You are on page 1of 3

HW 12 Huanhuan Shi 1.

Recreate the models in Spreng and Mackoy(1996) (1) Measurement model Based on Spreng and Mackoy(1996), I can first know the measurement model. Desires, Exp, Perf , Descon and Expdis are each measured by one indicator, the average score of multiple items. I fixed the factor loadings to be 1 and error variance of each indicator to be 0. Sat is measured by 4 indicators and I fixed one of the factor loadings to be 1; the same treatment for SQ. So the number of unknown parameters for the measurement model is 12: Factor loadings (): 5 Cov(, ): 7 (2) Latent variable model I first construct the model proposed by Oliver (1993), in which the desires congruency is not specified to influence overall satisfaction, expectations disconfirmation is not specified to affect perceived service quality, and expectations are not specified to affect perceived performance. To get the result in the paper, I need to set the error covariance of Descon and Expdis to be free and I get Minimum Fit Function Chi-Square = 273.79 , 51 df. This makes sense since this two construct are similar and very likely to be correlated. ( Model for 51 df.
DESCON=DESIRES PERF EXPDIS=PERF EXP SQ=DESCON SAT=SQ EXPDIS Set the Error Covariance of DESIRES and PERF to zero Set the Error Covariance of PERF and EXP to zero Set the Error Covariance of DESIRES and EXP to zero Set the Error Covariance of EXPDIS and DESCON free )

In this latent variable model, the number of unknown parameters is 15: Paths among endogenous latent variables : 3, Coefficients of exogenous latent variables on endogenous variables :4 Cov(, ): 3 Cov(, ): 5 Total data points: 12*13/2=78 df=7812-15=51 Spreng and Mackoy(1996) next test the model with 50 df by freely estimating the relationship between Exp and Perf. I can set the error covariance of EXP and PERF to be free to get Minimum Fit Function Chi-Square = 175.80, 50 df.

Because the path between Descon and SQ is not significant, Spreng and Mackoy(1996) test the model shown as Figure2 on page 209. I run the model with following codes and get Minimum Fit Function Chi-Square = 165.94, 49df. ( Model for 49 df.
DESCON=DESIRES PERF EXPDIS=PERF EXP SQ=DESCON SAT=SQ EXPDIS DESCON PERF=EXP Set the Error Covariance of DESCON and EXPDIS to free Set the Error Covariance of DESIRES and EXP to zero Set the Error Covariance of SQ and EXPDIS to zero)

The path estimates are as follows. Compared to Spreng and Mackoy(1996), the value of path estimates I got is different from theirs but the sign of the relationship is consistent. The relation between Desire and Descon, Exp and Expdis are both negative. The amount of variations explained in SQ and SAT is 38% and 67% respectively, which is consistent with the estimates in Spreng and Mackoy(1996).

2. Compare with a measurement model and improve the model I run the measurement model and let all relationships among latent variables be freely estimated. I got Minimum Fit Function Chi-Square = 46.56 (P = 0.16), df 38. This a saturated model with better fit. Although we dont want such model, given the big difference between Chi-Square 46.56 and 165.94, we may improve the model fit by looking into the modification suggestions. I can exclude some of the suggestions LISREL gives according to theoretical considerations, such as paths from SAT to Perf, SAT to Descon, SQ to Descon. The suggested paths that might be reasonable are Exp to Desire (Chi-Square decreases 15), Perf to SAT (Chi-Square decreases 17),

Perf to SQ (Chi-Square decreases 43.7), Exp to SQ (Chi-Square decreases 43.7). LISREL also gives suggestions adding error covariance among observable variables; I choose to consider them after I consider the possibility of adding paths. To consider the possible 4 paths, Exp to Desires is not reasonable since Desires is what students want to have and Exp is what students expect to have. There is no obvious justification that Exp causes Desire. It is more possible that Desire influence Exp. If adding path from Exp to SQ, ChiSquare decreases 43.7 and the path estimate is 0.47. However, this is hard to justify, because following the path Exp-> SA->SAT, we can see the increase in Exp leads to the increase in SAT. However, following the path Exp->Expdis->SAT, the increase in Exp leads to the decrease in SAT. I think the positive relation between Exp and SAT as suggested by the adding path is not reasonable. So I choose not to add the path from Exp to SQ. The suggestion of adding two paths from Perf to SAT (Chi-Square decreases 17), Perf to SQ (ChiSquare decreases 43.7) seems reasonable. The two new path estimates are positive, which is consistent with the theoretical consideration. By adding these two path, we are testing the direct effect of Perf on SAT and SQ. I decided to first add path from Perf to SQ and then add path from Perf to SAT. After adding path from Perf to SQ, Minimum Fit Function Chi-Square = 120.57, d.f. 48. Output still suggests adding path from Perf to SAT. By doing this, Minimum Fit Function ChiSquare decreases to 101.91, d.f. 47. The modification suggestions based on the model with 47 df include adding covariance between Exp and Desires and error covariance between SQ and Perf and other suggestions which is not sound. After setting the covariance of Exp and Desires free, I got Minimum Fit Function ChiSquare = 87.12, df.46. After finishing this step, there is no other reasonable suggestions LISREL give out. So I think this is the relatively best improvement I can do. In sum, the modifications Ive done are adding paths from Perf to SAT , from Perf to SQ and free the covariance of Exp and Desire. The structural equations model shows that the amount of variations explained in SQ and SAT is 47% and 69% respectively, which is also improved.

You might also like