Conditional Probability
Your digital team sends a fundraising appeal to 5,000 supporters. The email platform reports back with two numbers. A 25% open rate and a 3% click-through rate. The campaigner looks at that 3% and feels deflated. Three percent sounds like almost nobody engaged. But then a colleague pulls the data differently. "Of the people who actually opened the email, how many clicked the donation link?" The answer is 12%. Same campaign, same data, very different story.
The difference comes down to who you're counting, or more precisely, what you're dividing by. That 3% comes from dividing 150 clicks by all 5,000 recipients. But 3,750 of those recipients never even opened the email. They weren't choosing not to click. They weren't participating at all. When you narrow the question to "among the 1,250 people who opened, what fraction clicked?" you get 150 out of 1,250, which is 12%.
This is what statisticians call conditional probability, the probability of one event happening given that another event has already happened. The word "given" is doing all the work. "The probability of clicking" and "the probability of clicking, given that the person opened the email" are two fundamentally different questions. The first includes everyone. The second restricts attention to a subgroup. The condition changes the denominator, and the denominator changes everything.
You already do this intuitively every time you segment your data. When you say "our retention rate among online donors is 52%", you've placed a condition on the calculation. You're not looking at all donors. You're looking at online donors specifically, then asking a question within that group. When you report "program completion rates for participants who attended orientation", you've conditioned on orientation attendance. The probability of completion among all enrolled participants would be a different, likely lower, number.
If this sounds familiar, it should. In the previous entry on Bayes' theorem, the entire calculation depended on conditional probabilities. The wealth screening tool's 90% sensitivity was itself a conditional probability, measuring how likely a truly high-capacity donor was to be flagged. The 85% specificity was the same idea running in the other direction, measuring how likely a non-prospect was to be left alone. What Bayes' theorem actually does is flip the direction of the condition. It takes "probability of the test result given the truth" and converts it to "probability of the truth given the test result". That reversal is the entire insight, and conditional probability is the foundation it rests on.
This concept shows up everywhere in nonprofit work. Donation page conversion rates are conditional on reaching the page, so if your email-to-donation conversion seems low, the bottleneck might be the click, not the page itself. In program evaluation, completion rates conditioned on enrollment tell you about the program experience, but they hide the enrollment bottleneck. A program where 90% of enrollees complete but only 10% of eligible people enroll looks very different from one where 60% enroll and 60% complete, even though both produce a similar number of completions. In advocacy campaigns, the "share rate" on a petition thank-you page is conditioned on signing. Forty percent of signers sharing sounds powerful, but if only 5% of visitors signed, the true share rate across all visitors is just 2%.
Every percentage your organization reports carries a hidden condition baked into the denominator. The question is never just "what's the probability?" It's "the probability among whom?" Changing that answer changes the conclusion.
See It
Click the buttons to switch between all recipients and only openers. Watch how the same clicks produce a very different percentage when the denominator changes.
Reflect
Pick a key metric your organization reports regularly, whether it's a click rate, a completion rate, or a conversion rate. What condition is hiding in the denominator? What would happen to that number if you expanded the denominator to include everyone eligible, or narrowed it to a more specific subgroup?
When two teams in your organization report different numbers for what sounds like the same metric, is it possible they're conditioning on different populations? What would you need to align before comparing their results?