PENALTY SELECTION METHODS IN NETWORK MODELS ANNA WYSOCKI

  • Slides: 38
Download presentation
PENALTY SELECTION METHODS IN NETWORK MODELS ANNA WYSOCKI, M. S. awysocki@ucdavis. edu

PENALTY SELECTION METHODS IN NETWORK MODELS ANNA WYSOCKI, M. S. awysocki@ucdavis. edu

GAUSSIAN GRAPHICAL MODELS

GAUSSIAN GRAPHICAL MODELS

How are network models estimated?

How are network models estimated?

Ω= -1 Σ i i. 953 j. 01 k -. 03 l 0 j

Ω= -1 Σ i i. 953 j. 01 k -. 03 l 0 j . 01 . 957 -. 06 -. 02 k -. 03 -. 06 . 957 0 l 0 -. 02 0 . 952

Ω= -1 Σ i i. 953 j. 01 k -. 03 l 0 j

Ω= -1 Σ i i. 953 j. 01 k -. 03 l 0 j . 01 . 957 -. 06 -. 02 k -. 03 -. 06 . 957 0 l 0 -. 02 0 . 952

Ω= -1 Σ i i. 953 j. 01 k -. 03 l 0 j

Ω= -1 Σ i i. 953 j. 01 k -. 03 l 0 j . 01 . 957 -. 06 -. 02 k -. 03 -. 06 . 957 0 l 0 -. 02 0 . 952

Choosing lambda is extremely important. As different lambdas can correspond to different networks.

Choosing lambda is extremely important. As different lambdas can correspond to different networks.

PENELTY SELECTION • Cross-Validation. METHODS • Stability Approach to Regularization Selection (St. ARS) •

PENELTY SELECTION • Cross-Validation. METHODS • Stability Approach to Regularization Selection (St. ARS) • Rotation Information Criterion (RIC) • Extended Bayesian Information Criterion (EBIC)

PENELTY SELECTION METHODS

PENELTY SELECTION METHODS

PENELTY SELECTION METHODS i≠j

PENELTY SELECTION METHODS i≠j

PENELTY SELECTION • Cross-Validation. METHODS • Stability Approach to Regularization Selection (St. ARS) •

PENELTY SELECTION • Cross-Validation. METHODS • Stability Approach to Regularization Selection (St. ARS) • Rotation Information Criterion (RIC) • Extended Bayesian Information Criterion (EBIC)

PENELTY SELECTION • Cross-Validation. METHODS • Stability Approach to Regularization Selection (St. ARS) •

PENELTY SELECTION • Cross-Validation. METHODS • Stability Approach to Regularization Selection (St. ARS) • Rotation Information Criterion (RIC) • Extended Bayesian Information Criterion (EBIC)

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu &

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu & Feng, 2013) • EBIC + PC size (Kuisman & Sillanpaa, 2016) • Low-Dimensional • EBIC Performance (Epskamp, 2016; Epskmap, Borsboom & Fried, 2018) • EBIC vs non-regularized methods (Williams, Rhemtulla, Wysocki & Rast, 2018) • St. ARS vs. EBIC vs. RIC (Mohammadi & Wit, 2015)

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu &

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu & Feng, 2013) • EBIC + PC size (Kuisman & Sillanpaa, 2016) • Low-Dimensional • EBIC Performance (Epskamp, 2016; Epskmap, Borsboom & Fried, 2018) • EBIC vs non-regularized methods (Williams, Rhemtulla, Wysocki & Rast, 2018) • St. ARS vs. EBIC vs. RIC (Mohammadi & Wit, 2015)

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu &

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu & Feng, 2013) • EBIC + PC size (Kuisman & Sillanpaa, 2016) • Low-Dimensional • EBIC Performance (Epskamp, 2016; Epskmap, Borsboom & Fried, 2018) • EBIC vs non-regularized methods (Williams, Rhemtulla, Wysocki & Rast, 2018) • St. ARS vs. EBIC vs. RIC (Mohammadi & Wit, 2015)

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu &

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu & Feng, 2013) • EBIC + PC size (Kuisman & Sillanpaa, 2016) • Low-Dimensional • EBIC Performance (Epskamp, 2016; Epskmap, Borsboom & Fried, 2018) • EBIC vs non-regularized methods (Williams, Rhemtulla, Wysocki & Rast, 2018) • St. ARS vs. EBIC vs. RIC (Mohammadi & Wit, 2015)

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu &

• Expected Performance • CV Over-selects (Chetverikov et al. , 2017; Yu & Feng, 2013) • EBIC + PC size (Kuisman & Sillanpaa, 2016) • Low-Dimensional • EBIC Performance (Epskamp, 2016; Epskmap, Borsboom & Fried, 2018) • EBIC vs non-regularized methods (Williams, Rhemtulla, Wysocki & Rast, 2018) • St. ARS vs. EBIC vs. RIC (Mohammadi & Wit, 2015)

1. What is the best method for psychological data? 2. How does partial correlation

1. What is the best method for psychological data? 2. How does partial correlation size affect performance? 3. How does sparsity affect performance?

SIMULATION PROCEDURE • Conditions • Sample Size (100, 250, 500, …, 3000) • Number

SIMULATION PROCEDURE • Conditions • Sample Size (100, 250, 500, …, 3000) • Number of Variables (20) • Sparsity (. 10, . 20, …, . 90) • Partial Correlation Range (-. 1, . 1 ; -. 3, . 3 ; -. 5, . 5) • Performance Measures • False Positive Rate • Sensitivity

How does sparsity affect method performance?

How does sparsity affect method performance?

How does partial correlation size affect method performance?

How does partial correlation size affect method performance?

What is the best method for psychological data?

What is the best method for psychological data?

SUMMARY • Sparsity and partial correlation size matter • CV: Sensitive but with a

SUMMARY • Sparsity and partial correlation size matter • CV: Sensitive but with a cost • EBIC: Inconsistent Results • St. ARS + RIC most balanced results

THANK YOU! ANNA WYSOCKI, M. S. awysocki@ucdavis. edu

THANK YOU! ANNA WYSOCKI, M. S. awysocki@ucdavis. edu

Piotr Wysocki BIOGRAFIA Piotr Jacek Wysocki ur 10Piotr Wysocki BIOGRAFIA Piotr Jacek Wysocki ur 10
Penalty Shot Procedures Penalty Shot Procedures How DoPenalty Shot Procedures Penalty Shot Procedures How Do
FEDRAL vs STATE DEATH PENALTY The death penaltyFEDRAL vs STATE DEATH PENALTY The death penalty
Efficient Penalty Enforcement Techniques Efficient Penalty Enforcement TechniquesEfficient Penalty Enforcement Techniques Efficient Penalty Enforcement Techniques
Understanding and Avoiding Penalty Points What are PenaltyUnderstanding and Avoiding Penalty Points What are Penalty
PENALTY ENFORCEMENT Rogers Redding NCAA Rules Editor PENALTYPENALTY ENFORCEMENT Rogers Redding NCAA Rules Editor PENALTY
Penalty Kill Opening Remarks Penalty Killing is aPenalty Kill Opening Remarks Penalty Killing is a
Death Penalty DBQ EQ Does the death penaltyDeath Penalty DBQ EQ Does the death penalty
Penalty shooting competition Penalty shooting competition Each personPenalty shooting competition Penalty shooting competition Each person
DEATH PENALTY Mariana Guiza WHAT IS DEATH PENALTYDEATH PENALTY Mariana Guiza WHAT IS DEATH PENALTY
DEATH PENALTY JOLIEN VERBRUGGEN JEF VERCAMMEN Death PenaltyDEATH PENALTY JOLIEN VERBRUGGEN JEF VERCAMMEN Death Penalty
29 A Death Penalty No Death Penalty White29 A Death Penalty No Death Penalty White
Project Selection Models Project Selection Project selection isProject Selection Models Project Selection Project selection is
Transformation Methods Penalty and Barrier methods Study ofTransformation Methods Penalty and Barrier methods Study of
Anna Was wrde Anna denken Anna Wir mchtenAnna Was wrde Anna denken Anna Wir mchten
Network Selection Issues Pasi Eronen 1 Network SelectionNetwork Selection Issues Pasi Eronen 1 Network Selection
MODELS AND TECHNOLOGICAL DESIGN MODELS MODELS Models helpMODELS AND TECHNOLOGICAL DESIGN MODELS MODELS Models help
OBSERVATIONAL METHODS Research Methods Observation Methods Observational MethodsOBSERVATIONAL METHODS Research Methods Observation Methods Observational Methods
Natural Selection II RAPID SELECTION Intense selection rapidNatural Selection II RAPID SELECTION Intense selection rapid
Materials selection properties Materials selection properties Materials selectionMaterials selection properties Materials selection properties Materials selection
Personnel Selection Selection What is selection Using scientificPersonnel Selection Selection What is selection Using scientific
Selection Selection Learning Objectives Explain the Selection ProcessSelection Selection Learning Objectives Explain the Selection Process
Heroes Theme 5 Selection 1 Selection 2 SelectionHeroes Theme 5 Selection 1 Selection 2 Selection
Selection Controls Selection Controls A selection control presentsSelection Controls Selection Controls A selection control presents