Bias is self-perpetuating in lazy statistics
Don't be fooled by "what if this group really is more criminal" - apart from the main factor that others have mentioned, which is that socioeconomic factors are more predictive than race, you can easily envision how bias can set up a feedback loop.
Start with two groups, A and B, which actually have identical proportions of members who carry knives (or any other crime detectable by a search). Due to cultural bias, more people from group A are searched than from group B. Also due to bias, when charged, a greater proportion of members of group A are convicted than those from group B. Now take a look at the headline statistic of people convicted of knife offences - wow, people from group A are massively over-represented! We'd better make sure we focus our policing on them! Now not only do you have the support of those with (conscious or otherwise) bias against group A, you also have people with neither bias or the ability to properly check the validity of statistics.