Why does STAR filter out many paired end reads while they align fine as single end?

Question

[KDL]

I noticed that STAR filtered out a lot of reads when they were supplied in paired end fastq files but kept both of them when they were supplied as single end separately. When I used Tophat to run the same reads as paired end, they were both retained in the output. This happened to many reads in my data to the extent that Tophat output almost 2 times of what STAR did when I analyzed them as paired end (which is what they are). Noticeably these reads are of high quality and unique mappers too. The distance between R1 and R2 is not over the default cutoff. I'm new to STAR and get quite confused. Can anyone shed some light on this? TIA!