传统的中介作用检验程序(Baron and Kenny, 1986)新的改进

weiguozhong

Recently I came across a difficult situation on mediation using SEM for latent variables. My empirical results revealed that the predictor variable is not related to the ourcome variable. According to Kenny, Kashy, and Bolger (1998) it can still conduct mediation analysis. However, I find it difficult to interpret the results both logically and methodologically. Although the techniques conducting mediation analysis for latent variable is somewhat different, some puzzles still exist.

£
5 v& \* r2 `5 {4 z6 Z0 s) a2 |' |# ZHypothesis 1: X is positively related to Y (path c)

£
' k$ ?  c8 o& OHypothesis 2: M mediates the relationship between X and Y (X->M, path a, M->Y, path b, X->Y, path c').

To test the two hypotheses, four steps are suggested by Baron and Kenny (1986):

Step 1: estimate and test path c;

Step 2: estimate and test path a;

Step 3: estimate and test path b;

Step 4: see path c' whether equals to zero.

£
& c0 z" l$ w+ Q2 |; h1 Y0 o5 XNow my research question is somewhat different: X, M, and Y are all latent variables. According to Kenny (2008), for latent variable models, the total effect estimated from a model without the mediator (that is path c) is usually not comparable to the direct effect estimated from a model in which the mediator is included. It is then inadvisable to test the relative fit of two structural models, one with the mediator and one without.
4 M) V8 @( A- R- R8 U  URather c, the total effect, can be estimated using the formula of c' + a*b.

My question 1 is: if we only can use c' + a*b to see the total effect of X on Y for latent variables, how do we know the significance level of the effect (p value for c' + a*b) so we can determine whether the data support or reject Hypothesis 1.

£
' `0 A" ^9 C1 w# |The second question is whether all of the steps have to be met for there to be mediation. Kenny, Kashy, and Bolger (1998) showed that Step 1 is not required. Recently, one of my research projects comes across such situation. Results from SEM are as following:

	M	Y
SEM Model without mediator (M1)
X		.07(.91)
SEM Model with mediator (M2)
X	-.17**(-2.90)	-.06(-.51)
M		.39*(2.51)

Note: X, M, and Y are all latent variables; the number in the bracket is t-value

According to Kenny (2008), we cannot rely on M1 to test Hypothesis 1 and we must use the results of M2. So the total effect of X on Y = -.06 + .39*(-.17)
≈ -.13. Then, as my Question 1, how can we know the effect (-.13) significant or not?

If we suppose the effect (-.13) is insignificant, following Kenny, Kashy, and Bolger (1998), we can still conduct mediation analysis. Sobel test revealed that Z = -1.895 with p = .058. The percentage of indirect effect is .39*(-.17)/(-.13) = 51%. Thus, M partially mediated the effect of X on Y.

- e1 Q# Q/ f& g! T+ O' x) gMy question 2 is: if we can not determine the significance level of -.13, does the percentage of indirect effect really have informative meanings? That is, if we fail to support H1, we cannot rule out the possibility that X and Y has null effect. Then, the result that 51% of total effect is mediated really makes sense?

References

[1] Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic and statistical considerations. Journal of Personality and Social Psychology, 51, 1173-1182.

[2] Kenny, D. A. (2008). Reflections on Mediation. Organizational Research Methods. 11(2), 353-358.

[3] Kenny, D. A., Kashy, D. A., & Bolger, N. (1998). Data analysis in social psychology.
In D. Gilbert, S. Fiske, & G. Lindzey (Eds.), The handbook of social psychology (Vol. 1, 4th ed., pp. 233-265).
7 O. A7 r5 {$ v! W6 oBoston, MA: McGraw-Hill.

weiguozhong

目前针对这种中介作用（distal mediation; inconsistent mediation）的检验，有所运用的文章有：
' d. j, Q4 c" p# K[1] SCHNEIDER BENJAMIN, EHRHART MARK G. , MAYER DAVID M. , SALTZ JESSICA L. , and NILES-JOLLY KATHRYN. Understanding organization-customer links in
service settings. Academy of Management Journal, 2005, 48(6), 1017-1032.

6 Z9 D8 @+ d3 b* L[2] RICH BRUCE LOUIS, LEPINE JEFFREY A., and CRAWFORD EEAN R. Job engagement: antecedents and effects on job performance. Academy of Management Journal, 2010, 53(3), 617-635.

4 B8 c  K  V7 H" P. e[3] SRIVASTAVA ABHISHEK, BARTOL KATHRYN M., and LOCKE EDWIN A. Empowering leadership in management teams: effects on knowledge sharing, efficacy, and performance. Academy of Management Journal, 2006, 49(6), 1239 - 1251.
& i1 i  O6 k# j# J1 C$ t
- d2 p# M% }  y* C; d/ a[4] Homburg Christian, Wieseke Jan, and Bornemann Torsten. Implementing the Marketing Concept at the Employee–Customer Interface: The Role of Customer Need Knowledge. Journal of Marketing, 2009, 73 (July), 64–81.

[5] GALVIN BENJAMIN M., WALDMAN DAVID A., and BALTHAZARD PIERRE. Visionary communication qualities as mediators of the relationship between narcissism and attributions of leader charisma. Personnel Psychology. 2010, 63, 509 - 537.
: P8 J( i5 B' x* n
( v( u) T- Z" B. H% I8 _2 C[6] TIWANA AMRIT. Do bridging ties complement strong ties? an empirical examination of alliance ambidexterity. Strategic Management Journal. 2008, 29, 251–272.

[7] Bradley J. Alge, Gary A. Ballinger, Subrahmaniam Tangirala, and James L. Oakley. Information Privacy in Organizations: Empowering Creative and Extrarole Performance. Journal of Applied Psychology. 2006, 91(1), 221-232.

: I: C9 p0 y# e7 k" ]* E3 q5 K$ V[8] Christopher C. Rosen, Paul E. Levy, and Rosalie J. Hall. Placing Perceptions of Politics in the Context of the Feedback Environment, Employee Attitudes, and Job Performance. Journal of Applied Psychology. 2006, 91(1), 211-220.

weiguozhong

解决类似问题的方法论文章有：
$ G; x0 V0 }. F) r: ][1] MacKinnon David P. , Lockwood Chondra M., Hoffman Jeanne M., West Stephen G., and Sheets Virgil. A Comparison of Methods to Test Mediation and Other Intervening Variable Effects. Psychological Methods. 2002, 7(1), 83-104.

5 B$ W; O8 g( |, _[2] Shrout Patrick E.  and Bolger Niall. Mediation in Experimental and Nonexperimental Studies: New Procedures and Recommendations. Psychological Methods. 2002, 7(4), 422–445.
% g4 {1 l  L: P$ z7 R5 r+ g1 X
2 K* e2 H* ^7 f[3] Fletcher, T. D. Methods and approaches to assessing distal mediation. Paper presented at the 66th annual meeting of the Academy of Management, Atlanta, GA. 2006, August.

[4] Cheung Gordon W.  and Lau Rebecca S. Structural Equation Models Testing Mediation and Suppression Effects of Latent Variables: Bootstrapping With Structural Equation Models. Organizational Research Methods 2008, 11(2), 296 - 325.

[5] MacKenzie Scott B., Podsakoff Philip M., and Podsakoff Nathan P. Construct measurement and validation procedures in MIS and behavioral research: integrating new and existing techniques. MIS Quarterly, 2011, 35(2), 293-334. （最近的方法论总结文章）

chienhsin

回复 3楼 weiguozhong 的帖子

( J8 j8 k8 s9 b) sDear weiguozhong ,
+ X1 W1 |" T8 q+ `+ y$ [
Thank you for the sharing.
- Y2 F, N% H: F" B; k% c
Could you please send me (linchienhsin@gmail.com) all the papers you mentioned?
! c) P$ B7 T7 c  ~. v6 d
0 v0 o) `+ b" {4 l

# r* z0 I& Q9 N4 c; w: \( ^% `Thank you again.
   
4 ]- Q; y& h7 w* K2 u( A% k( s
5 W$ |. l" T# Y0 c. s

Kenneth

weiguozhong，
! U8 w, ^" K( ~, E( J# _第一、日后你要在这里提问，请用中文。这个圈子是给中国人用的。我打中文这么慢，还是坚持用中文沟通，是希望可以让“一切”中国学者都可以看这个圈子。如果用英文的话，我根本不需要花这么多时间来建立这个圈子。美国的RMNET不是一个很好的沟通渠道了吗？
2 O( d7 W" N4 h7 i$ b( u第二、c+a*b 当然可以用统计来验证，用bootstrapping就可以了。
第三、我是一个保守的人。如果X与Y不相关，我还是会说“根本没有关系可以给任何变量来中介”的。Jeff Edwards 与 Kenny 的观点是“统计学”的观点。我根本不同意他们用的“统计意义上的所谓中介”，所以对他们的结论不敢苟同。
第四、有了上面的第二点，你的第二个问题的前设就错误，所以不需要回答了。

weiguozhong

非常感谢Kenny的指点！！第一次在这里提问，所以不是很懂规矩，还请多多包涵！
我也是认为很奇怪，如果没有效应，何来中介之说。现在可以算是明白。看来中介的检验也自有派系之分。昨天又看到JCR（Journal of Consumer Research）上一篇文章，也是讲中介作用的“Reconsidering Baron and Kenny: Myths and Truths about Mediation Analysis”，也是用bootstrapping的方法来检验。综合两派来看，还是需要根据自己的理论研究立场选择自己的方法立场吧。