Choice modelling
An  overview  of  theory  and  development  in  individual  choice  behaviour  modelling  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
L.T.  Wittink  
BMI  Paper  
August  2011  
Supervised  by  Alwin  Haensel  
 
 
 
 
 
  
 
Table  of  contents  
 
Introduction                          4  
Choice  Theory                        6  
Framework                        6  
Rational  Behaviour                      8  
Discrete  and  Probabilistic  Choice  Theory                8  
Discrete  Choice  Theory                    8  
Probabilistic  Choice  Theory                    9  
Utility  Theory                     10  
Cardinal  Utility                   10  
Ordinal  Utility                   10  
Constant  Utility                   10  
Random  Utility                   11  
Expected  Utility                   11  
Stated  and  Revealed  Preference                 12  
Stated  Preference                   12  
Revealed  Preference                   13  
Exogenous-­‐,  Locational-­‐  and  Utility-­‐Based  Choice  Models         13  
  Exogenous-­‐based  Models                 13  
  Locational-­‐based  Models                 14  
Binary  Choice  Models                   14  
Logit  and  Probit                     18  
Linear  Probability  Models                 18  
  Probit                       18  
Logit                       19  
Estimation                     19  
Alternative  estimation  models               20  
Multinomial  Logit                     20  
  Multinomial  choice                   21  
  Multinomial  logit                   21  
  Estimation                     22  

  1  
Nested  Logit                       23  
  Multidimensional  choice  sets               24  
  Nested  Logit                     24  
  Estimation  of  Nested  Logit                 26  
  Higher  level  Nested  Logit  and  expansion  on  the  Nested  Logit  Model     27
  Cross-­‐Nested  Logit                   27  
  Estimation  of  Cross-­‐Nested  Logit               28  
Mixed  Logit                       29
  Estimation                     30
  Repeated  Choice                   31  
Latent  Class  Logit                     31  
  Estimation                     33  
Variations:  different  choice  models               36  
  The  Generalized  Extreme  Value  Model             36  
  Joint  Logit                     36  
  Multinomial  Probit                   37  
  Mixed  Probit                     37  
  Further  Research                   37  
Summary                       37  
Acknowledgement                     38  
References                       38  
 
   

  2  
 “As far as the laws of math refer to reality, they are not certain; and as far as they are
certain, they do not refer to reality.”
Albert Einstein (1879 – 1955)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

  3  
Introduction  
 
To  some  degree,  all  decisions  or  even  most  of  the  actions  we  take  in  life,  involve  choice.  
When  we  go  to  the  supermarket,  we  have  to  decide  on  the  way  to  travel.  When  we  are  at  
the  supermarket  we  have  to  choose  from  a  selection  of  vegetables  for  example.  When  we  
are   at   home   we   have   to   decide   on   what   to   cook   etcetera.   A   day   in   our   life   is   full   of  
sequences   of   choices   we   have   to   make.   But   not   just   us:   our   entire   fellow   society   goes  
through  similar  thought  processes.    
  The   fact   that   a   whole   population   goes   through   such   processes   makes   it   worth  
investigating.  Namely  if  it  would  be  possible  to  make  an  indication  on  the  behaviour  of  a  
population  on  certain  processes,  these  processes  could  be  adjusted  likewise.  If  it  would  
be  possible  to  discover  some  pattern  in  behaviour  or  even  better,  to  discover  a  certain  
demand   in   a   process,   it   would   be   able   to   adjust   to   these   discoveries.   This   could   be   of  
great  help  of  course.  Thurstone  (1927)  is  often  said  to  be  one  of  the  first  people  to  do  
research  in  individual  choice  behaviour  into  food  preferences.  He  is  considered  to  be  the  
first   to   describe   this   preference   with   some   sort   of   utility   function.   Nowadays   choice  
models   are   used   in   various   areas:   for   example   psychology,   transport,   energy,   housing,  
marketing,  voting  and  actually  many  more.  
Since  Thurstone’s  research  there  has  been  quite  some  development  considering  
choice  models.  As  in  most  fields  of  research,  a  new  topic  often  causes  more  research  and  
more   elaborate   research.   Since   the   late   1920s   new   models   have   been   developed,  
theories   have   been   adjusted   and   original   assumptions   can   be   avoided.   As   newer   models  
were   developed,   not   all   of   these   models   were   applicable   due   to   computational  
constraints.   Together   with   technological   advancements   in   society,   choice   models   that  
were  unusable  became  suddenly  became  usable.  Currently  models  that  were  not  usable  
thirty  years  ago  are  usable  and  computationally  possibilities  only  increase.  
In  this  paper  an  attempt  is  made  to  describe  the  theory  behind  choice  models  and  
after   that   the   actual   models.   The   theory   behind   the   models   can   be   considered   a  
framework,   some   sort   of   foundation   of   the   models.   To   be   able   to   understand   the   models  
and   understand   where   they   come   from   and   what   assumptions   are   made,   this   will   be  
discussed   first.   In   this   first   section   on   individual   choice   behaviour   this   framework   will  
be  discussed  and  explained.  Most  of  the  definitions  needed  for  the  choice  models  will  be  
given   here.   After   that   comes   the   section   on   choice   models,   where   the   most   important  
and   through   time   most   used   and   referenced   models   will   be   discussed.   Here   no  
derivations  will  be  given  –  for  derivations  of  the  models  the  reader  is  referenced  to  more  
extensive   literature   as   Ben-­‐Akiva   and   Lerman   (1985)   and   Train   (2003).   Finally   the  
paper   will   be   concluded   with   a   section   called   Summary,   comments   and  
acknowledgement.   Included   in   this   section   is   a   chapter   with   variations   on   the   section  
before.  Unfortunately  it  was  not  possible  to  include  all  these  models  in  a  more  elaborate  
way,   but   choices   had   to   be   made   on   what   literature   to   discuss.   The   models   that   are  
discussed   follow   on   each   other   and   are   instrumental   either   because   they   are   so   often  
referenced   and   important   in   the   development   of   new   models   or   because   these   models  
are  used  currently.  I  hope  the  reader  finds  this  paper  informative  and  insightful  and  in  
the   end   has   a   better   understanding   of   how   choice   models   work   and   how   they   have  
developed  over  the  years.  
 
 
 
 

  4  
  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Individual Choice Behaviour: Framework
“Go down deep into anything and you will find mathematics.”
Charles Schlichter (unknown)

  5  
Choice  Theory  
Observing  the  choices  of  one  individual  is  interesting,  but  when  statements  can  be  made  
about  a  larger  group  of  individuals,  or  even  a  whole  population,  then  really  something  
can  be  achieved.  We  could  therefore  state  that  we  are  not  just  interested  in  the  choices  
of  a  single  person,  but  rather  of  large  groups  of  individuals.  Think  of  market  demand  for  
some   kind   of   service   or   commodity.   Predicting   the   demand   can   be   done   by   modelling  
individual   choice   behaviour,   thus   with   the   use   of   choice   models.   This   chapter   will   be  
used  mostly  to  describe  principles  of  choice  theories  and  to  give  a  framework,  which  will  
be  useful  when  formulating  the  different  discrete  choice  models.    
  When   examining   the   behaviour   of   individuals,   in   theory   we   look   for   behaviour  
that   is   according   to   Ben-­‐Akiva   and   Lerman   (1985)   descriptive,   abstract   and   operational.  
Descriptive   so   that   the   theory   describes   how   individuals   actually   behave   and   not   how  
we   expect   individuals   to   behave.   We   would   also   like   to   formalize   their   behaviour,  
independent  of  specific  circumstances,  therefore  abstract.  At  last  we  look  for  operational  
behaviour,   meaning   that   it   results   in   actual   models   with   measurable   parameters   and  
variables,  or  at  least  parameters  and  variables  that  can  be  estimated.    
  However  there  is  no  choice  theory  that  satisfies  all  these  requirements.  There  are  
choice  theories  that  have  these  requirements  as  ideology,  though  different  models  differ  
in   the   level   of   detail   in   which   they   idealize   the   thought   process   behind   observed  
behaviour.   There   are   some   common   assumptions   though,   which   are   used   for   the  
different   models.   These   assumptions   will   be   described   as   a   framework   for   the   models  
that  will  be  described  later  on.    
 
Framework  
Ben-­‐Akiva   and   Lerman   (1985)   state   that   ‘a   choice   can   be   viewed   as   an   outcome   of   a  
sequential  decision-­‐making  process  that  includes  the  following  steps:’  
 
1. Definition  of  the  choice  problem  
2. Generation  of  alternatives  
3. Evaluation  of  attributes  of  the  alternatives  
4. Choice  
5. Implementation  
 
This   means   that   a   choice   is   not   viewed   as   a   single   choice   at   a   specific   time,   but   rather   as  
a  process.  An  example  would  be  the  way  someone  would  travel  to  work.  He  could  take  
the  bus,  go  by  car,  take  the  bike  or  walk.  Here  the  definition  of  the  problem  would  be:  
how   to   get   to   work?   The   alternatives   are   stated   above.   Now   the   choice   is   not   dependent  
on   the   alternatives   themselves,   but   rather   on   their   characteristics,   or   attributes:   how  
expensive  is  every  alternative?  How  much  time  would  it  take  for  every  alternative?  Is  it  
really   feasible   to   walk,   meaning   what   level   of   comfort   does   it   provide?   Eventually   the  
decision  maker  applies  some  decision  rule,  which  is  some  sort  of  calculation  to  select  the  
best   alternative.   In   order   to   define   the   process   above,   we   need   to   define   the   elements  
decision   maker,   alternatives,   attributes   of   alternatives   and   decision   rule.   Note   that   we  
consider   actual   decision-­‐making   process   here   –   choices   following   from   habit,   intuition  
or   imitation   or   any   other   form   of   behaviour   where   there   is   no   rational   process   are  
represented   as   a   choice   process   with   only   one   alternative.   Rational   behaviour   will   be  
discussed  later.  
  The   decision   maker   can   be   an   individual,   but   also   a   household,   a   family   or   an  
organization.   Because   in   this   case   we   consider   the   organization   as   an   individual   we  

  6  
abstract   the   internal   interactions.   In   this   case   we   are   not   as   much   interested   in   different  
individual   choices,   because   every   individual   or   family   has   different   interests   and  
backgrounds.  We  are  more  interested  in  predicting  aggregate  demand,  and  we  must  still  
treat  the  differences  in  decision-­‐making  behaviour  explicitly.  
  Luce   (1959)   defined   different   choices   in   a   situation   as   alternative   choices,   or  
simply  alternatives  and  every  choice  is  made  from  a  set  of  alternatives.  The  environment  
of   the   decision   maker   determines   the   universal   set   of   alternatives,   but   single   decision  
makers   do   not   consider   all   alternatives.   For   example,   when   one   goes   to   work   the  
alternative  to  take  the  car  to  work  could  be  excluded,  because  the  decision  maker  does  
not   own   a   car.   Therefore,   each   decision   maker   considers   not   the   universal   set   of  
alternatives,  but  a  subset  hereof.  This  subset  is  called  a  choice  set.  This  set  includes  the  
alternatives   that   are   feasible   and   known   during   the   decision   process.   Finally   all  
alternatives   should   be   mutually   exclusive,   the   choice   set   needs   to   be   exhaustive   –  
meaning  that  all  possible  alternatives  are  included  in  the  choice  set  –  and  the  number  of  
alternatives  in  the  choice  set  must  be  finite.  
  As   stated   before,   choices   are   not   made   on   based   on   the   alternatives   apart,   but  
rather   on   the   characteristics   or   attributes   of   the   alternatives.   Ben-­‐Akiva   and   Lerman  
(1985)  state  that  ‘the  attractiveness  of  an  alternative  is  evaluated  in  terms  of  a  verctor  of  
attribute  values.’  Attributes  can  be  ordinal  or  cardinal.    
  If   there   are   multiple   alternatives   in   a   choice   set,   the   decision   maker   needs   a  
decision   rule   to   choose.   The   decision   rule   describes   the   internal   process   used   by   a  
decision   maker   to   process   the   available   information   and   make   a   unique   choice.   Slovic  
(1977)  and  Svenson  (1979)  give  us  rules  that  can  be  classified  in  four  categories:  
 
1. Dominance  means  that  an  alternative  is  better  than  another  alternative  when  at  
least   one   attribute   is   better   and   all   other   attributes   are   no   worse.   In   most  
situations  this  does  not  lead  to  a  unique  choice.  It  is  more  often  used  to  exclude  
the  worse  alternatives  from  the  choice  set.  It  can  be  made  more  complex  by  using  
a   threshold:   one   attribute   is   only   better   if   the   difference   between   both  
alternatives  exceeds  a  certain  threshold.  
2. Another   decision   rule   concerns   a   level   of   satisfaction.   This   means   that   every  
attribute  of  an  alternative  must  assume  a  level  of  satisfaction.  This  level  is  set  by  
the  decision  maker  and  should  be  attainable.    
3. The   third   type   of   decision   rule   is   called   lexicographical  rules.   This   means   that   the  
attributes   are   ordered   by   importance.   The   decision   maker   chooses   the   attributes  
he   values   the   most.   If   attributes   are   qualitative,   all   alternatives   that   do   not  
possess  the  desired  quality  will  be  excluded  from  the  choice  set.  In  the  case  the  
decision  maker  cannot  make  a  decision  based  on  the  most  important  attribute,  he  
will   continue   to   try   and   make   a   decision   based   on   the   second   most   important  
attribute.  
4. The   last   type   of   decision   rule   assumes   that   a   vector   that   defines   an   objective  
function   expressing   the   attractiveness   of   the   attributes   of   an   alternative  
expresses  the  attractiveness  of  an  alternative.  This  attractiveness  is  referred  to  as  
the  utility.  The  utility  is  a  measure  that  the  decision  maker  tries  to  maximize.  In  
the  section  utility  theory  this  will  be  discussed  more  elaborately.    
 
Of   these   four   categories,   the   utility   has   been   used   most   in   recent   models.   We   mostly  
refer  to  the  utility  as  a  function  (existing  of  a  vector),  the  utility  function.  
 

  7  
Rational  Behaviour  
The  term  rational  behaviour  is  based  on  the  beliefs  of  an  observer  of  what  the  outcome  
of   a   decision   of   a   decision   maker   should   be.   It   only   seems   natural   that   different  
observers   have   different   beliefs,   therefore   it   seems   that   there   cannot   be   one   universal  
type  of  rational  behaviour.  Thus  rationality  is  not  really  a  useful  concept  when  applied  
to   individual   choice   behaviour.   The   concept   described   in   literature   is   one   opposing  
impulsiveness,   which   means   that   individuals   do   not   make   decisions   based   on   their  
variable   psychological   state   at   the   time   the   decision   has   to   be   made.   It   means   that   it  
follows   a   consistent,   calculated   decision   process.   This   does   not   mean   that   the   individual  
cannot  follow  his  or  her  own  objectives.  
  In  1957  Simon  described  the  distinction  between  what  he  called  a  perfect  and  a  
bounded  reality.  In  a  perfect  world  an  all-­‐knowing  individual  exists,  capable  of  gathering  
and   storing   large   quantities   of   data   and   perform   complex   calculations   on   these   data   and  
is  able  to  make  consequent  decisions  based  on  the  data.  Bounded  reality  recognizes  the  
bounded  capacity  of  humankind  as  problem  solvers  with  limited  information-­‐processing  
capabilities.    
  This  means  that  rationality  is  a  quite  ambiguous  concept.  It  is  therefore  necessary  
to  introduce  a  specific  set  of  rules,  to  be  able  to  use  the  concept.  Simply  said  this  means  
that  we  assume  that  a  decision  maker,  if  alternative  A  is  more  feasible  than  alternative  B  
will   choose   A   every   time   he   faces   that   same   decision.   And   if   alternative   A   is   more  
feasible  than  B  and  B  is  more  feasible  than  C,  the  decision  maker  will  prefer  A  to  C.  
 
Discrete  and  Probabilistic  Choice  Theory  
This  section  can  be  seen  as  an  expansion  of  the  section  about  the  framework  of  choice  
theory.   The   concepts   used   in   choice   theory   are   similar   to   the   concepts   in   Economic  
Consumer   Theory,   which   will   not   be   treated   in   this   paper.   The   view   on   demand   in  
Economic   Consumer   Theory   is   well   applicable   when   the   feasible   choices   have  
continuous   variables,   but   this   might   not   always   be   the   case.   In   discrete   choice   theory  
types  of  problems  are  better  described  as  discrete  bundles  of  attributes.  Furthermore,  in  
probabilistic   choice   theory   the   probability   that   a   decision   maker   chooses   a   certain  
alternative  can  be  provided,  which  makes  it  a  powerful  framework  when  working  with  
discrete  choice  situations.  
 
Discrete  Choice  Theory  
  Table  1  Choice  in  travelling  to  work      
 
Attributes          
 
 Alternatives   Travel  Time  (t)   Cost  (c  )   Comfort  (o)  
 
  Car   t1   c1   o1  
  Bus   t2   c2   o2  
  Walk   t3   c3   o3  
For   the   section   on   discrete   choice   theory,   consider   a   simple   example   that   is   used   in  
literature   (Ben-­‐Akiva   &   Lerman   1985,   Train   2003)   more   often.   Consider   a   decision  
maker  that  has  to  travel  to  work.  He  has  three  options:  take  the  car,  take  the  bus  or  walk.  
The  attributes  of  the  alternatives  are  travel  time,  travel  cost  and  comfort  (see  table  1).  
The  choice  will  have  the  utility  function  U  =  U(q1,  q2,  q3)  with  q  is  the  alternative  
chosen.  Obviously  the  decision  maker  can  only  choose  one  alternative,  this  qi  =  1  if  mode  
i  is  chosen  and  qi  =  0  if  otherwise,  for  all  i  in  the  choice  set  and  q1q2  =  q2q3  =  q1q3  =  0.  
Because  the  possibilities  U(1,0,0),  U(0,0,1)  and  U(0,1,0)  are  not  differentiable  we  apply  a  

  8  
maximization  on  the  function  with  the  attributes  as  parameters:  Ui  =  U(ti,  ci,  oi).  Now  we  
can  see  that  U1  >  U2  and  U1  >  U3  have  to  be  true  for  alternative  1,  taking  the  car,  to  be  
chosen.      
As   for   the   form   of   the   utility   function,   in   most   literature   an   additive   utility  
function  is  assumed:    
Ui  =  -­‐  β1ti  –  β2ci  +  β3oi,    
for   all   i   in   the   choice   set   and   βi   >   0   for   all   i.   With   this   formula   we   can   try   to   predict  
changes   in   U   for   different   numerical   values   for   the   parameters.   This   approach   to   the  
utility   function   is   called   revealed  preference   and   will   be   discussed   more   elaborately   in  
the  next  chapter.    
Lancaster  (1966)  defined  the  utility  function  Uin  =  U(xin),  where  xin  is  the  vector  of  
the   attribute   values,   for   every   alternative   i   by   every   decision   maker   n.   Ben-­‐Akiva   and  
Lerman   expand   this   formula   a   bit,   due   to   variability   in   population:   Uin   =   U(xin,  sn),   where  
sn   is   a   vector   of   different   characteristics   of   the   decision   maker,   for   example   income,   age,  
education  and  ethnical  background.  
 
Probabilistic  Choice  Theory  
In  probabilistic  choice  theory,  it  is  argued  that  we  cannot  approximate  human  behaviour  
by   deterministic   parameters.   It   seems   plausible   to   state   that   human   behaviour   has   a  
probabilistic   nature.   Furthermore,   it   can   be   argued   that   whilst   the   decision   maker   has  
knowledge   of   his   or   her   utility   function,   the   researcher   or   analyst   does   not   know   the  
exact  form.  Therefore  Train  (2003)  explains  about  the  term  representative  utility.  In  the  
section   about   the   framework   the   utility   function   of   the   form   U   was   introduced.   The  
decision   maker   chooses   the   alternative   if   Uin   >   Ujn   ∀   j   ≠   i,   where   j   are   the   different  
choices   from   the   choice   set   (Cn)   and   n   is   the   labelled   decision   maker.   Since   there   are  
aspects   of   utility   function   of   a   decision   maker   that   the   researcher   does   not   know,   we  
introduce   the   representative   utility   function   Vjn   =   V(xjn,   sn)   with   xnj   ∀   j,   again   the  
attributes  of  the  alternatives  and  sn  some  attributes  of  the  decision  maker.    
Because  V  depends  on  characteristics  the  researcher  cannot  know,  it  makes  sense  
that  Vjn  ≠  Ujn.  Train  states  that  the  utility  can  be  decomposed  as  Ujn   =  Vjn   +   εjn,  where   εjn  
captures  the  factors  that  affect  utility  but  are  not  known  to  the  researcher  and  therefore  
are  not  included  in  Vjn.  Simply  said   εjn  is  the  difference  between  Ujn  and  Vjn  and  could  be  
considered  an  error  term.  
It  seems  logical  that  if   εjn  are  factors  that  affect  the  utility,  but  are  not  known  by  
the  researcher,  the  form  of   εjn  is  unknown  as  well.  Therefore  these  terms  are  treated  as  
random.  The  joint  density  of  the  vector  of  these  ‘errors’  is  denoted  f(εn).    
Pin = Pr(! jn ! !in < Vin !Vjn "j # i)
 
= $ I(! jn ! !in < Vin !Vjn "j # i) f (! n )d! n .
!
The  first  part  is  due  to  the  probability  that  the  decision  maker  chooses  alternative  i.  In  
this  part  I(…)  is  the  indicator  function,  which  is  equal  to  1  if  the  statement  between  the  
parentheses   is   true   and   0   if   not.   This   is   a   multidimensional   integral   and   only   takes   a  
closed   form   for   specific   forms   of   the   density   function   f.   For   example   logit   and   nested  
logit   are   models   that   will   be   discussed   later   on,   but   those   have   a   closed   form   for   this  
integral.  Probit  and  mixed  logit  are  derived  differently  and  do  not  have  an  open  form  for  
this  integral.  They  are  not  calculated  exactly  but  they  are  evaluated  numerically.    

An  example  might  help  to  clarify.  Consider  the  example  used  in  the  last  section,  
but  this  time  without  the  attribute  comfort.  We  can  define  Vi   =   -­‐  β1ti   –   β2ci  with  i  =  {car,  

  9  
bus,   walk}.   Now   suppose   after   analysis   the   researcher   finds   that   Vcar  =   4,   Vbus   =   3   and  
Vwalk  =  3.  This  does  not  mean  that  the  decision  maker  will  choose  to  go  to  work  by  car.  It  
simply   means   that   by   observed   factors   it   seems   best   to   go   by   car,   but   there   are   still  
factors  that  are  unobserved  to  the  researcher.  The  probability  that  the  decision  maker  
walks  to  work  instead  of  taking  the  car  is  the  probability  that  the  unobserved  factors  for  
walking   are   sufficiently   better   than   those   for   taking   the   car.   As   stated   in   the   formula  
above  it  would  be:  

Pwalk  =  Pr(εcar    -­‐  εwalk  <  Vwalk  –  Vcar).    

 
Utility  theory  
Anand   (1993)   states   that   decision   theory   is   about   ‘choosing   the   act   that   is   best   with  
respect   to   the   beliefs   and   desires   that   an   agent   holds.’   He   states   that   utility   theory   helps  
achieve   this.   Loosely   said   decision   theory   is   about   maximizing   utility,   given   the  
‘attributes’  of  an  agent  or  rather  decision  maker.  Many  theories  exist  about  utility.  Ben-­‐
Akiva   and   Lerman   divide   utility   theory   into   two   possible   types:   constant   utility   and  
random   utility.   In   economics,   usually   the   difference   between   cardinal   and   ordinal   utility  
is  used,  which  are  in  fact  not  very  different  from  the  former  mentioned  types.    
 
Cardinal  utility  
Cardinal   utility   is   usually   considered   out-­‐dated   though,   as   is   the   constant   utility  
approach,   because   it   is   not   really   in   line   with   consumer   theory.   In   cardinal   theory   the  
magnitude   of   difference   between   utility   values   is   treated   as   behaviourally   significant.  
For   example   in   the   case   of   the   example   that   has   been   used:   if   taking   the   car   to   work  
takes   10   minutes,   taking   the   bus   takes   20   minutes   and   walking   takes   30,   one   can   say  
that   taking   the   car   is   as   much   better   than   taking   the   bus   as   is   taking   the   bus   than  
walking.  But  one  cannot  say  that  taking  the  car  is  twice  is  good  as  taking  the  bus.  From  
this   example   comes   forward   that   comparisons   between   alternatives   is   meaningless,  
because  there  is  no  good  way  to  interpret  differences  between  them.  Nowadays  instead  
of   cardinal   utility   often   preferences   are   used.   Stated   and   revealed   preference   will   be  
discussed  in  the  next  chapter.  
 
Ordinal  utility  
When  using  ordinal  utility  and  considering  the  example  above,  one  can  say  that  taking  
the   car   is   preferred   to   the   bus   to   walking,   but   one   cannot   state   anything   about   the  
strengths   of   these   preferences.   Thus   using   ordinal   utility   it   is   possible   to   capture  
ranking,   but   not   relative   strengths.   In   ordinal   utility   theory,   there   is   usage   of   a   utility  
function  to  capture  rank,  as  is  the  case  with  constant  and  random  utility.  
 
Constant  utility  
The  values  for  the  utilities  of  the  different  alternatives  are  fixed  in  this  approach.  Here  it  
is  not  the  case  that  the  decision  maker  chooses  the  alternative  with  the  highest  utility,  
but   it   is   assumed   that   there   are   choice   probabilities   involved.   These   probabilities   are  
defined  by  a  probability  density  function  (PDF)  over  the  different  alternatives,  with  the  
utilities   as   parameters.   Selecting   a   certain   PDF   in   this   approach   can   only   be   based   on  
specific  assumptions  with  respect  to  the  properties  of  choice  probabilities.  An  important  
property  of  this  approach  is  the  independence  from  irrelevant  alternatives  (IIA)  that  is  
noted  as  follows:    

  10  
 P(i | C n ) P(i | Cn )
= ,     with   i, j ! Cn " Cn  
P( j | C n ) P( j | Cn )
which   simply   said   means   that   removing   irrelevant   alternatives   from   the   choice   set   Cn,  
resulting  in  subset   Cn ,  has  no  influence  on  the  choice  probabilities.    
   
Random  utility  
Manksi   (1977)   formalized   this   approach,   which   is   more   in   line   with   consumer   theory  
than   the   constant   utility   approach.   The   observed   inconsistencies,   or   errors,   are   now  
viewed   as   to   be   result   of   observational   inaccuracies   on   the   researcher’s   side.   In   this  
approach  we  again  assume  that  a  decision  maker  tries  to  maximize  his  or  hers  utility,  as  
is   in   line   with   economic   consumer   theory.   But   –   as   stated   in   the   section   about  
probabilistic   choice   theory   –   the   researcher   does   not   know   the   utility   of   a   decision  
maker  with  full  certainty  and  therefore  they  are  treated  as  random  variables.  We  can  say  
that   the   researcher   defines   the   choice   for   a   specific   alternative   i   in   the   choice   set   as  
P(i | Cn ) = Pr(Uin ! U jn , "j # Cn ) ,    
as   stated   by   Ben-­‐Akiva   and   Lerman   (1985).   Here   we   assume   a   joint   PDF   for   the   set   of  
(random)   utilities,   because   a   logical   argument   can   be   made   about   the   underlying   source  
for  randomness  in  the  utilities.  Manski  (1973)  identified  four  sources:  
 
1. Unobserved   attributes:   the   vector   of   attributes   that   affects   the   decision   is  
incomplete  to  the  researcher.  Therefore  there  is  an  element  included  in  the  utility  
function  that  is  observationally  random,  thus  it  follows  that  the  utility  is  random  
as  well.  
2. Unobserved  taste  variations:   the   researcher   knows   all   attributes,   but   there   is   an  
unobserved  argument  that  is  unknown  to  the  researcher.  This  can  be  explained  
as   the   researcher   being   unknown   of   the   specific   taste,   of   preference,   of   the  
decision   maker.   The   variation   of   this   argument   is   unknown,   making   the   utility  
random.  
3. Measurement  errors:  the  attributes  of  the  alternatives  are  not  observable.  That  is  
why   the   researcher   estimates   the   attributes,   with   a   measurement   error  
accounting   for   probable   inaccuracy   in   the   measurement.   This   error   term   is  
unknown,  resulting  in  the  utility  becoming  random.  
4. Instrumental  variables:   the   true   utility   function   is   known,   but   some   elements   in  
the   vector   of   attributes   are   not   observable.   The   researcher   approaches   these  
variables   by   a   function   that   is   derived   by   the   relation   of   known   variables.   This  
means  the  utility  functions  contains  instrumental  variables,  which  are  in  nature  
an   expression   of   an   imperfect   relation   between   estimate   and   actual   attribute.  
Again  this  term  contains  a  random  error,  making  the  utility  random.    
 
Expected  utility  
Expected   utility   is   also   known,   especially   because   it   is   one   of   the   underlying  
assumptions   in   game   theory.   This   approach   deals   with   analysis   of   choices   in   risky  
projects.   This   is   one   of   the   oldest   utility   approaches,   as   it   was   formulated   in   1713   by  
Nicholas   Bernoulli   and   solved   in   1738   by   Daniel   Bernoulli.   Savage   (1954)   formulated  
the   subjective   expected   utility   theory,   which   is   a   more   up-­‐to-­‐date   work   and   was  
reviewed  by  Anand  (1993).  If  an  uncertain  event  has  a  number  of  possible  outcomes  zi  
all   have   utilities   U(zi),   and   there   is   a   subjective   probability   P(zi),   then   the   subjective  
expected  utility  would  be:  

  11  
 ! U(z )P(z ) .    
i i i

Savage   also   stated   eight   axioms   that   also   suit   well   with   the   other   utility   approaches.  
These  axioms  are:  
 
1. Completeness:  if  x  and  y  are  two  alternatives,  either  x  or  y  is  preferred.  Also,  x  and  
y  are  equally  desirable.  
2. Transitivity:  if  x  is  preferred  to  y  and  y  to  z,  then  x  is  preferred  to  z.  
3. Independence:  x  and  y  should  be  independent  of  each  other.  
4. Resolution  independence:  Preference  for  an  alternative  x  or  y  only  depends  on  the  
attributes  of  the  alternative  ex  ante.  
5. Expected   wealth   independence:  The  preference  for  an  alternative  depends  on  the  
chance  of  winning  and  not  on  the  size  of  the  stakes:  if  there  is  a  lottery  and  one  
can  choose  between  lottery  A,  which  has  a  15%  chance  of  winning  and  a  100  euro  
reward  and  lottery  B,  which  has  a  5%  chance  of  winning  and  a  1000  euro  reward,  
the  decision  maker  will  choose  to  participate  in  lottery  A.  
6. Minimal  strict  preference:  there  is  at  least  one  vector  of  attributes  that  is  strictly  
preferred  to  the  other  vectors  of  attributes.  
7. Continuity   in   probability:  very  unlikely  events  should  be  regarded  as  having  zero  
probability.    
8. Partial   resolution   independence:   if   the   attributes   of   x   are   preferable   to   the  
attributes  of  y  for  different  states,  then  x  is  preferred  to  y  if  one  of  the  states  is  
obtained.  
 
Stated  and  Revealed  Preference  
Now   the   main   framework   behind   choice   models   has   been   treated,   we   will   turn   to   When  
considering   data   involved   in   choice   models,   it   is   possible   to   divide   these   data   in   two  
distinct   types:   stated   preference   (SP)   data   and   revealed   preference   (RP)   data.     In   the  
following  two  sections  I  will  give  a  description  of  the  models  and  point  out  strengths  and  
weaknesses  of  the  both  data  types.  These  sections  will  not  contain  specific  methods  of  
both   approaches,   but   it   will   give   insight   into   these   approaches   and   how   they   differ   from  
each  other.  
 
Stated  preference  
According   to   Kroes   and   Sheldon   (1988)   SP   methods   refer   to   ‘a   family   of   techniques  
which   use   statements   of   individual   respondents   about   their   preferences’   in   a   set   of  
alternatives   to   estimate   utility   functions.   SP   data   are   collected   through   experimental  
situations   or   surveys   where   the   respondents   are   faced   with   hypothetical   choice  
problems.   For   example,   the   respondent   is   asked   to   choose   between   five   bikes.   In   this  
hypothetical  situation  only  these  five  bikes  exist.  The  response  is  the  stated  choice.    
  Another  way  to  describe  the  SP  approach  would  be  the  direct  approach,  because  
the  data  comes  directly  from  the  respondents,  the  hypothetical  decision  makers.  Due  to  
this  approach  the  data  does  not  describe  actual  behaviour,  but  it  describes  how  decision  
makers  state  they  would  behave  alike.  A  strong  point  from  this  data  is  that  it  can  give  an  
indication  how  respondents  would  behave  in  a  situation  that  currently  not  (yet)  exists.  
So  if  it  is  the  researcher’s  objective  to  examine  behaviour  for  example  in  a  product  that  
does  not  yet  exists,  the  SP  approach  would  be  suitable.  Also  SP  data  works  very  well  in  
data  that  contain  little  or  no  variation,  because  the  questionnaire  can  be  designed  in  a  
way  that  will  result  in  the  data  having  the  desired  variation.    

  12  
   The   main   disadvantage   of   SP   data   seems   obvious:   the   way   respondents   expect  
themselves   to   behave,   moreover   the   way   respondents   say   they   will   behave,   is   not   the  
way   they   actually   will   behave.   This   phenomenon   may   arise   because   the   respondent  
actually  does  not  know  how  they  would  respond  or  because  he  or  she  feels  it  is  expected  
of  them  to  respond  in  a  specific  way.    
 
Revealed  preference  
In  contrast  to  SP  data,  RP  data  relate  to  actual  behaviour.  It  is  called  RP  because  decision  
makers  reveal  their  preference  through  the  choices  they  make.  In  the  example  used  in  
the   SP   section   and   considering   the   RP   approach,   the   respondent   would   be   asked   what  
bike   he   or   she   bought   last,   instead   of   choosing   from   a   selected   set   in   a   hypothetical  
situation.   Therefore   we   can   state   that   purchasing   or   choosing   habits   reveals   preference.  
In  this  approach  utility  functions  are  defined  by  observing  behaviour.  
  Where   the   SP   approach   is   called   the   direct   approach,   the   RP   approach   is   called  
the   indirect   approach.   In   the   RP   approach   actual   behaviour   is   observed,   instead   of  
confronting   the   respondent   with   a   hypothetical   situation.   The   largest   advantage   of   RP  
data  is  that  the  data  represents  actual  choices.    
  The   downside   of   RP   data   representing   actual   choices   is   that   it   not   suitable   for  
situations   that   currently   not   exist.   Because   we   observe   behaviour,   there   is   too   much  
uncertainty   in   stating   behaviour   in   new   situations.   Of   course   approximations   can   be  
made,  but  SP  data  is  simply  better  suitable  for  these  situations.  Also  in  situations  with  
little  or  no  variation  RP  data  is  not  suitable,  because  relations  between  different  cannot  
be  estimated  well  without  variation.  
 
By  using  an  estimation  procedure  that  allows  the  relative  importance  of  the  attributes  to  
be   estimated   through   primarily   SP   data   and   at   the   same   time   allows   the   alternative-­‐
specific   constants   and   overall   scale   of   the   parameters   to   be   determined   by   RP   data  
(Train,  2003),  the  strengths  of  both  approaches  can  be  used.  This  will  not  be  discussed  
into   detail   in   this   paper,   but   Hensher   et   al   (1999)   and   Brownstone   et   al   (2000)   describe  
this  approach  for  respectively  logit  models  and  mixed  logit  models.  
 
 
Exogenous,  Locational  and  Utility-­‐based  Choice  Models  
Until  now  only  choice  models  based  on  utility  and  especially  random  utility  have  been  
considered.   The   utility   models   are   models   that   involve   a   set   of   alternatives,   a   decision  
maker   and   some   utility   function   that   describes   how   the   decision   maker   chooses   the  
most  attractive  alternative  to  them.  In  other  words,  the  decision  maker  bases  his  or  her  
decision   on   the   attributes   of   the   alternatives   and   chooses   the   most   preferable  
alternative  through  some  decision  process  described  by  the  utility  function.  Now  there  
are   different   types   of   models   besides   utility-­‐based   models.   These   models   will   not   be  
discussed  elaborately,  but  it  will  be  good  for  the  reader  to  be  aware  that  other  types  of  
models  exist.    
 
Exogenous-­‐based  Models  
Paul  Waddell  (1993)  investigates  whether  the  assumption  that  the  choice  of  workplace  
is   exogenously   determined   in   models   of   residential   location   is   true.   So   this   research  
builds  very  much  on  McFadden’s  (1978)  research.    Exogenous-­‐based  models  state  that  
choices   are   driven   by   outside   factors   and   are   therefore   very   different   from   utility-­‐based  
models,  where  the  decision  made  is  dependent  on  characteristics  of  the  decision  maker  

  13  
as  well  as  attributes  of  the  alternatives;  one  could  say  all  endogenous  variables.  There  
has   been   debate   whether   or   not   locational-­‐based   models   or   somewhat   exogenous,   since  
the   paper   by   Waddell.   Before   the   1990s   locational-­‐based   models   were   assumed   to   be  
exogenous,   as   residential   location   was   assumed   to   driven   by   mostly   workplace.   Even  
now   one   could   argue   if   this   makes   a   model   exogenous,   as   workplace   is   not   the   only  
deciding   factor   and   other   factors   might   not   be   exogenous.   In   his   paper,   Waddell  
reaffirms   many   of   the   influences   assumed   in   urban   economic   theory,   the   same   accounts  
for  the  assumption  on  the  relation  between  workplace  and  residential  location.    
 
Locational-­‐based  Choice  Models  
This   type   of   model   is   not   that   different   from   utility-­‐based   models   as   one   might   think.  
The  former  models  developed  in  the  1960s  (Alonso  1964,  Muth  1969)  originate  from  a  
model   called   the   monocentric   model   or   are   derived   from   gravity   model   (Lowry   1964)  
derivatives.   These   models   will   not   be   discussed   here,   but   they   do   have   a   very   important  
assumption   in   common:   workplace   choice   is   exogenous   in   determining   the   residential  
location   choice   of   households.   So   there   is   a   link   with   exogenous-­‐based   models.   Still   very  
often   researchers   reference   to   the   work   of   McFadden   (1978)   where   he   described   how  
utility-­‐maximizing   consumers   are   assumed   when   considering   residential   housing,   thus  
locational   choice.   In   this   type   of   choice   model   there   is   also   some   random   part   in   the  
utility.  McFadden  points  out  that  the  MNL  and  NL  model,  which  will  be  discussed  later  in  
this  paper,  are  very  usable  in  locational-­‐based  choice  modelling.    
  A   difference   to   some   extent   in   locational-­‐based   choice   modelling   is   that   the  
characteristics   of   alternatives   are   not   just   decided   by   their   own   attributes,   but   also   by  
some  external  attributes.  Think  of  climate,  image  of  the  location  or  employment  in  the  
area.   It   is   possible   that   alternatives   should   be   placed   in   a   Nested   Logit   model   which  
allows  for  overlap  (Cross-­‐Nested  Logit)  but  also  allows  for  classes  to  overlap,  as  houses  
belong  to  different  classes  but  also  are  defined  by  exogenous-­‐driven  factors.  McFadden  
(1978)   concludes   that   the   problem   of   modelling   disaggregate   choice   of   housing   location  
is   impractically   large.   So   to   a   certain   extent   locational-­‐based   choice   models   do   not   differ  
as  much  from  utility-­‐based  choice  models  as  the  foundation  is  similar.      
  Head   et   al.   (1995)   discussed   location   choice   through   the   example   of   Japanese  
manufacturing   investments   in   the   United   States.   They   state   that   firms   in   the   same  
industry  are  drawn  to  similar  locations  because  proximity  causes  positive  externalities.  
This   is   very   much   in   line   with   McFadden   who   stated   that   with   location   choice   much  
more   exogenous-­‐driven   variables   play   a   role.   Head   et   al.   also   state   that   chance   events  
can  have  a  lasting  influence  on  the  geographical  pattern  of  manufacturing  in  that  case.  
This  is  also  a  big  difference  with  utility-­‐based  modelling.  There  the  assumption  is  made  
that   in   the   largest   sense   rational   processes   take   place   and   therefore   if   the   same   decision  
maker   conducts   the   same   decision   processes,   similar   outcomes   will   take   place.   In   the  
case   of   location   choice   this   apparently   is   not   the   case,   as   choices   made   also   have   an  
effect  on  the  actual  alternatives.  
 
 
Binary  Choice  Models  
In  this  chapter  a  general  background  of  binary  choice  models  will  be  given,  that  will  be  
used  in  the  following  chapters  when  specific  models  will  be  discussed.    
As  stated  in  the  chapter  on  choice  theory,  we  have  a  decision  maker  facing  a  set  of  
feasible   discrete   choice   alternatives   and   he   or   she   will   select   the   alternative   with   the  

  14  
greatest  utility,  with  the  utility  a  random  variable  (r.v.).  As  in  random  utility  theory,  the  
probability  that  a  decision  maker  will  select  a  certain  alternative  is    
P(i | Cn ) = Pr(Uin ! U jn , "j # Cn )
.    
If   the   choice   set   Cn   exists   of   only   two   alternatives,   i   and   j,   we   have   a   so   called   binary  
choice   model.   In   this   case   we   can   state   the   probability   that   decision   maker  n   will   choose  
alternative  i  or  j  is:  
 
Pn (i) = Pr(Uin ! U jn ) P ( j) = 1! Pn (i)
 and   n .  
  Ben-­‐Akiva   and   Lerman   (1985)   describe   how   theory   described   in   the   last   chapter,  
random  utility  theory,  can  be  made  operational:  
 
1. Separating   the   total   utility   into   deterministic   and   random   components   of   the  
utility  function  
2. Specify  the  deterministic  component  
3. Specify  the  random  component  
 
Remember  that  in  the  section  of  probabilistic  choice  theory  we  stated  that  for  the  utility  
we  have  an  observed  part  and  an  unobserved  part,  we  also  called  disturbance:  
Uin = Vin + !in ,
U jn = Vjn + ! jn  
Here   Vin   and   Vjn   are   the   systematic   components.   In   the   chapter   on   choice   theory   V   is  
described   as   the   part   of   the   utility   that   can   be   observed   by   the   researcher.   These  
components   are   assumed   to   be   deterministic.   V   can   be   thought   of   as   the   means   of   U.   We  
can   shift   the   scale   of   measurement   by   transforming   both   Uin   and   Ujn   by   any   strictly  
monotonic   increasing   function.   Ben-­‐Akiva   and   Lerman   show   that   adding   a   constant   to  
both   utilities   has   no   affect   on   the   choice   probabilities.   It   does   change   Vin   and   Vjn,   but  
eventually  that  is  no  problem.  The  absolute  levels  of  V  and  ε  do  not  matter,  what  does  
matter   is   that   Vin !Vjn < !in ! ! jn .   Though   by   specifying   just   the   differences   instead   of  
individual  components  could  develop  binary  choice  models,  usually  each  utility  function  
is   specified   separately   for   the   sake   of   continuity.   There   exist   choice   models   with   more  
than  two  alternatives;  therefore  the  same  notation  is  used  for  binary  choice  models.    
  After   separating   the   utility   into   a   deterministic   and   a   random   part,   we   specify  
both   parts,   starting   with   the   deterministic   or   systematic   part.   As   V   does   not   just   depend  
on   the   underlying   attributes,   but   also   on   attributes   of   the   decision   maker,   we   can   define  
V   as   V(zin,  Sn),   as   is   similar   to   a   description   in   the   chapter   on   choice   theory.   Seeing   as  
these  two  vectors  z  and  S  are  actually  combined  to  describe  V,  we  define  a  new  vector  xin  
=  h(zin,  Sn),  with  h  is  some  vector-­‐valued  function.  Now  we  can  write  Vin  =  V(xin)  and  Vjn  =  
V(xjn).   Secondly,   a   functional   form   for   V   is   needed.   Because   we   would   like   it   if   the  
function  was  to  reflect  theory  about  how  the  elements  in  x  influence  utility  and  we  want  
the   function’s   parameters   to   be   estimated   easily,   most   researchers   have   chosen   to   use  
functions  that  are  linear  in  the  parameters.  If  we  define  β  =  [ !1, !2 , …, ! K ]  as  the  vector  
of  K  unknown  parameters,  we  can  write  
Vin = !1 xin1 + !2 xin2 +... + ! K xinK ,
 
Vjn = !1 x jn1 + !2 x jn2 +... + ! K x jnK .

  15  
for  both  observed  utilities.  One  important  characteristic  is  that  linearity  in  parameters  
does  not  mean  linearity  in  the  attributes  z  and  S.  This  totally  depends  on  the  form  of  h.  
Finally,  here  it  is  assumed  that  the  parameters   !1, !2 , …, ! K are  the  same  for  the  whole  
population.   It   is   possible   however,   that   market   segmentation   is   present.   Then  
!1, !2 , …, ! K  are  treated  as  r.v.’s  distributed  across  the  population.  
  Finally  we  need  to  specify  the  disturbances  to  obtain  an  operational  binary  choice  
model.   Where   in   the   last   paragraph   the   functions   for   V   were   depicted   separately,   now  
the  functions  for  the  disturbances  ε  will  be  depicted  in  the  most  convenient  way.  So  the  
differences   εjn   –   εin   are   discussed.   As   stated   before,   the   choice   probabilities   are  
unaffected  if  we  add  a  constant  to  both  disturbances.  Besides  this  it  also  will  not  make  
any   difference   if   the   mean   of   the   disturbance   is   shifted,   as   long   as   the   systematic  
component   is   shifted   by   the   same   amount.   From   this   follows   that   the   means   of   the  
disturbances  can  be  fully  represented  by  any  constant  without  loss  of  generality.  Usually  
it  is  assumed  that  all  disturbances  have  zero  means.  In  addition  to  the  mean,  the  scale  of  
the   disturbances   should   be   consistent   with   the   scale   of   the   functions   V.   As   for   the  
functional  form  of  the  distribution  of  the  disturbances,  it  does  not  make  sense  to  think  of  
the   distribution   of   the   ε’s   to   be   different   from   the   V’s   distribution.   Especially   since   the  
disturbances  reflect  the  different  sources  of  observational  error,  different  specifications  
of   V   will   lead   to   different,   fitting,   distributions   for   ε.  Because   there   are   many   different  
unobserved   factors   that   affect   the   overall   distribution,   it   is   hard   to   make   strong  
statements  about  this  distribution.  However,  nowadays  more  and  more  we  gain  insight  
about  what  is  included  into  the  disturbances.  
   

  16  
  

Choice models
“If people do not believe that mathematics is simple, it is only because they do not realize
how complicated life is.”
John von Neumann (1903-1957)

  17  
Logit  and  Probit  
Now  that  the  framework  of  choice  theory  in  general  and  binary  choice  models  has  been  
set,   we   are   now   able   to   talk   about   specific   models.   There   are   three   common   binary  
models:   the   linear   probability   model,   the   (binary)   logit   model   and   the   (binary)   probit  
model.  These  models  were  both  discussed  by  Thurstone  (1927)  to  some  extent.  
The  differences  between  these  models  are  based  on  the  assumption  that  is  made  
about  the  distribution  of  the  disturbances,  or  the  difference  between  the  disturbances.  
To   obtain   the   eventual   model,   the   choice   probabilities   can   be   derived   under   the  
assumption  about  the  disturbances.  
 
Linear  Probability  Model  
The   easiest   of   the   three   models   is   the   linear   probability   model.   In   this   model   the  
difference  in  the  disturbances  is  uniformly  distributed:   ! jn ! !in ~ Unif (!L, L) ,  where  L  >  
0.   The   difference   between   the   disturbances   is   defined   as   εjn   –   εin   =   εn,   with   density  
function  f(εn).  Here    
Pn (i) = Pr(! n ! Vin "Vjn ) .    
The  choice  probability  is  given  by  the  cumulative  distribution  function  of  εn.  When  V  is  
linear  in  its  parameters,  the  probability  function  is  linear  as  well  between  –L  and  L.    
According  to  Cox  (1970)  this  model  has  a  major  drawback:  unless  restrictions  are  
placed   on   the   β’s   (which   are   again   used   to   estimate   V),   the   estimated   coefficients   can  
imply  probabilities  outside  the  interval  [-­‐L,  L].  Therefore  the  logit  and  probit  models  are  
used   more   often.   Besides   this   drawback,   it   is   unrealistic   to   assume   the   interval   [-­‐L,  L],  
and  zero  probabilities  outside  this  interval.    
 
 Probit  
Another   way   to   view   the   disturbances   is   as   being   the   sum   of   a   large   number   of  
unobserved,   independent   constituents.   Due   to   the   large   number   and   the   central   limit  
theorem  the  disturbances  tend  to  be  normally  distributed.    
  Now   we   can   state   that   εin  and  εjn   both   have   a   normal   distribution   with   mean   zero  
and  variance   ! i2 and   ! 2j respectively.  Now  the  difference  between  the  disturbances  also  
has   a   normal   distribution   with   mean   zero   and   variance   ! i2 + ! 2j ! 2! ij = ! 2 .   When  
Vin = ! ' xin  and   Vjn = ! ' x jn
 
we  can  state  for  the  choice  probabilities:  
# ! '(xin " x jn ) &
Pn (i) = ! % ( ,    
$ " '
where  Φ  denotes  the  standardized  cumulative  normal  distribution.    
  The   choice   probability   here   only   depends   on   σ,   not   on   the   variance   of   either  
disturbance  or  the  covariance.  Moreover,  the  choice  for  σ  is  arbitrary,  as  rescaling  σ  or  β  
by  any  positive  constant  will  not  affect  the  choice  probability.  Usually  σ  =  1  is  chosen.  
  Of   course   the   assumption   on   normality   is   a   very   convenient   assumption,   as   it  
improves   possibilities   considering   calculations   compared   to   the   linear   probability  
model,   but   it   can   also   be   a   limitation.   Now   a   normal   distribution   is   required   for   all  
unobserved  components  in  the  utility.  Also  the  integral  for  the  choice  probabilities  has  
an   open   form   for   probit   models.   This   is   not   a   big   problem,   but   is   not   considered  
convenient  analytically.  
   

  18  
  
Logit  
Logit  models  are  very  much  alike  probit  models,  but  a  big  difference  is  that  the  integral  
for   the   choice   probability   has   a   closed   form,   which   makes   these   types   of   models  
analytically   more   convenient.   In   the   logit   model   it   is   assumed   that   εjn   –   εin   =   εn   is  
logistically   distributed.   The   logistic   distribution   approximates   the   normal   distribution,  
but  has  fatter  tails.  Under  this  assumption,  the  choice  probabilities  are:    
eµVin
Pn (i) = µV µV
e in + e jn .  
If  the  V  functions  are  linear  in  their  parameters,  the  choice  probabilities  can  be  derived  
onto  
1
Pn (i) = ! µ! '( xin !x jn )
.    
1+ e
Here   μ   is   a   positive   scale   parameter.   For   convenience   it   is   assumed,   as   has   been   done  
similarly  for  probit,  that  μ  =  1.  But  for  probit  σ  =  1  is  chosen,  which  corresponds  with  
var(εjn  –  εin)   =   1.   This   also   implies   that   the   scaled   logit   coefficients   are   !3 times   larger  
than  scaled  probit  coefficients.  
  Train  (2000)  describes  a  couple  of  characteristics  of  the  logit  model,  which  prove  
to  be  as  well  the  power  of  the  model,  but  in  some  sense  also  the  limitations.  Firstly  the  
logit   model   is   able   to   represent   systematic   taste   variation   very   well,   but   the   flipside   is  
that  it  cannot  represent  random  taste  variation.  Secondly,  if  the  unobserved  factors  are  
independent   over   time   in   repeated   choice   situations,   the   model   can   capture   the  
dynamics   of   repeated   choice.   On   the   other   hand   this   seems   restrictive   as   it   exhibits  
substitution  patterns.    
   
There   are   some   limiting   cases   for   all   three   models.   If   µ ! ", ! ! 0 or L ! 0 ,   Pn(i)   will  
become  1  if  Vin  –  Vjn  >  0  and  0  otherwise.  If   µ ! 0, ! ! " or L ! " ,  there  is  exactly  0.5  
probability  for  both  alternatives.    
 
Estimation  
For   both   logit   and   probit   models,   usually   maximum   likelihood   estimators   are   used   to  
estimate  the  parameters   !1, !2 , …, ! K from  a  (random)  sample  of  observations  from  the  
population.   An   indicator   variable   yin  is   constructed   and   defined   as   1   if   person   n   chose  
alternative  i  and  0  if  that  decision  maker  chose  alternative  j.  Also  two  vectors  xin  and  xjn  
are  constructed,  both  contain  all  K  values  of  the  relevant  variables.  Now  given  a  sample  
of  N  observations,  we  now  have  to  find  estimates   !ˆ1, !ˆ2 , …, !ˆK .    
  Now   we   consider   the   likelihood,   which   is   eventually   equal   to   the   probability   of  
the  observed  outcomes  given  the  parameter  values   !ˆ1, !ˆ2 , …, !ˆK .  Since  the  assumption  is  
that   they   are   drawn   at   random   from   the   whole   population,   we   can   state   that   the  
likelihood   of   the   sample   in   total   is   the   product   of   the   likelihood   of   all   individual  
observations.   Analytically   it   is   more   convenient   to   consider   the   logarithm   of   the  
likelihood   function,   denoted   as   log   L.   Now   we   can   write   the   likelihood   as:  
N
log L(!1, !2 , …, ! K ) = ![yin log Pn (i) + y jn log Pn ( j)] ,    
n=1
where   Pn(i)   is   a   function   of   !1, !2 , …, ! K .   Now   the   log   L   function   is   solved   to   maximize   it  
by   differentiation   with   respect   to   the   β’s   and   then   setting   the   partial   derivatives   to   zero,  

  19  
thus   we   solve   max   log   L( !ˆ1, !ˆ2 , …, !ˆK ),   while   we   seek   the   estimates   !ˆ1, !ˆ2 , …, !ˆK that  
solve  this  function.  
  Often   if   a   solution   to   the   first-­‐order   conditions   exists,   it   is   a   unique   solution.  
However   it   is   quite   possible   that   there   will   be   multiple   solutions   to   the   first-­‐order  
conditions.  Just  one  of  these  solutions  constitutes  the  maximum  likelihood  estimate.  The  
estimates   are   consistent   and   asymptotically   normal.   The   estimates   are   given   in   a   matrix  
of   the   second   derivatives   of   the   logarithmic   likelihood   function   with   respect   to   the  
parameters,  which  are  evaluated  at  the  true  parameters.  The  estimate  in  the  kth  row  and  
!2 L
the   lth   column   is   .   Since   we   do   not   know   the   actual   values   of   the   parameters  
!!k!!l
where   we   need   to   evaluate   the   second   derivatives   or   the   distribution   of   the   x   vectors,  
usually   an   estimated   variance-­‐covariance   matrix   that   is   estimated   at   the   estimated  
parameters   !ˆ1, !ˆ2 , …, !ˆK  and  the  sample  distributions  of  the  vectors  x  to  estimate  their  
actual  distribution.  Therefore    
N " 2
! [yin log Pn (i) + y jn log Pn ( j) %
( $
!!k!!l
'
n=1 # &! =!ˆ
 
is  used  as  a  consistent  estimator  of  the  actual  value.  
  As  for  the  computational  aspect  of  this  problem,  it  is  known  that  the  solutions  of  
the  first-­‐order  conditions  are  typically  non-­‐linear  and  the  use  of  a  computer  is  needed  to  
solve,  even  for  two-­‐variable  problems.  Ben-­‐Akiva  and  Lerman  (1985)  describe  how  the  
Newton-­‐Raphson   algorithm   can   be   used.   Briefly   described   the   algorithm   works   as  
follows:  firstly  an  initial  guess  for  the  parameters  is  made.  Then  the  function  of  second  
derivatives   around   the   parameters   is   made.   Then   the   linearized   form   received   after  
approximating  after  first-­‐order  condition  is  solved  and  finally  we  look  at  the  difference  
in  steps  between  the  new  approximations.  If  it  is  small  enough  (Ben-­‐Akiva  and  Lerman  
describe  different  criteria)  then  these  approximations  are  used.  Most  other  procedures  
are  similar,  but  with  different  steps  for  the  second  and  third  step.  
 
Alternative  estimation  models  
The   maximum   likelihood   estimation   procedure   is   used   mostly   for   the   logit   and   probit  
model.   For   the   linear   probability   model   the   least   squares   method,   as   is   more   common  
for   regression   models,   or   Berkson’s   procedure   is   used   more   often.   As   probit   and   logit  
are   the   main   models   to   be   discussed   here,   least   squares   and   Berkson’s   procedure   will  
not  be  discussed  in  this  chapter.    
 
 
Multinomial  Logit  
The   last   two   chapters   the   main   focus   has   revolved   around   binary   choice   models   and   the  
estimation   technique   behind   these   models.   In   most   decision   processes   however,   the  
number  of  alternatives  in  the  choice  set  is  not  limited  to  two.  This  type  of  model  is  called  
a   multinomial   choice   model.   Again,   the   choice   set   is   different   for   every   individual,   as  
each  individual  has  their  own  index  of  attributes  and  a  different  subset  of  the  universal  
set.  In  this  case  where  more  that  two  alternatives  can  be  chosen  the  derivation  of  choice  
models   and   estimation   models   are   more   complex   that   those   for   binary   choice   models.  
Instead   of   using   just   the   difference   between   the   disturbances,   we   now   need   to  
characterize  the  whole  joint  distribution  of  all  disturbances.    

  20  
   Different   types   of   multinomial   choice   models   exist.   It   is   possible   to   expand   the  
binary  logit  and  binary  probit  model  to  multinomial  models.  Dow  and  Endersby  (2003)  
compare   multinomial   logit   (MNL)   and   multinomial   probit   (MNP)   for   voting   research.  
They  state  that  the  MNL  model  is  preferable  to  the  MNP  model.  As  explained  in  the  last  
chapter,  the  logit  model  has  a  closed  form  integral  whilst  the  probit  model  has  an  open  
form.   Therefore   MNP   is   more   complex   than   MNL   and   it   could   give   some   estimation  
problems.   Burda   et   al.   (2008)   present   a   model   that   is   a   mix   between   MNL   and   MNP  
where   estimation   is   conducted   by   using   a   Bayesian   Markov   Chain   Monte   Carlo  
technique.  However  in  this  chapter  we  focus  specifically  on  the  multinomial  logit  model,  
thus  we  will  elaborate  and  expand  on  the  theory  treated  in  the  last  chapter.  
First   some   sort   of   background   will   be   painted   concerning   multinomial   choice  
theory.   Then   some   definition   of   the   MNL   model   and   its   characteristics,   strengths   and  
weaknesses  will  be  given  before  considering  estimation  models.    
 
Multinomial  choice  
As  stated  before,  every  decision  maker  has  as  his  or  hers  choice  set  some  subset  of  the  
universal   set,   and   every   decision   maker   can   have   a   different   subset.   Manski   (1977)   calls  
this   process   of   generating   a   subset   from   the   universal   set   “the   choice   set   generation  
process”.   However   for   the   researcher   it   makes   the   model   a   lot   more   complex   if   every  
individual  decision  maker  can  choose  a  different  choice  set  than  other  decision  makers.  
It   is   defined   that   every   choice   set   Cn   has   J n ! J feasible   choices.   Now   we   state   that   the  
probability  that  an  alternative  i  is  chosen  follows  directly  from  the  probability  described  
in  the  section  on  random  utility  theory:    
Pn (i) = Pr(Uin ! U jn , "j # Cn , j $ i) .    
Here   we   can   distinct   a   deterministic   and   a   random   component   in   the   utility:  
Pn (i) = (Vin + !in ! Vjn + ! jn , "j # Cn , j $ i) .    
Define f (!1n , ... , ! Jnn ) as   the   joint   density   function   of   the   disturbances.   Now   there   are  
different   ways   to   express   the   choice   probabilities   described   in   literature.   Ben-­‐Akiva   and  
Lerman   offer   three   ways   of   deriving   Pn(i).   The   most   insightful   way   is   to   reduce   the  
multinomial  problem  to  a  binary  problem,  as  we  discussed  this  earlier.  We  can  state  that    
Uin ! U jn, "j # Cn , j $ i  
and  from  this  follows  that  if  alternative  i  is  preferred  over  all  other  alternatives  j,  that  
also:  
Uin ! max U jn .    
j"Cn
j#i

Therefore  we  can  write  these  last  formulas  combined:  

$ '
Pn (i) = Pr &Vin + !in ! max(Vjn + ! jn )) .  
&% j"Cn
j#i
)(
As  Ujn  is  a  random  variable  the  maximum  of  Ujn  will  be  a  random  variable  as  well.  Now  
the  distribution  of  this  maximum  has  to  be  derived  from  the  underlying  distribution  of  
the  disturbances.  Where  for  the  MNP  model  for  example  this  is  quite  a  task,  there  it  is  
doable  for  the  MNL  model.  This  is  a  reason  that  MNL  is  a  much-­‐used  model.  
 
Multinomial  logit  
In  the  section  on  binary  logit  the  choice  probability  was  described  as  follows:  

  21  
 eµVin
Pn (i) = µV
eµVin + e jn .  
For  the  MNL  model  the  choice  probability  is  similar:  
eµVin
Pn (i) = .  
" eµVjn
j!Cn

These  formulas  are  equal  if  Jn  =  2  and  μ  =  1.  Here  it  is  evident  that  the  MNL  model  in  an  
extension,  a  development  on  the  binary  logit  model.  Also  in  the  function  it  can  be  seen  
that  it  is  a  proper  probability  mass  function,  as   0 ! Pn (i) ! 1  and  its  sum  over  all  i  is  equal  
to   1.   Here   again   Uin = Vin + !in , !i " Cn ,   with   the   disturbances   independently   and  
identically  distributed  (iid)  and  Gumbel-­‐distributed  with  location  and  scale  parameters  
η   and   μ.   As   with   binary   logit,   as   long   as   all   systematic   terms   of   the   utility   include   a  
constant,   it   is   not   restrictive   to   take   η   =   0.   To   calculate   the   probability   of   one   of   the  
alternatives  in  the  choice  set,  order  the  alternatives  so  that  alternative  i  =  1.  Now  we  still  
have   an   unidentifiable   parameter   μ,   but   it   is   common   to   set   this   parameter   to   a  
convenient  value,  usually  μ  =  1  is  used.  
  Obviously  a  big  advantage  of  the  MNL  model  is  that  is  able  to  analyze  a  choice  set  
consisting   of   multiple   alternatives,   as   is   possible   for   the   MNP   model.   Though   in  
literature   (Bunch   1991,   Alvarez   and   Michael   2001)   use   of   MNP   in   several   fields   is  
recommended,   MNP   cannot   work   optimally   with   a   large   number   of   observations.   The  
MNL   model   however   is   able   to   work   with   larger   datasets.   Both   models   do   not   work  
effectively   with   small   datasets.   Also   the   MNL   model   is   criticized   because   of   the  
independence   of   irrelevant   alternatives   (IIA)   property,   which   states   that   for   a   specific  
individual   the   ratio   of   the   choice   probabilities   of   any   two   alternatives   is   completely  
unaffected   by   systematic   utilities   of   other   alternatives.   This   is   closely   related   to   the  
assumption  that  all  disturbances  are  mutually  independent.  Ben-­‐Akiva  and  Lerman  state  
that  the  problem  does  not  per  se  lie  with  the  IIA  property,  but  rather  every  model  that  
has   an   underlying   assumption   that   the   disturbances   are   mutually   independent   state  
similar   results.   Dow   and   Endersby   (2003)   state   that   for   most   applications   the   IIA  
property  is  not  particularly  restrictive  and  for  most  applications  not  even  relevant.    
  Besides  these  strengths  and  weaknesses  there  are  two  limiting  cases  for  the  MNL  
model,  as  was  the  case  for  the  binary  models.  Firstly  if   µ ! 0  then:  
1
Pn (i) = , !i " Cn .   This   means   that   if   µ ! 0 the   variance   of   the   disturbances  
Jn
approaches  infinity,  so  the  model  will  not  provide  any  information.  That  means  that  all  
alternatives   are   equally   likeable.   The   other   limiting   case   is   when   µ ! " .   Now   the  
variance  of  the  disturbances  approaches  zero  and  the  model  becomes  deterministic.    
 
Estimation  
For   the   MNL   model   maximum   likelihood   is   commonly   used   for   estimation   of   the  
parameters   as   well.   For   the   maximum   likelihood   estimation   procedure   for   the   MNL  
model   there   are   no   big   differences   with   binary   logit,   but   their   computational   burden  
grows   with   the   number   of   alternatives.   McFadden   (1974)   showed   that   for   the   MNL  
model  has  some  special  properties  that  can  simplify  estimation  of  its  parameters  under  
certain  circumstances.    

  22  
   Again,   most   of   this   theory   is   expansion   on   the   section   on   maximum   likelihood  
estimation  for  the  binary  logit  model.  Again  let  yin  be  1  if  decision  maker  or  observation  
n  chose  alternative  i  and  0  otherwise.  We  write  the  likelihood  function:  
N
L = " " Pn (i) yin ,  
n=1 i!Cn

with  for  the  choice  parameters  or  as  Ben-­‐Akiva  and  Lerman  (1985)  state,  the  linear-­‐in-­‐
parameters  logit:  
e ! ' xin
Pn (i) = ! ' x jn .  

" e
j!Cn

As   in   the   section   on   binary   logit,   we   rewrite   the   likelihood   function   to   a   log   likelihood  
function:  
N
! ' x jn
log L = # # yin (! ' xin ! ln # e ).  
n=1 i"Cn j"Cn

When   setting   the   derivatives   of   the   log   likelihood   function   to   zero,   the   first-­‐order  
conditions  can  be  obtained:  
N

# # [y in ! Pn (i)]xink = 0, for k = 1, ... , K .  

n=1 i"Cn

This  can  be  rewritten  as  

 
1 N 1 N
" " in ink N " " Pn (i) xink ,
y x = k = 1, ... , K .  
N n=1 i!Cn n=1 i!Cn

This  shows  that  the  average  value  of  an  attribute  for  the  chosen  alternatives  equals  the  
average  value  predicted  by  the  estimated  choice  probabilities.  Moreover,  this  means  that  
if   an   alternative-­‐specific   constant   is   defined   for   the   alternative   i   at   the   maximum  
likelihood   estimates   the   sum   of   the   choice   probabilities   is   equal   to   the   number   in   the  
sample  that  chose  i.  All  properties  of  the  maximum  likelihood  estimation  of  binary  logit  
extend  to  the  MNL  model.  This  also  applies  for  the  computational  methods  that  are  used  
for  solving  the  system  of  K  equations.  
 
 
Nested  Logit  
Hensher  et  al.  (2005)  state  that  the  bulk  of  choice  behaviour  study  applications  do  not  
go  farther  than  the  simple  MNL  model,  because  of  the  ease  of  computation  and  because  
of  a  wide  availability  of  software  packages.  They  also  state  this  it  does  come  with  a  price  
in  the  form  of  the  IID  assumption  on  the  disturbances  and  the  IIA  property,  which  will  
are  violated  at  times.  The  nested  logit  (NL)  model  includes  a  partial  relaxation  on  both  
assumptions.  As  the  MNL  model,  the  NL  model  as  a  closed  form  solution  in  contrast  to  
for  example  the  multinomial  probit  (MNP)  and  mixed  logit  (ML)  model.  The  ML  model  
will  be  discussed  in  a  later  chapter.    
The  NL  model  is  a  so-­‐called  multidimensional  choice  model.  Many  choice  situations  are  
not  just  situations  where  a  decision  maker  has  to  choose  from  some  list  of  alternatives,  
but  where  the  set  of  alternatives  are  combinations  of  underlying  choice  dimensions,  for  
example  as  shown  in  figure  1,  where  a  NL  model  is  depicted  with  two  dimensions.  

  23  
  
 
 
 
 
 
 
 
 
 
     Figure  1.  An  example  of  the  structure  of  a  NL  model.  
 
It  is  not  the  case  that  the  MNL  model  cannot  be  used  as  a  multidimensional  model.  We  
can   distinct   two   cases   of   multidimensional   models:   multidimensional   choice   sets   with  
shared   observed   attributes   and   multidimensional   choice   sets   with   shared   unobserved  
attributes.   The   multidimensional   MNL   model,   called   the   joint   logit   (JL)   model   is   an  
example  of  the  former;  the  NL  and  MNP  models  are  examples  of  the  latter.  Before  it  is  
able   to   discuss   NL   models   properly,   it   is   necessary   to   have   a   look   at   multidimensional  
choice  sets  in  general  first.  
 
Multidimensional  choice  sets  
In   multidimensional   choice   theory,   every   decision   process   consists   of   more   than   one  
choice   set.   The   most   simple   example   is   probably   the   scenario   with   two   choice   sets:   C1  
and   C2.   Both   choice   sets   have   J1   and   J2   elements.   Therefore   the   choice   set  
C1 ! C2 (Carthesian product)  will   consist   of   J1 ! J 2  elements.   The   multinomial   choice   set  
for  a  decision  maker  n  will  be   Cn = C1 ! C2 " Cn* ,  where   Cn*  is  the  set  of  elements  that  are  
not   feasible   for   that   decision   maker.   This   obviously   also   goes   for   choice   sets   with   higher  
levels   of   dimensionality.   The   example   Ben-­‐Akiva   and   Lerman   give   is   a   choice   situation  
where   M   denotes   possible   modes   for   shopping   and   D   denotes   shopping   destinations.  
The   choice   set   becomes   M   ✕ D.   It   is   possible   however   to   add   extra   dimensions   as   for  
example  time  of  day  or  route.    
  The   difference   with   multinomial   choice   sets   is   that   elements   are   somewhat  
ordered,  meaning  that  elements  share  common  elements  along  one  or  more  dimensions.  
This   linkage   between   the   elements   makes   analysis   useful,   because   it   implies   that   for   a  
linkage  to  exist  either  some  of  the  observed  attributes  of  elements  in  the  choice  set  me  
be  equal  across  subsets  of  alternatives  or  this  may  be  true  for  some  of  the  unobserved  
attributes.  Consequences  of  the  former  will  result  in  the  JL  model;  the  latter  will  result  in  
the   NL   model.   Ben-­‐Akiva   and   Lerman   (1985)   state   that   the   results   for   multidimensional  
choice  situations  will  not  be  different  from  multinomial  situations,  as  long  as  elements  of  
the  choice  set  share  either  observed  or  unobserved  attributes.  
 
Nested  logit  
Train  (2003)  states  that  a  nested  logit  model  is  appropriate  when  the  choice  set  can  be  
partitioned   into   subsets,   or   nests,   in   such   a   way   that   two   properties   hold.   The   first   of  
these  being  that  the  IIA  property  holds  within  each  nest.  The  second  of  these  is  that  the  
IIA  property  does  not  hold  in  general  for  alternatives  in  different  nests.  The  NL  model  
can   be   derived   from   the   General   Extreme   Value   (GEV)   model,   but   that   will   not   be  
covered  in  this  paper.  

  24  
   So   when   designing   the   model,   one   should   note   that   the   two   properties   should  
hold.   A   way   to   test   this   is   by   removing   one   of   the   alternatives   from   the   choice   set.   If  
choice   probabilities   rise   equally   for   certain   alternatives,   these   would   fit   in   one   nest.  
Otherwise  they  would  have  to  be  in  two  different  nests,  because  the  IIA  property  does  
not  hold.  The  IIA  property  should  hold  within  each  nest  but  not  across  nests.    
  The  NL  model  is  consistent  with  utility  maximization  (Daly  and  Zachary  (1978),  
McFadden  (1978)  and  Williams  (1977)).  Let  the  total  choice  set  be  partitioned  in  K  non-­‐
overlapping   subsets   (nests)   B1,   .   .   .   ,   Bk.   The   utility   of   alternative   i   in   nest   BK   is  
Uin = Vin + !in ,   again   with   Vin  the   observed   part   of   the   utility   and   !in  the   random,  
unobserved   part.   The   NL   model   is   obtained   by   assuming   that   the   vector   of   disturbances  
has  a  cumulative  distribution  of  a  GEV  type  distribution:    
$ K$ ' k'
"
&
exp !#&& # e !!in / "k
)) ) .  
& k=1 % j"B ( )(
% k

This   is   a   generalization   of   the   distribution   that   is   underlying   to   the   logit   model.   The  
unobserved  utilities  are  correlated  within  nests.  For  any  alternatives  in  different  nests,  
the   correlation   is   zero.   In   the   function   above   the   parameter   !k  is   a   measure   of   the  
degree  of  independence  in  the  random  part  of  the  utility  among  the  alternatives  in  nest  
k.  A  higher  value  causes  greater  independence  and  thus  less  correlation.  When   !k = 1  for  
all   k,   the   GEV   distribution   becomes   the   product   of   extreme   values   terms   that   are  
independent,   as   !k = 1  represents   independence   among   all   alternatives.   This   means   that  
the  NL  model  reduces  to  the  MNL  model.    
  The  distribution  for  the  unobserved  components  proceed  the  choice  probability  
for  alternative  i  in  nest  Bk:  
!k #1

Pin =
eVin /!k (" j!Bk
e
V jn / !k
) !!
.  
"
K

!=1 (" j!B!

e
V jn / !!
)
If   k = ! ,  meaning  two  alternatives  are  in  the  same  nest,  the  factors  in  parentheses  cancel  
each   other   out   and   it   shows   that   IIA   holds.   For   k ! !  the   factors   in   parentheses   do   not  
cancel   each   other   out   and   IIA   does   not   hold.   Train   (2003)   shows   that   across   nests   some  
other   form   of   IIA   holds,   independence   from   irrelevant   nests   (IIN).   Therefore   in   a   NL  
model,   IIA   holds   for   alternatives   within   each   nest   and   IIN   holds   over   alternatives   in  
different  nests.    
  The   parameter   !k  can   be   different   within   different   nests,   because   correlation  
among   unobserved   factors   can   be   different   within   different   nests.   A   researcher   can,  
however,   constrain   the   !k ’s   in   different   nests   to   be   the   same.   This   would   indicate   that  
the   correlation   is   the   same   in   each   of   these   nests.  Testing   if   this   term   is   equal   to   1   for   all  
k  would  mean  testing  if  the  logit  model  were  appropriate.  For  the  model  to  be  consistent  
with   maximum-­‐utility   behaviour,   the   value   of !k  must   be   in   some   range.   If   !k  lies  
between  zero  and  one  the  model  will  be  consistent  with  utility-­‐maximizing  behaviour.  If  
!k  is  larger  than  one,  the  model  is  only  consistent  with  this  behaviour  for  some  range  of  
the  explanatory  variables  but  not  for  all  values.  A  value  smaller  than  zero  is  inconsistent  
with   utility-­‐maximizing   behaviour.   Kling   and   Herriges   (1995)   provide   tests   of  
consistency  of  NL  with  utility  maximization  in  the  case  that   !k > 1 .  In  reality   !k  does  not  
have   to   be   a   fixed   parameter,   as   every   decision   maker   has   different   correlations.   Bhat  

  25  
(1977)   describes   a   way   to   calculate   !k  based   on   a   vector   of   characteristics   of   the  
decision  maker  and  a  vector  of  parameters  that  need  to  be  estimated.    
 
The  choice  probability  as  given  before  is  still  quite  a  hard  formula  to  grasp.  It  is  possible  
to   express   this   choice   probability   in   a   different   fashion   without   loss   of   generality.   The  
observed  component  of  the  utility  function  can  be  distinct  in  two  parts:  
Uin = Wnk +Yin + !in .  
Here   Wnk   is   the   part   that   is   constant   for   all   alternatives   within   a   nest:   this   variable  
depends  only  on  variables  that  describe  nest  k,  therefore  they  differ  over  nests  but  not  
over  the  alternatives  within  a  nest.  Yin  depends  on  variables  that  describe  alternative  i,  
so   they   vary   over   alternatives   within   nest   k.   Finally   !in  is   the   unobserved   part   of   the  
utility.  Note  that  Yin  is  simply  defined  as  Vin  –  Wnk.  Now  the  NL  probability  can  be  written  
as   the   product   of   two   logit   probabilities.   The   probability   that   an   alternative   is   chosen  
can  be  written  as  the  product  of  the  probability  that  a  certain  nest  is  chosen  multiplied  
with  the  probability  that  an  alternative  within  that  nest  is  chosen:  
Pin = Pin|Bk PnBk ,  
where   Pin|Bk is  the  conditional  probability  that  given  an  alternative  in  nest  Bk  is  chosen  an  
alternative  i  is  chosen.   PnBk  is  the  marginal  probability  of  choosing  an  alternative  in  nest  
Bk.   Any   probability   can   be   written   as   the   product   of   a   marginal   and   conditional  
probability,   therefore   it   is   exactly   the   same   as   the   situation   before.   Now   both   can   take  
the  form  of  logits:  
eWnk +!k Ink
PnBk = K ,
!!=1 eWn! +!!In!
 
eYin /!k
Pin|Bk = ,
! eYjn /!k
j"Bk

where  
I nk = ln " eYin /!k .  
i!Bk

These   expressions   are   derived   from   the   choice   probabilities   stated   earlier.   Train   (2003)  
gives  the  derivation  by  algebraic  rearrangement.  It  is  customary  to  refer  to  the  marginal  
probability   as   the   upper   model   and   to   the   conditional   probability   as   the   lower   model.  
The  quantity  Ink  links  the  lower  and  upper  model  by  transferring  information  from  the  
lower  model  to  the  upper  model  (Ben-­‐Akiva  (1973)).  This  term  is  the  logarithm  of  the  
denominator  of  the  lower  model,  which  means  that   !k I nk  is  the  expected  utility  that  the  
decision  maker  obtains  from  the  choice  among  the  alternatives  in  nest  Bk.  The  formula  
for  the  expected  utility  is  the  same  as  the  utility  for  logit,  as  the  lower  and  upper  model  
are  both  logit  models.  Ink  is  often  referred  to  as  the  inclusive  utility  of  nest  Bk.  It  includes  
the  term  Wnk,  which  is  the  utility  the  decision  maker  receives  no  matter  what  alternative  
in  that  nest  he  chooses.  Added  to  that  is  the  extra  utility  he  or  she  obtains  by  choosing  
the  alternative  with  highest  utility  in  that  nest,   !k I nk .  
 
Estimation  of  nested  logit  
For  the  NL  model  the  same  applies  as  for  the  MNL:  its  parameters  can  be  estimated  by  
standard  maximum  likelihood  techniques:  

  26  
 N
yin
L = " " ( Pin|Bk PnBk ) ,  
n=1 i!Bk

thus  the  log  likelihood  becomes:  

N N
log L = " " yin ln Pin|Bk + " " ynk ln PnBk .  
n=1 i!Bk n=1 k!K

 
Train   describes   that   the   NL   model   can   also   be   estimated   in   a   sequential   fashion   in   a  
bottom   up   way:   the   lower   models   are   estimated   first.   Then   the   inclusive   utility   is  
calculated  for  each  lower  model.  Then  the  upper  model  is  estimated  with  the  inclusive  
utility   as   explanatory   variables.   However   Train   (2003)   also   describes   that   two  
difficulties   come   with   using   estimation   in   a   sequential   fashion.   Firstly   the   standard  
errors   of   the   upper   model   parameters   are   biased   downward   and   some   parameters  
appear  in  several  sub  models.  Estimating  the  lower  and  upper  model  separately  causes  
separate   estimates   of   whatever   common   parameters   appear   in   the   model.   Maximum  
likelihood   estimation   is   conducted   simultaneously   for   both   models;   therefore   the  
common  parameters  are  constrained  to  be  the  same  wherever  in  the  model  they  appear.  
It   is   stated   that   maximum   likelihood   estimation   is   the   most   efficient   estimation  
technique  for  the  NL  model.    

Higher-­‐level  nested  logit  and  expansion  on  the  NL  model  

Until  now  the  NL  model  has  been  discussed  at  a  two  dimensional  level,  also  known  as  a  
two-­‐level  nested  logit.  Here  there  are  two  levels  of  modelling:  the  marginal  probabilities  
and  the  conditional  probabilities.  In  some  situations  however,  a  higher-­‐level  NL  model  
might  be  appropriate.  The  choice  probabilities  of  three-­‐  or  higher-­‐level  NL  models  can  
be   expressed   as   a   series   of   logit   models.   The   top-­‐level   model   describes   the   choice   of   a  
nest,  then  it  describes  choices  of  subnets  to  a  certain  extend  and  eventually  the  lowest  
level   model   is   the   choice   of   alternatives   in   a   (sub)   model.   The   top   model   includes   an  
inclusive  utility  for  each  nest.    
  It   also   possible   that   an   alternative   is   a   member   of   more   than   one   nest.   The  
example   that   Train   gives   is   an   example   of   home-­‐to-­‐work   travelling.   Consider   four  
alternatives:   bus,   train,   drive   alone   and   carpooling.   Obviously   bus   and   train   can   be  
considered   public   transport   and   driving   alone   and   carpooling   can   be   considered   going  
by  car.  However,  carpooling  has  some  shared  unobserved  attributes  that  are  similar  to  
going   by   public   transport:   it   has   a   lack   of   flexibility   in   scheduling.   Now   it   is   possible  that  
this   alternative   is   placed   in   both   nests.   This   phenomenon   is   called   overlapping   nests.  
This  model  is  called  the  Cross  Nested  Logit  (CNL)  model.  Ben-­‐Akiva  and  Bierlaire  (1999)  
proposed   the   general   formulation   of   the   CNL   model.   The   CNL   model   is   also   called   the  
Generalized  Nested  Logit  model  (Wen  and  Koppelman  (2001)).    
 
Cross-­‐Nested  Logit  
According  to  Papola  (2003),  the  full  specification  consists  of  two  phases:  specification  of  
the  correlation  structure  and  the  identification  of  the  parameters.  Now  both  of  these  last  
variations  (higher  dimensions  and  overlap)  can  be  included  in  the  choice  probabilities  of  
the  NL  model,  leading  to  CNL  model  (or  Generalized  NL  model).  The  nests  are  labelled  
B1, B2 , …, BK .   Each   alternative   can   be   part   of   more   than   one   nest   and   also   to   varying  
degrees.  Therefore  allocation  parameter   ! ik  reflects  the  degree  to  which  alternative  i   is  
a  member  of  nest  k.  This  parameter  must  be  nonnegative  and  if  it  is  zero  it  means  it  is  
not   a   member   of   nest   k.   The   sum   of   over   all   nests   for   one   alternative   must   be   one.   Again  

  27  
we   have   parameter   !k  that   indicates   to   what   extent   alternatives   among   a   nest   are  
independent   and   a   higher   value   can   be   explained   as   greater   independence   and   less  
correlation.  Now  the  probability  that  decision  maker  n  chooses  alternative  i  is:  
!k )1

Vin 1/ !k # 1/ !k
&
! (!k ik e ) % ! j"B " jk e
$ k
V jn
( ) (
'
Pin = !!
.  
! (! (" V jn 1/ !!
)
K

!=1 j"B! j! e )
Now  if   ! ik  is  equal  to  one  for  all  alternatives  in  the  choice  set  and  enters  only  one  nest,  
we  get  the  same  choice  probability  as  for  the  two-­‐level  NL  model.  If   !k  is  equal  to  one  
for  all  nests  next  to  this,  the  model  becomes  a  standard  logit  model.  Also  for  higher-­‐level,  
overlapping  models  it  is  possible  to  decompose  the  model:  
Pin = ! Pin|Bk Pnk ,  
k
where  marginal  and  conditional  probabilities  are:  
1/ !k

" ( ! jk e
V jn
)
Pnk = j!Bk
!!
,  
" (" (" V jn 1/ !!
)
K

!=1 j!B! j! e )
1/ !k

Pin|Bk =
(! ik eVin )
.  
V jn 1/ !k
" ("j!Bk jk e )
 
The   model   was   first   used   by   Vovsha   (1997)   who   used   the   model   for   a   mode   choice  
survey  in  Israel.  The  model  is  appealing  because  it  is  able  to  capture  a  wide  variety  of  
correlation  structures.  The  CNL  model  has  a  closed  form,  as  it  is  derived  form  the  GEV  
model  and  an  expansion  on  the  NL  model.  As  shown  before,  it  is  in  some  ways  a  special  
case  of  the  standard  logit  model.  This  makes  the  CNL  model  analytically  interesting.    
  One  of  the  most  obvious  merits  is  the  ability  to  capture  complex  situations  where  
the   NL   model   cannot   handle   correlations,   because   it   does   not   allow   for   overlap.   Also   the  
open  form  of  the  PDF  makes  the  model  analytically  interesting.    
  A  disadvantage  of  the  model,  maybe  because  it  is  quite  a  new  model,  is  that  the  
issue  of  identification  still  remains  open.  There  are  different  estimation  techniques  and  a  
maximum  likelihood  estimator  can  be  identified.  However  if  the  model  is  over  specified  
the  speed  of  the  algorithm  may  decrease  significantly  or  not  even  perform  well  at  all.    
 
Estimation  of  cross-­‐nested  logit  
The  first  estimation  procedures  for  the  CNL  model  were  proposed  by  Small  (1987)  and  
Vovsha   (1997)   and   are   based   on   heuristics.   However,   currently   maximum   likelihood  
estimation   techniques   are   used.   Again,   these   techniques   aim   at   identifying   the   set   of  
parameters  that  maximize  the  probability  that  a  given  model  perfectly  reproduced  the  
observations   (Bierlare   (2001)).   The   objective   function   of   the   maximum   likelihood  
estimation  problem  for  the  CNL  model  is  a  nonlinear  analytical  function,  as  the  PDF  has  
a   closed   form.   Most   nonlinear   programming   algorithms   are   designed   to   identify   local  
optima  of  the  objective  function.  As  the  CNL  model  has  a  closed  form,  the  log  likelihood  
does  as  well:  

  28  
 ln L = " ln Pin|Cn ,  
n!sample

where  this  is  the  probability  that  alternative  i  is  chosen  by  decision  maker  n  and  Cn  is  the  
choice  set  for  that  specific  decision  maker.  Ben-­‐Akiva  and  Bierlaire  (1999)  give  a  more  
elaborate  derivation  of  the  log  likelihood  function.  Whatever  algorithm  is  preferred,  it  is  
instrumental   that   different   initial   solutions   are   used.   No   meta-­‐heuristics   can   provide   a  
global   optimum.   Ben-­‐Akiva   and   Bierlaire   also   give   a   number   of   steps   that   need   to   be  
taken.   First,   constraints   to   guarantee   the   model   validity   have   to   be   defined.   Then  
constraints   imposing   a   correct   intuitive   interpretation   could   be   important.   Finally  
normalization  constraints  are  necessary;  otherwise  the  model  would  not  be  estimable.    
 
 
Mixed  Logit  
According   to   McFadden   and   Train   (2000),   mixed   logit   (MXL)   is   a   highly   flexible   that   can  
approximate   any   random   utility   model.   The   limitations   stated   for   the   logit   model   are  
prevented   because   it   allows   for   random   taste   variation,   unrestricted   patterns   and  
correlation   in   unobserved   factors   over   time.   In   contrast   to   the   probit   model,   it   is   not  
restricted   to   normal   distributions   for   the   error   terms   and   together   with   the   probit  
model   it   has   been   known   for   years   but   has   only   become   applicable   since   simulation  
become  accessible.    
  MXL  models  can  be  derived  under  different  behavioural  specifications  and  each  
derivation  provides  a  different  interpretation.  The  MXL  model  is  defined  on  basis  of  the  
functional   form   for   its   choice   probabilities.   MXL   probabilities   are   the   integrals   of   logit  
probabilities  over  a  density  of  parameters:  
Pin = ! Lin (! ) f (! )d ! ,  
here  Lin  is  the  logit  probability  evaluated  at  parameters  β,  as  described  in  the  chapter  on  
MNL  and  f(β)  is  a  density  function.  If  the  utility  is  linear  in  β,  then  Vin(β)  =  β’xin.  Then  the  
MXL  probability  takes  its  usual  form:  
" ! ' xin %
e
Pin = ( $ ' f (! )d ! .  
$ ! e ! ' x jn '
# j &
This   probability   is   a   weighted   average   of   the   logit   formula,   but   evaluated   at   different  
values   of   β,   with   weights   given   by   f(β).   In   literature   a   mixed   function   is   a   weighted  
average   of   several   functions   and   the   density   that   provides   the   weights   is   a   mixing  
distribution.  MXL  is  a  mix  of  the  logit  function  evaluated  at  different  β’s  with  f(β)  as  the  
mixing   distribution.   Standard   logit   is   a   specific   case   where   the   mixing   distribution   is  
fixed  for  f(β)  =  1.  Then  the  formula  becomes  the  normal  choice  probability  for  the  MNL  
model.  The  mixing  distribution  can  also  be  discrete,  with  β  taking  a  finite  set  of  distinct  
values.   This   results   in   the   latent   class   model   that   will   be   discussed   later.   However   in  
most   cases   the   mixing   distribution   is   specified   as   continuous.   By   specifying   the  
explanatory   variables   and   density   appropriately,   it   is   possible   to   represent   any   utility  
maximizing   (and   even   some   forms   of   non-­‐utility-­‐maximizing   behaviour)   by   a   MXL  
model.   There   is   one   issue   though   concerning   notation.   There   are   two   sets   of   parameters  
in   a   MXL   model:   the   parameters   β   that   enter   the   logit   formula   and   have   density   f(β)   and  
the   second   set   parameters   that   describe   that   density.   Denote   the   parameters   that  
describe  the  density  of  β  by   ! ,  so  the  density  is  best  denoted  as  f(β| ! ).  The  mixed  logit  
choice   probabilities   are   not   dependent   on   the   values   of   β,   but   they   are   functions   of   ! .  
The   parameters   β   can   be   integrated   out.   In   that   sense   the   β’s   are   similar   to   the  

  29  
disturbances   in   the   sense   that   they   are   both   random   terms   that   are   integrated   out   in  
order  to  obtain  the  choice  probability.    
  One  of  the  positive  points  about  the  MXL  model  is  that  it  exhibits  neither  the  IIA  
property   nor   the   restrictive   substitution   patterns   of   the   logit   model.   The   ratio   Pin/Pjn  
depends   on   all   data,   including   alternatives   and   attributes   other   than   i   or   j.   The  
percentage  change  in  probability  depends  on  the  relation  between  Lin  and  Ljn,  the  logit  
probabilities  of  alternatives  i  and  j.  Besides  this,  Greene  and  Hensher  (2002)  state  that  
the  MXL  model  is  flexible,  even  though  it  is  fully  parametric.  Therefore  it  can  provide  the  
researcher   with   a   large   range   to   specify   the   individual   and   unobserved   heterogeneity.  
To   some   extent   this   flexibility   even   offsets   the   specificity   of   the   distributional  
assumptions.  They  also  state  that  with  the  MXL  model  it  is  possible  for  the  researcher  to  
harvest   a   rich   variety   of   information   about   behaviour   from   a   panel   or   repeated  
measures  data  set.  
   
Estimation  
Train  (2003)  states  that  MXL  is  well  suited  for  simulation  methods  for  estimation.  Utility  
is   again   Uin = !n' xin + "in  and   the   β   coefficients   are   distributed   with   the   mixing  
distribution   f(β| ! )   and   !  denotes   the   parameters   of   this   distribution   (mean   and  
covariance  of  β).  Now  the  choice  probabilities  are  
Pin = ! Lin (! ) f (! | " )d ! ,  
e ! ' xin
Lin (! ) = J ! ' x jn
.  
! j=1
e
Here   Lin   again   is   the   logit   probability.   The   probabilities   Pin   are   approximated   through  
simulation  for  any  given  value  of   ! .  According  to  Train  (2003)  this  is  conducted  through  
three  steps:  
 
1. Draw   a   value   of   β   from   the   mixing   distribution   f(β| ! )   and   label   it   β’   with   the  
superscript  r  =  1.  This  refers  to  the  first  draw.  
2. Calculate  the  logit  formula  Lin(βr)  with  this  draw.  
3. Repeat   step   1   and   2   often   and   average   the   result.   This   average   is   the   simulated  
probability:  
 
! 1 R
Pin = ! Lin (! r ),  
R r=1
!
where   R   is   the   number   of   draws.   Here   Pin  is   an   unbiased   estimator   of   Pin.   Its   variance  
decreases  as  R  increases  and  it  is  strictly  positive,  twice  differentiable  in  the  parameters  
!  and   the   variables   x   that   facilitate   the   numerical   search   for   the   maximum   likelihood  
function.  Train  denotes  this  simulated  log  likelihood  (SLL)  function  as  follows:  
N J
!
SLL = !! d jn ln Pjn .  
n=1 j=1

Here   djn   is   an   indicator   function,   it   is   equal   to   one   if   decision   maker   n   chose   alternative   j  
and  zero  otherwise.  The  maximum  simulated  likelihood  estimator  (MSLE)  is  the  value  of  
!  that  maximizes  the  SLL  function.  The  exact  properties  of  this  simulated  estimator  will  
not  be  discussed  in  this  paper,  but  can  be  found  in  Train  (2003)  and  Greene  (2001).    
   This  method  of  estimation  is  related  to  accept-­‐reject  (AR)  methods  of  simulation.  
This   AR   method   will   not   be   discussed   extensively,   as   the   MSLE   method   is   more   often  

  30  
used.  AR  simulation  can  be  applied  generally.  It  is  constructed  as  follows  (Train  (2003),  
Greene  (2001)):  
 
1. A  draw  of  the  random  terms  is  taken.  
2. The   utility   of   each   alternative   is   calculated   from   this   draw   and   the   alternative  
with  the  highest  utility  is  identified.  
3. Steps  2  and  3  are  repeated  often.  
4. The   simulated   probability   for   an   alternative   is   calculated   as   the   proportion   of  
draws  for  which  that  alternative  has  the  highest  utility.  
 
The   AR   simulator   is   also   unbiased   by   construction.   It   is   however   not   strictly   positive.  
Also   it   is   not   always   twice   differentiable.   It   seems   to   be   some   sort   of   step   function.  
Therefore  the  MSLE  method  is  more  often  used.    
 
Repeated  choice  
Each   sampled   decision   maker   easily   generalizes   this   model   for   repeated   choices.   The  
simplest   specification   treats   the   coefficients   that   enter   utility   as   varying   over   different  
decision  makers  but  constant  over  choice  situations  for  each  decision  maker.  Therefore  
utility  from  alternative  i  in  choice  situation  t  by  decision  maker  n  is   Uitn = !n xitn + "itn ,  with  
!itn  being   iid   extreme   value   over   alternatives,   time   and   decision   makers.   Now   consider   a  
sequence   of   alternative,   one   for   each   time   period:   i  =   {i1,   …   ,   iT}.   The   probability   that   the  
decision   maker   makes   this   sequence   of   choices   is   the   product   of   the   logit   formulas,  
conditional  on  β:    
T
" !n' xi tn %
e t '
Lin  (β)  =   ( $ .  
$ !n' x jt tn '
t=1 $ !
# j &'
The   !itn  terms   are   independent   over   time.   The   unconditional   probability   is   the   integral  
of  this  product  over  all  values  of  β:  
Pin  =   Lin (! ) f (! )d !.  
The   probability   is   simulated   in   a   similar   way   to   the   probability   with   just   one   choice  
period:  
 
1. A  draw  of  β  is  taken  from  its  distribution.  
2. The   logit   formula   is   calculated   for   each   period   and   the   product   of   the   logits   is  
taken.  
3. Repeat  steps  1  and  2  often.  
4. The  MSLE  is  the  average  of  the  simulated  probabilities.  
 
 
Latent  Class  Logit  
As  stated  in  the  last  chapter,  the  MXL  model  and  the  Latent  Class  Logit  (LCL)  model  are  
not  that  different.  Remember  the  usual  form  that  the  mixed  logit  probability  takes:  
" ! ' xin %
e
Pin = ( $ ' f (! )d ! .  
$ ! e ! ' x jn '
# j &
This   choice   probability   is   the   same   for   the   LCL   model,   but   instead   of   f(β)   being  
continuous,  the  mixing  distribution  is  discrete  with  β  taking  a  finite  set  of  distinct  values.  

  31  
Here   β  takes   M   possible   values   labelled   b1,   …   ,   bM   with   probability   sm   that   β  =   bm.   This  
model   has   been   popular   in   psychology   and   marketing   for   some   time   now   (Kamakura  
and  Russell  (1989)  and  Chintagunta  et  al.  (1991)).    
  Greene   and   Hensher   (2002)   state   that   the   LCL   model   is   similar   to   the   MXL   model  
by   McFadden   and   Train   (2001),   but   it   relaxes   the   requirement   that   the   researcher  
makes   specific   assumptions   about   the   (continuous)   distributions   of   parameters   across  
each  decision  maker.  Sagebiel  (2011)  states  that  the  unconditional  probability  to  choose  
a   certain   alternative   i   by   decision   maker   n   is   the   weighted   average   of   the   M   bm  
parameters:  
M
Pin = ! sm Pin|m ,  
m=1
with  Pin|m  being  the  conditional  logit  (CL)  probability  to  choose  alternative  i  that  belongs  
to  class  s:  
e(b1n xin +...+bkn xkn )
Pin|m = N .  
! e(b1n xin +...+bkn xkn )
n=1
The  parameters  hs    are  unknown,  but  can  be  estimated  with  a  MNL  model  with  help  of  a  
case-­‐specific  variable  vector  that  includes  age,  income  or  any  other  characteristic  of  the  
decision  maker.    
  The   underlying   theory   of   the   LCL   model   presumes   that   individual   choice  
behaviour  depends  on  the  observable  attributes  and  on  latent  heterogeneity  that  varies  
with   the   unobserved   factors   (Greene   and   Hensher   (2002)).   They   analyse   this  
heterogeneity  through  a  model  of  discrete  parameter  variation.  Here  decision  makers,  as  
in  the  analysis  by  Sagebiel  (2011),  are  sorted  in  a  set  of  Q  classes,  but  it  is  unknown  to  
the  researcher  which  class  contains  which  decision  makers,  even  if  the  decision  makers  
themselves   know   it.   There   are   bn   alternatives   by   decision   maker   n   in   Tn   choice  
situations.   As   is   clear,   Greene   and   Hensher   directly   explain   the   model   for   repeated  
choice.   Here   the   choice   probability   that   alternative   i   is   made   by   decision   maker   n   is  
choice  situation  t,  given  that  this  occurs  in  class  q:  
'
xitn !q
e
Pitn|q = Jn '
xitn !q
= F(i, t, n | q).  
! j=1
e
The   size   of   the   choice   set   and   the   number   of   observations   may   vary   per   decision   maker.  
Moreover,  the  choice  set  should  vary  per  choice  situation  as  well.  The  probability  for  the  
specific   choice   by   a   decision   maker   can   be   formulated   in   different   manners,   so   for  
convenience  can  be  formulated  in  different  manners,  so  for  convenience  ytn  denotes  the  
specific  choice  made,  so  the  model  provides  
Ptn|q (i) = Pr(ytn = i | class = q).  
Given   the   class   assignment,   it   is   assumed   that   Tn   are   independent   events   and   the  
contribution  of  decision  maker  n  to  the  likelihood  would  be  the  joint  probability  of  the  
sequence  yi  =  [y1n,  …  ,  yTn]:  
Tn
Pn|q = ! Ptn|q .  
t=1
Here  the  class  assignment  is  unknown.  Now  Hnq  is  denoted  as  the  prior  probability  for  
class  q  for  decision  maker  n.  For  Hnq  the  form  of  the  MNL  is  used:  
z' !
enq
H nq = Q zn' ! q
, q = 1, ... , Q, !Q = 0,  
! q=1
e

  32  
where   zn   is   a   set   of   observable   characteristics,   as   described   in   the   chapter   on   binary  
choice  models.   ! q  again  denote  the  parameters  for  class  q.    
  Greene   and   Hensher   (2002)   compared   the   MXL   model   and   the   LCL   model   for   a  
transport   situation,   with   the   objective   to   seek   understanding   of   the   relative   merits   of  
both   modelling   strategies.   They   state   that   it   is   not   possible   to   state   that   one   of   these  
models  is  unambiguously  preferred  to  the  other.  The  LCL  model  has  the  advantage  that  
it   is   a   semi   parametric   specification.   This   means   that   the   researcher   is   freed   from  
possibly   strong   or   unwarranted   distributional   assumptions,   in   their   case   about  
individual  heterogeneity.  For  the  LCL  model  the  same  applies  as  for  the  MXL  model,  it  
allows  the  researcher   to  harvest  a  rich  variety  of  information  on  behaviour  from  a  panel  
or  repeated  measures  data  set.  
 
Estimation  
Now   the   probability,   or   now   the   likelihood   is   the   expectation   over   classes   of   the   class  
specific  contributions:  
Q
Pn = ! H nq Pn|q .  
q=1

This  is  the  likelihood  for  one  decision  maker.  Now  the  log  likelihood  for  the  sample  is  
N N )Q # Tn &,
ln L = ! Pn = ! ln +! H nq % " Ptn|q (..  
n=1 n=1 +* q=1 $ t=1 '.-
Maximization   of   this   function   with   respect   to   the   structural   parameters   and   the   latent  
class   parameter   vectors   is   a   conventional   problem   in   maximum   likelihood   estimation.  
Greene   (2001)   discusses   the   mechanics   and   other   aspects   of   estimation,   as   it   is   a  
relatively  difficult  optimization  problem  in  comparison  to  other  situations.  The  choice  of  
a  good  starting  value  for  Q  is  crucial.  Testing  down  to  the  appropriate  Q  by  comparing  
the   log   likelihoods   of   smaller   models   is   not   a   proper   approach.   Roeder   et   al.   (1999)  
suggest  that  the  Bayesian  Information  Criterion  (BIC)  model  can  be  used:  
(model size)ln N
BIC(model)  =   ln L + .  
N
Now  we  now  parameter  estimates   ! q ,  the  prior  estimates  of  the  class  probabilities  are  
Ĥ nq .   Using   Bayes   theorem   (Bayes   and   Price,   1763)   it   is   possible   to   obtain   a   posterior  
estimate  of  the  latent  class  probabilities:  
P̂ Ĥ
Ĥ q|n = Qn|q nq .  
! P̂n|q Ĥ nq
q=1

The   maximum   value   of   this   formula   would   be   associated   by   a   strictly   empirical  

estimator   of   the   latent   class   within   the   class   that   the   individual   resides   in.   These   results  
can  be  used  to  obtain  posterior  estimates  of  the  individual  specific  parameter  vector:  
Q
!ˆ = ! Ĥ !ˆ .  
n q|n q
q=1

Also  this  data  can  be  used  to  estimate  marginal  effects  in  the  logit  model.  
 
 
 
 
 

  33  
  
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Summary, comments and acknowledgement
“I abhor averages. I like the individual case. A man may have six meals one day and none the
next, making an average of three meals per day, but that is not a good way to live.”
Louis D. Brandeis (1856 – 1941)
 
   

  34  
  
Variations:  different  choice  models  
Until   now,   the   most   instrumental   models   have   been   discussed.   However,   this   does   not  
mean   that   there   are   no   other   choice   models   known.   The   models   discussed   are   mostly  
models  that  are  used  very  often  and  are  instrumental  in  the  development  of  new  models.  
For  example,  the  standard  logit  models  are  not  used  that  often  anymore,  but  serve  more  
as   reference   models.   The   MNL   model   however,   is   still   used   in   some   fields.   Dow   and  
Endersby   (2003)   argue   that   in   political   voting   research   the   MNL   model   is   still   usable,   as  
it   is   easily   understandable   and   the   differences   in   results   are   very   small   and  
computationally  it  can  be  preferred  over  more  complex  models.  
  In  this  chapter  some  other  models  will  be  discussed  briefly.  The  reader  should  be  
aware   that   the   models   discussed   until   now   are   not   the   only   models   and   in   different  
areas  different  models  are  used.  Models  as  for  example  the  MNP  model  are  much  used,  
but  are  analytically  more  complex  and   did   not  fit   in  the   line   of   this   paper.   These   models  
are  worth  mentioning  however.  
 
The  Generalized  Extreme  Value  (GEV)  Model  
This  model  has  been  mentioned  before  in  this  paper.  The  GEV  model  is  actually  more  a  
class  of  models  than  a  specific  model.  The  MNL  model  and  the  NL  model  are  both  models  
that   are   part   of   the   GEV   family   of   models.   McFadden   (1978)   gave   a   generalization   of   the  
MNL  model  that  is  in  fact  the  GEV  model.  The  unifying  attribute  of  the  models  of  the  GEV  
family   is   that   the   unobserved   components   of   the   utility   for   all   alternatives   are   jointly  
distributed   as   a   generalized   extreme   value.   This   allows   for   correlations   over  
alternatives.   When   all   these   correlations   become   zero,   the   GEV   model   becomes   the  
standard  logit  model.  Currently  the  most  widely  used  member  of  the  family  are  the  NL  
and   CNL   models.   These   models   are   applied   in   for   example   energy,   transportation,  
housing  and  telecommunication.  Karlstrom  (2001)  showed  that  only  a  small  portion  of  
the  possible  models  within  the  GEV  classes  has  ever  been  implemented.  This  means  that  
the   full   capability   of   this   class   of   models   has   not   yet   been   fully   exploited   and   new  
research  can  be  used  to  further  investigate  the  potential  of  this  group  of  models.  
 
Joint  logit  
Regarding  multidimensional  choice  mostly  the  NL  and  CNL  models  have  been  regarded,  
models   that   are   based   on   multidimensional   choice   sets   with   shared   unobserved  
attributes.   The   Joint   logit   (JL)   model   on   the   other   hand   is   a   model   based   on  
multidimensional  choice  sets  with  shared  observed  attributes.  Here  there  are  observed  
factors   in   the   choice   set   that   are   shared,   therefore   overlap   in   some   sense.   As   in   the   MXL  
model,   marginal   and   conditional   probabilities   are   derived.   The   distribution   of   the  
disturbances   will   affect   the   form   of   the   conditional   choice   probabilities.   A   normal  
distribution   is   not   always   guaranteed.   Currently   joint   models   are   mostly   used   for  
combining  stated  and  revealed  preference  data.    
 
Multinomial  Probit  
As   the   MNL   model   is   an   extension   of   the   binary   logit   model,   exactly   that   way   is   the   MNP  
model   an   extension   of   the   binary   probit   model.   Now   the   vector   of   disturbances   is  
extended   from   two   to   Jn.   This   requires   for   the   model   to   have   a   solution   of   the   Jn   –   1  
dimensional  integral  for  evaluating  the  choice  probabilities.    
  The  concept  of  this  model  appeared  a  very  long  time  ago  in  writings  by  Thurstone  
(1927)   in   applications   of   mathematical   psychology.   Because   of   the   open   form   of   the  

  35  
integral   in   evaluating   the   choice   probability,   until   early   1980s   its   computational  
difficulty   made   it   unusable.   Since   the   rise   of   simulation   it   has   been   possible   to   use   the  
MNP   model   more   extensively.   Nowadays   the   MNP   model   is   used   in   different   areas.   Dow  
and  Endersby  (2003)  indicate  that  it  is  often  used  in  political  voting  and  in  psychology  
(Train  (2003)).    
  As   for   estimation,   this   is   a   bit   more   complicated   than   for   the   MNL   model,   as  
indicated   before.   Because   the   choice   probabilities   have   an   open   form   integral,   simple  
estimation   techniques   are   often   insufficient.   Geweke   (1996)   gives   an   explanation   of  
quadrature  methods,  which  approximate  the  integral  by  a  weighted  function  of  specially  
chosen  evaluation  points.  However  nowadays  not  many  non-­‐simulation  techniques  are  
used,  for  example  because  it  is  more  general.  Hajivassiliou  et  al.  (1996)  give  an  overview  
of  different  simulation  techniques.  The  most  straightforward  and  most  used  techniques  
are   accept-­‐reject   (AR),   which   was   briefly   noted   in   the   chapter   on   the   MXL   model,  
smoothed   AR   and   GHK   (after   Geweke,   Hajivassiliou   and   Keane)   that   can   be   combined  
with   maximum   likelihood   or   sampling.   Train   (2003)   describes   these   models   more  
elaborately.    
 
Mixed  Probit  
 Again   mixed   probit   (MXP)   can   be   seen   as   some   development   on   the   probit   model,   as  
was  a  similar  case  for  MXL  and  MNL.  A  constraint  for  the  MNP  is  that  all  random  terms  
enter  utility  in  a  linear  way  and  they  are  randomly  distributed  in  a  way  so  that  the  utility  
is  normally  distributed.  This  constraint  is  removed  in  MXP  models.  These  models  have,  
as  do  the  MXL  models  conditional  probabilities  and  the  integral  also  has  an  open  form.  
Therefore  the  model  has  long  run  times  for  the  GHK  simulator.    
  Train   (2003)   states   that   the   MXP   model   provides   a   way   to   avoid   some   of   the  
practical  difficulties  of  the  MXL  model,  as  representing  pure  heteroskedasticity  or  fixed  
correlation   patterns   among   alternatives.   In   the   MXP   estimation   is   made   instead   of  
specifying   numerous   error   components   in   the   MXL   model.   MXP   is   suitable   some   non-­‐
normal  random  terms,  which  is  not  possible  in  the  MNP  model.  However  MXP  is  more  
complex  and  computationally  more  burdensome.    
 
Further  research  
As   stated   before,   new   models   will   continue   to   arise   now   possibilities   have   opened   up  
with  the  advent  of  simulation.  In  2002  Ben-­‐Akiva  et  al.  wrote  a  paper  on  hybrid  choice  
models.  Burda  et  al.  (2008)  for  example  present  a  Bayesian  mixed  logit-­‐probit  model  for  
multinomial   choice.   It   is   expected   that   in   the   coming   years   more   of   the   hybrid   models  
will  surface  as  computationally  more  will  become  possible.  
 
 
Summary  
In   this   paper   the   reader   is   given   an   overview   of   theory   behind   choice   models.   Choice  
models   have   many   applications   in   society:   psychology,   transport,   energy,   housing,  
marketing,   voting   and   many   more   areas   make   use   of   these   models.   Since   the   1920s,  
when   Thurstone   (1927)   introduced   the   binary   logit   and   probit   model,   these   models  
have  developed  significantly.  
  Firstly   the   theory   behind   the   models,   choice   theory,   probabilistic   theory   and  
utility   theory   have   been   discussed.   Most   models   have   the   assumption   of   rational  
behaviour   and   utility-­‐maximizing   behaviour.   Also   the   framework   for   binary   models   was  
given.  

  36  
   In  the  second  part  of  the  paper  different  models  have  been  discussed.  Binary  logit  
and   probit   models   can   be   considered   predecessors   of   multinomial   logit   and   probit  
models.   In   a   similar   way   are   nested   logit   and   mixed   logit   models   derived   from   the  
multinomial   logit   model.   It   is   still   discernable   that   the   multinomial   logit   model   is   a  
special  case  of  both  the  nested  logit  model  and  the  mixed  logit  model,  as  with  the  right  
parameters  the  model  is  equal  to  the  multinomial  model.  For  the  nested  logit  model  it  
applies   that   both   the   marginal   and   conditional   probabilities   are   logits.   Also   the   cross-­‐
nested   logit   model   can   be   seen   as   an   extension   of   the   nested   logit   model,   as   it   allows   for  
overlap  between  nests.  The  latent  class  model  is  derived  from  the  general  extreme  value  
model,  as  applies  for  the  mixed  logit  model.  The  difference  between  these  models  is  the  
assumptions  underlying  and  mostly  the  difference  in  mixing  distribution.  In  this  part  for  
each   model   also   estimation   techniques   were   discussed.   For   two   models   that   were  
discussed  last,  the  latent  class  model  and  the  mixed  logit  model,  simulation  techniques  
are   necessary   because   of   the   open   form   of   the   integral   in   the   choice   probability.   This  
makes   these   models   more   complex   and   computationally   more   burdensome.   These  
models   however   do   avoid   some   of   the   assumptions   underlying   at   for   example   the  
multinomial  model.  
  In  the  third  and  last  section  a  short  overview  was  given  of  models  that  were  not  
discussed   elaborately   in   the   second   section,   but   were   worth   mentioning   due   to   their  
value  and  use  in  research.  The  general  extreme  value  model  is  often  referenced  and  the  
multinomial  probit  model  is  still  very  much  used.  
 
 
Acknowledgement    
My  sincere  thanks  goes  out  to  Alwin  Haensel,  who  besides  helping  me  with  my  questions  
put   me   on   the   path   to   writing   this   paper   by   making   this   subject   public   on   the   website   of  
the  VU.    
 
 
References  
 
Ai  C.  and  Norton  E.C.  (2003),  Interaction  terms  in  logit  and  probit  models  
Alvarez  R.M.  and  Nagler  J.  (1995),  Economics,  Issues  and  the  Perot  Candidacy:  Voter  
Choice  in  the  1992  Presidential  Election  
Alvarez   R.M.   and   Nagler   J.   (2001),   Correlated   disturbances   in   discrete   choice  
models:  a  comparison  of  multinomial  probit  and  logit  models  
Adamowicz   W.,   Louviere   J.   and   Williams   M.   (1994),   Combining   Revealed   and   Stated  
Preference   Methods   for   Valuing   Environmental   Amenities   –   Journal   of  
Environmental  Economics  and  Management  
Anand  P.  (1993),  Foundations  of  rational  choice  under  risk  
Bayes  T.  and  Price,  M.  (1763),  An  Essay  towards  solving  a  problem  in  the  doctrine  of  
chances  –  Philosophical  Transactions  of  the  Royal  Society  of  London  
Ben-­‐Akiva   M.E.   and   Lerman   S.R.   (1985),   Discrete   Choice   Analysis:   Theory   and  
Application  to  Travel  Demand  –  MIT  Press  

  37  
 Ben-­‐Akiva   M.   E.   and   Bierlaire,   M.   (1999),   Discrete   choice   methods   and   their  
applications   to   short-­‐term   travel   decisions   –   Handbook   of   Transportation  
Science  
Ben-­‐Akiva,   M.   E.,   McFadden   D.,   Train   K.,   Walker   J.,   Bhat   C.,   Bierlaire   M.,   Bolduc   D.,  
Boersch-­‐Supan  A.,  Brownstone  D.,  Bunch  D.S.,  Daly  A.,  De  Palma  A.,  Gopinath  
D.,   Karlstrom   A.   and   Munizaga   M.   (2002),   Hybrid   Choice   Models:   Progress  
and  Challenges  
Bierlaire  M.  (2001),  A  theoretical  analysis  of  the  cross-­‐nested  logit  model  
Collins  L.M.,  and  Lanza  S.T.  (2010),  Latent  class  and  latent  transition  analysis  for  the  
social,  behavioural  and  health  sciences  
Cox  D.R.  (1970),  Analysis  of  Binary  Data  
Dow   J.K.   and   Endersby,   J.W.   (2003),   Multinomial   probit   and   multinomial   logit:   a  
comparison  of  choice  models  for  voting  research  
Geweke  J.  (1996),  Monte  Carlo  simulation  and  numerical  integration  –  Handbook  of  
Computational  Economics,  Elsevier  Science  
Glasgow  G.  (2001),  Mixed  logit  models  for  multiparty  elections  
Greene  W.  (2001),  Fixed  and  Random  Effects  in  Nonlinear  Models  –  Working  paper  
Stern  School  of  Business,  Department  of  Economics  
Greene   W.H.   and   Hensher   D.A.   (2003),   A   latent   class   model   for   discrete   choice  
analysis:  contrasts  with  mixed  logit  
Hajivassiliou  V.,  McFadden  D.  and  Ruud  P.  (1996),  Simulation  of  multivariate  normal  
rectangle  probabilities  and  their  derivatives:  Theoretical  and  computational  
results  –  Journal  of  Econometrics  
Hensher  D.A.,  Rose  J.M.  and  Greene  W.H.  (2005),  Applied  Choice  Analysis:  A  Primer  
Karlstrom   A.   (2001),   Developing   generalized   extreme   value   models   using   the  
Piekands   representation   theorem   –   Working   paper   Infrastructure   and  
Planning  –  Royal  Institute  of  Technology  Stockholm  
Kreps  D.M.  (1988),  Notes  on  the  theory  of  choice  
Kroes   E.P.   and   Sheldon   R.J.   (1988),   Stated   Preference   Methods   –   Journal   of  
Transport  Economics  and  Policy  
Luce,  D.  (1959),  Individual  Choice  Behavior:  A  Theoretical  Analysis  
McCutcheon,  A.  L.  (1987),  Latent  class  analysis  
McFadden,   D.   (1978),   Modelling   the   choice   of   residential   location   –   Spatial  
interaction  theory  and  residential  location  
McFadden   D.   and   Train   K.   (2000),   Mixed   MNL   models   for   discrete   response   –  
Journal  of  Applied  Econometrics  
Papola  A.  (2003),  Some  developments  on  the  cross-­‐nested  logit  model  -­‐  Elsevier  
Stevens   T.   H.   (2005),   Can   Stated   Preference   Valuations   Help   Improve  
Environmental   Decision   Making?   –   Choices   Magazine,   A   publication   of   the  
American  Agricultural  Economics  Association  

  38  
 Train  K.  (2003),  Discrete  Choice  Methods  with  Simulation  
Thurstone  L.  L.  (1927),  A  law  of  comparative  judgment  –  Psychological  Review  
Varian  H.R.  (2005),  Revealed  Preference  
Vovsha,   P.   (1997),   Cross-­‐nested   logit   model:   an   application   to   mode   choice   in   the  
Tel-­‐Aviv  metropolitan  area  
Wassenaar   H.J.   and   Chen   W.   (2003),   An   Approach   to   decision-­‐based   design   with  
discrete  choice  analysis  for  demand  modeling    
Wardman   M.   (1988)   A   comparison   of   revealed   preference   and   stated   preference  
models  of  travel  behavior  –  Journal  of  transport  economics  and  policy    
Wooldridge  J.M.  (2002),  Econometric  Analysis  of  Cross  Section  and  Panel  Data,  MIT  
Press  
 

  39