BSEM actually relax a little the requirements for a good fit, especially if you have many small error correlations (typical if you use latent variables with a lot of indicators, as it's usually the case for questionnaires with many items).  

If you are getting a bad fit, then, and the syntax is correctly specified, the model quite simply doesn't fit the data.

Instead of torturing the data to get a good fit at all costs, turn back to your research questions: what important variables may be missing in your model? Are you using the right indicators for your latents? Are there paths were you expected an effect and that are actually indignificant? Are there anomalies in the data which might have prejudicated the result (e.g., missingness not at random)? Are there possible alternative explanations you can think of? 

As much as I sympathize with the despair you feel when your model doesn't fit, the solution is not to "get around" what the data is telling you. You cannot ignore the fit indices, because they are actually telling you that your research hypothesis is probably incorrect - or the model is misspecified. So I think the right action, in this case, is to turn back and consider the theory you started with and look for alternative explanations or ignored variables. That's what's at the heart of research, I think: looking for truth. As frustrating as it may sometimes be :)

BSEM actually relax a little the requirements for a good fit, especially if you have many small error correlations (typical if you use latent variables with a lot of indicators, as it's usually the case for questionnaires with many items).  

If you are getting a bad fit, then, and the syntax is correctly specified, the model quite simply doesn't fit the data.

Instead of torturing the data to get a good fit at all costs, turn back to your research questions: what important variables may be missing in your model? Are you using the right indicators for your latents? Are there paths were you expected an effect and that are actually indignificant? Are there anomalies in the data which might have prejudicated the result (e.g., missingness not at random)? Are there possible alternative explanations you can think of? 

As much as I sympathize with the despair you feel when your model doesn't fit, the solution is not to "get around" what the data is telling you. You cannot ignore the fit indices, because they are actually telling you that your research hypothesis is probably incorrect - or the model is misspecified. So I think the right action, in this case, is to turn back and consider the theory you started with and look for alternative explanations or ignored variables. That's what's at the heart of research, I think: looking for truth. As frustrating as it may sometimes be :)

Hello Ali.

As Marcello answered your question in full, just a small addition. On the one hand you are trying to prove the association but unfortunately on the other hand you need to acknowledge that what you have is something you should not play around with as you will be deviating from scientific, honest approach.

The only one fit index you may ignore (but still need to be reported) is chi-square.

Other fit indices such as AIC, BIC, RSMR, CFI, and TLI should be treated seriously.

Perhaps you may be able to develop alternative hypotheses to be tested and consequently achieve better model fit. Even then it should not be a "fishing expedition" to find satisfactory fit indices but well established ideas supported in past research.

All the best with your research.

You should deal with missing values before making new analysis of your data because they can cause several problems, especially if you are trying to use a program like AMOS. Maybe, you should visit this website http://statwiki.kolobkreations.com/wiki/Data_screening

Thank you Macerello and Witold, the point is all references mention that we can never get good chi square fit with very large sample so what is the alternative?

Thank you Silvia, I did multiple imutation but still the fit is bad