Appendix. Sector classification: Description of the reclassification procedure

Transcription

1 Appendix. Sector classification: Description of the reclassification procedure Note: The first part of this appendix reproduces what is in Appendix to the paper. The second part (which is not part of the published paper) explains the procedures we used in checking for misclassification biases and discusses the results. Since in the ISSP-NI survey there is no direct question about industry, we infer sector of employment from data on occupation. We use individual answers to two questions in the data set, one asking for occupation according to an international code (the 4-digit International Standard Classification of Occupations (ISCO) from 968 and from 988) and another one asking for occupation in terms of national codes. Individuals in each country give information about own occupation according to only one of the classifications (either ISCO 968 or ISCO 988 or according to a national classification). In particular, individuals occupations from the following countries are coded according to ISCO 968: Germany West, Germany East, USA, Austria, Norway, Bulgaria, New Zealand, Spain, Slovak Republic. The occupation codes of this group of countries are recoded all together. Respondents occupations from the following countries are instead coded according to ISCO 988: Hungary, Ireland, Czech Republic, Poland, Slovenia, Canada, Russia, Latvia. Again, we recode the occupation codes of this group of countries all together. Finally, respondents occupations for Great Britain, Sweden, the Philippines, Italy, Netherlands and Japan follow national occupation codes. Data from Great Britain, Sweden and the Philippines are recoded individually. The national occupation codes for Italy, Netherlands and Japan cannot be reclassified, since they are not detailed enough. We reclassify the occupation variables from the ISSP-NI data set in order to match the coding in the World Trade Analyzer (WTA) data set, containing world trade flows from 980 to 997. To classify industries, the WTA uses a slightly modified version of the Standard International Trade Classification (SITC), Revision 2. However, in the WTA -ROM, information is also available in a different format. Data is organized according to the 34 manufacturing industry basis used by the U.S. Bureau of Economic Analysis (BEA). This coding is quite similar to the U.S. Standard Industrial Classification. The WTA -ROM includes the annual bilateral trade values between all countries in the world in according to this 34-industry classification. We use the BEA classification to recode the occupation variables in the ISSP-NI data set and construct a new variable indicating the individual sector of employment. The 34 industries (plus one Non-manufacturing recoded as 35) are listed below (Table ). In order to obtain a more precise match between the ISSP-NI occupation data and the BEA industry codes, we base the recoding on a very detailed description of the correspondence between BEA codes and SITC Revision 2 (four-digit level) codes (we used Appendix D: SITC Revision 2 codes used by Statistics Canada, WTA, and the corresponding BEA 34 manufacturing industry codes, from World Trade Flows, , by Robert Feenstra). In addition to the 35 BEA industry codes, we create new codes as combinations of the original 35 codes. This is necessary since the occupation codes used in the ISSP-NI data set are not always detailed enough to be matched to any single BEA code. See an extract from the 968 ISCO classification in Table 2 and corresponding BEA codes, assigned by us, as an example of the reclassification.

2 2 For each of the 35 original BEA industries, we consider sector-specific exports and imports. For each new code, exports (imports) are obtained as sum of exports (imports) of the sectors used in the combination (so, for example, exports of sector 36, which is the combination of sectors 7 and 8, are set equal to the sum of exports of sectors 7 and 8). We then average both exports and imports over the years

3 3 Table : BEA (Bureau of Economic Analysis) 34 manufacturing industry codes. Grain, Mill and Bakery Products 2. Beverages 3. Tobacco Products 4. Other Food and Kindred Products 5. Apparel and Other Textile Products 6. Leather and Leather Products 7. Pulp, Paper and Board Mills 8. Other Paper and Allied Products 9. Printing and Publishing 0. Drugs. Soaps, Cleaners, and Toilet Goods 2. Agricultural Chemicals 3. Industrial Chemicals and Synthetics 4. Other Chemicals 5. Rubber Products 6. Miscellaneous Plastic Products 7. Primary Metal Industries, Ferrous 8. Primary Metal Industries, Nonferrous 9. Fabricated Metal Products 20. Farm and Garden Machinery 2. Construction, Mining, etc. 22. Computer and Office Equipment 23. Other Nonelectric Machinery 24. Household Appliances 25. Household Audio and Video, etc. 26. Electronic Components 27. Other Electrical Machinery 28. Motor Vehicles and Equipment 29. Other Transportation Equipment 30. Lumber, Wood, Furniture, etc. 3. Glass Products 32. Stone, Clay, Concrete, Gypsum, etc. 33. Instruments and Apparatus 34. Other Manufacturing 35. Non Manufacturing (natural resources, )

4 4 Table 2: Extract from 968 International Standard Classification of Occupations 968 ISCO BEA code Agricultural, animal husbandry and forestry workers, fishermen and hunters 60 Farm managers and supervisors 6000 Farm managers and supervisors, Farm managers and supervisors (general), Farm managers, 4 6 Farmers 600 Farmers, 4 60 General farmers, 4 62 General farmers (general), 4 65 Collective farmers, Specialised farmers, 4 62 Agricultural and animal husbandry workers 6200 Agricultural and animal husbandry worker, General farm workers, 4 62 Farm helpers (general), Farm hand, Field crop and vegetable farm workers, Orchard, vineyard and related tree and shrub crop workers Livestock workers Dairy farm workers Poultry farm workers Nursery workers and gardeners Farm machinery operators, Agricultural and animal husbandry workers, n.e.c., 4 63 Forestry workers 6300 Forestry workers Loggers Logger Forestry workers (except logging) NT 64 Fishermen, hunters and related workers 6400 Fishermen, hunters and related workers Fishermen Fishermen, hunters and related workers, n.e.c. 4 Production and related workers, transport equipment operators and labourers 70 Production supervisors and general foremen 7000 Production supervisors and general foremen m.v. 700 Production supervisors and general foremen (general) m.v. 7 Miners, quarrymen, well drillers and related workers 700 Miners, quarrymen, well drillers Miners and quarrymen 35 7 Quarrymen (general) Cutting machine operators (mine) Miners and related workers n.e.c Mineral and stone treaters 35

5 5 730 Well drillers, borers and related workers Metal processors 7200 Metal processors 7,8 720 Metal smelting, converting and refining furnacemen 7, Metal rolling-mill workers 7, Metal melters and reheaters 7, Metal casters 7, Metal moulders and coremakers 7, Metal annealers, temperers and case-hardeners 7, Metal drawers and extruders 7, Metal platers and coaters 7, Metal processers, n.e.c. 7,8 73 Wood preparation workers and paper makers 7300 Wood preparation workers Wood treaters Sawyers, plywood makers and related wood-processing workers Sawmill sawyers (general) Paper pulp preparers Paper makers 7 74 Chemical processers and related workers 7400 Chemical processers and related workers Crushers, grinders and mixers Cookers, roasters and related heat-treaters Filter and separator operators Still and reactor operators Petroleum-refining workers Chemical processers and related workers, n.e.c Spinners, weavers, knitters, dyers and related workers 7500 Spinners, weavers and related workers Fibre preparers Spinners and winders Weaving- and knitting-machine setters and pattern-card preparers Weavers and related workers Knitters Bleachers, dyers and textile product finishers Spinners, weavers, knitters, dyers and related workers, n.e.c Tanners, fellmongers and pelt dressers 7600 Tanners, fellmongers and pelt dressers Tanners and fellmongers Pelt dressers 6 77 Food and beverage processers 7700 Food and beverage processers 2,4 770 Grain millers 7720 Sugar processers and refiners Butchers and meat preparers Food preservers Dairy product processors Bakers, pastrycooks and confectionery makers Tea, coffee and cocoa preparers Brewers, wine and beverage makers 2

6 Food and beverage processers, n.e.c Tobacco preparers and tobacco product makers 780 Tobacco preparers Cigar makers Cigarette makers Tobacco preparers and tobacco product makers, n.e.c Tailors, dressmakers, sewers, upholsterers and related workers 7900 Tailors, dressmakers, sewers and rel. workers Tailors and dressmakers Tailor Fur tailors and related workers Milliners and hatmakers Patternmakers and cutters Sewers and embroiderers Upholsterers and related workers Tailors, dressmakers, sewers, upholsterers and related workers, n.e.c Shoemakers and leather goods makers 8000 Shoemakers and leather good makers Shoemakers and shoe repairers Shoe cutters, lasters, sewers and related workers Leather goods makers 6 8 Cabinetmakers and related woodworkers 800 Cabinetmakers and related woodworkers Cabinetmakers Woodworking-machine operators Cabinetmakers and related woodworkers, n.e.c Stone cutters and carvers 8200 Stone cutters and carvers 32

7 7 Appendix (cont.): Misclassification errors in observed comparative-advantage and comparative-disadvantage status The ultimate goal of reclassifying the occupation variables into a sector variable is to infer information on the comparative-advantage and comparative-disadvantage status of each individual. 2 In this appendix, we show that the occupation-based sector classification delivers a fairly good measure of the latter two variables. Indeed, for approximately 85% of individuals in three countries, using our reclassification on the occupation data we get the same information on comparative-advantage and comparative-disadvantage status as if we had used direct data on sector. We next adjust coefficient estimates for these three countries, taking into account the attenuation bias introduced by misclassification. Consider the following linear model of attitudes toward trade 3 : Trade Opinion = const. + X β + β + β + ε, () X 2 where X is a vector of observed covariates for individual i in country k, and:, if = 0, if M M < 0 for traded sector > 0 for traded sector or if non - traded sector, if = 0, if M M > 0 for traded sector < 0 for traded sector or if non - traded sector The variable ( ) represents the true comparative-advantage (comparativedisadvantage) status of the sector where individual i in country k works (true in the sense that it is based on direct information about the sector of employment). and are unobserved, since no information on individuals' sector of employment is included in the ISSP-NI data set. j j Let's assume that: E( ε ) = E( ε ) = E( ε ) = E( ε X ) = 0, where X refers to the jth covariate in X. While the true comparative-advantage (comparative-disadvantage) status cannot be observed, we can observe and, which are only imperfectly correlated with and. Let the observed comparative-advantage status be defined as follows: 4 We thank Jörn-Steffen Pischke and an anonymous referee for suggestions that prompted this analysis. 2 In other words, our goal is not to get the exact industry right for each respondent, but his comparative-advantage (comparative-disadvantage) status. 3 This analysis is based on Card (996)'s model of measurement errors. 4 is defined in analogous way, with parameters q and q 0.

8 8, =, with prob. q with prob. q 0, when < q =, when = 0 In other words, ( q ) represents the false negative rate, while q 0 is the false positive rate. 5 If we label the true fraction of individuals with comparative-advantage status π = E( ) = prob( = ), then the observed fraction of individuals with comparativeadvantage status is equal to p = E( ) = prob( = ) = = π q + ( π ) q0. Let's assume that the true comparative-advantage status is only a function of the observed comparative-advantage status: 6 = γ 0 + γ + η, with E( η ) = 0 (2) then substituting (2) into (), we obtain: Trade Opinion const. + X + β γ + β2γ + ( ε + βη + β η ) (3) = β X 2 The regression of Trade Opinion on const, X, and will give a consistent estimate of β γ, as the coefficient on. Thus γ represents the attenuation coefficient. In order to calculate the attenuation coefficients for the comparative-advantage and comparative-disadvantage variables, we use data from another ISSP data set, Social Inequality II (992), which contains information on both individuals' sector of employment and occupation. 7 In particular, for the United States, Social Inequality II has data on: individual sector of employment, according to the 980 Industrial Classification, and individual occupation, according to the digit International Standard Classification of Occupations ISCO. It is straightforward to aggregate the 980 Industrial Classification codes into the BEA industry codes, according to which data on sector exports and imports are presented in the World Trade Analyzer data set. In our empirical work using the ISSP-NI data set, we wrote a Stata program to recode the individual-occupation variable into an individual-sector variable (based on the BEA classification). Next we run this Stata program using data from Social Inequality II and construct the two observed variables and : 5 Note that E( ε ) = E( ε ) = 0 (see Card 996). 6 2 If the covariates in X have a significant impact on controlling for, and the R from a linear probability model of observed status on X is not low, the attenuation bias would be more pronounced. We don't find evidence of this in our empirical work. 7 We looked carefully at other ISSP data sets. Unfortunately, they either only have information on occupation or, if they have sector data, it is not detailed enough or, they cover more or less the same countries as Social Inequality II.

9 9, if M < 0, for traded occupation/sector = 0, if M > 0 for traded occupation/sector or if non - traded occupation/sector, if M > 0, for traded occupation/sector = 0, if M < 0 for traded occupation/sector or if non - traded occupation/sector where occupation/sector is the sector based on occupation data. We can then compare to the true variables and and. In particular, given all those individuals for which both sets of variables are defined, we can calculate the frequencies (in % terms) for both comparative-advantage status and comparative-disadvantage status. Note that we have information on industrial sector and occupation also for Germany and Austria. 8 We repeat our exercise focusing on these additional countries. Table 3a: U.S. (in %) 0 0 ( q0 )( π ) =79.38 q0 ( π ) =0.02 ( q ) π =4.56 q π =6.04 Table 3b: U.S. (in %) 0 0 ( q0 )( π ) =83.37 q0 ( π ) =3.53 ( q ) π =8.20 q π =4.90 Table 4a: Germany (in %) There are two industrial sector variables in Social Inequality II: Industrial sector I (V30) and Industrial sector II (V3).While the latter one is available for a much larger number of countries, it cannot be used for our purpose, since it is not detailed enough. Therefore we use Industrial sector I.

10 0 Table 4b:Germany (in %) Table 5a: Austria (in %) Table 5b: Austria (in %) Note that the sum of the two percentages in the main diagonal of each table (in bold) give the fraction of times in which true and observed variables are equal to the same value. Based on these percentages, we believe that the occupation-based sector classification delivers a fairly good measure of comparative-advantage and comparative-disadvantage status, for the following reasons:. The fraction of times in which "true" and "observed" variables are equal to the same value is quite high (across the three countries, it ranges between approximately 85% and 87% for comparative advantage, and between approximately 82% and 88% for comparative disadvantage); 2. In this analysis we stack the cards against ourselves, i.e. we assume that the industrial-sector information is "true" and that the occupation-based sector information is measured with error. However, it is lely that both sets of data are measured with error. In this case, the misclassification rates should be divided between the two classifications. 3. It is lely that most of the discrepancies originate on the industrial-sector classification side. The latter one is much less detailed than the occupation-based sector classification. The extract from the 968 ISCO classification in Table 2 shows the extent of detail of this classification. The estimated coefficients γ regression of on, are as follows: 9, from the regression of on, and γ, from the 9 The analysis for each country is based on country-specific industrial classifications, which naturally introduces cross-country variation in the magnitude of the attenuation coefficients.

11 Table 6: Attenuation coefficients γ U.S (.0262) Germany (.084) Austria (.039) γ (.0375) (.0292) (.0542) We can use these attenuation coefficients to adjust the coefficients, on comparativeadvantage and comparative-disadvantage status, for attenuation. 0 Table 7: Coefficient estimates Table 8: Coefficient estimates adjusted for attenuation U.S (.0966) Germany (0.365) Austria (.777) β γ 2 γ β (.446) (.353) (.587) U.S (.3004) Germany (.4532) Austria (.0453) β β (.2942) (.2320) (.2293) In this analysis we assumed a measurement-error model as in Card (996). One crucial assumption in this model is that the error term in (2) is uncorrelated with the observed comparative-advantage (comparative-disadvantage) status, i.e. E( η ) = 0 and E( η ) = 0. The most lely situation in which this assumption is violated is when the true variables are also a function of the covariates in X (Card 996, p. 959). As pointed out in footnote 5, if the covariates in X have a significant impact on, controlling for, and the 2 R from a linear probability model of observed status on X is not low, the attenuation bias would be more pronounced. However, we don't find evidence of this in our empirical work. Notice that, with this analysis, it is not possible to evaluate whether the attenuation bias varies across countries, since in Social Inequality II the individual-sector variable is based on country-specific industrial classifications. This naturally introduces cross-country variation in the magnitude of the attenuation coefficients. 0 The regressions in the following two tables include, as covariates: age, male, citizenship, years of education. Standard errors are in parentheses. significant at %; significant at 5%; + significant at 0%. Standard errors of coefficient estimates in Table 8 are derived from standard errors in Table 7 under the assumption that the attenuation coefficients are known.