Level 7 · Will doubling ad spend double orders?

What a correlation can say

r = +0.67 between daily spend and daily orders means: on days when spend is high, orders tend to be high too, and the relationship is moderate to strong. That’s all. Three different worlds produce the same number:

• spend brings in orders (what Jordan hopes);
• orders, or the expectation of them, bring in spend: the team raises bids when shoppers are out;
• something else lifts both, like the holiday season.

The number can’t tell these apart. A factor that moves both variables, like the season here, is a confounder.

Checking the confounder

You can see the season in the data. Spend jumps in November: paid spend alone went from $28,157.04 in October 2025 to $47,475.61 in November. Orders do the same, from 371 in October to 715 in November. Any two numbers that both follow the calendar will correlate, whatever their link.

A fairer look compares days within the same period, for example correlating spend and orders only within ordinary months, or comparing each day with the same weekday a week earlier. If the link mostly disappears, the calendar was doing the work.

Writing about it

In a conclusion, say “is associated with”, “goes together with” or “moves with”, and give the number. “Daily orders are associated with daily ad spend (r = 0.67), but both peak in the holiday season, so this doesn’t measure the return on spend.” That sentence is honest and still useful: it tells the reader what to test next.

To actually measure what spend does, you need variation you control: a holdout where some regions or weeks get less budget, or a planned test. That’s the step from “goes together” to “causes”.

Reading r, a quick recap

• The sign is the direction. -0.85 is a strong relationship in opposite directions, as strong as +0.85.
• The size is the strength: above 0.7 strong, 0.3 to 0.7 moderate, below 0.3 weak.
• r is not a percentage, and r = 0.67 doesn’t mean “67% more”.

Common mistakes

• Turning r into a forecast (“doubling spend adds 67%”).
• Forgetting a time-driven confounder whenever both series are daily or monthly totals.
• Hiding a weak or negative correlation. Saying what is not related is a finding too.