This is a complex issue.  Classical measurement their is based on reflective measurement, namely, that an overarching construct predicts the factors that comprise it.  There are several books on how to validate and test classical reflective measurements.  For an overview, see DeVellis, 2011: Scale Development: Theory and Applications. Formative measurement is a relatively new construct and has not gained wide acceptance in the scholarly literature with high impact factors. In general, formative measurements suggest the opposite -- that some constructs are too complex to be made up of specific facets.  Rather, the selection of those facets dictates the construct' so, how you measure something determines what you measure.  This is highly controversial in some circles because it is virtually impossible to test the reliability or validity of formative measurement models.  They are typically demonstrated using path analysis.  Critics of this form of measurement suggest that you can come up with a group of pretty much anything and say it is a formative measurement because there is no way to test it.  See Diamantopoulos, Riefler & Roth, 2008 (pro) and Howell, Breivik & Wilcox, 2007 (con) for example.

 Mariana, 

the choice would be formulate "a priori". If you think that "Psychological Empowerment-PE" exists independently of your measurement approach, than you are within a "realistic/reflective" epistemic framework. Once you have decided this, than you must decide a priori if you aim at a measurement or, at a check-list. If you think that PA is "multidimensional" construct, then you cannot aim at a "multidimensional" measure. Measure is, in itself, unidimensional (the abstract concept "from less to more" of a unique variabile). Let' assume you aim at a high-order, yet unidimensional instrument (say: a cumulative questionnaire). How to test reflective vs formative modeling? This depends on the "gold standard" against which you are comparing the two altrnatives.  You may 

a) aim at estimating the internal consistency of the measure

and/or

b) the "external" validity: say the capacity to predict  events, the correlation with scores from an established standard or so.

In my opinion, both are necessary.

The former approach can not provide "demonstration" of superiority. You may acheive highly consistent scales (even with the demandino Rasch analysis) , yet your variabile is  a mental"construct" of yours: it is made by your items (in other words, it is a illusion). Intentionally researchers often "costruct" scales with highly correlated (internally consistent) items, thus creating a self-realising prophecy. Yet, constency may apply only to that particular sample of obesrvations. 

The long-term field application of the instrument will only reveal how robust it is, with respect to other samples, other raters, other cultural contexts, good predictions of relevant events etc. Only "real" variabile, with itms that just "reflect" reality, are stable.

Hence, before embarking in the adventure, you should reason a lot  "a priori", based on previous knowledge and experience; don't aim at "cosmetici statistics" confriming you pet hypotheses!

I'm attaching some short publications highlighting the issue (see also cited references).

Hope this helps

Luigi Tesio

Mariana,

Here is a paper you should consider reading to organize your paper's comparison and analysis of the two model approaches.

• Formative versus reflective measurement models: Two applications of formative measurement (T. Coltman,T. M. Devinney, D. F. Midgley, S. Venaik, 2008)
• See attached link to access the paper.

Table 1 in the paper summarizes and outlines specific theoretical and empirical considerations when assessing formative versus reflective models while providing relevant literature for support.

Additionally, you should make sure that you are also using construct measurement guidelines detailed in the following:

• Conventional wisdom on measurement: A structural equation perspective (Kenneth Bollen, Richard Lennox, 1991/9)
• See attached link to Google Scholar for Kenneth Bollen

Abstract: The applicability of 5 conventional guidelines for construct measurement is

critically examined:(1) Construct indicators should be internally consistent for valid

measures,(2) there are optimal magnitudes of correlations between items,(3) the validity of measures depends on the adequacy with which a specified domain is sampled,(4) within-construct correlations must be greater than between-construct correlations, and (5) linear composites of indicators can replace latent variables.

Now, my humble opinion ... you are welcome to discard :)

While the exercise of comparing formative and reflective models is interesting, your proposed topic strongly falls in the realm of reflective model design.  You can go into the exercise treating each model type as alternative approaches, but in the end, you should come away with better discernment as to why reflective models are a better fit in psychometric research.

I hope this helps,

rg

I think that you will find support for the opinion offered as you get into the project.  My read of the responses to your question by Lia Nower and Luigi Tesio is that they also have concerns with the formative approach ... and certainly many thanks to the time and effort on their responses, given such a broad and detailed topic.  Good luck!

http://ro.uow.edu.au/cgi/viewcontent.cgi?article=1693&context=infopapers

http://scholar.google.com/citations?hl=en&user=eY3VhFcAAAAJ

The difference is Theoretical, if you assume that your construct is reflective it cant be formative. Is important not to rely only on statistics, but in the theoretical assumptions underlying the construct.

Just to add to all the great comments already... You might want to look at your construct in terms of orders.  there are models that have first order formative, second order reflective etc. and vice versa.  This would allow you to view the construct from both perspectives if applicable.  The attached article demonstrates this perspective.

Thank all of you for your answers!

You might be interested to look at a theoretically-driven approach to the use of formative measurement (in application to developmental risk)  explored in the paper below.  There's a gentle introduction and you can follow the top-down line of reasoning that justifies it in this instance: https://www.researchgate.net/publication/44610249_Measuring_the_combined_risk_to_young_children's_cognitive_development_An_alternative_to_cumulative_indices

Article Measuring the combined risk to young children's cognitive de...

Hi,

In case of FORMATIVE measurement models, could I calculate R-Squared for structural model and GoF for overal model?