Presentation is loading. Please wait.

Presentation is loading. Please wait.

Factor Analysis ( 因素分析 ) Kaiping Grace Yao National Taiwan University 2010-3-30.

Similar presentations


Presentation on theme: "Factor Analysis ( 因素分析 ) Kaiping Grace Yao National Taiwan University 2010-3-30."— Presentation transcript:

1 Factor Analysis ( 因素分析 ) Kaiping Grace Yao National Taiwan University 2010-3-30

2 “ Factor Analysis ” in MEDLINE

3 Classification on FA Exploratory factor analysis (EFA, 探索 性因素分析 ) Confirmatory factor analysis (CFA, 驗證 性因素分析 )

4 Exploratory factor analysis (EFA) Steps: 1.Communalities ( 估計共通值 ) 2.Number of Factors ( 決定因素的數目 ) 3.Estimation ( 估計因素負荷量 ) 4.Rotations ( 對因素做轉軸 ) 5.Interpretation ( 對結果作解釋 ) 6.Reports ( 報告 )

5 Confirmatory factor analysis (CFA) Steps: 1.Model specification ( 確立因素結構關係 ) 2.Identification ( 對參數找出單一組解 ) 3.Estimation ( 參數的估計 ) 4.Evaluation ( 檢驗資料與假設模式間的適合度 ) * Structural Equation Model (SEM, 結構方程模式 )

6 Purposes of EFA Developing a measurement scale : *construct validity (factor validity) *select items Test other hypotheses :

7 Data Reduction Methods Exploratory factor analysis ( 探索性因素分析 ) Cluster analysis ( 群聚分析 ) Multidimensional scaling ( 多元尺度分析 )

8 EFA History Spearman ’ s two-factor theory (1904, 1923, 1927) Holzinger & Swineford ’ s bi-factor model (1937) Thurstone (1921, 1931) : multiple factors Guilford (1956, 1967): Sternberg (1985, 1986): Gardner (1986):

9 Spearman ’ s two-factor theory g 1234 S1S1 S2S2 S3S3 S4S4 g factor specific factor

10 Holzinger & Swineford ’ s bi-factor model (1937)

11 EFA Under the assumption of having measurement error, EFA is to use fewer factors to interpret the correlations among many variables. Factor pattern matrix ( 4 vars , 2 factors ) F1F1 F2F2 X1X1 λ 11 λ 12 X2X2 λ 21 λ 22 X3X3 λ 31 λ 32 X4X4 λ 41 λ 42

12 Model of EFA i =1,…,P variables j =1,…,N subjects k =1,…,M factors (M << P) λ ik (PxM) factor loadings F kj (MxN) factor scores E ij (PxN) errors Observed score = Σ(factor loadings)x(factor score)+error assumptions : ε(F)=0; Cov(F)=I; ε(E)=0; Cov(E)=diagonal matrix; Cov(F,E)=0

13 EFA method r kk’ =φ kk’ = corr. between factor k & k’ = root mean square when no correlation among factors  == the goal of EFA is to minimize this value

14 F1F1 F2F2 X1X1 X2X2 X3X3 X4X4 λ 11 λ 21 λ 31 λ 41 λ 12 λ 22 λ 32 λ 42 E 1 E 2 E 3 E 4 φ 21 Est. Corr(X 1, X 2 ) =

15 Example of FA Holzinger & Swineford (1939)

16 Example purpose : item analysis and test development 生活品質 QOL 生理 PHY 心理 PSY 社會 SOC Factor 1Factor 2Factor 3 X10.550.110.20 :0.540.090.05 :0.420.010.11 X100.650.010.15 X11 :0.030.100.01 :::: X20 X21 :0.100.420.45 :::: X30

17 Geometric Representation F2 F1 θ 21 angle θ 21 --  correlation Vector length --  communality

18 1.Communalities ( 估計共通值, h 2 ) meaning : how much variance of each variable can be explained by each factor 每個變項之變異量能為各因素解釋之部分 Influence by the # of factors methods : Iterative method estimates: Thurstone ’ s: largest |r| Squared multiple correlation (SMC) : R 2

19 2.Number of Factors ( 決定因素的數目 ) 1 on diagonal: eigenvalue > 1 h 2 on diagonal: eigenvalue > 0 Scree test ML significant test Tucker-Lewis index Theory, integration, interpretability Parallel analysis Minimum average partial Bartlett ’ s test

20 3.Estimation/Extraction ( 估計因素負荷量 ) Principal component, PC Principal factor, PF Iterative principal factor, IPF [SAS: prinit = SPSS: principal axis factoring] Maximal likelihood, ML Image factor analysis Alpha factor analysis Minimum residual (Minres) ULS GLS Schmid-Leiman hierarchical factor analysis

21 4.Rotations ( 對因素做轉軸 ) Subjective vs. Objective rationale : factor indeterminacy Goal : simple structure Orthogonal rotations ( 正交轉軸 ) : varimax, quartimax, tanderm criteria, orthomax Oblique rotation ( 斜交轉軸 ) : oblimin, orthoblique, procrustes, promax

22 5.Interpretation ( 解釋結果 ) Naming : creativity, ingenuity, familiarity with the data Factor loadings ≧ 0.3 or 0.4 Most salient loadings Relative importance of variables Literature and theory

23 6.Reports ( 報告結果 ) Background info. : Sample size,composition, subject selection method Type of variables, descriptive statistics, correlation method methods : communalities, number of factors, estimation on factor loadings, rotations results : descriptive statistics of vars, correlation matrix, factor loadings, final communality estimates, inter-factor correlations

24 Good EFA Design Plan before data collection Hypothesize a factor structure based on theory An item should measure only one factor Aim at primary factors: do not analyze variables which belongs to different hierarchical levels Select suitable variables: normal,linear, reliable Choose appropriate sample Use proper correlations Use appropriate method on: communalities, # of factors, factor loadings, rotation Interpret the results, modify the hypothesis, cumulate study results Use other approaches Cross-validation, CFA

25 Problems of past studies (1) Wang, C.N., & Weng,L.J. (2002) : Evaluation on the applications of EFA in Taiwan : 1993~1999 。 Chinese Journal of Psychology, 44(2), 239-251 。 Investigate five journals on the areas of psychology, education, business management. 82/497(16%) papers use EFA , 154 times results : Item level (90%) Report on item distribution (0%) 5-point scale (34%) [No 14% ; 4-point 18% ; 7-point 19%] eigenvalue > 1 (29%) [No 53%] Principal component method (40%) [No 32% ; PF 22%] Varimax (26%) [No 33% ; orthogonal rot. 32% ; oblique rot. 31%]

26 Problems of past studies (2) Discussion : # of point of response scale and distribution Decision on # of factors Estimation on factor loadings Rotation methods Conclusion :


Download ppt "Factor Analysis ( 因素分析 ) Kaiping Grace Yao National Taiwan University 2010-3-30."

Similar presentations


Ads by Google