LISREL

Size: px

Start display at page:

Download "LISREL"

友莫
7 years ago
Views:

1 Nov, 06 LISREL (Linear Structural Relations) Path Diagram 0 Ctrl Control FlexSch ChldCare FSP 3 3 SelfRating 5 Job Satisfaction 5 SuprRating PeerRating 7 0 FamSS FSS 4 WFC 3 4 FmlyCare HousChor UnspWEnv WorkFlex 3 4 FSP = family-supportive policies; FSS = family-supportive supervisors; Control = employee control; WFC = work-family conflict; FlexSch = flexible work schedule; ChldCare = child-care assistance; FamSS = family-supportive supervisors; Ctrl = employee control; FmlyCare = family member caregiving; HousChor = household chores; UnspWEnv = unsupportive work environment; WorkFlex = work flexibility; SuprRating = supervisor rating. Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 )

2 Nov, 06 Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) Matrices for Measurement Models Measurement Models( 測量模式 ). Endogenous Variables: Y = Y Ctrl FmlyCare HousChor UnspWEnv WorkFlex SelfRating SuprRating PeerRating Control WFC JobSatisfaction Exogenous Variables: X = x FlexSch ChldCare FamSS FSP FSS 註 : 此處的 measurement models 為設定 full model syntax( 即本講義自我測驗的 syntax) 中的 matrix( 如 :LY and LX) 時使用的依據若要做 Anderson and Gerbing (988) 所建議的 two-step approach ( 先做 measurement model 再做 structural model) 0

3 Nov, 06 中的 measurement model, 則不需將與分開設定, 而是將與一起做成一個 model( 即上次上課的 CFA syntax), 以檢驗其 convergent validity and discriminant validity, 並與其他 alternative models 比較其是否為 fit 較佳的 model Matrices for the Structural Model Structural Model( 結構模式 ) = + + Control WFC JobSatisfaction 0 sym Control FSP WFC FSS JobSatisfaction 如何以 LISREL 執行模式適配度的檢測? ) 以 LISREL 中的 PRELIS, 利用你的原始資料檔, 產生一個 covariance matrix 或 correlation matrix ) 撰寫 LISREL 的 syntax 檔 ( 程式檔 ), 並存檔 ( 如 :wfc.ls8) 3) 執行 LISREL ( 此 3 點的相關細節請見 CFA 講義 : 如何以 LISREL 執行驗證性因素分析?) Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 3

4 Nov, 06 LISREL V8 Control Statement( 即 syntax 檔內容 ) ( 在寫 syntax 檔時不需隔行空白, 此為方便各位記筆記 ; 此 syntax 命名為 wfc.ls8) TI Work-Family Conflict DA NI= NO=000 MA=CM LA Ctrl FmlyCare HousChor UnspWEnv WorkFlex SelfRatg SuprRatg PeerRatg FlexSch ChldCare FamSS CM FI=C:\WFC MO NY=8 NX=3 NE=3 NK= LY=FU,FI LX=FU,FI TE=DI,FR TD=DI,FR GA=FU,FR BE=FU,FI PH=SY,FR PS=DI,FR LE Control WFC JobSat LK FSP FSS FR LY(3,) LY(4,) LY(5,) LY(7,3) LY(8,3) LX(,) BE(,) BE(3,) FI TE(,) TD(3,3) GA(3,) VA.0 LY(,) LY(,) LY(6,3) LX(,) LX(3,) PATH DIAGRAM OU SE TV MI SS SC 以上為完整之 syntax, 以下為解析 syntax 中每個設定 : Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 4

5 Nov, 06 Title Data TI Work-Family Conflict DA NI= NO=000 MA=CM Number of indicators Number of observations Matrix Covariance Matrix Label LA Ctrl FmlyCare HousChor UnspWEnv WorkFlex SelfRatg SuprRatg PeerRatg FlexSch ChldCare FamSS Covariance Matrix CM FI= C:\WFC File Number of Y (indicator Y 個數 ) Number of X (indicator X 個數 ) Number of η( 潛在內生變項個數 ) Number of ξ( 潛在外生變項個數 ) Model MO NY=8 NX=3 NE=3 NK= Lambda Y Lambda X Theta-epsilon Theta-delta LY=FU,FI LX=FU,FI TE=DI,FR TD=DI,FR Full Fix ( 全部參數不估計 ) Diagonal Free ( 僅估計矩陣對角線上的參數 ) Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 5

6 Nov, 06 Gamma(ξ 與 η 的關係矩陣 ) Beta(η 與 η 的關係矩陣 ) Phi(variance covariance matrix of ξ) Psi(variance covariance matrix of the residual of η) GA=FU,FR BE=FU,FI PH=SY,FR PS=DI,FR Full Free Full Fix Symmetric Free ( 僅估計對稱矩陣上的元素 ) Diagonal Free Label of η LE Control WFC JobSat Label of ξ LK FSP FSS 將模型中所假設的 matrix 作自由估計 (FR) 或是固定 (FI) FR LY(3,) LY(4,) LY(5,) LY(7,3) LY(8,3) LX(,) BE(,) BE(3,) FI TE(,) TD(3,3) GA(3,) Value( 給予固定的參數一特定值, 本例為.0; 要做 VA 設定之前, 該參數必須先被 fix 住 ) Output VA.0 LY(,) LY(,) LY(6,3) LX(,) LX(3,) PATH DIAGRAM OU SE TV MI SS SC Complete Standardized Solution: Both latent variables and indicators are standardized. Standard Error t-value Modification Index Standardized Solution: Only latent variables are standardized. Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 6

7 Nov, 06 LISREL/SIMPLIS 重要注意事項與經驗摘要在 LISREL/SIMPLIS 中的自由度 (Degrees of freedom): - 自由度 = 可能的自由度欲估計的參數數目 vv - 可能的自由度 = (v = 觀察變數 ; 即 observed variable 或指標 /indicator/measure) - 欲估計的參數 :,,, ( 欲估計的參數會出現在 output 檔案 ) 單一指標 (single indicator): - 當潛在變數只有一個指標時, 此指標稱之為單一指標 - 當樣本較小時, 可採每個潛在變數都只有一個指標, 目的是減少欲估計的參數數目 ( 如 : 指標的和 error variance) - 單一指標的 ( 因素負荷量 or factor loading) 和 error variance( 誤差變異 ) 的設定 : - : a. 未標準化資料 : (Barrick, Mount, & Strauss, 993) b. 標準化資料 : (Susskind, Kacmar, & Borchgrevink, 003; Williams & Hazer, 986) - error variance: a. 未標準化資料 : *(-)(Barrick, Mount, & Strauss, 993; Susskind, Kacmar, & Borchgrevink, 003; Tsai & Huang, 00) b. 標準化資料 :- (Williams & Hazer, 986) - 若無法得知指標的時, 可設為 0.85 有人建議 0.85 的假設值會比.00 的假設值好, 因為這樣的假設包含了 5% 的衡量誤差, 很符合變數無法百分之百衡量無誤的事實但是若為幾乎無衡量誤差的指標, 則可設, 如年齡等較客觀的變項卡方 ( ) 檢定 : 在一般的卡方檢定中, 卡方值 ( ) 愈小, 愈可能不顯著在 LISREL 中雖然也是如此, 但與一般概念不太一樣的是 : 當卡方值不顯著時, 表示模型適配良好然而, 卡方值對樣本數非常敏感 : 樣本數愈大, 卡方值愈容易顯著因此, 卡方值的判定並不能為適配度唯一的選擇解決方式 : 可參考其它判 Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 7

8 Nov, 06 斷模型適配好壞的指標, 這些指標並不會受樣本大小所影響, 如 : - RMSEA( 小於 0.08 為可接受的模型 ; 小於 0.05 更佳 ) - GFI CFI NFI ( 大於 0.9 為可接受的模型 ; 大於 0.95 更佳 ) - / df( 與自由度的比率 ; 此比率在左右為模型適配佳 ) 當 item-level 的 measurement model 適配不佳時, 可考慮使用 parceling, 即將 indicators 打包成一個個包裹, 以降低參數數目每個 latent variable 建議產生 3-5 個 parcels, 每個 parcel 建議 3-5 個 items, 但仍以題數而定 Parceling 的方法 : 若這些問項已經被分成 dimension, 則用 dimension 隨機 : 以亂數表取將 indicator 以 factor loading 由高排至低, 將高的搭配低的為一個 parcel, 如取最高與最低形成一個 parcel 則每個 parcel 的平均 factor loading 皆相似同一個 model 中, 不一定每一個 latent variable 皆需使用 parcel, 只取 item 較多的 variable 做即可 Measurement model and structural model 的 indicator 最好同是 item-level 或同是 parcel 處理不可 measurement model 為 parcel, 而 structural model 為 item-level 但可 measurement model 用 item-level, 而 structural model 用 parcel, 但是最好是再做一次用 parcel 的 measurement model 以確定其適配是否佳在寫程式中的 label 時,Y 觀測變數 (indicator) 要放在 X 觀測變數的前面, 且順序要和資料檔中變數的順序一樣否則會產生錯誤的相關係數矩陣或共變數矩陣 LISREL 的結果並沒有隱含因果關係, 不能因為 LISREL 架構圖類似有因果關係, 就此推導, 因為 LISREL 也是使用迴歸的概念, 和迴歸一樣, 要做因果推導必需有一些條件, 如 :X 先測量,Y 後測量注意程式中的英文拼字, 拼錯字常是結果跑不出來的原因!! 控制變數不一定要畫在 path diagram 裡, 除非是重要的控制變數 Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 8

Nov, 06 Cudeck, R., & Henly, S. J. (99). Model selection in covariance structures analysis and the problem of sample size: A clarification. Psychological Bulletin, 09, 5-59.

9 Nov, 06 Cudeck, R., & Henly, S. J. (99). Model selection in covariance structures analysis and the problem of sample size: A clarification. Psychological Bulletin, 09, 此文章提及 :) 樣本愈大, 或 ) 模型愈繁複 ( 見表中 q k, 其為 number of parameters), 則各種 discrepancy( 如 : 觀察資料共變數和理論共變數的差距 ) 愈小, 但其沒有考慮到簡效 (parsimonious) 簡效的概念為 :LISREL 模型中每多拉一條線 ( 多估計一條線 ), 雖然會增加 fit, 且 χ 值降低, 但是也會喪失個自由度哪些線應該增或減, 可用 χ 差異檢定 (χ difference test) 來驗證, 若原始模式的 χ 值和增一條線的模式的 χ 值沒顯著差異, 仍以原始模式比較好, 因為合乎簡效原則 Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 9

10 Nov, 06 LISREL 結果檔 Fit 指標解析 Goodness of Fit Statistics v( v ) # ofparameters Degrees of Freedom = 0 Minimum Fit Function Chi-Square = 5.4 (P = ) Normal Theory Weighted Least Squares Chi-Square = 4.68 (P = ) Estimated Non-centrality Parameter (NCP) = 4.68 (n-)*fit min 90 Percent Confidence Interval for NCP = (3.76 ; 33.7) Minimum Fit Function Value = Population Discrepancy Function Value (F0) = Percent Confidence Interval for F0 = (0.03 ; 0.) Root Mean Square Error of Approximation (RMSEA) = Percent Confidence Interval for RMSEA = (0.036 ; 0.) P-Value for Test of Close Fit (RMSEA < 0.05) = 0.4 F min Expected Cross-Validation Index (ECVI) = Percent Confidence Interval for ECVI = (0.8 ; 0.8) ECVI for Saturated Model = 0.0 ECVI for Independence Model =.57 Chi-Square for Independence Model with Degrees of Freedom = Independence AIC = Model AIC = Saturated AIC = Independence CAIC = 76.4 Model CAIC = 44.4 Saturated CAIC = 86.7 Normed Fit Index (NFI) = 0.96 Non-Normed Fit Index (NNFI) = 0.95 Parsimony Normed Fit Index (PNFI) = 0.46 Comparative Fit Index (CFI) = 0.98 Incremental Fit Index (IFI) = 0.98 Relative Fit Index (RFI) = 0.93 Critical N (CN) = Root Mean Square Residual (RMR) = 0.03 Standardized RMR = Goodness of Fit Index (GFI) = 0.98 Adjusted Goodness of Fit Index (AGFI) = 0.93 Parsimony Goodness of Fit Index (PGFI) = 0.35 Or RMSR Or SRMR Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 0

11 Nov, 06 References: Anderson, J. C., & Gerbing, D. W. (988). Structural equation modeling with practice: A review and recommended two-step approach. Psychological Bulletin, 03, Barrick, M. R., Mount, M. K., & Strauss, J. P. (993). Conscientiousness and performance of sales representatives: Test of the mediating effects of goal setting. Journal of Applied Psychology, 78, Susskind, A. M., Kacmar, K. M., & Borchgrevink, C. P. (003). Customer service providers attitudes relating to customer service and customer satisfaction in the customer-server exchange. Journal of Applied Psychology, 88, Tsai, W. C., & Huang, Y. M. (00). Mechanisms linking employee affective delivery and customer behavioral intentions. Journal of Applied Psychology, 87, Williams, L. J., & Hazer, J. T. (986). Antecedents and consequences of satisfaction and commitment in turnover models: A reanalysis using latent variable structural equation methods. Journal of Applied Psychology, 7, 9-3. Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 )

12 Nov, 06 LISREL 自我測驗 : 針對以下模式 :. 模仿 LISREL 講義第一頁, 標出 path diagram 裡的所有符號. 用課程網站上的 LISRELEx.lsf 資料檔處理以下 : ) 定義變項 ) 查看資料 3) 產生 covariance matrix ( 以上 3 步驟請見上週 CFA 講義 ) 3. 寫出 measurement models 和 structural model 的矩陣 4. 寫出執行 LISREL 的程式內容 (syntax 檔 ), 可以 based on wfc.ls8 做修改 5. 執行 LISREL 6. 這個模式與資料適配的程度如何? 需要做修正嗎? ProcP ProcI ProcL Locus LocusL RecomL Recom DisL DecFair DecAgr Note. LocusL = Locus of control; ProcL = Procedural justice; DisL = Distributive justice; RecomL = Recommendation to others; Locus = Locus of control; ProcP = Selection procedural fairness; ProcI = Selection instrument fairness; DecFair = Decision fairness; DecAgr = Decision agreement; Recom = Recommedation to others. PS. Feel free to relax/free the path specification (e.g., ), make assumptions, or make any other changes. Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 )

13 Nov, 06 LISREL 自我測驗結果與 Modification Index. How to conduct convergent and discriminant validities using SEM? 見課程講義 PowerPoint 檔 Quantitative Research Methods and Application/Grad/NTU BA/ 06 Aichia Chuang( 莊璦嘉 ) 3

<4D F736F F D20BDD7A4E5A4BAA4E5BB50A5D8BFFD2E646F63>

<4D F736F F D20BDD7A4E5A4BAA4E5BB50A5D8BFFD2E646F63> 附錄一最終模型 SIMPLIS 語法 Title: 建模樣本 Observed Variables: x1-x22 covariance matrix from file final01cov means from file final01means sample size: 2181 latent variables: interest instrument anxiety selfconcept