In multi-environment trial (MET) analysis you can use GGE or GREG biplot graphic method. With biplots you can determine similarity of environments, stability of genotypes and many more. For more info see:

Yan W, Kang MS (2002) GGE Biplot Analysis: A Graphical Tool for Breeders, Geneticists, and Agronomists. CRC Press, London

Kang MS, Gauch HG (1996) Genotype-by-Environment Interaction. CRC Press, London

Data on same set of entries may differ over environment (year and/ or season and/ or location) and so each set of data has specific value of correlation and it may varied over environment. Pooled analysis may give overall magnitude and direction of correlation. Environments may be treated as replications for analysis.

Rameshwer

You did't provided the information about the trial that was conducted at different location/environment/year. In this case may be likely

1. If the trial was conducted at the same location within the specific state/zone for different years, first check the GxE interaction. If interaction is significant then the performance is due to the effect of Environment. Then the correlation will be strong and unidirectional.

2. If the trial was conducted at the different location within the specific state/zone. GxE interaction should be tested and should be significant theoretically. Pooled correlation may be differ but not the direction.

The different environment can help estimating  an average performance, but the performance can differ. The best way will be conducting the trial at same place for multiple year or multi-location to to draw the trust-ability of the data and potential of the performance.

In multi-environment trial (MET) analysis you can use GGE or GREG biplot graphic method. With biplots you can determine similarity of environments, stability of genotypes and many more. For more info see:

Yan W, Kang MS (2002) GGE Biplot Analysis: A Graphical Tool for Breeders, Geneticists, and Agronomists. CRC Press, London

Kang MS, Gauch HG (1996) Genotype-by-Environment Interaction. CRC Press, London

Rameswar -

The changing of signs is of interest.  In regression this could be caused by changing from a simpler model to one with collinearity, but I wonder what is happening with your data?  Are sample sizes small so that such cases are essentially uncorrelated, or does something change from year to year? 

Not an area of expertise for me, but Marcin seems familiar with this, and I think graphics are generally more helpful than many statistics such as r, or R, or certainly the misused/misunderstood p-value.  (I do encourage graphics, but principal components can make interpretation less clear.)  

If you figure out what is happening, perhaps you might post your findings for general interest.

Cheers - Jim

Dear Rameswar

I think the case that you described reflect the interaction between genotype and environment , especially cross G×E interaction, so I agree with the comment of Marcin

Good Luck