Mohammad, did you standardize your variables? This could be one of the reasons. Another reason could be that you included some paths recursive paths.

HTH

Karin

Karin, I used correlations to calculate the path coefficients. However, the variables used in the estimation of correlations were not standardized.

Mohammad, I am not quite sure how you calculated the path coefficients. Did you use regression analysis, or a SEM program? Could you please provide more information?

Regards, Karin

I used R to calculate path coefficients. All we have to do is give correlations file in the script and it comes with the output.

Mohammad , there may be some estimation problems. Reduce your model to a smaller model by using only 3 variables and check whether these results are plausible. Then add the next variable, check the results etc. I guess that there may be a problem with the specification of your model. There may also be a suppressor effect. Check that the correlations are not too high.

Regards, Karin

Dear Dr. Tahir,

Factors genotype showed no significant effect in ANOVA is certainly not variation. If used to calculate the correlation coefficient, can be misleading. Based on my experience, if these traits are also included in the path analysis (base the estimation of correlations), is also likely to generate -1 < direct and indirect > 1. So not only multicollinearity that contribute to case. 

I do not know for sure, since multicollinearity was already taken in consideration and some variables were remove there is not much more to do. I remember that when we obtain a variable as the product of two other variables of interest (as we do in sugarcane with TCH and Pol to obtain TPH) it may cause an artificial correlation between the obtained variable and the variables used to obtain the artificial one. That kind of correlations may inflate the relative importance of the estimated variable at the phenotypic level. It is what Kang et al, 1991, called spurious associations produced by artificially created relationships. I once had a correlation coefficient greater than one and the reason was the artificially created variables.

I don't know what method you are using to calculated the correlations but it sounds like you have a scaling problem (one or more of your variables are on wildly different scales) or the co-linearity problem is so bad that you are close to dividing by zero in the calculations. You could also have a coding error, extreme outliers or just bad data elements somewhere in your data set.

Good luck, Brian

Brian, my variables are definitely on different scales and they do have outliers. Do you think they could affect correlations? And after we have removed highly correlated variables from pairs?

Yes Brian. Correlations are just fine. It is the direct and some indirect effects which are more than 1.

in case of path coefficients, unstandardized path coefficients tend to be greater than one, but we use standardized coefficients in interpretations and also while quoting in research articles. possible reason for unstandardized path coefficients  is mostly use of different scales of measurements in studies.

standardized path coefficients can be greater than one s. here the explanation 

http://www.ssicentral.com/lisrel/techdocs/HowLargeCanaStandardizedCoefficientbe.pdf

Great Article.  Thanks.

follow!!

If standardized path coefficient can be larger than 1, how should we explain the result? For example, I got a standardized path coefficient larger than one--for example--1.5, can I say that for each increment in variable A is associated with 1.5 unit change in variable B?

I would like to ask the same question. If you got the answer, please share Zhijuan Zhang Zhijuan Zhang

Follow

Standardized coefficients greater than 1.pdf (statmodel.com)

https://www.statmodel.com/download/Standardized%20coefficients%20greater%20than%201.pdf

the answer is explained clearly in this chapter by Joreskog...thanks to him :)