Endogenous Switching: Endogenous Selection and Sample Selectivity

Một phần của tài liệu The Economics of Trade Secrets: Evidence from the Economic Espionage Act (Trang 260 - 264)

CHAPTER 6 THE DETERMINANTS OF TRADE SECRET INTENSITY

6.8 Endogenous Switching: Endogenous Selection and Sample Selectivity

 

The  use  of  prosecution  data  for  economic  analysis  is  rife  with  challenges.    One   major  challenge  is  that  of  self-­‐selectivity  within  the  sample.    As  a  whole,  this  can   be  referred  to  as  endogenous  switching  and  includes  both  endogenous  selection   and  sample  selectivity.    In  this  study,  the  data  are  restricted  to  only  those  cases   in  which  a  trade  secret  was  stolen,  the  theft  was  detected,  the  theft  was  reported   to  the  FBI,  the  FBI  referred  the  case  the  district  attorney  and  the  case  reached   court.    This  series  of  steps  means  that  the  EEA  cases  represent  a  small  portion  of   the  wider  population  of  trade  secrets,  and  even  stolen  trade  secrets.    Endogenous   variables  influence  the  observed  data  and  the  inclusion  of  observations  in  the   observed  data.    Hence,  a  strictly  OLS  regression  will  fail  to  account  for  this   sample  selectivity  or  endogenous  selection.      

 

This  section  of  the  study  uses  methods  to  correct  for  this  endogenous  switching   problem  including  the  Heckman  Correction  and,  while  not  strictly  an  

endogenous  switching  correction,  Truncated  Regression.      

 

                                                                                                               

355  Note  that  the  Stata  code  used  in  this  case  was  user-­‐generated  by  Jann  (2010)  and  thus  does   not  conform  to  the  standard  Stata  output.    In  this  case,  no  p-­‐values  are  calculated.  

6.8.1 Heckman  Correction    

A  method  to  correct  for  endogenous  switching  is  that  of  a  sample  selectivity   correction.    The  sample  is  sample  selected  in  that,  of  all  trade  secrets,  only  those   that  have  been  stolen  are  observed.    However,  when  working  with  the  Heckman   correction356  for  sample  selectivity,  the  procedure  needs  missing  values  on   which  to  base  its  analysis.    Hence,  the  use  of  a  complete  data  set  (i.e.  one  that  has   had  all  missing  values  replaced)  is  inappropriate  as  the  model  reverts  to  an  OLS   analysis.  

 

In  order  to  examine  the  sample  selectivity  of  TSI,  we  will  proceed  by  examining   the  original  data  set  (i.e.  before  adjusting  for  missing  values.)    However,  the  data   set  is  solely  for  those  cases  that  concluded  in  prosecution  and  no  information  is   available  for  unprosecuted  cases.    The  missing  values  are  due  to  incomplete   information  in  terms  of  the  availability  of  the  value  of  the  trade  secrets,  details   regarding  the  victim  firm,  ongoing  cases  and  other  complications  with  data   collection.    Thus,  the  data  set  will  only  allow  the  Heckman  correction  to  account   for  missing  information  and  not  the  sample  selection  concern  of  the  decision  to   prosecute.    Table  6-­‐15  below  reports  on  the  results  of  a  Heckman  correction   applied  to  the  data  using  the  logs  of  firm  size  (vsales)  and  a  valuation  of  the  trade   secret  (xref)  as  the  variables  in  the  selection  model.  

 

Table  6-­15:  Heckman  Correction  of  Log-­linear  model  with  Sectoral   Dummies;  Selection  Model  Based  on  Sales  and  Xref    

(no  model  of  missing  values)    

                                                                                                               

356  Also  known  as  the  Heckman  selection  model,  or  Heckman’s  estimator  for  sample  selection,  the   Heckman  correction  calculates  expected  value  of  the  error,  known  as  the  Inverse  Mill’s  Ratio   (IMR),  and  then  uses  it  as  a  regressor  in  the  linear  outcome  model.    The  IMR  is  the  ratio  of  the   probability  density  function  over  the  cumulative  distribution  function  of  a  distribution  and  is   calculated  using  a  probit  model.    See  Greene  (1993)  for  further  details.  

   

However,  a  number  of  problems  appear  with  the  results  in  Table  6-­‐15.      One  is   that  the  model  overall,  and  virtually  all  of  the  coefficients,  are  not  significant.    

The  other  is  that  ρ  is  equal  to  1,  which  indicates  that  the  sample  is  not  

conforming  to  the  Heckman  assumptions.    The  use  of  the  Heckman  correction  is,   therefore,  inappropriate.    Additionally,  the  sample  size  has  dropped  to  16  

making  the  analysis  rather  weak.    In  this  case,  the  Heckman  correction  does  not   further  the  analysis.  

6.8.2 Truncated  Regression    

Another  method  of  correcting  for  sample  selection  is  the  similar  concept  of   truncation.    Truncation  assumes  that  we  do  not  observe  variables  below  or   above  a  certain  level.    The  sample  is  truncated  in  that,  of  stolen  trade  secrets,   only  those  reaching  a  certain  minimum  value  of  value  reach  the  court.    That  is,   that  the  FBI  likely  investigates  only  those  trade  secrets  whose  value  exceeds  a  

minimum.    The  FBI’s  Reporting  Theft  checklist357  asks  victims  to  place  the  value   of  their  trade  secret  within  a  range.    As  discussed  in  Chapter  3,  FBI  Assistant   Direct  Chip  Burrus  “likened  the  FBI’s  current  fraud-­‐enforcement  policies  –  in   which  losses  below  $150,000  have  little  chance  of  being  addressed  –  to  ‘triage.’  

Even  cases  with  losses  approaching  $500,000  are  much  less  likely  to  be  accepted   for  investigation  than  before  9/11.”358    While  there  is  no  public  document  

supporting  this  triage  policy,  anecdotal  evidence  suggests  that,  in  practice,  it   exists.    If  this  were  the  case,  we  would  expect  that  a  truncated  regression  would   improve  the  analysis.  

 

The  use  of  missing  value  analysis  is  permitted  in  truncated  regression.359    Using   the  truncreg  method  in  Stata,  and  the  mean  inputted  for  missing  values,  we   get  the  following:  

Table  6-­16:  Truncated  Regression  for  Log-­linear  Model  with  Sectoral   Dummies  

  These  results  are  close  to  those  presented  in  Table  6-­‐4,  which  is  comforting  but,   in  this  way,  the  truncation  analysis  does  not  add  to  our  earlier  analysis.      

 

                                                                                                               

357  www.justice.gov/criminal/cybercrime/reportingchecklist-­‐ts.pdf.    

358  Shukovsky,  Paul  et  al  (2007.)  

359  Truncated  regression  assumes  that  the  observed  cumulative  density  function  is  a  truncation   of  the  standard  normal.    In  this  case,  $150,000  could  be  level  at  which  the  observations  are   truncated.    Taking  this  truncated  standard  normal  density  function,  Truncated  regression   performs  a  ML  estimate  with  a  normalized  density  function.  

Additionally,  it  suggests  that  the  FBI  is  not,  in  fact,  conducting  the  alleged  triage   discussed  earlier  and  in  Chapter  3.    This  is  supported  further  by  a  case  studies   analysis.    The  analysis  suggests  that  the  FBI  is  seeking  prosecution  for  a  wide   range  of  values  of  trade  secrets.    For  example,  in  the  case  of  Genovese,360  the  FBI   chose  to  prosecute  the  theft  of  Microsoft  source  code,  which  the  defendant  sold   for  $20.    The  decision  to  prosecute  a  theft  which  the  defendant  valued  at  a  mere  

$20  is  likely  a  case  of  the  FBI  strategically  choosing  to  prosecute  a  single   defendant  with  the  intent  of  dissuading  other  would-­‐be  thieves.    Thus,  in  the   absence  of  a  truncation,  the  truncreg  method  fails  to  enhance  the  analysis.  

Một phần của tài liệu The Economics of Trade Secrets: Evidence from the Economic Espionage Act (Trang 260 - 264)

Tải bản đầy đủ (PDF)

(328 trang)