61
APPENDIX A – Tables & Figures
65 Table A.3.1: Industrial Composition of Sampled Firms
NACE Rev. 2 Industry Codes Sample* Netherlands**
C Manufacturing Nr. Pct. Nr. Pct.
C10 Manufacture of food products 137 14,56%
4500 9,71%
C11 Manufacture of beverages 1 0,11%
C12 Manufacture of tobacco products 7 0,74% 20 0,04%
C13 Manufacture of textiles 22 2,34%
2680 5,78% C14 Manufacture of wearing apparel 7 0,74%
C15 Manufacture of leather and related products 5 0,53% 210 0,45% C16 Manufacture of wood and of products of wood and
cork, except furniture; manufacture of articles of straw and plaiting materials
18 1,91% 2005 4,32% C17 Manufacture of paper and paper products 37 3,93% 425 0,92% C18 Printing and reproduction of recorded media 46 4,89% 6485 13,99% C19 Manufacture of coke and refined petroleum
products 7 0,74% 35 0,08%
C20 Manufacture of chemicals and chemical products 102 10,84%
815 1,76% C21 Manufacture of basic pharmaceutical products and
pharmaceutical preparations 7 0,74%
C22 Manufacture of rubber and plastic products 49 5,21% 1010 2,18% C23 Manufacture of other non metallic mineral
products 39 4,14% 1615 3,48%
C24 Manufacture of basic metals 20 2,13% 290 0,63% C25 Manufacture of fabricated metal products, except
machinery and equipment 111 11,80% 7825 16,88%
C26 Manufacture of computer, electronic and optical
products 65 6,91%
18450 39,79% C27 Manufacture of electrical equipment 32 3,40%
C28 Manufacture of machinery and equipment n.e.c. 128 13,60% C29 Manufacture of motor vehicles, trailers and
semi-trailers 18 1,91%
C30 Manufacture of other transport equipment 35 3,72%
C31 Manufacture of furniture 26 2,76%
C32 Other manufacturing 22 2,34%
C33 Repair and installation of machinery and
equipment 0 0,00% 941 100,00 % 46365 100,00 %
** Indicates number of and percentage of firms per manufacturing industry, for the entire 2007 sample. ** Indicates number of and percentage of firms per manufacturing industry, for the entire population of
66 Table A.3.2: Overview of main variables
Variable Name (type) Description
Dependents: productivity
LP Labour productivity: operating revenue per employee LPVA
Labour productivity: value added per employee, in which value added is: Profit for period + Depreciation + Taxation + Interests paid + Cost of employees
Independents: structural firm characteristics
TAS Total Assets, Ths. EUR
TFASpEMPL Tangible Fixed Assets per employee, Ths. EUR
AGE Age of the firm (2007)
AGE2 Age of the firm (2007) - squared
CGMpEMPL Cost of goods sold (incl. depreciation of those costs), or cost of material inputs to production, per employee, Ths. EUR CEMPLpEMPL Cost of all employees (incl. pension costs), per employee,
Ths. EUR
independents: controls of systemic differences
IND10 - IND33 23 dummies, 1 per 2-digit industry code, NACE Rev. 2 C10-C33
independents: multinational status
FOR More than 50% of shares in hands of foreign shareholder(s) MNE * Dummy, firm owns foreign subsidiaries
FORMNE Dummy, foreign owned firms owns foreign subsidiaries DOMMNE Dummy, domestic owned firm owns foreign subsidiaries FORNMNE Dummy, foreign owned firms owns no foreign subsidiaries DOMNMNE Dummy, domestic owned firm owns no foreign subsidiaries MNEC0 - MNEC2 * 3 dummies on ownership of foreign subsidiaries *
FORMNEC0 - FORMNEC2 3 interaction dummies on ownership of foreign subsidiaries by foreign owned firms DOMMNEC0 - DOMMNEC2 3 interaction dummies on ownership of foreign subsidiaries by domestic owned firms NFC0 - NFC02 * 3 dummies on ownership of foreign subsidiaries *
FORNFC0 - FORNFC02 3 interaction dummies on ownership of foreign subsidiaries by foreign owned firms DOMNFC0 - DOMNFC02 3 interaction dummies on ownership of foreign subsidiaries by domestic owned firms
US Foreign owner resides in the US
EU Foreign owner resides in an EU-member country
Notes:
67 Table A.3.3: MNE-status classifications
MNE MNEC NFC
MNE = 0 NF = 0 MNEC0 = 1 NF = 0 NFC0 = 1 NFC = 0
MNE = 1 NF > 0
MNEC1 = 1 NF <= ND NFC1 = 1 NFC = 1 MNEC2 = 1 NF > ND NFC2 = 1 NFC > 1 - NF indicates number of majority-owned foreign subsidiaries of parent
- ND indicates number of majority-owned domestic subsidiaries of parent
- NFC indicates the number of foreign countries in which a parent has majority owned subsidiaries
Table A.3.4: Unknown Ownership / Subsidiary data
Total Sample(a) Sub-sample 1(b) Sub-sample 2(c) Sub Sample 3(d) Sub Sample 4(e)
number percent
of total number
percent
of total number percent of total number
percent
of total number percent of total
total 941 822 749 790 822
MNENK(1) 394 41.87% 339 41.24% 288 38.45% 333 42.15% 0 0.00%
OWNNK(2) 219 23.27% 165 20.07% 139 18.56% 156 19.75% 0 0.00%
ININFO(3) 532 56.54% 448 54.50% 389 51.94% 434 54.94% 0 0.00%
NOINFO(4) 451 47.93% 392 47.69% 351 46.86% 379 47.97% 0 0.00%
(1) MNENK : ownership of subsidiaries not known (2) OWNNK : nationality of the parent's owner not known
(3) ININFO : either nationality of the parent's owner or ownership of subsidiaries not know (4) NOINFO : neither nationality of the parent's owner nor ownership of subsidiaries know
(a) total sample of 941 firms
68 Table A.3.5: Firms per industry and ownership/subsidiary segment
NACE Rev. 2 Industry Codes
DOMNMNE FORNMNE DOMMNE FORMNE
C Manufacturing
C10 Manufacture of food products 11 4 30 15
C11 Manufacture of beverages 0 0 0 0
C12 Manufacture of tobacco products 0 2 0 3
C13 Manufacture of textiles 3 1 8 3
C14 Manufacture of wearing apparel 0 1 2 2
C15 Manufacture of leather and related
products 0 0 0 2
C16 Manufacture of wood and of products of wood and cork, except furniture; manufacture of articles of straw and plaiting materials
2 0 0 0
C17 Manufacture of paper and paper
products 8 1 4 3
C18 Printing and reproduction of recorded
media 4 1 13 2
C19 Manufacture of coke and refined
petroleum products 0 1 1 2
C20 Manufacture of chemicals and
chemical products 2 7 17 23
C21 Manufacture of basic pharmaceutical products and pharmaceutical
preparations
1 1 1 0
C22 Manufacture of rubber and plastic
products 2 2 8 8
C23 Manufacture of other non metallic
mineral products 4 7 5 6
C24 Manufacture of basic metals 0 1 2 7
C25 Manufacture of fabricated metal products, except machinery and equipment
3 4 17 13
C26 Manufacture of computer, electronic
and optical products 5 1 7 16
C27 Manufacture of electrical equipment 2 3 8 4
C28 Manufacture of machinery and
equipment n.e.c. 10 7 24 23
C29 Manufacture of motor vehicles, trailers
and semi-trailers 0 2 3 3
C30 Manufacture of other transport
equipment 5 1 4 3
C31 Manufacture of furniture 0 0 4 4
C32 Other manufacturing 0 0 0 0
C33 Repair and installation of machinery
and equipment 0 0 0 0
62 47 158 142
** Indicates number of and percentage of firms per manufacturing industry, for the entire 2007 sample.
69 Table A.3.6: Firm counts in (sub-) sample(s) per segment
Full Sample(a) Sub-Sample 1(b) Sub-Sample 2(c ) Sub-Sample 3(d) Sub-Sample 4(e ) Included 941 822 749 790 822 Excluded . 119 192 151 0 MNE 380 342 325 326 342 NMNE 167 141 136 131 480 FOR 346 321 301 312 321 DOM 376 336 309 322 501 DOMNMNE 62 53 50 51 289 FORNMNE 47 42 42 39 191 DOMMNE 158 149 145 142 212 FORMNE 142 130 123 124 130 MNEC0 167 141 136 131 480 MNEC1 126 110 103 104 110 MNEC2 254 232 222 222 232 DOMMNEC0 62 53 50 51 289 DOMMNEC1 53 49 47 47 50 DOMMNEC2 105 100 98 95 99 FORMNEC0 47 42 42 39 191 FORMNEC1 35 33 32 31 33 FORMNEC2 107 97 91 93 97 NFC0 167 141 136 131 480 NFC1 141 127 118 123 127 NFC2 239 215 207 203 215 DOMNFC0 62 53 50 51 289 DOMNFC1 53 51 48 49 51 DOMNFC2 105 98 97 93 98 FORNFC0 47 42 42 39 191 FORNFC1 55 50 48 49 50 FORNFC2 87 80 75 75 80
(a) total sample of 941 firms
(b) excludes all firms reporting employment under 5 (c) excludes all firms reporting employment under 20
70 Figure A.4.2: Cumulative distribution plot of Sub-Sample 1
0 .2 .4 .6 D e n s it y 0 5 10 15 20 lnLPN
Kernel density estimate kdensity lnLP for DOMNMNEs kdensity lnLP for FORMNEs kdensity lnLP for DOMMNEs
Figure A.4.3: Cumulative distribution plot of Sub-Sample 1
0 .2 .4 .6 .8 1 C u m u la ti v e O b s e rv a ti o n s 0 5 10 15 20
Labour Productivity, natural logarithm (lnLP)
cDOMNMNE (53 obs) cDOMMNE (149 obs)
71 Table A.4.4: Main Results on Labour Productivity – dismissed model (Sub-Sample 1)
Dependent: lnLP # Independents (1) (2) (3) (4) (5) 1. lnTAS 0.07373*** 0.03446 0.02961 0.02961 0.02376 (0.02346) (0.03180) (0.03180) (0.03180) (0.03153) 2. lnTFASpEMPL 0.34509*** 0.51488*** 0.52445*** 0.52445*** 0.52774*** (0.04603) (0.07229) (0.07316) (0.07316) (0.07320) 3. AGE -0.00198** -0.00095 -0.00263*** -0.00263*** -0.00254*** (0.00078) (0.00244) (0.00080) (0.00080) (0.00078) 4. AGE2 -0.00001 (0.00001) 5. FOR 0.09749 0.49510** (0.11661) (0.19336) 6. MNE -0.42442*** -0.17026 (0.12950) (0.17270) 7. FORMNE -0.52824** -0.20340 (0.24205) (0.18445) 8. DOMMNE -0.17026 (0.17270) 9. FORNMNE 0.49510** (0.19336) 10. DOMMNEC1 -0.28074 (0.21274) 11. DOMMNEC2 -0.10674 (0.18676) 12. FORMNEC0 0.50138** (0.19493) 13. FORMNEC1 -0.24367 (0.20043) 14. FORMNEC2 -0.17660 (0.20403)
15. IND controls yes yes yes yes yes
17. Constant 7.68142*** 6.20273*** 6.19505*** 6.19505*** 6.25977*** (0.60406) (0.93156) (0.93368) (0.93368) (0.91600) Observations 782 363 363 363 363 R-sqrd 0.33951 0.45768 0.46362 0.46362 0.46516 R-sqrd (adj.) 0.31767 0.41571 0.42212 0.42212 0.42032 P > F # 1-4 0 0 0 0 0 P > F # 5-14 n.a. 0.00213 0.000223 0.000223 0.000945 P > F # 15 4.08e-06 1.93e-09 4.86e-07 4.86e-07 4.67e-07 P Reset(1) 0.000536 0.0662 0.0517 0.0517 0.0594 P Reset(2) 7.79e-08 0.000591 0.000154 0.000154 0.000101
P Jarque-Bera 0 0 0 0 0
72 Table A.4.7: Correlation Matrix 1 Sub-Sample 1
(obs=513) 1 2 3 4 5 6 7 1 lnLP 1 2 lnTAS 0,2693 1 3 lnTFASpEMPL 0,4837 0,4142 1 4 lnCGMpEMPL 0,9176 0,2664 0,467 1 5 lnCEMPLpEMPL 0,7054 0,1518 0,4649 0,6105 1 6 AGE -0,0303 0,1689 0,0836 -0,0625 0,0162 1 7 AGE2 -0,0419 0,1082 0,0674 -0,0851 -0,0086 0,8819 1
Table A.4.8: Correlation Matrix 2 Sub-Sample 1
73 Table A.4.9: Correlation Matrix 3 Sub-Sample 1
(obs=259) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 lnLP 1 2 lnTAS 0,214 1 3 lnTFASpEMPL 0,5295 0,3314 1 4 lnCGMpEMPL 0,9356 0,2035 0,5059 1 5 lnCEMPLpEMPL 0,6912 0,0573 0,4944 0,5856 1 6 AGE -0,0885 0,1017 0,0679 -0,1209 -0,0261 1 7 AGE2 -0,0803 0,0496 0,0678 -0,1408 -0,0319 0,9118 1 8 DOMMNEC0 -0,0046 -0,2794 0,0004 -0,0395 0,0361 -0,0886 -0,0044 1 9 DOMMNEC1 -0,095 -0,1396 -0,0276 -0,1242 0,0278 0,1071 0,0892 -0,1484 1 10 DOMMNEC2 -0,1 0,1622 -0,1036 -0,0893 -0,0627 0,0327 -0,0138 -0,2285 -0,2406 1 11 FORMNEC0 0,1864 -0,0567 -0,0554 0,1928 0,0739 0,017 0,0188 -0,1227 -0,1292 -0,1989 1 12 FORMNEC1 0,0279 0,0305 0,0763 0,0274 -0,0035 0,11 0,0934 -0,1172 -0,1234 -0,19 -0,102 1 13 FORMNEC2 0,0342 0,1675 0,1106 0,0672 -0,0318 -0,129 -0,1219 -0,2375 -0,25 -0,3849 -0,2067 -0,1974 1 14 US 0,1538 0,1738 0,0683 0,191 -0,0042 0,0051 -0,015 -0,1557 -0,1639 -0,2524 0,0862 0,1774 0,3175 1 15 EU -0,0406 -0,0386 0,0354 -0,0411 -0,0459 0,0336 0,0138 -0,2039 -0,2147 -0,3305 0,3212 0,2835 0,2474 -0,2252 1
Table A.4.10: Correlation Matrix 4 Sub-Sample 1
74 Table A.4.11: Auxiliary Regressions Sub-Sample 1
R-squared Adj. R-squared lnTAS 0,3850 0,3075 lnTFASpEMPL 0,6410 0,5958 lnCGMpEMPL 0,6162 0,5678 lnCEMPLpEMPL 0,5838 0,5313 AGE * 0,1265 0,0203 AGE2 * 0,1084 0,0003
* Auxilary regression of both age variables were run without the other age variable on the right hand side of the equation
Table A.4.12: Heteroskedacity-tests for all models and specifications, Sub-Sample 1
model p-value specification (column nr) (1) (2) (3) (4) (5) LP (model 2) White 0.0000 0.0000 0.0000 0.0000 0.3717 BP 0.0000 0.0265 0.0275 0.0299 0.5763 LPVA (model 3) White 0.0000 0.0000 0.0000 0.0000 0.0000 BP 0.0000 0.0000 0.0000 0.0000 0.0000 - 'White' refers to White's test for heteroskedacity of the residuals.
75 Table A.4.13: 2SLS with Instrumental Vars., LP- and LPVA model (Sub-Sample 1)
LP-model (2) LPVA model (3)
# Independents (1) (2) (3) (4) 1. lnTAS -0.01393 0.01781 0.13992* 0.12378** (0.02577) (0.02485) (0.07656) (0.06040) 2. lnTFASpEMPL 0.11442*** -0.01051 0.00462 -0.06698 (0.04335) (0.03177) (0.08721) (0.05766) 3. lnCGMpEMPL 0.76939*** 0.68268*** (0.03912) (0.03415) 4. lnCEMPLpEMPL 0.31920*** 0.97360*** (0.04874) (0.08776) 5. AGE -0.00321** -0.00235* -0.00004 -0.00022 (0.00149) (0.00137) (0.00099) (0.00065) 6. AGE2 0.00001** 0.00001** (0.00000) (0.00000) 7. FORMNE(x) 0.38393 0.01452 -0.76913 -0.35580 (0.45143) (0.36343) (1.08686) (0.91574) 8. DOMMNE(x) 0.21841 -0.22901 -1.70120 -0.94870 (0.50130) (0.43316) (1.17242) (0.92770) 9. FORNMNE 0.18691 -0.04717 -0.66186 -0.29210 (0.39610) (0.32749) (0.94416) (0.75011)
10. IND controls yes yes yes yes
11. Constant 3.08873*** 1.18615*** 13.34161*** -0.60209 (0.45822) (0.39043) (0.79398) (1.22875) Observations 288 260 277 277 R-sqrd 0.90656 0.93015 R2 0.53573 R-sqrd (adj.) 0.89686 0.92168 AdjR2 0.48539 P > F # 1-6 0 0 0 0 P > F # 7-9 0.366 0.353 0.00477 0.0114 P > F # 11 0.925 0.308 0.301 0.207 P Reset(1) 0.270 0.815 0.266 0.392 P Reset (2) 0.0160 8.36e-06 0.996 0.246 P Jarque-Bera 0 0 0.0406 0
Robust standard errors in parentheses; significance levels *** p<0.01, ** p<0.05, * p<0.1
(x)
76 Table A.4.14: Other Sample Results on Productivity, LP model
Sample: Sub-Sample 2 Sub-Sample 3 Full Sample
Dependent: lnLP lnLPC lnLP # Independents (1) (2) (3) (4) (5) (6) 1. lnTAS 0.01786 0.01881 0.01557 0.01771 -0.39046*** -0.17005** (0.01427) (0.01417) (0.01262) (0.01386) (0.09637) (0.08121) 2. lnTFASpEMPL 0.08367* -0.00697 0.01902 -0.01169 0.88223*** 0.15946 (0.04466) (0.03171) (0.02724) (0.03056) (0.19461) (0.11476) 3. lnCGMpEMPL 0.73700*** 0.69215*** 0.74414*** 0.67668*** 0.55055*** 0.01400 (0.03804) (0.03474) (0.02835) (0.03360) (0.18716) (0.17269) 4. lnCEMPLpEMPL 0.24559*** 0.24324*** 1.58641*** (0.06181) (0.06036) (0.32508) 5. AGE -0.00287*** -0.00300*** -0.00248*** -0.00294*** -0.01118* -0.00910** (0.00099) (0.00093) (0.00092) (0.00091) (0.00582) (0.00438) 6. AGE2 0.00001*** 0.00001*** 0.00001*** 0.00001*** 0.00003* 0.00003** (0.00000) (0.00000) (0.00000) (0.00000) (0.00002) (0.00001) 7. DOMMNEC1 -0.14422 -0.06178 -0.11419 -0.08689 -0.25301 -0.17620 (0.10223) (0.08209) (0.08526) (0.08482) (0.65945) (0.60389) 8. DOMMNEC2 -0.10214 -0.07824 -0.11346 -0.07964 0.02824 -0.05703 (0.08179) (0.07337) (0.07266) (0.07398) (0.49059) (0.42599) 9. FORMNEC0 -0.15452 -0.02134 -0.09511 -0.07510 -0.53270 -0.58868 (0.12044) (0.09340) (0.08211) (0.08574) (0.59703) (0.46557) 10. FORMNEC1 -0.20388** -0.07518 -0.17931** -0.08488 -0.65306 -0.28287 (0.08262) (0.07415) (0.07616) (0.07311) (0.49803) (0.43724) 11. FORMNEC2 -0.15289* -0.09990 -0.12248 -0.08677 -0.33005 -0.06893 (0.08465) (0.07561) (0.07583) (0.07564) (0.56428) (0.45399)
12 IND controls yes yes yes yes yes yes
13. Constant 3.50219*** 1.85404*** 3.86525*** 1.64247*** -5.61338** -16.29677*** (0.46328) (0.55167) (0.35076) (0.50722) (2.47865) (3.15931) Observations 280 254 276 253 303 273 R-sqrd 0.89984 0.91155 0.88843 0.90056 0.59295 0.70328 R-sqrd (adj.) 0.88822 0.89965 0.87528 0.88712 0.54972 0.66649 P > F # 1-6 0 0 0 0 6.77e-09 4.08e-09 P > F # 7-11 0.215 0.804 0.333 0.884 0.433 0.278 P > F # 12 7.25e-05 0.00546 3.02e-08 0.0303 0.000103 0.0848
P Reset(1) 0.174 4.27e-08 1.89e-07 2.90e-06 0 0
P Reset (2) 0.271 2.16e-07 6.87e-07 4.80e-06 0 0
P Jarque-Bera 0 0 0 0 0 0
77 Table A.4.15: Other Sample Results on Productivity, LPVA model
Sample: Sub-Sample 2 Sub-Sample 3 Full Sample
Dependent: lnLPVA lnLPVAC lnLPVA
Independents (1) (2) (3) (4) (5) (6) 1. lnTAS 0.08756** 0.10348*** 0.03730* 0.05278*** -0.00494 0.09655*** (0.03498) (0.02825) (0.02200) (0.01726) (0.03759) (0.02719) 2. lnTFASpEMPL 0.09219 -0.02004 0.15850** 0.06195 0.59334*** 0.01934 (0.08639) (0.06816) (0.06815) (0.04621) (0.08260) (0.06810) 3. lnCEMPLpEMPL 0.97435*** 0.98316*** 0.96020*** (0.08931) (0.10156) (0.06582) 4. AGE -0.00100* -0.00073 -0.00112** -0.00067* -0.00221*** -0.00094** (0.00060) (0.00045) (0.00050) (0.00037) (0.00075) (0.00048) 5. DOMMNEC1 -0.18419 -0.17355* -0.16233 -0.13496 -0.21900 -0.18391** (0.12654) (0.09464) (0.12037) (0.08931) (0.18535) (0.09268) 6. DOMMNEC2 -0.24341 -0.24420** -0.13797 -0.14518 0.02749 -0.17605 (0.15605) (0.12044) (0.13968) (0.11135) (0.21095) (0.12401) 7. FORMNEC0 0.27364 0.12692 0.20803 0.07533 0.36934 0.14903 (0.20114) (0.15788) (0.16965) (0.12230) (0.25877) (0.15614) 8. FORMNEC1 -0.19007 -0.20770 0.01431 -0.11403 -0.23077 -0.18464 (0.19316) (0.15713) (0.15032) (0.12227) (0.22911) (0.15529) 9. FORMNEC2 -0.19395 -0.18317 -0.08866 -0.14556 -0.15804 -0.14605 (0.15111) (0.11595) (0.12760) (0.10743) (0.21632) (0.11481)
10. IND controls yes yes yes yes yes yes
111. Constant 13.57576*** 1.71942 9.02723*** -0.94547 4.69912*** -0.97172** (0.66152) (1.23763) (0.78698) (1.03617) (1.03294) (0.49307) Observations 277 277 272 272 294 294 R-sqrd 0.39067 0.65461 0.34518 0.64033 0.58269 0.86366 R-sqrd (adj.) 0.32188 0.61405 0.27273 0.59889 0.53860 0.84868 P > F # 1-4 0.00370 0 0.00249 0 0 0 P > F # 5-9 0.106 0.144 0.117 0.152 0.0610 0.139 P > F # 10 1.88e-10 0.000335 0.627 0.0431 0.000767 0.0320 P Reset(1) 0.147 0.0275 0.00223 0.0602 0 0.907 P Reset (2) 0.339 0.0322 0.00225 0.0983 0 0.516 P Jarque-Bera 0 0 0 0 0 0
78 Table A.4.16: US/EU Effects on Labour Productivity (All Samples)
Sample: Main Sample (Sub-Sample 1) Sub-Sample 2 Dependent: lnLP lnLP # Independents (2) (3) (5) (6) 1. lnTAS 0.00480 0.01785 0.01690 0.01846 (0.01456) (0.01454) (0.01469) (0.01430) 2. lnTFASpEMPL 0.11620*** 0.00392 0.08459* -0.00675 (0.04145) (0.03114) (0.04510) (0.03179) 3. lnCGMpEMPL 0.75115*** 0.69457*** 0.73484*** 0.69188*** (0.04051) (0.03586) (0.03823) (0.03684) 4. lnCEMPLpEMPL 0.27678*** 0.24684*** (0.05160) (0.06149) 5. AGE -0.00273*** -0.00301*** -0.00274*** -0.00302*** (0.00102) (0.00094) (0.00100) (0.00093) 6. AGE2 0.00001*** 0.00001*** 0.00001*** 0.00001*** (0.00000) (0.00000) (0.00000) (0.00000) 7. DOMMNEC1 -0.14503 -0.07733 -0.14863 -0.06089 (0.10221) (0.08311) (0.10271) (0.08241) 8. DOMMNEC2 -0.09531 -0.07342 -0.10415 -0.07655 (0.08742) (0.07331) (0.08256) (0.07346) 9. FORMNEC0 -0.08530 -0.05209 -0.07245 -0.03145 (0.13472) (0.12329) (0.12966) (0.12287) 10. FORMNEC1 -0.13445 -0.10644 -0.11966 -0.08656 (0.12962) (0.11841) (0.12689) (0.12091) 11. FORMNEC2 -0.10883 -0.12496 -0.09741 -0.10854 (0.11677) (0.10043) (0.11217) (0.10136) 12. US -0.05445 0.03293 -0.05183 0.01828 (0.08773) (0.07984) (0.08875) (0.08091) 13. EU -0.12156 0.01845 -0.10744 0.01015 (0.10803) (0.08950) (0.10656) (0.09070)
14. IND controls yes yes yes yes
15. Constant 2.12857*** 1.42104*** 2.48491*** 1.84866*** (0.42307) (0.46238) (0.51425) (0.57456) Observations 287 259 280 254 R-sqrd 0.91781 0.92938 0.90038 0.91157 R-sqrd (adj.) 0.90781 0.91938 0.88793 0.89877 P > F # 1-6,12,13 0 0 0 0 P > F # 7-11 0.807 0.764 0.781 0.819 P > F # 12 0.00332 0.00183 0.0153 0.00556 P Reset(1) 0.0666 1.10e-05 0.212 2.53e-08 P Reset (2) 0.0824 2.16e-07 0.269 1.24e-07
P Jarque-Bera 0 0 0 0
79 Table A.4.17: US/EU Effects on Labour Productivity (Other Samples)
Sample: Sub-Sample 3 Full Sample
Dependent: lnLP lnLP # Independents (8) (9) (11) (12) 1. lnTAS 0.01545 0.01760 -0.39661*** -0.14687* (0.01271) (0.01402) (0.09894) (0.07460) 2. lnTFASpEMPL 0.01930 -0.01151 0.90885*** 0.13517 (0.02742) (0.03080) (0.19651) (0.10603) 3. lnCGMpEMPL 0.74268*** 0.67594*** 0.55504*** -0.02721 (0.02969) (0.03527) (0.19714) (0.16230) 4. lnCEMPLpEMPL 0.24280*** 1.77767*** (0.05979) (0.28706) 5. AGE -0.00243*** -0.00292*** -0.01134** -0.00841** (0.00092) (0.00091) (0.00560) (0.00397) 6. AGE2 0.00001*** 0.00001*** 0.00003* 0.00003** (0.00000) (0.00000) (0.00002) (0.00001) 7. DOMMNEC1 -0.11573 -0.08762 -0.22713 -0.16249 (0.08539) (0.08506) (0.65906) (0.59801) 8. DOMMNEC2 -0.11482 -0.08030 0.06443 -0.03649 (0.07290) (0.07404) (0.49492) (0.43574) 9. FORMNEC0 -0.06637 -0.06189 -0.91819 -0.97420 (0.11705) (0.11965) (0.82087) (0.61502) 10. FORMNEC1 -0.14941 -0.07125 -1.09685 -0.75504 (0.11632) (0.11769) (0.81665) (0.62361) 11. FORMNEC2 -0.10158 -0.07705 -0.61314 -0.36175 (0.10171) (0.10054) (0.78447) (0.57800) 12. US -0.02435 -0.01058 0.68230 0.81185 (0.08579) (0.08235) (0.75759) (0.52086) 13. EU -0.03557 -0.01678 0.37794 0.39932 (0.08731) (0.08644) (0.58642) (0.38526)
14. IND controls yes yes yes yes
15. Constant 3.88782*** 1.65609*** -10.23774*** -18.02907*** (0.37559) (0.51769) (2.73805) (2.72128) Observations 276 253 302 272 R-sqrd 0.88852 0.90058 0.60194 0.73948 R-sqrd (adj.) 0.87436 0.88612 0.55623 0.70460 P > F # 1-6,12,13 0 0 2.26e-08 1.46e-10 P > F # 7-11 0.585 0.919 0.642 0.199 P > F # 12 4.58e-08 0.0289 0.00565 0.0409
P Reset(1) 1.21e-07 1.39e-06 0 0
P Reset (2) 4.32e-07 2.62e-06 0 0
P Jarque-Bera 0 0 0 0
80 Table A.4.18: Results with NFC classification, LP model
Sample: Sub-Sample 1 Sub-Sample 2 Sub-Sample 3 Full Sample Dependent: lnLP lnLP lnLPC lnLP # Independents (1) (2) (3) (4) 1. lnTAS 0.01869 0.01788 0.01835 -0.19910** (0.01509) (0.01441) (0.01413) (0.09654) 2. lnTFASpEMPL 0.00639 -0.00386 -0.00998 0.14069 (0.03136) (0.03197) (0.03105) (0.11649) 3. lnCGMpEMPLN 0.69597*** 0.69425*** 0.67885*** 0.02447 (0.03365) (0.03471) (0.03374) (0.17630) 4. lnCEMPLpEMPL 0.29635*** 0.24120*** 0.23980*** 1.59316*** (0.05093) (0.06335) (0.06176) (0.33131) 5. AGE -0.00289*** -0.00292*** -0.00290*** -0.00899** (0.00091) (0.00090) (0.00088) (0.00441) 6. AGE2 0.00001*** 0.00001*** 0.00001*** 0.00003* (0.00000) (0.00000) (0.00000) (0.00001) 7. DOMNFC1 -0.04547 -0.03360 -0.05635 -0.37454 (0.08955) (0.08919) (0.09124) (0.50043) 8. DOMNFC2 -0.09362 -0.09526 -0.09902 0.11721 (0.07205) (0.07110) (0.07155) (0.47689) 9. FORNFC0 -0.03698 -0.02349 -0.07709 -0.57005 (0.09452) (0.09306) (0.08535) (0.46495) 10. FORNFC1 -0.14360** -0.13876* -0.11908 -0.18775 (0.07263) (0.07454) (0.07435) (0.52756) 11. FORNFC2 -0.07526 -0.07337 -0.07469 0.01520 (0.07833) (0.07767) (0.07787) (0.42942)
12 IND controls yes yes yes yes
13 Constant 1.14842*** 1.86799*** 1.62157*** -15.76348*** (0.43119) (0.55781) (0.51814) (3.20498) Observations 260 254 253 273 R-sqrd 0.93967 0.91222 0.90096 0.70490 R-sqrd (adj.) 0.93177 0.90042 0.88757 0.66832 P > F # 1-6 0 0 0 5.97e-09 P > F # 7-11 0.422 0.445 0.664 0.266 P > F # 12 0.00224 0.00499 0.0304 0.0903
P Reset(1) 6.41e-06 3.85e-08 2.18e-06 0
P Reset (2) 4.22e-08 1.36e-07 1.36e-06 0
P Jarque-Bera 0 0 0 0
81 Table A.4.19: Effect of unknown FOR/MNE status in LP model
Dependent : lnLP # Independents (1) (2) (3) 1. lnTAS 0.104*** 0.00915 0.0193 (0.0284) (0.0131) (0.0120) 2. lnTFASpEMPL 0.343*** 0.0726*** -0.00320 (0.0462) (0.0222) (0.0174) 3. lnCGMpEMPL 0.771*** 0.666*** (0.0238) (0.0219) 4. lnCEMPLpEMPL 0.322*** (0.0346) 5. AGE -0.00180** 0.000372 -0.000784 (0.000782) (0.000380) (0.000807) 6. AGE2 4.37e-06* (2.51e-06) 7. MNE -0.259** -0.0517 -0.0640 (0.108) (0.0592) (0.0415) 8. FOR 0.130 -0.0521 0.0130 (0.0860) (0.0456) (0.0407) 9. MNENK 0.0325 -0.0617 -0.0379 (0.106) (0.0644) (0.0458) 10. OWNNK -0.0490 0.0634 0.0809 (0.107) (0.0574) (0.0534) 11. Constant 7.154*** 2.501*** 0.816*** (0.695) (0.366) (0.296) Observations 782 576 513 R-squared 0.353 0.879 0.889
82 Table A.4.20: Main Results with assumed FOR/MNE characteristics, LP model
Dependent: lnLP # Independents (1) (2) (3) (4) (5) (6) 1. lnTAS 0.00613 0.01085 0.01182 0.01182 0.01050 0.01886 (0.01228) (0.01303) (0.01288) (0.01288) (0.01373) (0.01303) 2. lnTFASpEMPL 0.07379*** 0.07348*** 0.07189*** 0.07189*** 0.07125*** -0.00298 (0.02214) (0.02192) (0.02171) (0.02171) (0.02182) (0.01726) 3. lnCGMpEMPL 0.76762*** 0.77167*** 0.77296*** 0.77296*** 0.77292*** 0.66377*** (0.02348) (0.02359) (0.02357) (0.02357) (0.02347) (0.02273) 4. lnCEMPLpEMPL 0.32483*** (0.03472) 5. AGE 0.00036 -0.00087 -0.00085 -0.00085 -0.00074 -0.00082 (0.00037) (0.00084) (0.00083) (0.00083) (0.00084) (0.00079) 6. AGE2 0.00001* 0.00001* 0.00001* 0.00001* 0.00000* (0.00000) (0.00000) (0.00000) (0.00000) (0.00000) 7. AFOR -0.07972** -0.11250** (0.03613) (0.05186) 8. AMNE -0.00963 -0.04095 (0.04014) (0.05646) 9. AFORMNE 0.06729 -0.08616 (0.06753) (0.05706) 10. ADOMMNE -0.04095 (0.05646) 11. AFORNMNE -0.11250** (0.05186) 12. ADOMMNEC1 -0.06544 -0.06768 (0.08731) (0.06543) 13. ADOMMNEC2 -0.04103 -0.06126 (0.05364) (0.04854) 14. AFORMNEC0 -0.10857** -0.02439 (0.04495) (0.04026) 15. AFORMNEC1 -0.15855*** -0.05397 (0.05753) (0.06052) 16. AFORMNEC2 -0.05191 -0.04603 (0.05785) (0.05083)
17 IND controls yes yes yes yes yes yes
18. Constant 2.52410*** 2.41058*** 2.39183*** 2.39183*** 2.41873*** 0.81593*** (0.33879) (0.34721) (0.34528) (0.34528) (0.33891) (0.29203) Observations 576 576 576 576 576 513 R-sqrd 0.87778 0.87904 0.87920 0.87920 0.87952 0.88875 R-sqrd (adj.) 0.87223 0.87285 0.87278 0.87278 0.87265 0.88134 P > F # 1-6 0 0 0 0 0 0 P > F # 7-16 FprobMS 0.0883 0.143 0.143 0.0652 0.809 P > F # 17 0.631 0.979 0.788 0.788 0.804 0.187 P Reset(1) 0.00380 0.00277 0.00311 0.00311 0.00379 2.36e-06 P Reset (2) 3.67e-07 8.31e-07 9.53e-07 9.53e-07 2.24e-06 1.46e-07
P Jarque-Bera 0 0 0 0 0 0
83
APPENDIX B – Further Information
B.2.1 - Literature background
While our interest in this paper lies with the effects of international investment, much of economists’ knowledge on international investment finds it roots in advances on
international trade. In his seminal work, Krugman1(1979) formalised a theoretical framework that embodied a dramatic departure from the classic Heckscher-Ohlin world of trade between countries to a world of trade between firms2. A vast amount of empiric and theoretic contributions followed in response, and became known as ‘new trade theory’. The development of new trade theory also propelled new advances in theories on the multinational enterprise (MNE)3. That was an inevitable consequence, since trade and investment can be viewed as alternative foreign market entry strategies. Two different models of the multinational firm4 evolved from this. ‘Horizontal’ MNEs they set up new production facilities abroad to serve these local markets, essentially ‘duplicating’
everything except headquarter activities (Barba Navaretti & Venables, 2004). ‘Vertical’ multinationals, however, split up production of their goods across countries in distinct stages that require different relative factor inputs. Horizontal investors have made a ‘proximity-concentration trade-off’ (Brainard, 1993): they have decided that being close to the foreign market outweighs the costs of duplicating production facilities. Vertical investors, however, have decided that they can extract most benefits from favourable factor prices in the host country, accepting trade costs as they are. Horizontal MNEs are therefore known as ‘market seeking’, while vertical MNEs are ‘factor seeking’ (Barba Navaretti & Venables, 2004).
Both the horizontal and vertical model have been extensively tested at the industry level, and especially the horizontal model has found considerable support. Vertical models tend to perform less strongly, however. When these do find support, they tend to do so mostly for ‘North-South’ tests while rarely for samples of similar countries. In balance, therefore, horizontal investment appears to be dominant on a global scale (Helpman, 2006). This is found for example by Markusen & Maskus (2002). More ambiguous results, however, are found increasingly, e.g. by Amiti (1998). Such findings lead academic to say that
“evidence points to a growing importance of more complex integration strategies by
1
On December 8, 2008, Krugman received the Nobel Prize in Economics “for his analysis of trade patterns and location of economic activity” as first set up in his 1979 paper and developed in subsequent works over following the years. Source: http://nobelprize.org/nobel_prizes/economics/laureates/2008/index.html 2
This was possible because the Krugman’s assumptions are radically different from those in classic models. Krugman assumes increasing returns to scale and monopolistic competition on the supply side (each firm producing its own unique variety of a product), and a love of variety at consumers – the demand side. Classic models assume constant returns to scale and perfect competition.
3
See Faeth (2009) for an extensive historic review of theories on MNEs and foreign direct investment.
4
The terms ‘multinational’, ‘~ enterprise’ and ‘~ firm’ are used interchangeably in this paper. A more strict definition would be to use the term ‘enterprise’ to indicate only the very highest level in ownership
84 multinational corporations” (Helpman, 2006: 599), meaning a combination of both
horizontal and vertical motives5(Bronzini, 2008).
The framework provided by the different MNE-models, in a world under the assumptions of new trade theory, has shown to be powerful in explaining the size and direction of trade and investment flows between countries. Most notably, it provides the tools to explain observations of increasing stocks flows of FDI (Foreign Direct Investment) between similar, developed countries. That is because trade and investment can now “arise independently of any pattern of comparative advantage” (Markusen, 1995: 169)6.However, its power lies only at the industry-aggregate level. This is, of course, an inevitable result of the fact that all determinants considered in the models are on the country- and industry-level: market sizes, factor costs, trade costs, and industry-wide economies of scale. These lie outside of reach for the individual firm. The fact that different firms may make different decisions on whether or not to serve foreign markets, as well as by which entry mode (exporting or investing), is simply not modelled to be a possibility. Or in the words of Castellani & Zanfei (2007: 158) “Brainard’s model does not predict which firms do which activity within sectors”7.
B.3.1 - Sampling Strategy
A cross-sectional sample of highly detailed micro-data was constructed on 941 firms, all operating in The Netherlands in 23 manufacturing industries in the year 2007. The required information was retrieved from the Amadeus database (full version on DVD) of Bureau Van Dijk Electronic Publishing8. This database gives access to standardised information on millions of public and private European firms, on issues such as: financial statements and ratios; descriptive information (e.g. firm activities); firm shareholders; and subsidiary ownership.
At the outset, the aim was to collect a valid sample of the Dutch manufacturing sector. Therefore, the following actions were taken regarding the sampling strategy. Firstly, it was made sure that the level of measurement of all dependent and explanatory variables is identical for every firm. Therefore only consolidated figures from the highest ranking parent firm within Dutch ownership networks are taken9. This resolves two issues. Firstly, it prevents ‘double counting’ of parents and their domestic subsidiaries – as these latter may report financial information independently. Secondly, it prevents possible erroneous
5
An example of a complex model would be the ‘knowledge capital’ model tested by Markusen & Maskus
(2002). In this ‘hybrid’ model both the horizontal and vertical model are nested. The authors find that this
model does not perform better than the horizontal model, but it does outperform the vertical model.
6
Markusen (1995) is hereby referring to the fact that under classic Heckscher Ohlin assumptions, only
vertical trade and FDI may occur, and horizontal trade or investment is not possible.
7
In this citation ‘activity’ refers to a firm’s choice between exporting, producing abroad, or neither.
8
Using the full version on DVD instead of the web-based version makes it possible to draw samples of significantly higher quality. The DVD version’s database includes far more privately owned and smaller sized firms than the web-based version. By using the DVD version we therefore enhance the width of our sample and so reduce to a minimum any possible sampling bias against such smaller/private firms.
9
85 qualification of firms as being ‘domestic’ when their direct parent is located in The Netherlands, while that parent in turn is actually foreign owned. The second action taken to improve validity of the sample relates to a special problem presented by firms of the ‘financial holding’ type. In some ownership networks the reporting entity at the top is registered as operating in financial services even though de facto it owns mainly manufacturing subsidiaries. So for our research purposes it could be included in the sample10. In sampling for this paper, however, it was chosen to exclude all such firms and their networks. This was done because it is practically near to impossible to prevent measurement errors in manual case-by-case analysis of hundreds of firms11: one is bound to in- or exclude some firms incorrectly, or double-count others. Moreover, there is no compelling theoretic argument why manufacturing firms headed by a ‘financial holding’ would be different in structure or performance from those with other ownership network structures. As a result, exclusion of these firms is not expected to bias empiric results. As said, the sample is confined to the manufacturing sector. This was done for two reasons. Firstly because production in manufacturing firms can more easily be connected to ‘classic’ capital and labour inputs, on which information is largely available. In service industries, for example, intangible factors (such reputation and knowledge) are expected to play a bigger role, on which information is much more scarce. Secondly, by choosing to focus on manufacturing industries, this paper’s empiric results will be comparable to several influential papers in the field12. An overview of the distribution of the sampled 941 firms over 23 manufacturing industries13 can be found in Table A.3.1 in the appendix. This table also compares the industrial-composition of the sample to that of the Dutch economy. The overall picture is that the sample is roughly in line with the Dutch
economy, though it over-weighs some of the smaller industries and under-weighs several larger ones.
Finally, for practical purposes it was chosen to restrict the sample in two other ways. Firstly, all data is from the year 2007. This year was chosen as it is the most recent year for which Amadeus has richly detailed information available. Secondly, some minimum data requirements were demanded. Of every firm, at minimum was required that figures on operating revenue and the number of employees for 2007 are known. This ensures that at least a basic calculation of labour productivity can be made for every firm.
B.3.2.2 - Employment & Productivity Censoring
A special remark should be made regarding employment. Upon inspection of these figures, it quickly became clear that the 941-firm sample contains large variation in employment numbers reported by firms: this ranges from single digit numbers to as much as 175000. In principal, such increased variety in the x’s is a desirable sample property. It increases the sampling range and so may increase the regression sum of squares (Carter
10
An example of this is Heineken Holding N.V. This entity is classified as a financial company under NACE Rev. 2 K.64.xx., though ultimate owner of an network of firms that concerns mainly manufacturing.
11
Exclusion of these firms reduced the sample to 941 from roughly 200 more otherwise.
12
See Literature Review and Table A.2.1, Appendix A
13
86
Hill et. al., 2008: 30), i.e. the ‘explained’ part of the regression variance. On the other
hand, however, close inspection revealed that for a non-negligible amount of firms such low-levels of employees seem plainly incorrect. It appears that some firms report ‘1’ employee or just a few while it is plainly unrealistic that this is the truth. Some of these employment numbers are obviously erroneous data-entries in Amadeus, others can be explained due to the fact that the entity observed is of the holding-type. These holdings were not yet excluded by our sampling strategy, as these holdings (contrary to the already excluded ones) are correctly categorized into a manufacturing industry. While the
holding legally may indeed employ only a handful of people, it may still report
consolidated data that includes all of its subsidiaries. As a result, our labour productivity variables will contain many erroneously high observations. To deal with this problem, two samples were extracted from the original 941 firms. In the first, named sub-sample 1, all firms that report under five employees are excluded. This reduces the sample to 822 firms14. In the sub-sample 2 only 749 firms remain, as all firms reporting under 20 employees are excluded15. The decision on where to ‘draw the line’ about which firms to exclude may be influential on our empirical results. In terms of size, our sample and sub-samples are somewhat smaller than is typical in cross-sectional firm-productivity analysis.
As an alternative to picking a valid sub-sample from the total of 941 firms, we may also consider trimming our dependent variables (thus, omit certain observations after
measurement). Upon examining descriptive information, we decided to create a sub-sample in which we omitted the top and bottom 2 percent of the LP and LPVA
distributions, contingent on the fact that sub-sample 1 was already selected16. This set of firms that remain then has been named sub-sample 3.
3.4 - Methodological Issues with OLS
As with any economic set of relationships when translated econometrically, we need to be wary that the most adequate model is chosen. Examining changes in the coefficient of determination (better known as R-squared) as our models’ specifications change is a first step. This statistic reflects the ratio of a model’s explained variance of the dependent to its total variance. Comparing ordinary R-squared to its adjusted variety is important too, as the latter penalizes the researcher for adding extra explanatory variables17. Therefore, both will be calculated for each regression. Choosing the right specification also entails
14
Prohibited that no other restrictions are made on the dataset by other variables in the sample.
15
As an alternative solution, we considered collecting employment information from a different source than Amadeus, namely from publicly available annual reports on company websites. Aside from the fact that this is a time-intensive job to do, another problem arised: not all firms in our sample that report such low employment figures in Amadeus do have this information publicly available. Specifically, this is often the case for the many small private-owned companies included in our sample. Therefore, we could not resolve the problem of erroneous reported employment levels in this way.
16
We decided that the full-sample already included to many erroneously high LP and LPVA observations. Censoring these productivity variables for the full sample caused great difficulty in finding an adequate ‘cut-off point’ for which percentage of extreme observations to exclude.
17
87 preventing possible problems from omitting important variables or adding irrelevant ones. The former mistake results in (a) biased coefficient(s) if the omitted variable is correlated with an included one. The latter mistake increases standard errors of correctly included variables, thus may result incorrect rejection of otherwise significant independents
(Carter Hill et. al. 2008). The solution to preventing these problems is examining
correlations/covariances between (possible) independents and observing the changes in coefficients, standard errors and p-values when changing the model’s specification. Furthermore, for our research purposes we are especially interested whether several groups of variables together are significant determinants of firm productivity. Most specifically, we want to now if our dummies indicating multinational status are important determinants relative to a firm’s structural characteristics and industry controls. For this purpose, we will perform F-tests for these different sets of independents. Finally, our models’ adequacy depends on whether the correct functional form is chosen. Therefore, two Ramsey RESET-tests18 will be done for each specification of every model.
Our above expressed concerns and solutions regarding our models’ adequacy are ways in which to be as certain as possible that the first two fundamental assumptions of multiple OLS regression hold (Carter Hill et. al., 2008: 111; Brooks, 2002: 56 & 145). There are, however some more assumptions underlying OLS that need to be taken account of. Firstly, a model’s sample should not exhibit heteroskedacity. This is the phenomenon that “variances of the all observations are not the same” (Carter Hill et. al. 2008: 198).
Preferably, “all observations come from probability density functions with the same variance” (Ibid), in which case our random dependent and random errors are
homoskedastic. Cross-sectional models using input-output data, like our models, are well-known for exhibiting heteroskedacity (Carter Hill et. al., 2008). To discover whether our (sub-) sample(s) given (a) certain model(s) suffer(s) from this problem, we will perform White’s tests and Breusch-Pagan tests19. If the(se) test(s) indeed indicate heteroskedacity, the least squares should try to resolve this. In the most fortunate case we can identify the cause of heteroskedacity, namely one or more or our independent variables. For example, perhaps firm size causes heteroskedacity, as we have a broad sampling range of this firm characteristic. If the cause is succesfully identified, we can then use this knowledge in a generalised least squares (GLS), or use a weighted least squares (WLS) procedure, or correct for the heteroskedastic partition (in case one of our segmentation dummies is the cause of heteroskedacity). In the less fortunate case that we cannot find the cause of heteroskedacity, we can use White’s heteroskedacity-consistent standard errors, also known as ‘robust’ errors. Although using this technique means that there will be another least squares estimator with a smaller variance, at least it will make sure that we can rely on the returned coefficients and standard errors (Carter Hill et. al., 2008: 201-202). As a results we will still be able to draw correct inferences on our hypotheses.
18
RESET tests are done by running auxiliary regressions in which polynomials of predicted values of the dependent are included as independents is the equation. The practice of adding a squared transformation of the predicted y-values is commonly known as RESET(1). RESET(2) adds a another transformation of predicted y-values to the power of three to that. If t- and/or F-tests on these artificial variables are then significant, this shows that possibly the functional form may be improved.
19
88 Secondly, we should be assured of the independence of the errors of our observations. Only if these errors are uncorrelated with one another, the dependent variable’s
observations are also independent. Given the fact that our research is cross-sectional of nature, we do not expect this to be a problem in our estimations. In longitudinal or panel data research this problem is more likely to occur, and know as autocorrelation.
According to Brooks (2002) autocorrelation may also occur in cross-sectional samples “in a spatial sense, if there is a regional dimension to [the dependent’s value] that is not captured by the model”20 (Brooks, 2002: 177). Since The Netherlands is only a small country, this should not occur in our data, however.
Thirdly, OLS estimation with multiple regressors demands that our independent variables are not too strongly linearly related to each other. This is known as the problem known of multicollinearity. In its most extreme case certain variables are exact linear functions of others. The reason for this happening is that “economic variables may move together in systematic ways” (Carter et. at. 2008: 153). In our case, collinearity is most likely to exist between our variables on firm-specific characteristics other than those indicating its multinational status (especially between size and input variables). As a result, their standard errors will be higher than when in isolation. This reduces their significance and makes it difficult to retrieve an accurate estimation of their individual contribution to productivity. Therefore we will be conservative with dropping such variables, so as to prevent the negative effects of omitting relevant variables. To detect if indeed collinearity exists in our (sub-) sample(s) we will present correlation matrices. A second thing we will do, just to get an extra indication of the (non-) severity of potential collineairity, is to run auxiliary regressions. In these we replace our productivity measure by an independent, and keep the rest of the independents on the right hand side. We will do this for every independent21, and report on R-squared and adjusted R-squared for each of these auxiliary regressions.
Another thing to we need to be as sure of as possible is that our independents are
exogenous to the model. It may be, however, that some of our independents are somehow correlated to the error term. This is know as the problem of endogeneity. In our case it may be that labour productivity is jointly or simultaneously determined with its firm-specific characteristics and/or its multinational status. There may exist a feedback
relationship between these components (Carter et. al., 2008: 276). If so, this would result in the so-called ‘simultaneous equations bias’: the coefficient(s) of the endogenous variable(s) will be biased. For example, we might suspect that in the a firm experiences a productivity ‘shock’ (not captured by other explanatory variables in the model) managers will very quickly this, and respond by adjusting their factor inputs (e.g.. capital and intermediate materials)22. Inputs and productivity are then simultaneously determined in our sample, making inputs at least partly endogenous in our models. The same reasoning may apply to our dummies measuring multinational status. Idiosyncratic productivity
20
Insert in brackets replaces “bank profitability”, which is the example used by Brooks (2002) in the text.
21
Except for our industry dummies.
22
89 shocks (i.e. shocks unique to the firm) may lead to foreign investment, thus to firms ‘being’ multinational. Thus our ownership and foreign subsidiary dummies may also be subject to a feedback relationship with productivity, thus be simultaneously determined. To detect any possible endogeneity problems we need to find suitable instruments to the suspected endogenous variable(s). Suitable instruments are variables that are strongly correlated to the endogenous variable, but not to the dependent variable. Once such instruments are found, we can execute the (Hausman23) tests for endogeneity. In case endogeneity of our instrument independent(s) is indeed indicated, we will run a two-stage least squares (2SLS) regressions in which our endogenous independents are replaced by their instruments.
The final assumption of OLS is that the residuals are normally distributed about their mean (which should be zero). If they are, then so are the observed y-values (i.e. the dependent). Compliance to this assumption is desirable, but not necessary for the Gauss-Markov theorem to hold: if the before discussed assumptions are valid, already the regression estimated coefficients are the best linear unbiased estimators (BLUE) of their true values (Carter Hill et. al., 2008: 32). The central limit theorem then, states that the estimators will have an approximate normal distribution if the sample if sufficiently large. In that case “violation of normality is virtually non-consequential’ (Brooks, 2002: 182). Given that our sub-samples and different specifications of models (2) and (3) all have observation in the range of 250 – 500, we believe we need not worry about negative consequences from non-normality. Nevertheless we will report probability values of Jarque-Bera statistics for the residuals of each specification.
Aside from ensuring that our models are adequate and also comply with other OLS assumptions, we aim to examine the sensitivity or ‘robustness’ of our sample to minor changes. As discussed before, we constructed several sub-samples from our original 941-firm sample (as can be seen in Table A.3.6). Sub-sample 1 excludes all 941-firms from the full sample that report under 5 employees. Sub-sample 2 excludes those reporting under 20 employees. Sub-sample 3 excludes the top and bottom 2% percent of the productivity distribution of sub-sample 1. Finally, sub-sample 4 takes the original sample, makes assumptions on unknown ownership or subsidiary information, and subsequently takes the same 5 employee cut-off point as sub-sample 1.
B.4.3 - Diagnostic Checks
As part of our diagnostic checks we firstly want to assess the adequacy of our models. Examining changes in R-squared and adjusted R-squared, as well differences between the two, already goes a long way. In general, the reader may notice that the rate of explained
23
90 variation differs between our models. The dependent of model (3), labour productivity in terms of value added, is least explained by its regressors, while labour productivity in the model that approximates TFP-effects (2) is best explained. Moreover, ordinary and adjusted R-squared remain close together. Especially for model (2) this is the case, indicating that the choice of included variables is perhaps best of all in that model. Also revealing to look at, are the rows that show the F-probabilities of different groups of independents24. In each specification of both models the groups of structural firm characteristics are highly significant determinants of labour productivity differences at the 1% level. The same is true for the group of industry control dummies. For the groups of ownership and subsidiary dummies the picture is a little different, however. In both models their F-statistic indicates that they are insignificant as a group in certain specifications (especially advanced ones). This reinforces our general finding that, as more and more structural differences between firms are taken into account, less and less of between-firm productivity differences can be uniquely attributed to differences in their multinational status. Finally, a look at the p-values from our RESET tests suggest25 that our functional form in models (2) and (3) is likely to be not optimal. This questions whether the productivity – firm relationship is linear in its parameters altogether.
Unfortunately, Ramsey’s RESET test “presents the user with no guide as to what a better specification might be” (Brooks, 2002: 195). Multiple experiments with adding
polynomials of independents have been tried, none of which seemed to help. Thankfully, RESET tests on model (3)’s specifications did turn out favourable results.
Another assumption of OLS we need to be certain that holds, or else our found coefficients are not the best linear unbiased estimators of productivity, is that of non-collinearity between the independents. Tables A.4.7 to Table A.4.10 (see Appendix A) are correlation matrices of different sets of the independent variables and the dependent LP26 in sub-sample 1. Most of interest are the correlations between our variables that indicate firm-specific characteristics. As we suspected, correlations are highest amongst the factor input variables (capital intensity, cost of goods or materials, and cost of employees). While their values, in the range of 0.4 to 0.6 do warn us to be aware of (possible consequences of) their interdependencies27, they are not high enough to result in failure of the OLS procedure 28. The other correlation matrices are interesting mostly because they show us how the correlations between these input variables change when more demands have been put on the data with respect to ownership and subsidiary
24
i.e. the probability that the tested F-statistic lies within the confidence interval of Ho: the group of variables is not significant.
25
The p-values should be interpreted inversely from the p-values of our F-tests.
26
The variables that are not dummies are presented in the logarithmic transformation, like in our econometric models.
27
These possible consequence are the following: the standard errors of our coefficients may be too large, causing certain coefficients to be insignificant, while they are in fact not; the separate effects of correlated variables may be hard to distinguish; and, “estimators may be very sensitive to the addition or deletion of a few observations, or the deletion of an apparently insignificant variables” (Carter Hill et. al., 2004: 154) 28
91 information29. These indeed change somewhat, but the correlations remain to be of no critical concern30. To further supplement our analysis of whether linear relationships exist between our independents, we ran a series of auxiliary regressions. Table A4.11
(Appendix A) reports the R-squared and adjusted R-squared values for estimating such ‘artificial’ equations, in which the independents measuring structural firm characteristics are put on the left-hand side. The highest (adjusted) R-squared value does not exceed 0.60, and so remains below the 0.80 level which serves as a rule-of-thumb critical level
(Carter Hill et. al., 2008). This confirms us in our assessment that collinearity in our
sample is at noticeable but acceptable levels. The fact that the correlations between our inputs variables are in the medium region may explain a certain aspect of our regression results. Namely, it explains why some of these structural characteristics become
insignificant in advanced specifications that include more of that type of variables. For example, it may explain why lnTAS is insignificant in the LP-model and lnTFASpEMPL in the LPVA-model.
Another problem to evade in OLS is heteroskedacity of the standard errors. To find out whether heteroskedacity occurs in our models, we performed White test’s and Breusch-Pagan tests on the dependents of all specifications of all models. Table A.4.12 shows the outcomes: heteroskedacity seems present in almost every regression. The only instance in which it is not present is in specification 6 of model 4, i.e. the approximate TFP-effects model that includes average costs of labour (CEMPLpEMPL) as a determinant. To take account of the heteroskedacity problem, ‘robust’ or White’s standard errors were adopted for every regression. Tables 4.4 – 4.6 therefore only show these results. As a consequence we can be confident that the interpretation of our coefficients as discussed in the previous sub-section (4.2) is not erroneous. We also tried to identify the cause(s) of
heteroskedacity, our prime suspect being firm size. If such a cause could be found, we could then use generalised least squares (GLS) or weighted least squares (WLS)
technique. Such a solution would then have been preferred to using White standard errors. A clear cause(s) of heteroskedacity could not be found, however, leaving us with only the possibility to use White standard errors.
Preferably, the distribution of residuals equals the normal distribution. Graphical analysis of the residual distributions as well as calculating skewedness, kurtosis (peakedness) and Jarque-Bera statistics of the residuals pointed out that in our case they are not normally distributed. The bottom row of each regression-output table shows the p-values of the Jarque-Bera test, which tests the hypothesis that the errors are normally distributed. The p-values are always zero. The residuals are somewhat skewed but especially highly peaked31. Given that we had already observed some skewedness and much peakedness in the distribution of LP (see Figure 4.2, A.4.2 and A.4.3), we are not surprised to find this32. To solve this, “it is not obvious what should be done” (Brooks, 2002: 182) Fortunately,
29
Notice that, as a result, the amount of observations drops.
30
A common rule of thumb is to not include independent variables that have a correlation of over 0.6.
31
Skewedness over the various models and specification ranges from -0.2 to -0.8, kurtosis ranges from 7.0 to 8.8. Full range of figures are available on request.
32
92 “for sample sizes that are sufficiently large, violation of normality is virtually
non-consequential” (Ibid). We feel confident enough to say that with us that is case, thus that we need not adjust our interpretations in section 4.2.
The final issue we need to address is that of possible endogeneity of our independents, i.e. possible correlations of these explanatory variables with the error term due to the fact that they are simultaneously determined with our dependent. We suspect that either some of our structural firm characteristics or some of our ownership/subsidiary dummies are endogenous. The possible endogeneity of a variable, however, can only be tested once one has strong and valid instruments: variables that correlate strongly with the suspected endogenous independent but little with the dependent. Once one has found such
instruments, endogeneity of the independent that arouses suspicion is tested by the Hausman test. Regarding the structural firm characteristic we could not find any suitable instruments33. Regarding the ownership/subsidiary dummies, it also proved hard to find suitable instruments for every firm segment. However, for both the dummies FORMNE and DOMMNE we did succeed, by using non-sample information. We found three instruments. The first is a dummy which indicates whether the firm’s ultimate owner is publicly listed at any stock exchange worldwide, which we named LISTED ( = 1 if listed). The second is firm’s total number of subsidiaries and affiliates owned34, both in The Netherlands and abroad: NSUBSAFS. The third is the total number of foreign countries in which the firm owns at least one majority-owned subsidiary: NFCNTRS. These instruments make economic sense, since MNEs are likely to be both larger (captured by NSUBSAFS and NFCNTRS) and more often publicly quoted (captured by LISTED) as they require more capital. . Furthermore, the instruments were successfully tested35 on both their strength and validity of overidentifying restrictions36. According to the
Hausman tests, FORMNE and DOMMNE are indeed endogenous using these instruments. Therefore, we felt confident to present our new results of a two-stage least squares (2SLS) estimation procedure in Table A4.13 in Appendix A. Specification (3) of the original model is run twice for both the LP- and LPVA model: once without CEMPLpEMPL included, once when it is included. The new results in Table A4.13 are remarkable, but in line of our previous findings: no significant relationships between our dummies
measuring a firm’s multinational status and its labour productivity remain. This means that neither the nationality of a firm’s owner, nor its ownership of foreign subsidiaries, and not even the interactions of these properties of a firm matters. Moreover, not even differences between industries seem to matter anymore, judging by the F-statistics in the rows below in the table. Thus, once mitigating possible endogeneity problems, only
33
The reader is invited to review Tables 4.7-4.10, which are the correlation matrices of all our variables. The author of this paper also considered as instruments information on firm characteristics. Namely, information that was collected, though which were not included in the final sample and econometric models (2) and (3). Such data is known as non-sample information to construct instrumental variables.
34
This means that ownership participations of 10-50% are also included, instead of only majority stakes.
35
The test on instruments’ strength was done by an auxiliary regression in the form of Adkins & Carter Hill
(2008: 261). This showed that F > 10 and individual t-tests > 3.3. The NR^2 test statistic for
overidentifying restrictions scores are well within any reasonable confidence range, making our instruments also valid apart from strong.
36
