Dropping indicator(s) in the formative model would change the meaning of the construct, since each indicator variable represents a dimension of the construct in a formative relationship. Dropping a dimension just because it does not load well (“loading” for reflective; “weights” for formative) may have other consequences, because formative indicators are assumed to be exhaustive and definitive in defining a construct. Thus, indicators with low weights are not dropped even if they are insignificant (page 103, reference 1). In fact, model fit assessment in PLS-SEM differs for formative models as compared to reflective models. For example, using composite reliability, Cronbach’s alpha, and AVE for assessing fit of formative models is not appropriate. Hair et al. (2012) mentioned this to be a widespread error in reporting PLS-SEM models. In addition, high multicollinearity among the indicators in formative models is highly unusual as the indicators represent a complete set of the separate dimensions which compose the factor in a formative relationship. Thus the dimensions are not expected to correlate highly. (IF they do, it’s also ok as long as literature supports so). For details on how to check measurement fit for formative models, please see pages 73-79 of reference # 1. Hope it helps.

References:

1. Garson, G. D. (2016). Partial Least Squares: Regression & Structural Equation Models. Asheboro: Statistical Associates Publishing.

2. Hair, J. F., Sarstedt, M., Pieper, T. M., & Ringle, C. M. (2012). The Use of Partial Least Squares Structural Equation Modeling in Strategic Management Research: A Review of Past Practices and Recommendations for Future Applications. Long Range Planning, 45(5–6), 320–340. http://doi.org/10.1016/j.lrp.2012.09.008

I agree with Dr. Rahman. You can do data cleaning to eliminate response based errors.

Thank you Dr. Tsiotsou, Dr. Rahman and Dr. Tiwari for your inputs.