What if I Get an Error Message in the Analysis That Says “Iteration Limit Reached”?

The default iteration limit in the analysis function of AMOS is 50. If you get an error message that says the iteration limit has been reached and the analysis did not run to completion, it usually means that you are having an issue with the error term of an indicator or construct.You can increase the iteration limit to see if the analysis will run to completion.To do so, you need to go into the Analysis Properties function  and then go to the “Numerical” tab at the top of the pop-up window. In the iteration limit option, you can change this to 1,000 and then run the analysis again. Even if the iteration limit was not reached, more than likely you still have a problem in the model. Often when you have a model that reaches an iteration limit of 50, increasing the iteration to 5,000 will still not fix the problem. If you receive an error mes- sage about the iteration limit, it is a good idea to go into the output and look at the “Notes for model” link.This will usually give you a good idea where the problem is in your model. If you have a negative error variance, this can cause the iteration limit message to appear.Your best bet is to find the specific indicator or construct that is having issues and remove error covari- ances or remove the indicator all together.

In My Opinion: Covariance-Based SEM vs. Variance-Based SEM

Variance-based or partial least squares SEM (PLS-SEM) has become more popular because of its flexibility and potential problem-solving capabilities to address the chal- lenges that are present with covariance-based SEM. For instance, PLS-SEM can be used with small samples, uses weighted scores for greater explained variance, can handle non-normal data, and is best used for “exploratory” research. While at face value this seems like a great alternative to covariance-based SEM, if you look deeper, PLS-SEM is a problematic alternative. First, PLS-SEM does not calculate any model fit statistics. One of the primary advantages to SEM is the ability to assess a whole structural model. This is done via the fit statistics that compare the estimated covariance matrix to the observed covariance matrix. In PLS-SEM, model fit is ignored, which means model misspecification may be present, and it is simply not being assessed. This seems espe- cially reckless with large and complex models.

Second, PLS-SEM is presented as an especially valuable technique for exploratory research compared to covariance-based SEM, which is confirmatory in nature (Peng and Lai 2012). Before a study takes place, the researcher should a priori have an idea of how the constructs should influence one another and what indicators are needed to capture the constructs.The theory used by the researcher should be the basis of the hypotheses between constructs. In essence, a study should either support or fail to support the hypotheses that were outlined a priori. This should be consistent whether you are using PLS-SEM or covariance-based SEM. PLS-SEM has been used as an “exploratory” attempt to understand a model but has been often categorized as data driven modeling (Ronkko et al. 2016) where relationships are simply added or subtracted based on the data and not theory. This data driven approach has the possibility of capitalizing on chance and would be advisable only if a second sample could confirm the findings of the first model. Third, PLS-SEM allows the researcher to vary the weights of indicators in a construct.

This is done to increase the reliability of the construct, but recent research has found that it has no relative impact on a construct’s reliability and often inflates correlations between constructs (Ronkko et al. 2016). Fourth, PLS-SEM is lauded as a program that will work with small samples compared to covariance-based SEM that requires a larger sample. The appropriate sample size for a SEM model should be based on the complex- ity of the model and not the software. Sample size should be based on the necessary power to find an effect. For those looking for more information on this topic, McQuitty (2004) does a good job of outlining the necessary sample size needed based on complex- ity of the model. A simple model with few degrees of freedom could be grossly under- powered with a small sample. A small sample can bias results with a covariance-based SEM approach, but previous research has found the PLS-SEM was just as problematic in regards to a small sample size bias (Chumney 2013; Reinartz et al. 2009).

Fifth, PLS-SEM uses Ordinary Least Squares (OLS) regression, which has the same assumptions about normality of data as covariance-based SEM. Previous research has even noted that PLS has no more benefits of handling non-normal data than other covariance- based programs (Dijkstra 2015). Lastly and most concerning, PLS-SEM is widely known to produce inconsistent and biased estimations (Ronkko et al. 2016).The research by Reinartz et al. (2009) found that estimations were biased from 6% to 19% based on the strength of the path estimate. Overall, the evidence against using PLS-SEM is strong and convincing. In my opinion, I would encourage you to stick to a covariance-based SEM approach.

Source: Thakkar, J.J. (2020). “Procedural Steps in Structural Equation Modelling”. In: Structural Equation Modelling. Studies in Systems, Decision and Control, vol 285. Springer, Singapore.