Picture the two-way table as a building with rooms. A joint relative frequency compares one room with the whole building: of all 200 students, 18 play sports and have a job, so 18/200=9%. A conditional relative frequency first walks into one wing and closes the door behind it. Inside the sports wing there are only 30 students, and 18 of them have jobs, so the share is 18/30=60%. Same cell, different whole.
Doing this for each wing makes the groups comparable. In the non-sports wing, 34 of 170 students have jobs, which is 20%. Now the comparison is fair, 60% against 20%, even though the non-sports group has more job-holders by raw count (34 against 18). If the conditional shares were equal in every wing, the two variables would show no association; the bigger the gap, the stronger the link.
Which wing you enter matters. "Of the athletes, what share have jobs?" divides by the sports row, 30. "Of the students with jobs, what share are athletes?" divides by the job column, 52, and gives 18/52≈35%. Both are conditional relative frequencies, but they answer different questions, and swapping them is one of the most common errors in reading data. The words right after "of" or "given that" name the denominator.
Think of it like this. A conditional relative frequency works like a zoom lens on the table: you frame one row or one column, and that frame becomes the whole picture, so its total goes underneath. Zoom on a different row or column and the same cell becomes a different percentage.
Core idea. The denominator is the group you condition on: pick one row or column, divide its cells by that group's own total, and compare those shares across groups to see whether two categorical variables are associated.