20
Event sponsored by Affinnova All copyright owned by The Future Place and the presenters of the material For more informa=on about Affinnova visit h>p://www. affinnova.com/ For more informa=on about NewMR events visit newmr.org Advanced Quant Techniques July 14, 2011 An Introduc4on to Hierarchical Bayes Ray Poynter, The Future Place

Ray poynter advanced quant - 2011 - 3

Embed Size (px)

Citation preview

Page 1: Ray poynter   advanced quant - 2011 - 3

Event  sponsored  by  Affinnova  All  copyright  owned  by  The  Future  Place  and  the  presenters  of  the  material  For  more  informa=on  about  Affinnova  visit  h>p://www.  affinnova.com/  

For  more  informa=on  about  NewMR  events  visit  newmr.org  

Advanced  Quant  Techniques  July  14,  2011  

An  Introduc4on  to  Hierarchical  Bayes  

Ray  Poynter,  The  Future  Place  

Page 2: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

An Introduction to Hierarchical Bayes

Ray Poynter The Future Place

Page 3: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Why HB?

Page 4: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Traditional View To solve a linear equation

y = mx + c 1 unknown, 2 observations

To solve a quadratic y = ax2 + bx + c 2 unknowns, 3 observations

To solve a 2 variable linear equation Y=b0+b1x1 + b2x2

2 unknowns, 3 observations

Page 5: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

HB?

≈ HB ≈ Magic

e.g. 40 unknowns 12 observations

Page 6: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

A  company  has  three  drivers,  Andy,  Bob,  and  Chris.  I  tell  you  one  of  them  had  a  crash  today  If  you  had  to  guess  who,  you’d  have  a  ⅓  chance  

But,  if  I  say  Andy  and  Bob  have  never  had  a  crash  And,  Chris  has  had  fiQeen  crashes  already  this  year  You  might  start  to  think  it  was  Chris  

With  thanks  to  Rich  Johnson  at  Sawtooth  SoQware  for  the  example  

Page 7: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Introducing  Bayes  

Reverend  Thomas  Bayes  1702  -­‐  1761  

Page 8: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Bayes  Theorem  

)()(*)|()|(

YPXPXYPYXP =

This  is  much  clearer  with  an  example,  based  on  a  nasty  disease  Bayesi=s,  which  is  fortunately  very  rare,  only  1  in  1000  people  have  it.  

Page 9: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Bayesi=s  Example  

•  1-in-1000 have Bayesitis - chance is just 0.1%

•  The test is very reliable: –  If you have it the test will be positive in 99% of cases –  If you don’t have the it, chance of a negative result is 95%

•  You are tested and get a positive result, what is the chance that you have the disease?

Just  under  2%  

Page 10: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Alterna=ve  View  of  Bayesi=s  Imagine  we  have  a  popula=on  of  1  million  

Have  the  disease?  P=0.1%  

Posi=ve  Test?  P=99%  

Yes,  n=990  

No,  n=10  

Posi=ve  Test?  P=5%  

Yes,  n=49,5

00  

No,  n=940,500  

50,490  receive  a  posi=ve  result.  990  are  real  posi=ves,  1.96%  of  all  posi=ve  results.  

Page 11: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Hierarchical  Bayes  

In the context of choice data

Page 12: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Conceptual  Guide  to  HB  2 level model is assumed – the Hierarchy in HB

–  Upper level - the part worths for a respondent are drawn from a multivariate normal distribution

–  Lower level - the logit assumption that the probability of choosing x is its transformed utility divided by sum of all the utilities

Markov Chain Monte Carlo Simulations –  Very computer intensive

Great ability to produce individual level results from choice data

–  even with many parameters and relatively few choices

Page 13: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Comparison of Hit Rates with Disaggregation

Summary of Six Commercial Studies

Aggregate Logit

Hierarchical Bayes Improvement

Study One 75.8 99.5 23.8Study Two 24.8 79.5 54.7Study Three 60.5 62.6 2.1Study Four 61.2 79.3 18.1Study Five 59.2 78.8 19.6Study Six 71.9 68.1 -3.8

Note,  the  last  study  was  par=al  profile  .  .  .    Source:    Pinnell  (2000)  

Does HB Work?

Page 14: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Why  Did  We  Need  HB?  

Allows us to estimate individual utilities for DCM studies

– The missing piece

Opens up other cases where sparcity of data meant individual results not available:

– MaxDiff scaling – Bundled choice modeling

Page 15: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

HB

Page 16: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

The  Impact  of  HB  on  DCM  Data  If data for a respondent is ‘typical’, the HB contribution to their results is small

–  If the respondent was an outlier, the contribution of HB to their solution is more

Data from groups of people who’ve been through the HB are normally good predictors of the group solution

– An individual’s post HB data is not necessarily a good predictor of that individual’s values

Page 17: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

HB is very useful

But it’s not magic

Page 18: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Thank you

Ray Poynter The Future Place

Page 19: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Q & A

Ray  Poynter  The  Future  Place  

Andrew  Jeavons  Survey  Analy=cs  

Page 20: Ray poynter   advanced quant - 2011 - 3

Speaker Ray Poynter, The Future Place, UK NewMR Advanced Quant Techniques, July 14, 2011

Ray Poynter

Read  Ray’s  blog  @  h>p://thefutureplace.typepad.com/  Follow  Ray’s  tweets  at  @  h>p://twi>er.com/raypoynter    Connect  with  Ray  on  LinkedIn  @  h>p://uk.linkedin.com/in/raypoynter    Find  out  about  Ray’s  book  at  @  h>p://bit.ly/cmFnbo