For example, we might want to know the impact of getting canvassed---when someone knocks on your door---on voter turnout. You might be tempted to just let a campaign conduct its normal canvassing program and compare people who were successfully treated and those who weren’t. But people who were reached might be different than others who weren’t in ways other than just getting canvassed: for example, they might be older (so they’re more likely to be home), or they’re friendlier. These qualities (age, personality, etc.) can impact the outcome we’re trying to measure (e.g., turnout), so we don’t know if the difference between the two groups is due to differences in age/personality or because of the canvass itself. When you randomly assign people to get canvassed or not, by contrast, you can ensure that the only difference between the groups is the treatment itself.