After running the PLS algorithm, I find that for one construct, I have to delete two items in order for the AVE to go above 0.5.
Originally, the construct contain five items. After deleting two items, it remains with only three items.
Is it ok to delete two items? Is there any rule of thumb on the maximum number of items you can delete for a construct?
Is it ok to delete 2 items for one construct with 5 items?
Hello,
How are you?
Actually you can use only one single item for a reflective construct (read Hair et al., 2014), but you wont have AVE and composite reliability.
So, if you improve your AVE (>0,50) and composite reliability (>0,708) deleting those two indicators, it´s good.
In my investigation i used 2 reflective indicators...
i dont remember where i read that 3 indicators is good.
Deleting formative indicators is the problem.
Hope this helped someone...
How are you?
Actually you can use only one single item for a reflective construct (read Hair et al., 2014), but you wont have AVE and composite reliability.
So, if you improve your AVE (>0,50) and composite reliability (>0,708) deleting those two indicators, it´s good.
In my investigation i used 2 reflective indicators...
i dont remember where i read that 3 indicators is good.
Deleting formative indicators is the problem.
Hope this helped someone...
-
bmckinley1
- PLS Junior User
- Posts: 8
- Joined: Thu Aug 07, 2014 3:32 pm
- Real name and title: Bryan McKinley, PhD candidate (effective agile teams) @ Capella University, Professional Cloud-based Software Developer
Re: Is it ok to delete 2 items for one construct with 5 item
I've seen several articles (PRed) ignore the second-order AVE for reflective-reflective models when using SmartPLS. Instead, they manually calculate the average of squared loadings (normalized) for a repeated indicators approach.
For example, if you have one second-order construct (ignoring it's large nomological net for brevity) with three first-order reflective dimensions (see Edwards or Wetzels et al.), you have a reflective-reflective model.
If you have 5 indicators per first-order construct, then 15 second-order indicators are reused or repeated in the second-order construct. In my situation, the average of squared loadings of the second-order construct is greater than the AVE reported by SmartPLS3. Why the difference? How does SmartPLS3 calculate AVE in the reflective-reflective situation. The difference is substantial in terms of the number of indicators retained in my measurement model.
Here's an MIS Quarterly snippet that refers to the AVE calculation (MacKenzie et al., 2011): Note: Hair's MV statistics texts are often cited in other PRed contexts.
For second-order latent constructs with reflective indicators
(see Figure 3, Panel C), it is also important to examine the
convergent validity of the set of first-order sub-dimensions as
reflective indicators of the second-order construct. As noted
by Edwards (2001, p. 163), this can be assessed using the
multivariate coefficient of determination, R2...(For an equivalent formula for R2
m, see Edwards 2003.)
Alternatively, AVE could be calculated for the second-order
construct by averaging the squared multiple correlations for
the first-order sub-dimensions (or averaging the square of
each sub-dimension’s completely standardized loading on the
second-order construct). In either case, values greater than
.50 mean that, on average, a majority of the variance in the
first-order sub-dimensions is shared with the second-order
latent construct. (p. 313)
MacKenzie, S. B., Podsakoff, P. M., & Podsakoff, N. P. (2011). Construct measurement and validation procedures in MIS and behavioral research: Integrating new and existing techniques. MIS Quarterly, 35(2), 293–334.
For example, if you have one second-order construct (ignoring it's large nomological net for brevity) with three first-order reflective dimensions (see Edwards or Wetzels et al.), you have a reflective-reflective model.
If you have 5 indicators per first-order construct, then 15 second-order indicators are reused or repeated in the second-order construct. In my situation, the average of squared loadings of the second-order construct is greater than the AVE reported by SmartPLS3. Why the difference? How does SmartPLS3 calculate AVE in the reflective-reflective situation. The difference is substantial in terms of the number of indicators retained in my measurement model.
Here's an MIS Quarterly snippet that refers to the AVE calculation (MacKenzie et al., 2011): Note: Hair's MV statistics texts are often cited in other PRed contexts.
For second-order latent constructs with reflective indicators
(see Figure 3, Panel C), it is also important to examine the
convergent validity of the set of first-order sub-dimensions as
reflective indicators of the second-order construct. As noted
by Edwards (2001, p. 163), this can be assessed using the
multivariate coefficient of determination, R2...(For an equivalent formula for R2
m, see Edwards 2003.)
Alternatively, AVE could be calculated for the second-order
construct by averaging the squared multiple correlations for
the first-order sub-dimensions (or averaging the square of
each sub-dimension’s completely standardized loading on the
second-order construct). In either case, values greater than
.50 mean that, on average, a majority of the variance in the
first-order sub-dimensions is shared with the second-order
latent construct. (p. 313)
MacKenzie, S. B., Podsakoff, P. M., & Podsakoff, N. P. (2011). Construct measurement and validation procedures in MIS and behavioral research: Integrating new and existing techniques. MIS Quarterly, 35(2), 293–334.