dumping margin of 19 percent across 1997 to 2005, meaning below the cost of production, with some 3 million Mexican corn producers affected and estimated losses to Mexican corn producers totaling 6.6 billion dollars over nine years. Mexican agricultural employment fell from about 8.1 million in the early 1990s to about 6 million in the first quarter of 2006. Polaski (2006) attached an explicit caveat to that decline: not all of the decline can be attributed to the agreement, and the peso devaluation pushed agricultural trade in the opposite direction.

Q: Why do NAFTA job loss estimates differ so much?

The estimates differ because the methods answer different questions. General equilibrium models estimate economy-wide net effects and reallocate labor by construction, so such models cannot show displacement. Local labor market econometrics estimate relative effects on exposed groups and places, comparing workers sorted by the protection enjoyed before the agreement. Deficit accounting converts a bilateral trade balance into jobs through fixed labor-content coefficients and produces the largest figures, including the 682,900 jobs that Scott (2011) reported for 1994 to 2010. Administrative certification counts adjudicated cases only and supply a documented floor, the 525,094 workers compiled from Department of Labor records in Hufbauer and Schott (2005). A figure quoted without a method states a size and withholds the claim that gives the size meaning.

Q: Did the peso crisis confuse NAFTA results?

The peso crisis confounded every measurement drawn from the first years of the agreement. In December 1994 the incoming administration of President Ernesto Zedillo devalued the peso and then allowed the currency to float. Mexican gross domestic product fell 6.2 percent in 1995, inflation ran about 52 percent in 1995, and unemployment peaked at 7.6 percent in August 1995, figures reported by the International Monetary Fund in Finance and Development in June 1998. A United States led support package of about 50 billion dollars followed, and growth resumed at 5.2 percent in 1996 and 7.0 percent in 1997. Mexico ran a trade surplus with the United States in 1995 because the crisis crushed Mexican import demand, an artifact of the crisis rather than an effect of the agreement.

Q: Was NAFTA politically bigger than its measured effect?

The political scale of the agreement exceeded the measured scale for structural reasons that the ratification debate displayed in advance. Ross Perot warned during the 1992 campaign of a giant sucking sound from production moving south, and the House passed the implementing bill on November 17, 1993, by 234 to 200, while the Senate passed the bill on November 20, 1993, by 61 to 38. The gains that the research later measured arrived in diffuse form, as Hufbauer and Schott (2005) estimated consumer gains of about 5.5 billion dollars per year spread across millions of purchases. The losses arrived in concentrated and visible form, in named plants and named towns. Concentrated losses organize political pressure more readily than diffuse gains, so the argument remained larger than the aggregate estimates.

Q: How did NAFTA affect US manufacturing employment overall?

At the national level, the aggregate literature found a small effect. Burfisher, Robinson, and Thierfelder (2001) surveyed the post-agreement studies and concluded that the agreement had little effect on the United States labor market, and Caliendo and Parro (2015) modeled a welfare gain of 0.08 percent for the United States rather than a national employment collapse. Manufacturing contained the exposed industries within that national picture, and Hakobyan and McLaren (2016) found severe relative wage losses for less educated workers in the most protected industries. Certification records also documented substantial manufacturing displacement between 1994 and 2002. The aggregate finding therefore describes a manufacturing sector that absorbed concentrated losses in exposed branches while national employment totals moved little under the agreement itself.

Q: What happened to trade between the United States and Mexico under NAFTA?

Trade between the two countries expanded rapidly across the first decade of the agreement. Burfisher, Robinson, and Thierfelder (2001) reported that Mexico’s share of United States imports rose from 6.9 percent in 1993 to 11.5 percent in 2000, and two-way trade grew several-fold in nominal terms across the first decade. Caliendo and Parro (2015) modeled increases in trade within the bloc of 41 percent for the United States and 118 percent for Mexico. Burfisher, Robinson, and Thierfelder also concluded that trade creation greatly exceeded trade diversion, especially in intermediate goods, meaning that much of the expansion ran through components and materials that firms bought for further production rather than through a rerouting of purchases toward higher cost suppliers inside the bloc.

Q: Did NAFTA lower prices for American consumers?

Hufbauer and Schott (2005), in NAFTA Revisited: Achievements and Challenges, estimated consumer gains of about 5.5 billion dollars per year across the first nine years of the agreement, amounting to about 49.5 billion dollars in total. The estimate carries the qualifier that the authors attached to the estimate: the figure is an approximation, the period is the first nine years, and the gain described is a consumer benefit as Hufbauer and Schott assessed the benefit. Gains of that character arrive as slightly lower prices spread across millions of purchases, which explains why a benefit measured in billions of dollars in aggregate could remain invisible to any individual household. The consumer estimate complements the modeled welfare percentage rather than replacing the modeled estimate.

Q: What is NAFTA-TAA, and how did it differ from regular trade adjustment assistance?

NAFTA Transitional Adjustment Assistance was an agreement-specific component of Trade Adjustment Assistance, created by the implementing act for workers whose job loss was attributable to trade with Mexico or Canada. A group qualified where increased imports from either partner country contributed to the separations or where production had shifted to either partner country, a causal test that tied eligibility to the two partner countries alone. Worker groups petitioned the Department of Labor, the department investigated, and certification unlocked income support, training, and relocation allowances. Workers at supplier firms secondarily affected were covered under a commitment in the Statement of Administrative Action. The Trade Act of 2002 merged the agreement-specific program into the general Trade Adjustment Assistance program and repealed the separate subchapter.

Q: How did NAFTA affect jobs and wages in Canada?

The marginal change for Canada was the smallest of the three partners, because liberalization between Canada and the United States was largely complete under the bilateral free trade agreement of 1988. The principal tariff change between the two largest partners had therefore already occurred six years before the trilateral agreement entered into force on January 1, 1994. Caliendo and Parro (2015) modeled a welfare change for Canada of minus 0.06 percent, a figure in the range of a rounding difference, alongside a modeled increase in trade within the bloc of 11 percent, also the smallest of the three countries. Canada’s stake in the trilateral text ran through access to Mexico and through the investment and dispute chapters rather than through tariff change with the United States.

Q: Why did border assembly employment in Mexico grow so fast?

Border assembly employment grew from a large established base rather than from a new industry. The maquiladora program dated from 1965, and employment in the sector stood at 546,600 in December 1993, according to INEGI, the Mexican statistics institute. Polaski testified in 2006 that the sector added about 800,000 jobs between January 1994 and its peak in early 2001, and the sector reached about 1.3 million workers in 2000. The agreement contributed through a specific channel: the tariff cuts of the pact stimulated the apparel subsector of the maquiladoras specifically, according to Polaski. The expansion had limits set from outside the pact, and plants shed about 125,000 jobs through January 2006 under pressure from the United States recession of 2001 and from competition with China.

Q: How much did American corn exports to Mexico increase under NAFTA?

United States corn exports to Mexico roughly quadrupled across the first decade of the agreement. The import share figures state the same change from the Mexican side of the border: imports rose from about 7 percent of Mexican corn consumption to about 34 percent. Wise (2010) found that American corn was exported to Mexico at an average dumping margin of 19 percent across 1997 to 2005, meaning below the cost of production. Mexico never enforced the tariff-rate quota that the agreement permitted on corn imports, and Procampo payments were the transition support offered in place of border protection for Mexican producers.

Q: Did NAFTA cause the long decline of American manufacturing?

The chronology of the research record assigns the larger shock to a later and separate policy change. Permanent normal trade relations with China arrived in 2000, and China acceded to the World Trade Organization in December 2001, and the shock that followed across American manufacturing localities in the decade after 2000 was the larger of the two shocks. Hakobyan and McLaren (2016) built a design that separates the two effects, using Census data for 1990 to 2000 and controlling for import competition from China, and the study found the agreement effect separate from the China effect. The aggregate literature also found national effects for the trilateral agreement that were small. Attributing the whole post-2000 manufacturing decline to the 1993 agreement contradicts both the chronology and the research design that measured the earlier agreement.

Q: How do economists separate the effects of NAFTA from other causes?

Each research design separates the agreement through a different device. General equilibrium models build a counterfactual economy in which the tariff reductions never occurred and compare the modeled outcome against the economy with the reductions in place, as Caliendo and Parro (2015) did for the three signatory countries. Local labor market econometrics measure exposure from the tariff protection that existed before the agreement and compare exposed workers and places against unexposed counterparts, as Hakobyan and McLaren (2016) did with Census data for 1990 to 2000, with a control for import competition from China inside the design. Survey research weighs the accumulated studies against each other, as Burfisher, Robinson, and Thierfelder (2001) did. Administrative counts avoid modeling altogether and rest on adjudicated cases.

Q: Did NAFTA affect wages outside manufacturing?

Hakobyan and McLaren (2016) traced an effect beyond the plants that traded goods directly. Workers in local service industries in exposed places were also affected, through local demand: a protected plant that lost orders or reduced earnings lowered the spending available to retailers, restaurants, repair shops, and personal service businesses in the same town, and the wages in those businesses depended on the incomes of the households around the plant. No tariff had protected a service job of that kind, and no import competed with the job directly. The wage penalty for residence in a vulnerable locality, 8 percentage points of slower wage growth for a high school dropout across 1990 to 2000, captured that place-based transmission. A separate study by the same authors (2017) found married blue-collar women in exposed places among the hardest hit.

Q: Did NAFTA increase migration from Mexico to the United States?

Philip Martin of the University of California, Davis (2005) estimated that the unauthorized Mexican population in the United States rose from 2.5 million in 1995 to 4.5 million in 2000. Martin’s estimates establish the size of the increase across the second half of the 1990s. Attribution of the increase to the agreement is a separate question, and the research record treats that attribution as disputed. How much of the rural outflow the agreement caused, as against the peso crisis of December 1994 and longer trends in Mexican demography, migration networks, and American labor demand, is genuinely disputed among researchers. The careful statement therefore holds the two parts apart: a large measured increase in the unauthorized population occurred during the agreement’s first years, and the share of that increase caused by the agreement remains unsettled.

Q: How long did the NAFTA tariff phase-outs take to complete?

Tariff phase-outs under the agreement ran on staged schedules, and the longest schedules ran as long as fifteen years for the most sensitive agricultural goods. The implementing act was signed on December 8, 1993, and the agreement entered into force on January 1, 1994, so a fifteen year schedule for a sensitive agricultural category reached completion fifteen years after entry into force rather than on the first day of the pact. The staged structure matters for measurement: a study of the first five years measured industries that had completed a phase-out alongside industries whose reductions had barely begun, while a study across a longer window captured more of the completed schedule. The duration of the schedules is one reason estimates built on different periods describe different stages of the same legal change.