The Gains from International Trade: Empirical Evidence Using the Method of Instrumental Variables

Lecture Note 13 – The Gains from International Trade: Empirical Evidence Using the Method of Instrumental Variables David Autor, MIT and NBER 14.03/14.003 Microeconomic Theory and Public Policy, Fall 2016 1 1 Introduction: Measuring the Causal Effect of Trade on GDP (James Feyrer, 2009) Using data from the Penn World Tables, Figure 5 of James Feyrer’s 2009 paper, “Trade and Income—Exploiting Time Series in Geography” shows that countries that experienced rising trade volumes between 1960 and 1995 also experienced rising GDP. Is this relationship causal, or does it simply stem from rich countries trading more? As 14.03/003 students understand, economic theory clearly predicts that trade increases national income, since it expands the set of goods and services a country can consume. (Note that “income” in this discussion refers to real income – the purchasing power of residents in a country – rather than nominal income in terms of local currency.) But this theoretical prediction is difficult to test because it’s hard to conduct a credible experiment. We cannot readily manipulate the trade flows of various countries to study the impact this has on their national incomes. Figure 5: Average Per Capita GDP Growth versus Trade Growth 1960-1995 SGP Average per capita GDP growth 1960−1995 −2 0 2 4 6 KOR THA CYP ROM JPN MYS IDN PRT CHN MUS IRL PAK SYC GRCITAISR NOR BRA MAR SYR ISL FIN FRA EGY PAN IND USA LKA NLD TUR CHL CAN DOM COL DNK AUS SWE ECU FJIGNBGBR MEX JOR PNG GTM KEN NZL PHLCRI URY GHA ZAF ARG NAM BGD GUY JAM PER SLV TZA GMB HND MRT CIV CMR TGO GINBEN BRB SEN MDG ESP NIC MOZ −4 GNQ 0 2 4 6 Average trade growth 1960−1995 8 10 Courtesy of James Feyrer. Used with permission. source: Penn World Tables 6.2, IMF Direction of Trade database. Second, the simple instrument described in Equation (8). The former instrument Let’sgravity formalize this estimates point. Applying our familiar causal framework, we first wouldstage. like to uses model to maximize the predictive power of the measure the causal effect ofwhich trade on country as follows:in average distance to trading The latter instrument, is just the jdifference partners by air and by sea, has the advantage of simplicity. It does not require any γ = YjT − YjA , estimation to construct and it has aj simple interpretation. Figure 6 shows the first stage relationships between trade growth and the prewhere γj is the causal effect of trade on Y in country j ( γ stands for “gain from trade”) dicted change in trade from the gravity model estimations and the difference in air while Y T and Y A signify counterfactual national income according to some income measure and sea distance described in Equation (8). The relationship between the instru(e.g., income per capita) under Autarky and Trade, respectively. Note that the T and A ments andhere actual trade growth is very The stage superscripts denote counterfactuals, so strong. they serve theF-statistic role of theon0 the and first 1 subscripts is 29 for the change in predicted trade and 15.5 for difference between air and sea 2 relationship between the growth in per distance. Figure 7 show the reduced form capita GDP and the two instruments. More formally, I run the IV regression of trade growth versus GDP growth in- which we used in the first few lectures of class. The Fundamental Problem of Causal Inference says that we can never directly observe γj : we cannot observe income per capita for country j under both Autarky and free trade simultaneously. To uncover the true γj , one standard solution would be to contrast incomes of trading and non-trading countries. We could form     γ̂ = E Y T |T = 1 − E Y A |T = 0 , where T ∈ {0, 1} denotes whether or not a country is open to free trade and hats denote that we are using our data to estimate expectations (taking sample means). As you all learned when you studied for the first midterm, γ̂ is an unbiased estimate of γ only if the following holds:     E Y T |T = 1 = E Y T |T = 0 ,     E Y A |T = 1 = E Y A |T = 0 . That is, the autarkic economies would have the same income per capita as the trading countries if they opened to trade, and vice-versa for the trading countries if they became autarkic. (As we discussed earlier in the semester, a good shorthand term for this assumption is exchangeability: if the experimenter had exchanged the treatment and control groups prior to performing the experiment, she would have have obtained the same causal effect estimate.) Are these assumptions plausible? Would countries that trade a lot be similar to countries that trade very little, absent these differences in trading? Probably not. The extent to which a country trades is an endogenous outcome that is very likely correlated with other factors that directly affect income per capita. A few possible factors: 1. Countries that are rich for other reasons might trade more because they can afford to import more goods from overseas. 2. Countries that pursue sound economic policies (that raise income) may also choose to pursue trade (another sound economic policy). 3. Countries that are rich in natural resources may trade because there is high world demand for their goods, but these countries might have been rich due to their copious endowments even in the absence of any trade. One should therefore be very skeptical of any “causal inference” that naively compares the incomes of trading and non-trading countries. In point of fact, countries that trade more are on average wealthier, but this correlation need not be causal. 3 2 Using the method of Instrumental Variables (IV) to measure causal effects 2.1 Looking for experiments in strange places What we need is an “experiment” that exogenously raises or lowers trade in some group of countries. In past class examples, we’ve used both “natural” or “quasi-” experiments (e.g. the NJ minimum wage change, the rollout of cell phones in Kerala, India) and randomized experiments (e.g. the Jensen-Miller rice subsidy) to isolate exogenous variation in the treatment variable of interest. In the case of free trade, such experiments are difficult to find. Even policy changes that open or close a country to trade (for example, war, natural disaster, revolutionary overthrow) are potentially suspect: they are quite likely to induce other economic and policy shocks in addition to trade shocks that also directly raise or lower real income. This means that even a difference-in-differences design – the “gold standard” from our first few lectures – fails to meet the “parallel trends” assumption that we discussed, and therefore fails to give us a reliable estimate of γj . (Refresher: Recall that a DD identifies the effect of a policy under the assumption that the treatment and control groups’ outcomes would have evolved in parallel absent the policy change. If a war closes off trade but also destroys the national economy of country j, then we suspect that country j would have evolved very differently from it’s neighbors even absent the closure of trade). We need a new – and even cooler – technique to uncover causal effects in this setting. This subtle and powerful approach to identify causal effects is the method of Instrumental Variables (IV). IV is frequently referred to by the name of the statistical procedure conventionally used to implement it, Two Stage Least Squares (2SLS), and in this class we will use these two terms synonymously. Here’s the idea: we are interested in measuring the effect of trade on income. Since trade is endogenous, we are reluctant to draw any causal inferences from the observed correlation between trade and income. And we haven’t yet found a difference-in-differences design that passes the smell test for parallel trends. • Assume now that there is some third, exogenously assigned variable, Z ∈ {0, 1} that affects the extent to which countries trade. • Assume further that we have reason to believe that Z has no effect on national income, except potentially through its effect on trade. • Under these assumptions, Z may serve as an “instrument” that exogenously manipu4 lates trade, allowing us to study trade’s effect on income. Economists would say that Z is a valid instrumental variable for analyzing the causal effect of trade on income. 2.2 The Feyrer strategy James Feyrer’s 2009 paper, “Trade and Income: Exploiting Time Series in Geography,” proposes an ingenious IV approach for analyzing the causal effect of trade on national per capita income. His insight is that, historically, most trade between non-contiguous countries occurred by sea. As the cost of air freight fell over the last four decades, a substantially larger share of trade was transported by airplane rather than ship. The impact of this cost reduction was not uniform across different pairs of trading partners. For country pairs connected by a direct sea route (e.g., Spain and Brazil), the declining cost of air freight is not particularly important: it reduces transport time but not necessarily transport cost. For country pairs that are connected by a highly indirect sea route however (e.g., Japan and the Western Europe), the reduction in the cost of air freight means that traded goods will potentially have to travel a much shorter distance by air than sea. This makes trade much cheaper for these country pairs. This insight underlies Feyrer’s empirical approach: As air freight gets cheaper, countries that have a high value of their “Air-Sea Distance Difference” (ASDD)—that is, the air distance to their trading partners relative to their sea distance to their trading partners—will experience a large increase in trade volumes. By contrast, trade flows among countries that have small or zero ASDDs will not be greatly affected. S Here’s how ASDD is defined. Let Djk be the sea distance between countries j and k A S A and Djk be the air distance. Let ASDDjk = Djk − Djk . If country j and k have nothing between them but water, then their sea and air distances will be the same, meaning that ASDDjk = 0. If they are separated by land masses that a cargo ship must circumnavigate, then ASDDjk > 0. Now, define the average ASDD for each country j as the trade-volume weighted ASDDjk for all of its trading partners k. Specifically, P ASDDj = k
S A Djk − Djk × Tjk P , Tjk k where Tjk is the trade volume between j and k (in dollars, for example) in 1960. Note that T P jk captures the historical importance of country k relative to other countries in country k Tjk j’s historical trading patterns. So ASDDj measures the overall change in trading costs for 5 country j thanks to the advent of the airplane, assuming country j followed it’s historical trading patterns. If Feyrer’s hypothesis is correct, then trade flows will rise differentially between countries with relatively high ASDDjk as air freight gets cheaper. Moreover, if ASDDjk exclusively affects a country’s economy via its effect on trade, then cross-country variation in ASDDj provides a kind of natural experiment for studying the causal effect of trade on income: as the cost of air freight falls, countries with high ASDDj should begin to trade more than countries with low ASDDj , which will in turn allow us to study the effect of trade on national incomes. You may object: ASDD is not the only determinant of changing trading patterns. For example, the U.S. began trading extensively with China in the 1990s but was trading extensively with Japan decades earlier. Clearly, the ASDD gap between the US-China and US-Japan ASDD is trivial, so the falling cost of air freight cannot be the cause of rising China trade. That’s correct! But that’s not a problem for the IV approach. ASDD need not be the only determinant of trade. What we need is: 1. ASDD has a measurable, direct causal effect on trade. This is called the first stage. This is directly testable. 2. ASDD does not plausibly affect national income through any other channel but trade. Restated, trade is the exclusive channel by which ASDD affects national incomes (if at all). This is called the exclusion restriction. The exclusion restriction is not directly testable. Thus it deserves extra scrutiny whenever we are designing or evaluating an IV strategy. 2.3 Setting the stage for IV Figure 1 of Feyrer (2009) shows that air freight came to encompass a substantial share of U.S. trade between 1965 and 2005, while Figure 3 documents that countries’ trading volumes became substantially more sensitive to air distance between 1960 and 1995 and, simultaneously, substantially less sensitive to sea distance. 6 Air Freight Share of Trade Value 10 20 30 40 50 60 Figure 1: Air Freight Share of US Trade Value (excluding North America) 1960 1970 1980 year Imports 1990 2000 Exports source: Hummels (2007), ppCourtesy 133. of James Feyrer. Used with permission. Figure 3: The Change in Elasticity of Trade with Respect to Sea and Air Distance over Time -1.5 consumer electronics. Overall about 40 percent of goods in these two categories are transported by air. Goods in HS 71, made up of jewelry and precious metals and -1.0 stones, are predominantly transported by air. The remainder of the categories fall into a few general areas. The majority of pharmaceuticals and organic chemicals -0.5 goods such as watches, works of art, and leather goods are often travel by air. Luxury transported by air. A substantial value in apparel (over 15 percent) is transported 1940 1950 1960 1970 1980 1990 2000 by air though the0.0majority of apparel is transported by sea. Table 2 lists the top 20 countries by value of imports into the US by air. There is substantial variation 0.5 amongst US trading partners in the proportion of trade by air. Sea Air Japan shipped only 27 percent by air and China only 13 percent by air. Singapore, source: Coefficients from regression table column 2. Used with permission. Courtesy of 9 James Feyrer. Malaysia, andrepresents the Philippines exportsEstimates to the US by Each point the coefficientshipped on (sea orthe air) majority distance overofa 5their year interval. How can we use this information about the changing relationship between ASDDareand from a gravity withplot country fixed effects. air. Figure 2 istoamodel scatter showing the percentage exports sent the US by air trade volumes find the causal effect of trade on income? of That’s where the to subtlety comes Error bars represent plus or minus two standard errors for each coefficient. versus thevalidity log of of gdp workerwill in rest 1960. is no significant between in. The ourper approach on There three pillars, which we willrelationship discuss in turn: income per worker in 1960 (before the advent of air freight) and the percentage of 1. Balance of treatment and control groups: Observations with different values for the The Change in Elasticity of Trade with Respect to Sea and Air Distance tradeFigure by air4: in 2001. instrumental variable have similar counterfactual outcomes.. over Time -1.5appendix) lists the top overall importers to the US, their share Table 8 (in an 2. First stage relationship: There is a causal effect of the instrumental variable on the of imports to the-1.0US by air and the HS4 category with the highest value of goods endogenous variable transported by air to the US. The primary air export varies quite a bit from country -0.5 to country. Many of the Asian countries 7export computers and parts to the US by 1940 1950 1960 1970 1980 1990 2000 air. European countries export chemicals and pharmaceuticals to the US by air. 0.0 Many developing countries export precious metals and jewelry to the US by air. 3. Exclusion restriction: It is plausible that the instrumental variable affects the outcome variable only through its effect on the endogenous variable Now, imagine that we have a set of potentially comparable countries that differ according to whether they have High ASSD (A = 1) or Low ASDD (A = 0). (Note that A take the place of Z on pages 4-5.) In our example: a) the endogenous variable of interest is a country’s trading volume; b) the instrumental variable is the country’s ASDD; and c) the outcome variable is the country’s GDP. 2.3.1 Condition 1: Balance of treatment and control groups As with our previous techniques for causal inference, our treatment and control groups be comparable—that is, they must have have balanced counterfactual outcomes. • Let Yjt equal the GDP of country j in time t. • Imagine that there are two time periods, t = {0, 1}, and that in the early period (t = 0), traded goods travel exclusively by sea, whereas in the latter (t = 1), traded goods can travel by air or by sea. • Let ∆Yj equal the change in GDP in country j between t = 0 and t = 1. Note that this paper focuses on how changes in trade affect changes in income, rather than how levels of trade affect levels of income. Theoretically, both designs could uncover the effect of trade on income. • For each country, imagine two potential outcomes  ∆Yj ∈ ∆Yj1 , ∆Yj0 , where ∆Yj1 is the change in GDP in j if A = 1 and ∆Yj0 is the change in GDP in j if A = 0. • Of course, each country j is either one type or the other (ASDD is either High or Low, A = 1 or A = 0). So, we will never observe both ∆Yj1 and ∆Yj0 (i.e., the fundamental problem of causal inference, FPCI). Thus ∆Yj1 and ∆Yj0 are counterfactuals of one another. • Assuming balance of the treatment and control groups means that we believe in ex- 8 changeability:     E ∆Yj1 |A = 1 = E ∆Yj1 |A = 0     E ∆Yj0 |A = 1 = E ∆Yj0 |A = 0 . If the countries with high ASDD were somehow assigned low ASDD, their GDP growth would be the same as the the countries that actually have low ASDD, and vice versa if the low ASDD countries were somehow assigned to have high ASDD. We can’t completely prove exchangeability by looking at data, but by comparing observable characteristics of countries with A = 0 and countries with A = 1 we can make a plausible case that exchangeability is reasonable. 2.3.2 Condition 2: There is a causal effect of the instrumental variable on the endogenous variable For our proposed Instrumental Variables approach to be valid, it must be the case that ASDD has a causal effect on the amount that countries trade. This is called the “first stage” relationship by econometricians. The existence of a first stage relationship is verifiable as a statistical matter. (Though as always, correlation does not imply causality. More on this below.) • Write Tjt as the trade volume (in dollar terms, for example) of country j in year t. • Again, imagine two counterfactual states for each country j, one in which it has Low ASDD (A = 0) and the other if it has High ASDD (A = 1). • We know that between 1965 and 1995, air transport got considerably less expensive overall and simultaneously the air volume of U.S. trade increased considerably (Figure 1). • Define the counterfactual change in trade volume between 1965 and 2005 in each country under ASDD ∈ {0, 1} as  ∆Tj ∈ ∆Tj1 , ∆Tj0 • We require the following: ∆Tj1 ≥ ∆Tj0 ∀ j, In words, country j 0 s trade volume must increase by more between time 0 and 1 if ASDD is High than if ASDD is low. 9 • Due to FPCI, this assumption is also not testable. We only see countries in one state—ASDD is High or Low—or another. • However, we can test one necessary but not sufficient condition for the validity of this relationship, which is: E [∆Tj |A = 1] > E [∆Tj |A = 0] . That is, the average growth in trade in the A = 1 countries must be greater than in the A = 0 countries. • We can check this empirically by verifying that: 1 nA=1 × X ∆Tj > j,A=1 1 nA=0 × X ∆Tj , j,A=0 where nA=1 is the number of countries with A = 1 and similarly for nA=0 • Figure 6 of Feyrer suggests that this relationship holds in the data. 2.3.3 Condition 3: Exclusion restriction • A valid instrumental variable must also satisfy an “Exclusion Restriction.” The exclusion restriction says that the instrumental variable (here ASDD) must only affect the outcome variable of interest (here GDP) indirectly through its effect on the intermediating endogenous variable of interest (here, Trade). • If we do not find it plausible that ASDD only affects national income through its impact on trade, we cannot rely on any measured relationship between distance and income to help us uncover the causal effect of trade on income. • Conversely, if we find it plausible that ASDD only affects national income through its impact on trade, we can interpret the measured relationship between distance and income as reflecting (though not identical to) the causal effect of trade on income. • The exclusion restriction can be expressed formally as follows: E [∆Yj |∆Tj = k, A = 1] = E [∆Yj |∆Tj = k, A = 0] , where k is some constant. • This equation says that if were were to hold trade in country j constant at a given level k, ASDD would have no effect on GDP—since its entire effect operates through influ10 encing trade. Holding country j 0 s trade constant at level k, GDP of j is independent of ASDD. • The exclusion restriction must be plausible or the IV strategy is a non-starter. However, this postulate is not testable. We cannot directly manipulate ASDD for a given country. Moreover, if we could, this manipulation would also affect Tj (under our hypothesis above). Thus, we cannot verify that ASDD only affects a country’s GDP through its effect on trade. • If we believe that ASDD affects GDP through some other mechanism (e.g., ASDD increases a country’s air traffic, and the smell of burning jet fuel makes citizens happier and more productive, raising GDP), then using ASDD as an instrumental variable for trade will not allow us to isolate the causal effect of trade on GDP. 2.3.4 The smell test for our conditions If we tentatively accept the conditions above, the empirical analysis proceeds as follows: 1. We check that trade grows by more in ASDD = 1 than ASDD = 0 countries between times t = 0 and t = 1: Ê [∆Tj |A = 1] > Eˆ [∆Tj |A = 0] or, the same expression written differently: 1 nA=1 × X ∆Tj > j,A=1 1 nA=0 × X ∆Tj j,A=0 If this inequality is satisfied, then A is a candidate instrument for T . If this inequality is not satisfied, then our assumption that [∆Tj |A = 1] > [∆Tj |A = 0] ∀ j is false. Verifying the inequality above does not prove that the assumption is true. But rejecting it would demonstrate that the assumption is false, and therefore we will not want to use our proposed IV strategy. 2. If we pass this test, we can next test whether GDP rises by more over time (between time t = 0 and t = 1) in ASDD = 1 versus ASDD = 0 countries. The hypothesis that trade raises income implies that E [∆Yj |A = 1] > E [∆Yj |A = 0] . If trade raises GDP, then the fact that trade rises by more in A = 1 than A = 0 countries implies that GDP also rises by more in A = 1 than A = 0 countries. 11 If both of these relationships are verified in the data, we may be correct to conclude that trade has a positive causal effect on national income. But we would not yet have an estimate of the size of this effect. Instead, we would have an estimate of the causal effect of ASDD on trade, and another estimate of the causal effect of ASDD on income. That’s close, but not quite what we’re after. We need to take one more step. 2.4 Estimating the causal relationship using the method of Instrumental Variables 2.4.1 The parameters we can grab from the data • Our goal is to estimate the causal effect of trade volumes on GDP. Let’s write this as: E [∆Y |∆T ] = α + γ∆T, (1) where γ denotes the causal effect of trade on GDP. This is the parameter we’d like to estimate. • We found that ASDD is correlated with the change between 1960 and 1995 in the extent that a country trades, and given our balance assumptions above, we view this correlation as causal: π1 = E [∆T |A = 1] − E [∆T |A = 0] > 0 • We compare the change in the incomes of ASDD High and Low countries. π2 = E [∆Y |A = 1] − E [∆Y |A = 0] . Here, π2 is the causal effect of ASDD (not trade) on GDP. • That’s a start, but we have not yet estimated γ, the causal effect of trade on GDP. If we had exogenous (as good as randomly assigned) variation in the change in trade that countries experienced, we could simply estimate equation (1) above, and γ̂ would be our causal effect estimate. • We cannot do that because the variation in trade that we observe is endogenous. Naively regressing ∆GDP on ∆T will tell us about the correlation between trade and GDP, but it will not provide an unbiased estimate of γ. 12 • It turns out that we can infer this causal relationship using the observed causal relationships between (1) ASDD and ∆T, and (2) ASDD and ∆Y . 2.4.2 Using those parameters to construct a causal estimate Putting the pieces together: • Causal effect of ASDD on Trade: E [∆T |A = 1] = α1 + π1 (2) E [∆T |A = 0] = α1 E [∆T |A = 1] − E [∆T |A = 0] = π1 • Causal effect of ASDD on GDP growth: E [∆Y |A = 1] = α2 + π2 (3) E [∆Y |A = 0] = α2 E [∆Y |A = 1] − E [∆Y |A = 0] = π2 • Substituting (2) and (3) into (1) gives us the expression for the causal effect of ASDD on GDP growth: E [∆Y |A = 1] − E [∆Y |A = 0] = π2 = γ (E [∆T |A = 1] − E [∆T |A = 0]) = γ × π1 By implication π2 = γ × π1 . • Thus, our estimate of π2 is closely related to the causal effect of trade on GDP (γ) in equation (1) above. They only differ by a scalar: π2 = γ × π1 . • Combining our two causal effects , π1 and π2 , we can estimate the causal effect of trade on income: π2 π1 × γ E [∆Y |A = 1] − E [∆Y |A = 0] = = =γ E [∆T |A = 1] − E [∆T |A = 0] π1 π1 • We thus estimate the causal effect of trade on income by taking the ratio of the two causal effects: the causal effect of ASDD on GDP growth (π̂2 ) and the causal effect of 13 ASDD on trade growth (π̂1 ) . This ratio gives us γ̂, our Instrumental Variables (IV) estimate of the causal effect of trade on GDP. • Intuitively, we are comparing incomes among potentially similar countries that have different ASDD0 s. This comparison gives us the causal effect of ASDD on income growth (π̂2 = γ × π1 ). We convert this relationship into an estimate of the causal effect of trade on income by re-scaling the GDP growth difference between high and low ASDD countries by the causal effect of ASDD on trade growth. • [A bit of history: The IV method was developed in 1928 by the economist, P.G. Wright, who wanted to measure the causal effect of supply changes on the price of flaxseed. He used weather shocks as an exogenous source of variation in supply of flaxseed. Instrumental Variables has become central to causal empirical analysis in economics within the last two decades.] 3 Findings The main figures in the Feyrer paper tell the story. You should understand how each of these figures contribute to the empirical case. See figures in the following order : 1. Figure 1: Air freight shares to the U.S. 2. Figure 3: Change in elasticity of trade with respect to Sea and Air distance over time 3. Figure 2: Air imports to the US versus 1960 GDP per capita 4. Figure 6 panel B (right-hand side): Air and Sea Distance Differential (ASDD) versus Average Trade Growth 1960-1995 5. Figure 7 panel B (right-hand side): ASDD and per capita GDP growth, 1960-1995 14 Figure 2: 2001 Air Imports to the US versus 1960 GDP per capita IRL 1 1 Figure 2: 2001 Air Imports to the US versus 1960 GDP per capita Proportion of Exports to US by Air 2001 .2 .4 .6 .8 Proportion of Exports to US by Air 2001 .2 .4 .6 .8 BWA BEN CHE BRB MYS PHL MLI CAF MRT COM CHE BRB IRN MYS PHL TWN KOR BOL PRT MARHKG IND UGA ISR ISR SGP SGP NER CAF ZAF GBR BEL FRA URY ITA IRN NLD DNK SWE AUT AUS JPN CRI FIN ISL THA CPV CYP SYC AUS ZWE MDG TUR MUS PER ESP PAN LUX BRA GRC ROM GMB ARG GNB LKA GIN JOR CHL COL DOMPRY PAK CHN IDN ZMB SYR NIC NOR NZL EGY TGO RWA ECU FJI SLV GTM BDI PNG HNDGUY BGDSEN JAM CAN GHA MOZ MEX LSO BFA NAM TTO CIVCMR MWI ETH VEN GAB AGO NGATCD TZA MLI MRT COM NPL KEN ZAF TWN KOR BOL PRT MARHKG IND UGA 0 0 BEN NER GBR BEL FRA URY ITA NLD DNK SWE AUT AUS NPL 6 9 JPN 10 CRI FIN ISL KEN7THA CPV 8 CYP ln(real GDP per worker 1960) SYC AUS ZWE MDG TUR MUS PERESP PAN LUX BRA GRC ROM GMB ARG Tables GNB – US Imports ofLKA GIN Bureau Merchandise 2001, Penn World CHL JOR COL PRY DOM SYR PAK CHN IDN ZMB EGY NIC NOR NZL TGO RWA ECU SLV GTM FJI BDI PNG BGD JAM SEN HND CAN GHA MOZ GUY NAM MEX LSO BFA CIVCMR TTO MWI ETH VEN GAB AGO NGATCD TZA source: US Census IRL BWA 6.1. Table7 2: Top 20 Countries for 9US Imports by 8 10 Air ln(real GDP per worker 1960) Air Import Value Percent Courtesy of James Feyrer. Used with permission. Country (billion $) 2001, by Air Penn World source: US Census Bureau – US Imports of Merchandise Japan 34.1 26.9% KOR KOR Figure 6: First Stage: 1960-1995 versus Instruments UKActual Trade Growth 21.5 52.0% Germany 17.8 THA 30.2% ESP ESP Ireland 16.8 90.7% SGP France 14.2 47.0% JPNCHN MEX MEX Taiwan 14.0 41.9% PRT PRT IRL IRL South Korea 13.4 37.9% ISR ISR IDN ITA ITA 13.3 TUR 59.3% PAN Malaysia FRA YS FRA TUR MYS GRC China 13.0 GRC12.7% NLD NLD USA CRI ECU PNG Singapore 11.5 76.8% CAN NORFIN CAN NOR BRA HND FIN PHL ROM ROM BGD Canada 9.8 4.5% CHL MUS DNK DNK CYP CYP ISL GNQ SWE SWE AUS Italy 9.5 JOR GBR 39.7% GTMNAM GBR COL JOR MRT SYC CIV TGO9.4SYR 78.3% DOM SYR Israel GNB EGY EGY SLVFJI BEN ARG URY ZAF MAR IND Switzerland 6.8 71.1% PAK CMR GIN NZL PER source: IMF Direction of Trade database, author’s calculations. Philippines 6.5 57.2% KEN BRB JAM Mexico 5.3 4.0% GMBLKA NIC Belgium 4.9 48.6% TZA India 4.1 41.7% MOZ GUY Thailand 3.9 26.7% GHA SEN MDG Netherlands 3.7 38.8% 4 5 3.8 4 4.2 Imports of Merchandise 4.4 source: US Census Bureau – US 2001. 4.6 6 Tables 6.1. 10 10 10 ge: Actual Trade Growth 1960-1995 versus Instruments KOR KOR THA THA ESP SGP JPN CHN MEX JPN CHN 8 Average trade growth 1960−1995 Average trade growth 1960−1995 0 2 4 6 8 4 6 8 SGP ESP MEX Table 2: Top 20 Countries for US Imports by Air IRL ISR MYS FRA TUR IDN PAN ITA GRC USA PNG ECU NLD CRI PRT Average trade growth 1960−1995 4 6 PRT IRL ISR IDN PAN ITA FRA TUR GRC MYS CRI ECU PNG USA NLD Air Import Value Percent Country (billion $) by Air Japan 34.1 26.9% UK 21.5 52.0% Germany 17.8 30.2% Ireland 16.8 90.7% France 14.2 47.0% Figure 7: Reduced Form: Average PerDistance CapitaDifference GDP Growth 1960-1995 versus Inth 1960−1995 Air and Sea Taiwan 14.0 41.9% struments South Korea 37.9% Courtesy of James Feyrer. Used with permission. 7 13.4 ade database, author’s calculations. Malaysia 13.3 59.3% China 13.0 12.7% Singapore 11.5 76.8% Canada 9.8 4.5% Italy 9.5 39.7% Israel 9.4 78.3% 15 source: Switzerland Penn World Tables 6.2, author’s calculations. 6.8 71.1% 6.5 versus 57.2% m: Average Per Capita Philippines GDP Growth 1960-1995 InCAN NORFIN PHL HND BGD CHL MUS CYP ISL GNQ SWE AUS GTM GBR NAMCOL MRT SYC CIV TGO DOM SYR GNB EGY SLV BEN ARG FJI URY ZAF IND PAK MAR CMR GIN NZL PER BRA ROM DNK JOR CHL COL MUS AUS GTMNAM MRT SYC CIV DOM GNB SLVFJI ARG URYZAF GIN MAR NZL PER BRB JAM NIC 2 BRBKEN JAM GMB NIC LKA BRA HND FIN ROM DNK SWE JOR GBR SYR TGO EGY BEN IND CMR PAK KEN TZA MOZ GUY GHA 5 3.8 4 4.2 Air and Sea Distance Difference 4.4 4.6 0 2 3 4 Average predicted trade growth 1960−1995 GHA SEN MDG 0 SEN MDG 2 GNQ BGD CYP ISL GMBLKA TZA MOZ GUY 1 CAN NOR PHL SGP SGP 6 KOR 6 KOR CYP Average per capita GDP growth 1960−1995 −2 0 2 4 JPN BRB MUS IDN CHNMYS ROM PRT SYC BRA AUS FJI NZLZAF ESP PAK GRC ISR NOR SYR MAR FIN ISLEGYFRA USA PAN IND CAN TUR CHLCOL LKA DOM SWEGBR ECU GNB MEX JOR PNG GTM KEN CRI PHL URY GHA ARG GUY BGD JAM PER NAM TZA HND SLV MRT CIV GMB CMR BEN TGO GIN IRL ITA DNK NLD SEN MDG NIC MOZ THA CYP Average per capita GDP growth 1960−1995 −2 0 2 4 THA BRB IDN MUS MYS PRT JPN ROM CHN ESP IRL GRCPAK ITA ISR NOR BRA SYR MAR FRA ISL EGY PAN IND USA LKA TUR NLD CAN CHL DOM COL DNK AUS SWE GBR ECU FJI GNB MEX PNG GTM KEN JOR CRI NZLURYZAF GHAPHL ARG GUY BGD JAM PER NAM SLV HND MRT CIV TZA GMB CMR BEN TGO GIN SYC FIN SEN NIC MDG MOZ −4 GNQ −4 GNQ 1 2 3 4 Average predicted trade growth 1960−1995 5 3.8 4 4.2 Air and Sea Distance Difference 4.4 4.6 Average predicted trade growth 1960−1995 Air and Sea Distance Difference source: IMF Direction of Trade database, author’s calculations. rm: Average Per Capita GDP Growth 1960-1995 versus In6 GRC ISR NOR SYR AR FIN ISLEGYFRA CAN TUR SWEGBR MEX JOR IRL ITA DNK NLD D LV 6 6 BRB JPN SYC BRA NZLZAF MYS PRT ROM JPN ROM CHN THA KOR CYP JPN ROM ESP CHN BRBIRL IDN MYS PRT MUS ESP IRL SYC GRCPAK GRCPAK ITA ITA ISR NOR NOR ISR BRA NOR ISRSYR SYR BRA SYR MAR MAR FRA FRA FIN ISLEGY ISL EGY PAN IND USA PAN IND USA LKA MAR TUR NLD CAN TUR CAN CHLCOL LKA CHL DOM DOM FRA COL DNK NLD DNK AUS SWEGBR SWE ISLGNB GBR FIN ECU ECU EGY FJI GNB PAN MEX MEX IND USA LKA TUR JOR JOR PNG PNG NLD GTM GTM KEN CRI KEN CAN CHL CRI DOM NZLURYZAF PHL URY COL GHA GHAPHL DNK ARG ARG NAM GUY GUY BGD BGD JAM JAM PER PER AUS NAM SWE GBR SLV SLV TZA HND HND MRT MRT CIV CIV TZA GMB GMB ECU FJI CMR CMR GNB BEN BEN MEX TGO TGO GIN GIN JOR SEN SEN PNG GTM KEN CRI URY MDG MDG PHL NIC NZL NIC GHA ZAF ARG GUY BGD JAM PER NAM MOZ MOZ SLV HND MRT CIV TZA GMB GNQ GNQ CMR BEN TGO GIN BRB MUS AUS FJI CYP IDN MUS Average per capita GDP growth 1960−1995 −2 0 2 4 ESP THA CHNMYS IDN PRT ESP PAK GRC SYC IRL ITA FIN −4 PRT SGP KOR Average per capita GDP growth 1960−1995 −2 0 2 4 ROM THA CYP SGP −4 CYP Average per capita GDP growth 1960−1995 −2 0 2 4 KOR PN NMYS Figure 7: Reduced Form: AverageSGP Per Capita GDP Growth 1960-1995 versus InKOR struments 1 2 3 4 Average predicted trade growth 1960−1995 5 3.8 4 SEN 4.2 Air and Sea Distance Difference 4.4 4.6 MDG source: Penn World Tables calculations. NIC6.2, author’s MOZ GNQ −4 22 3.8 4 4 4.2 Air and Sea Distance Difference 4.4 4.6 5 owth 1960−1995 Courtesy of James Feyrer. Used with permission. s 6.2, author’s calculations. Jim Feyrer was kind enough to make a special table exclusively for 14.03/14.003 that shows the key results in a format that complements the analytic tools presented above. The  Effect  of  Trade  Growth  on  Per  Capita  GDP  Growth,  1960  -­‐  1995 22 Dependent  Variable Trade  Growth OLS (1) GDP  Growth Reduced  Form (3) GDP  Growth 0.55 [0.070]** Air  Sea  Distance  Difference Constant First  Stage (2) Trade  Growth -­‐0.50 [0.35] Instrumental Variables (4) GDP  Growth 0.75 [0.16]** 5.30 [1.35]** 4.00 [1.04]** -­‐17.71 [5.65]** -­‐14.72 [4.37]** -­‐1.37 [0.74]~ 76 0.12 76 0.407 Observations 76 76 R-­‐squared 0.464 0.142 Robust  standard  errors  in  brackets +  significant  at  10%;  *  significant  at  5%;  **  significant  at  1% Courtesy of James Feyrer. Used with permission. • The first column shows the Ordinary Least Squares (OLS) relationship between the change in GDP and the change in trade at the country level during 1960 - 1995 for 76 16 countries: Column (1) : ∆ ln GDPj,60−95 = α + β1 ∆ ln T radej,60−95 + ej . The point estimate of 0.55 implies that a 1% rise in trade is associated with a 0.55% rise in GDP (an elasticity of 0.55). You should not view this relationship as causal. • The second and third column show the relationship between ASDD and trade growth (column 2) and GDP growth (column 3). Column (2) : ∆ ln T radej,60−95 = α0 + π1 ASDDj + e0j , where Feyrer estimates that π̂1 = 5.30 • And Column (3) : ∆ ln GDPj,60−95 = α00 + π2 ASDDj + e00j , where π̂2 = 4.00. • Recall that π̂2 = γ × π1 . Hence, we can calculate the causal effect of trade on GDP as: γ̂ = π̂2 4.00 π1 × γ = = = 0.75 π1 π̂1 5.30 • This is exactly what Feyrer obtains in Column 4: Column (4) : ∆ ln GDPj,60−95 = α000 + γ∆Tj∗ + e000 j , where γ̂ = 0.75. I’ve denoted the change in trade in this equation with an asterisk (∆Tj∗ ) because this is not the endogenous trade variable available in the data. Rather, it is the exogenous component due to ASDD, which is found in column 2 of the Feyrer table. • Thus, our causal estimate of the effect of trade on GDP is that a one percent rise in trade raises GDP per capita by three-quarters of a percentage point. • We’ll talk further about this evidence (both its strengths and limitations) in class. 17