According to the 2005 Tannhaus formative questionnaire, not only should the correlation of one-variable questions and their divergence from other variable questions not be considered, like the reflective questionnaire, but one-variable formative questions should to check for multicollinearity. And this is exactly the opposite of the concept of Cronbach's alpha in reflective questionnaires. An indicator called VIF(variance inflation factor) examines this. Haier 2010 believes that if the VIF is less than 0.5 for each question, the multicollinearity is confirmed, and Tenenhaus believes that if this value is less than 0.2, the situation is excellent. Which do you consider?
Hello Sara,
The lower bound for a VIF value is 1, which would be the case if the target variable was uncorrelated with all other variables in the set. So, I don't see how VIF thresholds of 0.5 or 0.2 make any sense.
Could you perhaps be referring to tolerance instead in your query? Or is it the case that among the users of PLS-SEM, there is some ad hoc definition of VIF that is different from the usual, linear model definition?
VIF = 1 / (1 - R^2j) for variable j (= 1 / tolerance)
Tolerance = 1 - R^2j ( = 1 / VIF)
Good luck with your work.
Sara Hamidi, a formative construct including more than one indicator/item will directly correspond to a multiple regression model. Thus, you can use VIF thresholds (e.g., 2.5) typically used for multiple regression models.
Squared root of VIF value for an item tells us how many times the standard error of a coefficient/weight has increased as a result of its correlation with the remaining indicators/items in the equation. There is an inverse relation between tolerance and VIF values. A tolerance value of 0.4 corresponds directly to a VIF value of 2.5 (i.e., 1/0.4). As such, tolerance values above 0.4 and VIF values less than 2.5 are signs of no serious multicollinearity among the indicators/items.
Incidentally, you may have mixed up VIF with Tolerance.
Cronbach's alpha, calculated for dichotomies or variables that take only two values (for example, for true / false answers), is identical to the so-called Kuder-Richardson-20 formula for the reliability of summary scales. In either case, since reliability is actually calculated based on the consistency of all questions in the total scale, the reliability factor calculated in this way also refers to internally consistent reliability.
Removing an item using "alpha if item removed" can lead to "alpha inflation" when it is reported that reliability at the sample level is higher than reliability at the population level. It can also reduce reliability at the population level. The exclusion of less reliable items should be based not only on a statistical basis, but also on a theoretical and logical basis. It is also recommended to split the entire sample in two and cross-validate.
Regards, Sergey Viktorovich Pushkin
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS