270 likes | 357 Views
This wrap-up notes on Analysis of Variance (ANOVA) covers extensions like Intraclass Correlation, Latin Squares Design, and Random Effects Models. Learn about the application, advantages, and disadvantages of these advanced techniques in experimental design.
E N D
Analysis of Variance Wrap-Up Notes 46-511 Fall, 2007
Learning Objectives • Understand a couple of useful extensions of ANOVA • Repeated Measures Designs • Intraclass Correlation • Latin Squares Design • Random Effects Designs • Hierarchical Designs
Intraclass Correlation Coefficient (ICC) • Based on within-subjects design • Columns = Judges / Raters • Rows = Stimulus • Coefficient reveals degree of rater agreement or reliability Three experts watch job candidates in an assessment center and provide ratings of performance:
Use RM ANOVA • Assume experts are random sample
RM ANOVA & Nuissance Effects • Carryover effects • Fatigue/boredom effects • Methods of addressing • Averaging Out • Randomization/Counterbalancing • Partitioning • Latin Squares/Counterbalancing
Randomization • Advantages • Disadvantages
Latin Squares Designs • Objective: to partition out variance due to nuissance factor(s) • Example: we want to know the effects of four computer screen layouts on a vigilance task • Thus, our design is…
Latin Squares Design (cont’d) If we use (Latin) letters to denote the different treatments: The Latin Squares Solution becomes:
Between Subjects Latin Squares Designs • Objective: to accommodate multiple between subjects factors while reducing sample size needs / balanced fractional replication. • Example: we want to know the effects of… • A: Display-color (4 colors), • B: Display-size (4 sizes), and • C: Layout of information (4 levels) on computer operators. • On accuracy in a vigilance task • A 3-way factorial design requires 4 x 4 x 4 = 64 conditions. • Latin squares design requires 16 conditions
The Latin Squares Solution • A & C main effects derived in the usual way • B main effects derived off diagonal (B1, etc.).
Issues with Latin Squares designs • Strict Assumption • No interaction effects • Calculating F • By hand, fairly straight forward • In SPSS… • Planning these designs quickly becomes complicated • Algorithms for generating cell arrangements • Extensive tables (Fisher & Yates, 1953; Cochran & Cox, 1957)
Random Effects Model • What are random effects? • One-Way • N-Way • Repeated Measures • Characteristics • Power • Interpretation • Calculation
1-Way Example • A computer screen must contain 20 different pieces of information • Does the placement of the 20 items make a difference in perception & processing by operators? • There are 20! Permutations (2.43 x 1018) • What to do?
Random Effects Solution • Randomly select a subset of categories to see if there is a main effect • Generalize findings to entire “population” of layouts Five layouts are chosen randomly (Factor/IV), number of operator errors recorded (DV)
Results • For one-way designs • No difference in ANOVA calculations • Main difference = ability to calculate variance components… Variance Component =
Two-Way Random Effects Model… • For example… • Assume our 2-way example (anxiety by task difficulty) was a random effects model • That we had chosen our levels of difficulty & anxiety at random • Why would we have done this?
Results • Error Term • Drawbacks of this design • Caveat on SPSS
Hierarchical/Nested Models • Sometimes interested in the effects of more than one factor, but are unable to fully cross the factors. • May be due to… • Experimental content – different levels of B need to be associated with different levels of A • Constraints in how data are collected (e.g., organizational structure) • Can occur with between, within, or mixed designs
Example 1: Experimental Manipulation • Inducing False Memories • Inducing mundane false memories (e.g., getting lost in a grocery store) • More extraordinary memories (house fire when a child)
Characteristics of this design • B is nested under A • B is treated as a random factor • No interaction effects • Other possibilities • Completely between • Completely within • Mixed (A between, B within)
Example 2: Structural Constraints • Wish to test two different types of employee interventions to reduce turnover • Type of intervention (Factor A): • Level 1: intervention involving greater employee involvement in decision making • Level 2: intervention involving different array of benefits, compensation & training • Location (Factor B): • Each intervention must be implemented at entire work locations, thus 10 work locations are selected for A1 and 10 different ones for A2.
Example 3: Example 1 made more complicated • Inducing False Memories • Inducing mundane false memories (e.g., getting lost in a grocery store) • More extraordinary memories (house fire when a child) • Wish to cross with gender
Issues with Nested Designs • Cannot obtain information about interactions • Most software (e.g., SPSS, SAS) will allow for analysis, but not straight forward or easy • Other, newer regression based methods for hierarchical designs available, and may be more appropriate
Final Thoughts • Experimental Methods / ANOVA can be very flexible • Strike a balance between complexity and elegance • Important sources on ANOVA; • Howell, D. C. (2002). Statistical methods for psychology (5th ed.). Pacific Grove, CA: Duxbury.* • Keppel, G. & Wickens, T. D. (2004). Design and analysis: A researcher’s handbook (4th ed.). Englewood Cliffs, N. J.: Prentice-Hall • Kirk, R. E. (1995). Experimental design: Procedures for the behavioral sciences (3rd ed.). Monterey, CA: Brooks/Cole • Tabachnick, B. & Fidell, (2001). Computer-assisted research design and analysis. Pearson/Allyn & Bacon.* • Winer, B. J., Brown, D. R., & Michels, K. M. (1991). Statistical principles in experimental design, 3rd ed. New York: McGraw-Hill * Newer editions available