Due to missing data, I did imputation in order to access bootstrapping in AMOS .Now I have 3 different model indices fit and outputs for the same data as shown below. Which one do you think is accurate? What can I do for the data?
1-The model fit indices before imputation are CMIN/DF: 1.743; NFI: .972; RFI: .860; IFI: .988; TLI: .935; CFI: .987 & RMSEA: .061 (PCLOSE: .296). P value in regression weights in most variables are not significant. Most of hypotheses are not supported in the study.
2-The model fit indices after imputation CMIN/DF: 14.786; NFI: .974; RFI: .906; IFI: .976; TLI: .912; CFI: .976 & RMSEA: .083 (PCLOSE: .000). P in regression weights in most variables were significant. Also most of hypotheses are supported. The result are well and in the line of literature review, but the problem is related to high CMIN/DF: 14.786 and RMSEA: .083.
3- The model fit indices after imputation without control variables (age and tenure didn’t mention in the model this time) are: CMIN/DF: 4.221; NFI: .996; RFI: .971; IFI: .997; TLI: .978; CFI: .997 & RMSEA: .040 (PCLOSE: .795). P in regression weights in most variables are not significant. Most of hypotheses are not supported.
Note: (Total number of questionnaires are 200. The study consists of six Observed, endogenous variables; five Observed, exogenous variables (two of these variables refer to Age and Tenure as control variables) and six (e) unobserved variables).
Thanks a lot for your help
Shaghayegh Ghaffaripour Before reaching on a conclusion, you need to specify how many missing values were present.
Also as far as your data size is concerned, although at many places, 200 sample size seems to be ok for SEM, but many researchers argue that the data size should be close to a ratio of 1 item to 10 observation.
@ Shaghayegh Ghaffaripour ,
Here, the first point is :
" how many missing data have been imputed is to be mentioned".
As this number increases, the accuracy of the results decreases. Thus this point is very important one.
The second point is;
" the forms of the three fitted regression models are to be mentioned clearly ".
Right now, it is possible to comment that the data size (200) is reasonable for the validity of the findings of the analysis provided the number of missing data imputed is not large.
The P value can be different for different models. Conclusion on the accuracy of a model can be drawn on the basis of the significance of the P value.
The issue is with the sample data. If you have imputed all missing data you may end up with wrong results. You need to work on that and make sure that the imputation is done for those missing data worth manipulating and others may be rejected.
Model -1 provides a correct CMIN/DF < 3.0 but should be insignificant (>0.05). If your GoFIs are satisfactory within threshold fit indices, the p-values can be tested with Boolenstine p-values for more than 0.05.
Also, try with your observed values for the test of hypotheses then add the control variable in second stage.
Hope it helps.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression