Data Science

Narrow the question first

A request to "take a look at the data" isn't a question yet. The work only starts once it's narrowed down to something data can actually answer. "Why have customers been dropping off lately" can't be answered as it stands, but "is the share of last month's new sign-ups who never came back in their first week higher than it used to be" can.

Narrowing the question also settles what you'll do once you have an answer. If the outcome wouldn't change your next move either way, that's an analysis you didn't need to run. Skip this check and it's easy to spend weeks building a table nobody ends up using.

Most of the time goes to gathering and cleaning

Real-world data never arrives as a tidy table. Rows go missing, some show up twice, units get mixed, and timestamps come from clocks set to different zones. It's common for the same person to appear under one name at sign-up and a different one at checkout.

This stage looks tedious, but it decides the outcome. A handful of bad values can drag an average off course, and duplicate rows can inflate results that look strong but aren't. However good the model bolted on afterward, a broken ingredient still makes for a broken answer.

Cleaning also means checking how the data came to exist in the first place. If only people who opened the app ever show up in the records, that data can't support a conclusion like "people open the app often."

Look before modeling

Before reaching for a model, there's a step where you plot the data and just look at it. Averages and totals alone hide too much. Two groups can share the same average while telling completely different stories — one clustered together in the middle, the other split into two camps at opposite ends.

This is where odd values and unexpected clusters surface: a row where someone's age reads 200, a record that goes quiet on one particular weekday, users who split cleanly into two distinct camps. These discoveries, arriving before any answer does, are what shape the next question.

The model is the last tool you reach for

A model becomes worth reaching for once you have to guess what happens next. Before that point, if a simple comparison or a basic split already answers the question, that's enough — and it's easier to explain, too.

The spot where people slip most often here is mixing up correlation with causation. Two numbers moving together doesn't mean one is causing the other. Ice cream sales rising alongside drowning incidents doesn't make ice cream the cause; a hidden third factor, like hot weather, is usually driving both.

Confirming an actual cause takes a different approach: splitting people into groups under different conditions and comparing the outcomes directly — showing one group a changed screen, say, and measuring what happens against a group that never saw it.

The job isn't done until it's communicated

Once the analysis wraps up, what's left is persuasion. What was found, what to do about it, and how far the conclusion can be trusted all need to be delivered together.

The people who know numbers best never skip that last part. They flag that the sample was small enough to wobble, or that something specific to this one stretch of time got mixed in. One confident conclusion is worth less, in practice, than a clear line drawn between what's solid and what isn't.