Data Analyst interview

The five things they actually test, the questions that come up, and the answers worth having ready.

They are not testing whether you know SQL. They are testing whether you can be trusted with a number.

The five they will test

SQL joins
Data cleaning
A/B tests
Dashboards
Telling the story

SQL joins

Inner vs left is the easy half. The half they are listening for: a join can duplicate your rows, and if you did not notice, every number downstream is wrong.

Inner vs left, in one card

INNER

Only rows that match on both sides

Silently drops what does not

Use when a missing match means "not relevant"

LEFT

Every row on the left, matched or not

Nulls where there was no match

Use when a missing match is itself the finding

Cleaning, in the order they expect

Nulls — decide drop, fill or flag, and say why
Duplicates — find the key that should have been unique
Outliers — keep them until you can explain them

0.05

The p-value everyone quotes

And the one they will ask you to defend. "Significant" means the result is unlikely under the null — not that it matters, and not that it is big.

The join that duplicates rows

sql

select o.id, o.total, i.sku
from orders o
join items i on i.order_id = o.id
-- One order, four items → four rows.
-- sum(o.total) is now 4x the truth.

Dashboards

One question per chart. Nobody reads twelve KPIs
If a chart has no decision attached, cut it
Label the axis in the unit the reader thinks in

Number, so what, recommendation. Every single time.
— The shape of every good analyst answer

Your join made total revenue four times bigger. What happened?

AThe left table had duplicate keys
BThe right table had several rows per key
CThe join type should have been inner
DNulls were counted as zero

One order with four items becomes four rows, and every sum over the order table is now counted four times. Aggregate before you join, or join to something already at one row per order.

What does "statistically significant" actually promise?

Only that a result this extreme is unlikely if nothing were going on. It says nothing about the size of the effect, and nothing about whether it is worth acting on.

Five to ask them back

How do you know a dashboard is being used? · Who owns a metric when it breaks? · What was the last analysis that changed a decision? · How much of the job is cleaning? · What does good look like at six months?

Now do it with yours.

Paste your own job description and Claude writes the deck on your phone.