Waqar--Are those standardized or unstandardized estimates in your path diagram? And, I am assuming that you used a fixed factor variance strategy to identify your CFA model, right?

The measurement model is wrong. the haywood cases have happened when your factor liading is greater than 1.0. Another issue is negative factor loading. You need to do sone reverse coding.

No offense meant here but are you sure path analysis is the choice technique for your research question. I have found that chemical reactions are. mostly deterministic. If you would give your research question this might help other chemists to better understand your approach. Best wishes, David Booth

BTW I realize that Sewall Wright invented this method for mammalian genetic analysis. But.he approached that from the position that genes assort.independantly, one of Mendel's laws not as a molecular biologist. If am confused please advise me on where I went wrong. Thanks very much, David Booth

Thank you for replying.

Sarah D. Newton yes these are standardized estimates. I used eigen value criteria as well as rotated factor loading, and removed those factors which cross loaded. I wanted to use 3 factor loadings however I was losing important parameters. Then i decided to use 4.

I am new to SEM and I am not sure if i am doing it right. My data is continuous and I scaled all parameters between 0.1 - 0.9.

Initially, I wanted to keep my 3 domains separate (physical, chemical and topographical properties). However, while I performed PCA, all independant variables in 3 domains got mixed. So, I am thinking to do a path analysis instead of EFA and CFA.

My hypothesis is "topography affects soil physical and chemical parameters which in turn impact N2O emissions". So, I want to calculate direct and indirect impact of all parameters within these 3 domains using path analysis. However, as I have 13 parameters, the path analysis is being so complicated. I want to reduce the number of parameters for each domain.

Zainudin Awang Thank you for your reply. I am new to SEM and still learning. For now, my biggest challenge is to decide which factors are more important than others.

Can you please suggest me what I can do in this case.

I can share more about my research question and data

Thanks

Dear David Eugene Booth yes the path analysis is right choice. I have seen various manuscripts in my field using SEM both EFA/CFA and path analysis.

Examples are here.

Thanks

Zainudin Awang Thank you for the reply. Yes i have negative factor loadings. Please see the attached figure. Can you please suggest me ow to deal with these values.

For modern variable selection see attached and references. Best, David Booth

Thank you all

Waqar Ashiq

I am late in responding to your query and not sure whether you have decided.

Before addressing your queries I would like to share that the first step is to judge and evaluate whether the model that one has tested is valid (correct). It is not necessary that all the models that can be estimated are valid/correct. There are several ways to judge this. As Zainudin Awang has pointed out the possibility of some potential issues ( likely haywood case) as the standardized estimate (factor loading) in measurement model exceeds 1. In principle a variable can at best explain 100% of the variance and thus the value should theoretically not exceed 1. A Google search will help you to learn the criteria of specification of a correct model. If I remember correctly, Barron and Keny's page also provide a list of criteria for different models.

Once we are sure that our model is theoretically sound and empirically correct then we can address the relative contribution (or most contributing path/model).

The path coefficients (the standardized one) are like beta coefficients and thus the magnitude of the values can help to judge which variable/path (direct or indirect) is contributing most to the criterion variable.

However, if one is interested in judging which model (out of the two or more alternative path models) is best model then one can sue chi-square difference test, difference in the CFI values etc. If the improvement is significant or beyond the prescribed cut-off then one can infer that one model is better than the other.

All said, I would suggest to use the latent factor scores (rather the aggregate score) in the path model as I can see the factors that you have got is similar to bipolar factors (some variables loaded positively while others negatively).

I am not sure whether I have understood your query but tried to answer what I comprehended.

With best wishes

Rakesh