Different postings I have read seem to answer the questions of whether or not the following can be entered into PLS path models and/or if this can be done in SmartPLS:
Dichotomous categorical variables (both as IVs and DVs)
Nominal categorical variables (both as IVs and DVs)
Ordinal variables (both as IVs and DVs)
Can each be used in PLS path modeling? Can they be entered as would continuous variables in SmartPLS, if so? How should they be handled, if not?
I would much appreciate someone either answering these questions or directing me to a resource/resources where the correct answer to these questions already exists.
Dichotomous, Nominal, and Ordinal Variables
Dichotomous, Nominal, and Ordinal Variables
Thanks!
Anthony
Anthony
Dichotomous/binary variables
Hi Anthony,
I have similar questions and already some answers/hints. Maybe we can get all answers by merging our knowledge and receiving input from others as well.
In my model there are 3 independent constructs:
- one with (6) dichotomous indicators
- one with (3) metrically scaled indicators
- one with (3) ordinally and (2) metrically scaled indicators (could be generalized to 5 ordinal-scaled indicators)
... and 2 dependent constructs:
- one with (3) metrically scaled indicators
- one with (3) ordinally scaled indicators
For the dichotomous case:
According to Temme/Kreis/Hildebrandt (2005, PLS Path Modeling - A Software Review, p. 7): "... All programs at present expect that the indicators of the latent variables are continuous, or – for instance in the case of rating scales with 5 or more answer categories – approximate a continuous scale. In addition, binary exogenous variables can be included in the analysis ..."
Consequently, SmartPLS should be able to cope with binary (independent/exogenous) indicator variables. To my judgment, however, 2 questions remain:
1) Can binary indicators be mixed with other scale types within one (construct) measurement model? (Irrelevant for my model at the moment, but still important to know)
2) Is there any further adjustment (e.g. to the PLS algorithm) necessary in SmartPLS for including binary indicators? Or is okay to calculate the raw data right away?
I'd also be grateful, if somebody could reply to our questions! (For other scale types, please see my next post.)
Cheers,
Jörg
I have similar questions and already some answers/hints. Maybe we can get all answers by merging our knowledge and receiving input from others as well.
In my model there are 3 independent constructs:
- one with (6) dichotomous indicators
- one with (3) metrically scaled indicators
- one with (3) ordinally and (2) metrically scaled indicators (could be generalized to 5 ordinal-scaled indicators)
... and 2 dependent constructs:
- one with (3) metrically scaled indicators
- one with (3) ordinally scaled indicators
For the dichotomous case:
According to Temme/Kreis/Hildebrandt (2005, PLS Path Modeling - A Software Review, p. 7): "... All programs at present expect that the indicators of the latent variables are continuous, or – for instance in the case of rating scales with 5 or more answer categories – approximate a continuous scale. In addition, binary exogenous variables can be included in the analysis ..."
Consequently, SmartPLS should be able to cope with binary (independent/exogenous) indicator variables. To my judgment, however, 2 questions remain:
1) Can binary indicators be mixed with other scale types within one (construct) measurement model? (Irrelevant for my model at the moment, but still important to know)
2) Is there any further adjustment (e.g. to the PLS algorithm) necessary in SmartPLS for including binary indicators? Or is okay to calculate the raw data right away?
I'd also be grateful, if somebody could reply to our questions! (For other scale types, please see my next post.)
Cheers,
Jörg
Ordinal variables
Hi Anthony,
Based on my previous post I want to share my info regarding ordinal (categorical) indicator variables:
I don't know if such indicator data can be used in SmartPLS, but most probably a plug-in (if it exists) is needed. I'm quite sure that the standard SmartPLS algorithm is not sufficient for receiving optimal results in this case, because a precondition/assumption of the standard algorithm is to use at least interval-scaled data.
However, there is an article by Jörg Betzin ("PLS-Pfadmodellierung mit kategorialen Daten"; in: "Handbuch PLS-Pfadmodellierung" by Bliemel/Eggert/Fassott/Henseler, 2005) which describes a way to include categorical data. (So far I couldn't find any online source of the article.) Betzin suggests using Correspondence Analysis (CorA) instead of Principal Component Analysis (PCA) in order to adapt the PLS algorithm (the outer approximation part for estimating measurement models) for categorical data. According to his introduction PCA strives to minimize a loss function which is only sensible in the case of metrically scaled data. CorA could be an option (1) to transform categorical data in such a way that (2) the loss function can still be applied. The goal is to transform the categorical data for calculating a correlation matrix. In analogy to PCA principal components are then extracted from that matrix.
He mentions the SPSS procedure "HOMALS" (available via SPSS add-in "Categories") which can be used to conduct CorA. However, even if one calculates CorA outside of SmartPLS, how to import it into SmartPLS? As far as I know, SmartPLS "only" accepts raw data for import, and no correlation matrix, for instance.
In my opinion, there are 2 possible ways for including ordinal-scaled data:
A) SmartPLS allows for such data by the aid of a plug-in --> Does it exist? If yes, where is it available from?
B) Calculate scores for such indicators outside of SmartPLS (e.g. by CorA in SPSS, see above) and find a way to import the results into SmartPLS where it can be merged with other data in the model. --> Is there a proven way to do that? Is it allowed to mix independent constructs (with ordinal-scaled indicators) with other types of independent constructs in the same model?
I strongly assume that we are not the only ones dealing with this challenge. I'd be grateful for any hint in this area!
(For nominal variables I have no information, because I haven't considered this scenario so far.)
Cheers,
Jörg
Based on my previous post I want to share my info regarding ordinal (categorical) indicator variables:
I don't know if such indicator data can be used in SmartPLS, but most probably a plug-in (if it exists) is needed. I'm quite sure that the standard SmartPLS algorithm is not sufficient for receiving optimal results in this case, because a precondition/assumption of the standard algorithm is to use at least interval-scaled data.
However, there is an article by Jörg Betzin ("PLS-Pfadmodellierung mit kategorialen Daten"; in: "Handbuch PLS-Pfadmodellierung" by Bliemel/Eggert/Fassott/Henseler, 2005) which describes a way to include categorical data. (So far I couldn't find any online source of the article.) Betzin suggests using Correspondence Analysis (CorA) instead of Principal Component Analysis (PCA) in order to adapt the PLS algorithm (the outer approximation part for estimating measurement models) for categorical data. According to his introduction PCA strives to minimize a loss function which is only sensible in the case of metrically scaled data. CorA could be an option (1) to transform categorical data in such a way that (2) the loss function can still be applied. The goal is to transform the categorical data for calculating a correlation matrix. In analogy to PCA principal components are then extracted from that matrix.
He mentions the SPSS procedure "HOMALS" (available via SPSS add-in "Categories") which can be used to conduct CorA. However, even if one calculates CorA outside of SmartPLS, how to import it into SmartPLS? As far as I know, SmartPLS "only" accepts raw data for import, and no correlation matrix, for instance.
In my opinion, there are 2 possible ways for including ordinal-scaled data:
A) SmartPLS allows for such data by the aid of a plug-in --> Does it exist? If yes, where is it available from?
B) Calculate scores for such indicators outside of SmartPLS (e.g. by CorA in SPSS, see above) and find a way to import the results into SmartPLS where it can be merged with other data in the model. --> Is there a proven way to do that? Is it allowed to mix independent constructs (with ordinal-scaled indicators) with other types of independent constructs in the same model?
I strongly assume that we are not the only ones dealing with this challenge. I'd be grateful for any hint in this area!
(For nominal variables I have no information, because I haven't considered this scenario so far.)
Cheers,
Jörg