f / v ln 0.351ln oil DIST.0063ln v

Transcription

1 Data Appendix. More details about Data Sources and Construction For the irm data, the irm identiier in FirmStat is derived rom the register Old Firm Statistics or the period and rom General Firm Statistics or the period These two registers in combination allow us to track the same irm during the entire period despite the structural break in 999. For our irm characteristics variables, the number o employees is rom FirmStat and is calculated as the number o ull time equivalent workers in a year. Capital stock, measured as the value o land, buildings, machines, equipment and inventory is rom the Accounting Statistics register. Gross output (net o taxes) is rom the VAT register. Firm level skill intensities are computed using the educational attainment records o individual workers in IDA which are then aggregated to the irm level using the matched worker irm link (FIDA). For the worker data, we use annual hours which is common in the literature (e.g. Christensen et al. 2005). A concern is that annual hours do not capture overtime work. For a portion o our sample in 2006 we have data or overtime work. A wage rate including overtime is correlated 0.86 with our main wagerate variable, and overtime hours are uncorrelated with oshoring (0.05 or the ull sample and 0.07 or the subsample o high skilled workers). This suggests that our results are unlikely to be driven by the issue o overtime work. We measure labor market experience as actual time in employment since 964. Other worker level inormation regarding union membership and marriage are also derived rom the IDA database. We experimented with breaking low skilled workers into two subgroups, medium skilled (those with a vocational education, deined as the inal stage o secondary education that prepares students or entry into the labor market) and very low skilled (those with the equivalent o high school education or less). We obtained very similar results. Our trade data, the Foreign Trade Statistics Register, consists o two sub systems, Extrastat (trade with non EU countries) and Intrastat (trade with EU countries). Extrastat has close to complete coverage as all extra EU trade lows are recorded by customs authorities. Intrastat does not have complete coverage because irms are only obliged to report intra EU trade i the annual trade value exceeds a threshold. In 2002 the thresholds were DKK 2.5 million or exports and DKK.5 million or imports. Ater merging data on manuacturing workers, irms, and trade lows, we have 2.8 million workerirm year observations. We then trim our sample as ollows. Since we have annual data we cannot investigate the changes in wage or employment status at weekly, monthly or quarterly requencies. Thus we drop all the worker irm year observations o which the employment relationship, or job spell, lasts or a single year (about 200,000 observations). We also drop all the workers whose skill level changes in our

2 sample period (about 35,000 observations), in order to get a clean identiication o how the eects o oshoring vary across skill groups. We next drop the irms with ewer than 50 employees and less than 0.6 million DKK in imports, which corresponds to average annual wages or two manuacturing workers. This eliminates another 600,000 observations. This de minimis restriction eliminates rom our sample very small irms who in some cases have imputed balance sheet variables and are more likely to all below the reporting thresholds or intra EU trade data. For data on occupational characteristics, The occupation variable in IDA is based on a Danish version o the International Standard Classiication o Occupations (ISCO 88) developed by the International Labour Oice (ILO). We map the O*NET data into the ISCO 88 classiication system using the crosswalk at the National Crosswalk center tp://tp.xwalkcenter.org/download/xwalks/. For nonroutine tasks we use the principal component o mathematical reasoning (O*NET task id.a..c.), response orientation (.A.2.b.3), gross body coordination (.A.3.c.3), mathematics (2.A..e), thinking creatively (4.A.2.b.2), and organizing, planning, and prioritizing work (4.A.2.b.6). For routine tasks we use manual dexterity (.A.2.a.2), inger dexterity (.A.2.a.3), multilimb coordination (.A.2.b.2), processing inormation (4.A.2.a.2), and evaluating inormation to determine compliance with standards (4.A.2.a.3). For social sciences we use the principal component o 2.C. (2.C..a, 2.C..b, etc.), 2.C.6, 2.C.7, 2.C.8, 2.C.9, 2.C.4.e, and 2.C.4.. For natural sciences we use 2.C.2, 2.C.3, 2.C.5, 2.C.4.b, 2.C.4.c, 2.C.4.d, 2.C.4.g, 2.C.0, and 2.A... For communication and language we use 4.A.4.a, 2.B.,.A..a, 4.C..a.4, 4.C..b., 2.A..a, 2.A..b, 2.A..c, and 2.A..d. For on the job hazards we use 4.C.2.c, 4.C.2.b., and 4.C.2.e.. 2. Construction o the transport cost instruments The Danish trade data do not contain inormation on transportation costs paid by irms. To construct transportation costs we proceed in two steps. First, we use data on ad valorem shipping costs taken rom US sources to estimate costs as a unction o transportation mode, product weight/value, uel prices, and distances shipped. Second, we construct itted cost measures using these same variables that are speciic to Danish irms. During our pre sample years, 42% o Danish imports by value arrive by sea, 20% by air, 37% by truck, and % by rail. For sea and air transport we employ data on transportation costs taken rom US Imports o Merchandise data or the sample period. For sea transport we estimate w / v ln 0.35ln oil DIST.0063ln oil * DIST v t c t c

3 where c indexes exporters, k indexes HS6 products, t = year, = transportation charge, v = value o shipment, and w = weight in kg, DIST = distance in 000km measured to the nearest US coast. For air transport we estimate w / v ln 0.209ln JETFUELt DISTc.08ln JETFUELt * DISTc v Note that this generates air shipping costs that are higher in levels, more sensitive to uel prices, and more sensitive to the interaction between uel prices and distance. Also, jet uel prices, while correlated with crude oil prices, can vary rom year to year as a unction o dierences in reining capacity and availability o high grade crude suitable or distilling light uels. For rail and truck transport we draw on transportation costs taken rom the US Transborder Surace Freight data, which reports US state to Canadian province lows at the HS2 level monthly rom For truck transport we estimate wspkt / v ln 0.862ln oil.2ln DIST 0.373ln oil *ln DIST v spkt spkt t sp t sp spkt Here, sp reers to a state province pair. Note that we use log distance rather than levels as it provides a better it to the land based data. For rail transport we estimate wspkt / v ln 0.079ln oil 0.90ln DIST 0.224ln oil *ln DIST v spkt spkt t sp t sp spkt With the exception o the rail cost unction (which represents only percent o our sample), these estimates are broadly consistent with estimates in the literature. We then take the coeicients rom this regression to construct the costs that would ace a Danish irm with similar shipment characteristics. This is speciic to each input purchased. Oil prices and distance are the same or all irms. We use data on transport mode used and weight/value ratio or all irms purchasing a particular c k input; however to avoid introducing endogeneity we use pre sample inormation in both variables. We construct transport costs or each input rom the itted equation as exp( / v ) and aggregate over inputs using the share o each input in pre sample trade or each irm. To understand the source o variation generated by this approach, note that inputs travel dierent distances, have dierent bulk (product weight/value), and use dierent transport modes. Over time there are shocks to the level cost o each transport mode as a unction o technological change and input prices. Further, oil prices luctuate substantially in our sample, alling or 4 years and then rising sharply, as we show in Figure A. Shocks to oil prices dierentially aect costs depending on which mode is used and how ar goods travel.

4 3. Measures o oshoring: details 3. Input Output Tables An alternative approach to identiy imported inputs, commonly used in the literature, employs input output (IO) tables. Under the proportionality assumption (see Feenstra and Jensen 200 or limitations) that all irms in an industry use the same inputs and in the same ratios and that imports and domestic supplies have the same market share, one can use IO table input coeicients interacted with shocks to trade costs to generate industry time variation in the desirability o oshoring. We do not employ this method or three reasons. One, we employ industry time ixed eects in the estimation in order to control or demand shocks, and this eliminates all variation that can be exploited using an IO table. Two, unlike the literature we see the actual inputs purchased by irms and the data strongly reject the assumption o common input usage within an industry. Three, because the data indicate signiicant withinindustry variation across irms in both inputs and source countries, we can use exporter product time variation in our instruments to better explain changes in oshoring. Nevertheless, the IO tables provide a useul additional check on the irm level data. The most disaggregated Danish IO table is at the industry level, covering 57 manuacturing industries, and does not distinguish inputs by source. This is in contrast to our country product disaggregation, where we have over 3,500 distinct inputs. We make use o the import matrix o the Danish IO table or two cross checks. First, are there instances where our trade data says that product k is an imported input or irm j, but the IO table disagrees (i.e. the industry o k is not an imported input or the industry o j)? This occurs or only 2 percent o cases. As another cross check on our irm level data, we construct a hypothetical import matrix o the IO table using our trade data and compare it with the import matrix o the oicial Danish IO table. They are not identical, because the oicial IO table employ the proportionality assumption. Nevertheless, there is a broad correspondence between the inputs used by our irms and what we seem in the IO table. The input shares o our constructed import matrix have a 0.73 correlation with the oicial IO table. 3.2 Machinery Imports Imports o machinery are potentially problematic in terms o interpretation. Access to oreign technology embodied in machinery imports may aect labor demand and wages (e.g. Hanson and Harrison 999) but through a dierent channel than oshoring o material inputs that could have been produced by the irm. While we do not take a strong stand that we can completely separate the eects o oshoring

5 material inputs versus technological change embodied in machinery imports, we do want to distinguish where such eects are likely to appear in our analysis. The HS system classiies most types o machinery in HS84, Nuclear reactors, boilers, machinery etc, and HS85, Electric machinery etc; sound equipment; TV equipment. Our broad oshoring measures include imports o HS 84 and HS 85 or all irms, and this represents 6.9% o imports. Our narrow oshoring measure excludes machinery imports or all irms except or those who also produce machinery or sale. For irms that produce machinery or sale, narrow oshoring could potentially include machinery imports. The question or these irms is whether imports within HS 84, 85 represent machinery itsel or parts or machinery. As an example, consider the ive largest irms selling in HS 843, Pumps or liquids. The top three import categories are HS 843 itsel, which could be machinery, and HS 8483, Transmission shats, bearings, gears, and HS 848, Taps, cocks, valves which are clearly parts. We ound similar results or the top ive irms in HS 848 and HS 8482, Ball or roller bearings. At more disaggregated levels o data it is possible to distinguish machinery rom parts o machinery. Looking over all irms and imports we ranked the value share or each six digit product within HS 84. Table A lists the top 20 products, comprising 59% o the imports o HS 84. All are parts, and not machinery itsel. The largest HS6 import that is clearly a machine and not parts o a machine is HS , Packing or wrapping machinery It ranks 34 th on the list and its share in imports is 0.007%. The results are similar or HS85. Thereore, even in those HS categories where machinery imports are concentrated, actual machinery accounts or a small share o total imports. 4. Robustness o Results to Sample Selection To gauge the eects o our sample selection criteria, Table A2 compares the summary statistics o our estimation sample with the ull sample. The ull sample is the collection o workers and irms in the manuacturing sector we have beore we implement our sample selection criteria. The numbers under the heading Estimation Sample are identical to those in Table. As compared with the ull sample our estimation sample has slightly larger irms that employ slightly more experienced workers with slightly higher wages, but the dierences are very limited. A related concern is that we include only the irms in the years in which they both import and export. I a large raction o irms import/oshore but do not export, then our estimation sample may not be representative o the Danish manuacturing sector. Table A3 breaks down the trading irms (i.e. the irms that import/oshore, or export, or both) in the ull sample (as deined or Table A2 above) by employment and trade categories, and shows the shares in employment (upper panel), output (middle panel) and import value (lower panel). Table A3 shows that oshoring only irms account or 2% o

6 employment,.6% o output and 0.7% o import value among all trading irms; in comparison, the irms that both oshore and export account or 86.6% o employment, 90% o output and 99% o import value. These results suggest that the oshoring only irms are a small raction o our Danish data. In Table A4, we present the results o within job spell wage regressions with the oshoring only irms added into the sample (columns 4). The number o observations (.98 million) is very similar to our estimation sample (.93 million, as in Table 5), consistent with the results o Table A2. Since the oshoring only irms have export values o 0 we replace the two instrumented export variables in our main estimation with the un instrumented variables o log(export value + ) (columns and 3), and export values as a share o output (columns 2 and 4). The coeicient estimates or oshoring and its interaction with high skill dummy are similar to Table 5. In Table A4 we also present the results o within job spell wage regressions or a balanced panel o irms that are in the sample in all years (columns 5 8). The balanced panel sample has 40%, or roughly 800,000, ewer observations than in our estimation sample. Despite this reduction in sample size, the coeicient estimates or oshoring and its interaction with high skill dummy are again similar to Table 5. However, the coeicient estimates or log exports are about twice their size than in Table 5, and there is a larger dierence between the eects o exports on high and low skill wages, and a smaller dierence or high and low skill earnings. These are likely because we are unable to it the log export regression as well in the irst stage IV. To be speciic, the WID instrument or export has a smaller coeicient estimate than in Table 4 and is not statistically signiicant when irm controls are included. When irm controls are not included, the WID instrument has a similar coeicient estimate to Table 4 but the WES instrument has a larger coeicient estimate. In addition, the F statistics or the instruments are also smaller than in Table 4 (2.59 and 0.39, respectively, vs and in Table 4). These are likely due to the reduction in sample size. The other irst stage IV regressions are similar to Table 4 (the irst stage IV results are available upon request). 5. Additional Robustness Exercises To save space, we show the ollowing two robustness exercises in Table A5. The other robustness exercises mentioned in note 34 o the text are available upon request. In the text, we have emphasized narrow oshoring (imports purchased in same industry categories as the irm s sales) because these are more likely to be inputs the irm could have produced itsel. In our next robustness check we use broad oshoring (all import purchases by the irm) instead. We ind much larger eects o oshoring on wages, and more pronounced dierences across skill types. A possible explanation is that broad oshoring includes inputs o all types and is thereore more likely to capture the

7 eect o technological change operating through imports o machinery. Further, the estimation with irm controls yields a much larger wage drop than the estimation without irm controls. This is consistent with the view that the productivity eect, as distinct rom the labor substitution eect, can be seen more clearly when imported inputs are dierent rom those made by the irm. Next, one may be concerned that irm level shocks originating within Denmark may have general equilibrium consequences or product prices in exporting and importing partners i Denmark is responsible or a large share o trade. We experiment with dropping trade lows where Denmark is responsible or more than % o trade with that partner and product. Results are very similar.

8 Theory Appendix Generalizing the Production Function F To generalize our production unction, equation (), we have Y A K C, where = jt jt jt jt jt jt jt C L M,,2,,F index types o labor, /, and F In words, the production unction is Cobb Douglas in capital (whose share is α) and composite inputs C (whose share is α ). Each composite input C is produced with imported inputs M and type labor L using CES technology with the substitution elasticity σ >. σ may vary across labor types. Each labor type can be a skill group or an occupation, and dierent labor types enter into the production unction symmetrically. We irst show that (A) ln C c ln M ( c )ln L c jt o jt o jt where c 0, c are constants and 0 < c 0 <. Proo Drop the subscripts j,, and t, and let y = ln(l/m). Then C = M ( e ) and lnc = lnm + g(y), y where g(y) = ln( e ). The irst order Taylor approximation or g(y) is g(y) = g(y 0 ) + g (y 0 ) (y y 0 ), y where y 0 is a constant, and g (y 0 ) = y0 e y0 e lies between 0 and or all values o y 0. Let c 0 = g (y 0 ) and c = g(y 0 ) y 0 g (y 0 ) and we have equation (A). QED. Similar to equation (2) in our paper, the marginal product o type labor is F ( ) jt jt jt jt 2 jt MPL A K L C C. Taking the log o MPL and using equation (A) we obtain ln MPL ln[( ) A K L ] c ln L ( ) c ln L F jt jt jt 2 0 jt 0 jt F [( )( c ) ( c )]ln M jt.

9 I c 0 = c 0 or all = 2,..,F, the coeicient or lnm jt in the expression or lnmpl simpliies to F [ ]( c ) ( )( c ), where the equality uses F. Thereore, an increase in M jt increases the demand or type (type ) labor i /σ α < 0 (/σ α < 0). This condition is analogous to what we have in section II.A. Since σ diers across labor types, this condition also suggests that an increase in M jt may increase the wage or some labor types (those with small σ ) but decrease the wage or the other types (those with large σ ). Derivation o Equation (3) Assume that irm j aces the ollowing supply curve or unskilled labor, (A2) w, ( ) L S L jt c L jt, where w L, jt is the unskilled labor wage or irm j in year t and Ls is the elasticity o supply or unskilled labor. Equations (2) (in the text) and (A2) imply that the response o unskilled wages to oshoring (holding output constant) is (A3) b ln w ( / ) c, L, jt 0 L, S LM, Kconstant ln M jt L, S L, D where c 0 (0,) is a constant and, 0 is the elasticity o labor demand implied by equation (2). LD Equation (A3) says that blm, 0 i / ( ), which is the same condition under which oshoring lowers labor demand. I labor supply is perectly elastic, Ls, then shocks to labor demand will result in employment changes but not wage responses. A similar demonstration shows that oshoring raises skilled labor wages and exporting raises wages or both skilled and unskilled workers. To derive equation (3), assume that each unskilled worker i has productivity h ijt in year t and h exp( x ), where x it represents observable worker characteristics (e.g. experience), is a ijt it ij vector o coeicients, and ij represents unobservable ability that is speciic to the worker irm match. Unskilled workers are the same up to the productivity term, so that worker i receives wage (A4) wl, ijt wl, ithijt. A similar expression governs high skill labor wages. Then it is straightorward to derive equation (3) by solving or log w Lijt, using equations (A2), (A4) and equation (2) in the text.

10 When the production unction has multiple types o labor we can carry out the analyses in the same way. In particular, to derive the counterpart o (A3), let, S > 0 be the labor supply elasticity or type labor. Then the wage elasticity or type labor is ln w ( ) c, ln M, jt 0, S K and L constant jt, S, D where [ ( c )( )] 0 is the demand elasticity or type labor and c is as deined in 0, D 0 equation (A). This expression is analogous to (A3). The Productivity Eect Finally, we use Figure A2 to illustrate the eects o oshoring on unskilled wage, with and without the productivity eect. LS is the supply curve or unskilled labor. Suppose that unskilled labor and imported inputs are highly substitutable; i.e. /( ). An increase in oshoring shits the unskilled labor demand curve rom LD 0 to LD, holding constant physical capital, K jt. This is the direct wage eect o oshoring. As the increase in oreign inputs makes the irm more proitable and the irm increases the use o all inputs in response, there is a secondary shit o the unskilled labor demand curve, rising rom LD to LD 2. This is the productivity eect o oshoring and it tends to increase unskilled wage. I the direct eect dominates the productivity eect, LD 2 lies between LD and LD 0. We now calculate the wage elasticity o unskilled labor inclusive o the productivity eect. We assume that irm j takes the rental rate or capital, r t, as given, and that irm j increases capital input, K jt, until its marginal revenue product equals the rental rate r t, or that r A K H C t jt jt jt jt jt, which ln K jt lncjt implies that c0 0, where 0 < c 0 < is the same as deined in (A3). ln M ln M jt jt Using this expression and equation (2) we can show that b ln w c, ln ( ) * jt 0 LS, LM, * M jt L, S L, D where c is the elasticity o unskilled labor demand inclusive o the productivity eect. Comparing * 0 LD, 0 this expression with (A3) we show that b L,M < b L,M *; i.e. the productivity eect tends to increase the wage or unskilled labor. Instrumental Variables Strategy Index Danish irms by j, years by t, exporting countries by c, products (measured at the HS 6 level) by k, and destination countries other than Denmark by d. For ease o exposition, assume that irm j only

11 imports a single product k rom a single destination country c (the case o multiple product x country is similar). Firm j s production unction is given by equation () in the text, but we re write irm j s imported inputs as. j (A5) M b q ( ), /, jt ck ck ck In equation (A5), ck is a preerence shiter or a given exporter product that is common across importers. This can be thought o as quality, or in cases where there are multiple irms providing product k in exporter c, this can be interpreted as variety. In addition, there are preerence weights, each irm j. As we note in Section I.C, a eature o the Danish data is that or a given irm j, a small number o inputs, and the modally occurring case is that or a given c,k, j b ck, that are idiosyncratic to j bck is positive or j bck is positive or only one Danish irm. By the production unction () and equation (A5), the value o irm j s imported inputs equals (A6) j j j bcket j j p q m ( ) ( p ) ( ), P b p w j P Constant c-k Supply Denmark t t ck L, jt char. Trans. Cost Firm-j Demand char. In equation (A6) j E t is the output value o irm j, which relects demand or j s outputs, and j Pt represents the CES price index or j s imported inputs and low skilled wage. Equation (A6) says that imports o a particular input ck can rise because there is a shock to supply characteristics (price, quality, variety), shocks to transport costs, or because overall demand or inputs rises. We want to isolate the component o the irm s changing import demand that arises rom shocks to supply characteristics or to transport costs. We will measure transportation costs directly. We next discuss how we identiy shocks to supply characteristics. One possible approach is to use highly disaggregated gravity equations to separately identiy shocks to supply and demand characteristics. This is similar in spirit to Redding and Venables (2004), except that they use aggregate bilateral trade in a single cross section and extract aggregate supply characteristics by exporter country. To see how this approach works, suppose that input demand in the rest o the world is o a similar orm to demand in Denmark, though we write preerence weights as idiosyncratic to a destination d. Imports into destination d are then

12 (A7) m b E ( )( )( ), dl, dl, d d ck t p l dl, P Constant c-k Supply country-d t char. Trans. Cost country-d Demand char. where the summation is over the irms l in destination country d. Assume that there is no idiosyncratic component to demand (i.e. i b b c), then equation (A7) can be rewritten as d kt d, l, (A8) ln d d ln d m kt, where ln p dl, d d Et and kt bkt. l dl, P In principal, one could estimate (A8) or every product and time period, using vectors o origin and d destination ixed eects to capture the supply characteristics and demand characteristics kt. Ater netting o the demand and trade cost components, one collects a vector o ixed eects t and expressing over time changes or a given country x product, we have shocks to exporter supply. We do not pursue this approach or two reasons. One, this decomposition requires that demand and supply shocks be linearly separable, which is not the case i destination d has unusually strong preerences or the output o origin c, i.e. i d b varies over exporters c. Our data or Denmark show that the idiosyncratic component to demand is a salient eature o the data. That is, two Danish irms within the same industry purchase very dierent input bundles. Adding the additional dimensions o variation across importing country and including dierences in cross industry composition likely make this idiosyncratic component to demand even more important. Two, while Redding and Venables (2004) employs aggregate bilateral trade in a single cross section, we need the supply characteristics or each exporting country by HS6 product (over 5000 in total) in each year. The typical exporter ships an HS6 product to a small and changing number o destinations each year (i.e. the large majority o bilateral trade lows are zero or a given HS6 product). This means that the average conditional value o bilateral trade,, is extremely sensitive to variation in the number o destinations, itsel endogenous to the time varying supply and demand characteristics o interest. Tackling these two issues is beyond the scope o our paper and so we employ a dierent approach. As we discussed in the text, our approach is based on the variable WES (world export supply). Using (A8) we can derive the expression or WES by summing import demands over all destinations worldwide other than Denmark,

13 (A9) b E WES m ( ) ( p ) ( ), dl, dl, d d ck t ddenmark ddenmark l dl, P Constant c-k Supply t char. Rest o World Trans. Cost and Demand Char. WES can change over time because o changes in supply, trade costs, or demand. (A9) and (A6) imply that: (A0) b E b E log m log ln ln( ) ln{ ( ) }. j j d, l d, l j ck t d ck t WES j ddenmark l d, l Pt Pt Firm-j speciic unobserved variables. Equation (A0) motivates the speciication o our irst stage IV regression. Using irm j characteristics as controls we regress irm j s imports o inputs on WES and transportation costs, hoping to capture movements in supply characteristics, p in exports or exporter c and product k over time, and movements in, that are invariant to destinations and give rise to changes. (When ocusing on instruments or exporting at the irm level, we ollow a similar strategy. That is, we sum imports over all sources or a given destination to capture movements in demand that are invariant to sources and give rise to changes in imports or a given importer c and product k over time.) We can now use the explicit orm o the imports expression to describe threats to identiication. The main concern or our instruments is that a worldwide shock to demand or product k will simultaneously aect world export supply and input demand or Danish irm j; i.e. in equation (A0), the unobserved variable is correlated with WES and the irm j speciic variables. Further, i there is overlap between the inputs that irm j uses and the products that it sells, this demand shock also aects the desirability o irm j s products and thereore labor demand, and the wage. For example, rising demand or consumer electronics aects both cell phone manuacturers and producers o memory chips, and potentially the wages o workers employed by cell phone makers. We have addressed these concerns in the text. To recap:. We experiment with omitting industries with obvious shocks to demand occurring in this period (housing, electronics) and obtain similar results. 2. We incorporate an industry time ixed eect in all the wage speciications. Any time varying shock to demand at the industry level (e.g. electronics) is absorbed by the ixed eect, leaving only shocks that are idiosyncratic to irms. 3. We incorporate irm level exports, instrumented by world import demand in the wage regressions. This variable is used speciically to address the possibility that there might be idiosyncratic shocks to product demand aecting the wage equation. World import demand, constructed in a manner

14 symmetric to world export supply, captures movements in demand in the destination country that are invariant to sources. I there is a common shock to demand, it will aect WID, and exports. This control or shocks to demand is signiicant in the Danish context, where the average irm exports 45% o its output. 4. We experiment with excluding rom our estimates any exporter products where Denmark represents a large share o worldwide demand or the product. Because Denmark is small, we obtain similar results.

15 Figure A Fuel prices over time year crude jetuel Figure A2. The Eects o Oshoring on Unskilled Wage Wage LD LD 2 LD 0 LS Direct Eect, Holding K jt constant Productivity Eect, K jt increases Employment

16 Online Appendix or The Wage Eects o Oshoring: Evidence rom Danish Matched Worker-Firm Data, by David Hummels, Rasmus Jørgensen, Jakob Munch, and Chong Xiang, September 203. Table A The Top 20 HS6 Products in HS 84 (Machinery) Product Cumul HS6 Description Share Share GEARS; BALL OR ROLLER SCREWS; GEAR BOXES, ETC 9.8% 9.8% 8439 PARTS OF PUMPS FOR LIQUIDS 8.8% 8.6% TAPS COCKS ETC F PIPE VAT INC THERMO CONTROL NESOI 6.3% 24.8% SPARK IGNITION RECIPROCATING INT COM PISTN ENG PTS 5.3% 30.% PTS F TAPS ETC F PIPE VAT INC PRESS & THERMO CNTRL 4.3% 34.4% ENGINE AND MOTOR PARTS, NESOI 3.2% 37.6% MARINE COMPRESS IGNIN COMBUSTION PISTON ENGINE ETC 2.2% 39.8% CENTRIFUGAL PUMPS, NESOI 2.2% 4.9% REFRIGERATOR FREEZER AND HEAT PUMP PARTS NESOI.8% 43.7% BALL BEARINGS.8% 45.5% VALVES F OLEOHYDRAULIC OR PNEUMATIC TRANSMISSIONS.5% 47.0% PARTS FOR HARVESTER, GRASS MOWERS, SORTING EGG ETC.5% 48.5% PTS OF MACH/MECHNCL APPL W INDVDUL FUNCTION NESOI.4% 49.9% PARTS OF MACH OF CH 84, NESOI,IND PREP FOOD,DRINK.4% 5.3% MACH F MANUF OR FINISH NONWOVENS;HAT BLOCKS; PARTS.4% 52.7% PARTS AND ATTACHMENTS NESOI FOR DERRICKS ETC..3% 54.0% PARTS & ACCESSORIES FOR ADP MACHINES & UNITS.2% 55.3% MACH & MECHANICAL APPL W INDIVIDUAL FUNCTION NESOI.2% 56.5% COMPRESSORS USED IN REFRIGERATING EQUIPMENT.2% 57.6% MACHINE PARTS WITH NO ELECTRIC FEATURES NESOI.% 58.8%

17 Online Appendix or The Wage Eects o Oshoring: Evidence rom Danish Matched Worker-Firm Data, by David Hummels, Rasmus Jørgensen, Jakob Munch, and Chong Xiang, September 203. Table A2 Summary Statistics or the Full Sample and the Estimation Sample Full sample Estimation sample Obs Mean Std. dev. Obs Mean Std. dev. Hourly wage 2,800, ,950, Log hourly wage 2,797, ,950, Log gross output 2,797, ,950, Log employment 2,800, ,950, Log capital per worker 2,779, ,950, High skill 2,800, ,950, Experience 2,739, ,950, Union 2,79, ,950, Married 2,739, ,950, Notes: The ull sample is the collection o workers and irms in manuacturing we have beore we implement our sample selection criteria. The estimation sample is the one we use in our paper, and the summary statistics or the estimations sample are identical to those reported in Table. All variables are calculated over the distribution o worker year observations, which means irm characteristics such as output are repeated as many times as there are worker years or that irm.

18 Online Appendix or The Wage Eects o Oshoring: Evidence rom Danish Matched Worker-Firm Data, by David Hummels, Rasmus Jørgensen, Jakob Munch, and Chong Xiang, September 203. Table A3 Employment, Output and Import Value Shares by Trade and Employment Categories Number o Employees Trade categories Oshoring and exporting Exporting only Oshoring only % in employment % 6.7% 0.6% % 4.7%.4% 00% % o output % 4.7% 0.4% % 3.6%.2% 00% % o imports % 0 0.3% % 0 0.4% 00% Notes: The sample consists o all the trading irms in the ull sample. Trading irms either import/oshore, export, or both. The ull sample is the same as in Table A; i.e. the collection o irms we have beore we implement our sample selection criteria.

19 Data, by David Hummels, Rasmus Jørgensen, Jakob Munch, and Chong Xiang, September 203. Table A4 Wage Regressions with Oshoring Only Firms and or the Balanced Panel Sample Oshoring Only Firms Balanced Panel Dependent Variable: log hourly wage log hourly wage log hourly wage log labor income () (2) (3) (4) (5) (6) (7) (8) Log oshoring 0.038** 0.035** *** *** 0.079** *** *** *** [ 2.36] [ 2.32] [ 5.7] [ 5.69] [ 2.27] [ 3.32] [ 4.72] [ 4.98] Log oshoring x high skilled *** *** *** *** *** *** *** *** [.70] [.9] [0.49] [0.62] [6.99] [7.07] [5.09] [5.55] Log exports 0.09*** 0.082*** *** 0.099*** [5.] [6.85] [4.22] [7.90] Log exports x high skilled ** 0.035** [ 2.07] [ 2.] [0.62] [0.30] Log (exports + ) [.57] [0.80] Log (exports + ) x high skilled [.55] [.07] Share, exports/output 0.000** *** [ 2.57] [5.57] Firm Controls No No Yes Yes Yes No Yes No Observations,976,883,976,883,976,883,976,883,24,449,24,449,24,443,24,443 Number o job spell ixed eects 388, , , ,575 9,653 9,653 9,653 9,653 R squared Notes: Clustered (irm year) t statistics in brackets. *** p<0.0, ** p<0.05, * p<0.. In columns () (4) the sample is our main estimation sample plus the oshoring only irms. The variables log (exports + ), log(exports + ) x high skilled, and exports/output are not instrumented. Columns (5) (8) are the same speciication as Table 5, but the sample is a balanced panel o irms that are in the sample or the entire period.

20 Data, by David Hummels, Rasmus Jørgensen, Jakob Munch, and Chong Xiang, September 203. Table A5 Additional Robustness Exercises Dependent variable: Log Hourly Wage Log Hourly Wage Robustness check: I. Broad oshoring FE IV II. DK dominant lows removed (%) FE IV Log(oshoring) 0.064*** ** [ 2.58] [ 2.3] [ 2.59] [ 3.50] Log(oshoring) x high skilled 0.3*** 0.322*** *** *** [8.93] [9.32] [7.32] [7.35] Log(exports) *** *** *** *** [6.32] [5.] [2.88] [4.47] Log(exports) x high skilled *** *** [ 4.05] [ 3.8] [.03] [.7] First stage IV F statistics: log oshoring x high skill log exports x high skill Other irm level controls Yes No Yes No Obs,950,896,950,896,928,599,928,599 No. job spells 384, , , ,035 R Notes: Table A5 presents the results rom worker level Mincer regressions, using log hourly wage as the dependent variable. All speciications include job spell, industry year and regional ixed eects. Log oshoring, log exports and their skill interactions are instrumented using world export supply (WES), world import demand (WID) and transport costs. Robust T statistics in brackets. Standard errors clustered at irm year levels. *** p<0.0, ** p<0.05, * p<0..