PGS prediction differ across different samples

I have a polygenic score for baldness estimated based on the summary statistics from the UK Biobank. The polygenic score that I use is constructed purely for European descendants, and this score is highly predictive for one of the outcomes (income) in the UKBB male sample that I am interested in. However, when I replicate this analysis using the Health and Retirement Study, I found the opposite results (reverse sign), even it is not significant.

 

My understanding is that I restrict people to non-Hispanic white males in HRS, and thus HRS should have the same results like that in the UKBB. Even though I understand that UKBB is much larger than the HRS sample and therefore, may yield insignificant results in HRS, I don't know why the sign of the coefficient is totally reversed.  

 

Anyone has any comments why this happen?