Missing value treatment w/ casewise replacement (=deletion)

[Johannes Kotte]

Dear fellow researchers, dear SmartPLS developers

the data set I am using contains quite a bit of missing values. I read that there a re two ways of dealing with it - mean replacement and casewise replacement - and that casewise replacement really is casewise deletion. In case of casewise deletion, how exactly does it work in SmartPLS? I see 2 ways to deal with missing values:

Way 1) Observations with one or more missing values are completely eliminated from the dataset. In cases like mine, this would drive down sample size tremendously.

Way 2) Observations with missing values are excluded from the calculation of a construct/variable, if there are missing values for this construct. But they are used for calculating other variables, where no values are missing. In other words: observations are being used for all constructs where the data is complete (no missing values) and are being ignored for constructs where the data is not complete.

As to my understanding, the PLS-Algorithm iteratively runs through the model step by step, i.e. it first calculates the scores of the constructs in the outer model (one by one) and then calculates the scores of and the relationships between the variables in the inner model. So I would presume that SmartPLS uses way 2) when using casewise deletion. Is that correct? If not, how does SmartPLS handle the problem?

Thanks in advance for your help!
Johannes

PS: Sorry for cross-posting (viewtopic.php?t=360) but the other thread was already pretty old so I figured it might be better to open a new one.

[Hengkov]

Hi,
I also interest this issue.
Using simulation study with SmartPLS, Parwoll and Wagner (2012, p. 543) conclusion mean replacement leads to more reliable result than casewise deletion. For futher readings, I recommend you check some books below:

Alison, P. 2002. Missing data. Sage publications.
Enders, C.K. 2010. Applied missing data analysis. Guilford Press.
Little, R. J., and Rubin. 1987. Statistical analysis with missing data. John Wiley & Sons.
Parwoll, M., and Wagner. R. 2012. "The impact of missing values on PLS model fitting" in Gaul, W et al., (Eds.) Challenges at the interface of data analysis, Spinger.

Best Regards,

[Johannes Kotte]

Dear Hengky,

in the meantime I found out that SmartPLS deletes the observations completely [alternative 1)] . So I guess there's still room for improvement :-D

In their book "A Primer on Partial Least Squares Structural Equation Modeling (PLS-SEM)" Hair et al. (2013) provide recommendations for dealing with missing values: eliminate all observations with >15% missing values in all the indicators used. Afterwards check the percentage of missing values per indicator (not per observation as in the first step). If there are indicators with >5% missing values, use casewise replacement (observation is deleted completely), otherwise mean replacement


Best
Johannes

Book: http://www.sagepub.com/books/Book237345 ... uctsSearch

[Hengkov]

Hi,

OK. I think this issue need future investigation, because not easy to understand. B-)
Btw, in WarpPLS not option for dealing missing values and I test many date never found missing. ;-)

Regards,