What about new developed constructs in PLS?
Posted: Fri Feb 12, 2016 8:35 am
Dear all,
I have an important and urgent question related to validating new constructs in the PLS context, and I really appreciated it if any one can assist me. Please provide me your comments and answers to my questions. Your assistance is highly appreciated.
I have read a lot about (1) exploratory factor analysis (EFA), including both common factor analysis and principle component analysis (PCA) and (2) confirmatory factor analysis (CFA).
What I understood from my intensive reading can be condensed into the following points, and please correct me if I am mistaken:
1. Researchers should use EFA before CFA if they have developed a new construct, which has not been tested by prior literature. (What I mean by a new construct is that although researchers might derive the construct from prior theoretical literature, the new construct has not been used by empirical studies in the way in which it was developed by researchers.) AM I right?
2. After exploring the structure of the new construct by using EFA, researchers should use CFA to confirm and validate the new construct with a new sample or part of their sample that has not been included when using the EFA test (i.e., half of the sample is used for EFA, and the another half is used for CFA). (This means that researchers should not conduct both EFA and CFA with the whole sample.) AM I right?
3. As all PLS constructs (reflective and formative) are composites, principle component analysis (PCA) is the applicable version of EFA. (This means that researchers should use PCA instead of common factor analysis in SPSS or other programs when evaluating the new construct before entering it into the measurement model (i.e., CFA in PLS). The new construct should be entered into the measurement model (CFA in PLS) based on the results of the PCA.) AM I right?
My case is:
1. I have developed a new construct, PC with 12 items.
2. I have derived this construct from prior theoretical literature. There is no study that has used the same combination of items that I have used to measure PC.
3. The construct is formative (i.e., most of the items are not interchangeable).
4. It is possible that the construct can be modeled as a second order construct with three or four first order constructs, but there is no strong theory that suggests which items should belong to each of the first order constructs. To be honest, I conducted PCA in SPSS, and found that PC has four factors (based on the Eigenvalue above 1 criterion) or one factor (based on Parallel analysis and MAP test criteria). It is difficult, however, to determine the number of factors based on the scree plot criterion (in particular, the one factor and four factors solutions are both possible).
I have tested both solutions (one factor solution and four factors solutions) in PLS, however, the results are not very good in relation to validating the first order formative constructs and the second order formative construct (i.e., PC). There were some instances of negative weights and weights above 1. I had to use Mode A or equal weights to solve these problems.
5. My sample size is 200, which means that it is difficult to divide it by 2 to run EFA for one half and CFA for the another half.
- Now, what is the best course of action regarding modeling PC in the measurement model in PLS?
Action 1: Should I conduct PCA before PLS and then using the PCA results to model PC in the measurement model?
This alternative is based on the fact that:
1. PC is a new construct.
2. PC is formative.
However, my sample size is relatively small to conduct both PCA in SPSS and CFA (in PLS).
Action 2: Should I conduct CFA only?
However, how can I know whether PC is a first order construct or second order construct since it is a new developed construct?
Should I model it in the model based on my own judgment and logical reasons?
I have an important and urgent question related to validating new constructs in the PLS context, and I really appreciated it if any one can assist me. Please provide me your comments and answers to my questions. Your assistance is highly appreciated.
I have read a lot about (1) exploratory factor analysis (EFA), including both common factor analysis and principle component analysis (PCA) and (2) confirmatory factor analysis (CFA).
What I understood from my intensive reading can be condensed into the following points, and please correct me if I am mistaken:
1. Researchers should use EFA before CFA if they have developed a new construct, which has not been tested by prior literature. (What I mean by a new construct is that although researchers might derive the construct from prior theoretical literature, the new construct has not been used by empirical studies in the way in which it was developed by researchers.) AM I right?
2. After exploring the structure of the new construct by using EFA, researchers should use CFA to confirm and validate the new construct with a new sample or part of their sample that has not been included when using the EFA test (i.e., half of the sample is used for EFA, and the another half is used for CFA). (This means that researchers should not conduct both EFA and CFA with the whole sample.) AM I right?
3. As all PLS constructs (reflective and formative) are composites, principle component analysis (PCA) is the applicable version of EFA. (This means that researchers should use PCA instead of common factor analysis in SPSS or other programs when evaluating the new construct before entering it into the measurement model (i.e., CFA in PLS). The new construct should be entered into the measurement model (CFA in PLS) based on the results of the PCA.) AM I right?
My case is:
1. I have developed a new construct, PC with 12 items.
2. I have derived this construct from prior theoretical literature. There is no study that has used the same combination of items that I have used to measure PC.
3. The construct is formative (i.e., most of the items are not interchangeable).
4. It is possible that the construct can be modeled as a second order construct with three or four first order constructs, but there is no strong theory that suggests which items should belong to each of the first order constructs. To be honest, I conducted PCA in SPSS, and found that PC has four factors (based on the Eigenvalue above 1 criterion) or one factor (based on Parallel analysis and MAP test criteria). It is difficult, however, to determine the number of factors based on the scree plot criterion (in particular, the one factor and four factors solutions are both possible).
I have tested both solutions (one factor solution and four factors solutions) in PLS, however, the results are not very good in relation to validating the first order formative constructs and the second order formative construct (i.e., PC). There were some instances of negative weights and weights above 1. I had to use Mode A or equal weights to solve these problems.
5. My sample size is 200, which means that it is difficult to divide it by 2 to run EFA for one half and CFA for the another half.
- Now, what is the best course of action regarding modeling PC in the measurement model in PLS?
Action 1: Should I conduct PCA before PLS and then using the PCA results to model PC in the measurement model?
This alternative is based on the fact that:
1. PC is a new construct.
2. PC is formative.
However, my sample size is relatively small to conduct both PCA in SPSS and CFA (in PLS).
Action 2: Should I conduct CFA only?
However, how can I know whether PC is a first order construct or second order construct since it is a new developed construct?
Should I model it in the model based on my own judgment and logical reasons?