COJO SNPs with non-significant p values but small pJ

Hi I used COJO to perform stepwise model selection (--cojo-slct) with --cojo-p 5e-8 on a meta-analysis summary statistics. I found that ~10-15% SNPs selected had non-significant p values (p > 5e-8) but with joint model p value < 5e-8. For a few SNPs, the GWAS p values and joint model p are even very different. I wonder what the problem might be?

Thank you very much!