What is PredictiveAbout Partial Least Squares?Galit Shmueli Otto KoppiusUMD Erasmus
Barclay, Higgins & Thompson Technology Studies, 1995
Explanatory PLS: test causal hypothesesPath coefficientsloadingsweightsInitialize: = avg(y1,…,yq)= f(x1,…xp)= avg( 1x1,…, ...
Biased coefficients(Dijkstra , J Econ 1983; Chin et al., ISR 2003)Under-estimatedpath coefficientsOver-estimatedloadingswe...
PLS path modeling is “causal-predictive”Jöreskog & Wold, 1982Fornell & Bookstein, 1982Barclay, Higgins & Thompson, 1995
PLS RegressionPLS Path Models
Explanatorystatistical model:Theory-based,empirical testing ofcausal hypothesesExplanatory power:Strength ofrelationship i...
Measuring Predictive PowerHoldout data setPrediction errorsniii yynRMSE12ˆ1 ni iiiyyynMAPE1%100ˆ1
Predictive Aspects of PLS?Not in PLS softwareNot in published research
Note: Cross-validation(blindfolding)Explanatory context• Validate measurementmodel, structural model• Estimate communality...
How to predict holdout datawith PLS?
Example: TAMGefen & Straub, CAIS 2005PUPEOUUSE
TAM: Predictive Power?Training: n=70 Holdout: n=30RMSE = 2.1MAPE = 38%
Can we do better with aPredictive Model?
Iterative estimationNode-level errorsNeural NetworkHackl & Westlund TQM 2000 ; Hsu, Chen & Hsieh TQM 2006
PLS Neural NetHypothesis testing Purely predictiveCausal theory No causal theoryLatent nodes importantPath coefficients ma...
Predictive Performance(on holdout)Method RMSEPLS (reflective) 2.10Neural Net 1.94Linear Regression 1.84PLS (formative) 2.1...
What is Predictive About Partial Least Squares?
What is Predictive About Partial Least Squares?
Upcoming SlideShare
Loading in...5
×

What is Predictive About Partial Least Squares?

820

Published on

Presentation by Galit Shmueli at the Sixth Symposium on Statistical Challenges in eCommerce Research (SCECR), Austin TX, 2010

Published in: Education, Technology
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
820
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
0
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide
  • All reflective
  • Formative on left; reflective on right
  • Transcript of "What is Predictive About Partial Least Squares?"

    1. 1. What is PredictiveAbout Partial Least Squares?Galit Shmueli Otto KoppiusUMD Erasmus
    2. 2. Barclay, Higgins & Thompson Technology Studies, 1995
    3. 3. Explanatory PLS: test causal hypothesesPath coefficientsloadingsweightsInitialize: = avg(y1,…,yq)= f(x1,…xp)= avg( 1x1,…, pxp)yi = f( )Last: = f( )iterate
    4. 4. Biased coefficients(Dijkstra , J Econ 1983; Chin et al., ISR 2003)Under-estimatedpath coefficientsOver-estimatedloadingsweights
    5. 5. PLS path modeling is “causal-predictive”Jöreskog & Wold, 1982Fornell & Bookstein, 1982Barclay, Higgins & Thompson, 1995
    6. 6. PLS RegressionPLS Path Models
    7. 7. Explanatorystatistical model:Theory-based,empirical testing ofcausal hypothesesExplanatory power:Strength ofrelationship instatistical modelPredictive model:Empirical methodfor predicting newobservationsPredictive power:Ability toaccurately predictnew observations
    8. 8. Measuring Predictive PowerHoldout data setPrediction errorsniii yynRMSE12ˆ1 ni iiiyyynMAPE1%100ˆ1
    9. 9. Predictive Aspects of PLS?Not in PLS softwareNot in published research
    10. 10. Note: Cross-validation(blindfolding)Explanatory context• Validate measurementmodel, structural model• Estimate communality,redundancy
    11. 11. How to predict holdout datawith PLS?
    12. 12. Example: TAMGefen & Straub, CAIS 2005PUPEOUUSE
    13. 13. TAM: Predictive Power?Training: n=70 Holdout: n=30RMSE = 2.1MAPE = 38%
    14. 14. Can we do better with aPredictive Model?
    15. 15. Iterative estimationNode-level errorsNeural NetworkHackl & Westlund TQM 2000 ; Hsu, Chen & Hsieh TQM 2006
    16. 16. PLS Neural NetHypothesis testing Purely predictiveCausal theory No causal theoryLatent nodes importantPath coefficients matterHidden layers – utilityWeights – for predictionLinear relationships Nonlinear relationshipsRestricted paths All possible pathsbetween layers
    17. 17. Predictive Performance(on holdout)Method RMSEPLS (reflective) 2.10Neural Net 1.94Linear Regression 1.84PLS (formative) 2.18Predicts “3”Lots of data?

    ×