A well established questionnaire measuring interpersonal relations is modified to measure interorganizational relations. 11 parameters of the relation are determined by 4 items each. Items are removed until the best possible $\alpha$ per scale is achieved (.58 – .81). This results in some scales with 3 items, some scale with 4 items.
A confirmatory factor analysis is applied to calculate the estimates of the items and verify the internal structure of the questionnaire.
As expected with some $\alpha$ as low as .58 and .61 the CFA does not fit very well. Since it is "kind of" a new scale, is it ok to use the predictions for further modelling and report them?
The fit indices are (bad):
Number of observations 251
Estimator ML Robust
Minimum Function Test Statistic 2475.330 2029.585
Degrees of freedom 854 854
P-value (Chi-square) 0.000 0.000
Scaling correction factor 1.220
for the Satorra-Bentler correction
Model test baseline model:
Minimum Function Test Statistic 5718.372 4583.881
Degrees of freedom 946 946
P-value 0.000 0.000
Full model versus baseline model:
Comparative Fit Index (CFI) 0.660 0.677
Tucker-Lewis Index (TLI) 0.624 0.642
Loglikelihood and Information Criteria:
Loglikelihood user model (H0) -18173.075 -18173.075
Loglikelihood unrestricted model (H1) -16935.410 -16935.410
Number of free parameters 180 180
Akaike (AIC) 36706.150 36706.150
Bayesian (BIC) 37340.731 37340.731
Sample-size adjusted Bayesian (BIC) 36770.110 36770.110
Root Mean Square Error of Approximation:
RMSEA 0.087 0.074
90 Percent Confidence Interval 0.083 0.091 0.070 0.078
P-value RMSEA <= 0.05 0.000 0.000
Standardized Root Mean Square Residual:
SRMR 0.146 0.146
Edit:
The real question is probably not only whether to continue with estimates based of a bad fit, but rather whether the scale is new or not.
Edit2:
Barrett (2007) recommends to not rely on "approximate fit indices" at all.
The criterion used for "fit" is actually an abstract concept in the
majority of SEM models. It is clearly not predictive accuracy. In fact
whether models "approximately fit" with an RMSEA of 0.05 or 0.07 is
a literally meaningless scientific statement.
Bibliography
Barrett, P. Structural equation modelling: Adjudging model fit. Personality and Individual Differences, 2007, 42, 815-824 [PDF]
Best Answer
Some immediate comments:
1) I would not predict from this model, since it has bad fit (CFI should be >.95 and RMSEA <.05).
2) It appears you are exploring and confirming on the same data set. This is capitalizing on chance and generally not recommended. It is surprising that your CFA still has bad fit. If you explore on the same data, CFA fit often is much better.
3) You could try exploratory FA or stepwise CFA with modification indices (Lagrange multipliers) to find a better fitting model. I would do this instead of relying on a bad index like $\alpha$ (it is not a good estimator of scale reliability, in fact it is just one estimator, which may be too optimistic about reliability).
4) In any way, I would recommend cross-validating your model. This means splitting your data set in two, explore on one (EFA, modification idices), and confirm (CFA) on the other. This way you avoid capitalizing on chance in scale exploration.