The Statistics You Actually Need for Data Science

McTaba runs coding and data bootcamps in Kenya. This guide is written to be useful whether or not you ever join one.

Imagine a statistics lecture that starts with a coin toss and ends in a wall of Greek letters. You do not need that wall. A working analyst reaches for five ideas again and again, and this page shows each on a small example you can redo with a pen. For the larger picture, see our data science career map for Kenya.

Idea	The question it answers	Kenyan picture
Mean and median	What is typical?	Salaries in a small company
Spread	How much does it vary?	Daily sales at one stall
Sampling	Can I trust a small group?	Surveying 200 shoppers
A/B test	Did the change work?	Two versions of a USSD menu
Bayes	How likely is it, given the alert?	A fraud alert on M-Pesa style data

If you want these taught alongside code, with weekly feedback, the Data Science to AI Engineering programme covers them in its maths and statistics phase, with Bayes for fraud alerts and A/B testing. No coding, data or maths background needed.

For example, imagine nine staff at a small Nairobi firm, with made-up monthly salaries in thousands of shillings: 35, 38, 40, 42, 45, 48, 50, 52 and 400. The 400 is the owner.

• Mean: add them up (750) and divide by nine. About 83 thousand.
• Median: put them in order and take the middle one. The fifth is 45 thousand.

Eight of nine people earn less than the mean. The mean got pulled by one big number. The median stays close to what most people take home. When data has a long tail, as incomes, prices and mobile money transfers do, report the median first.

Now spread. Two stalls can both sell a median of 5,000 shillings a day. One sells between 4,500 and 5,500 every day. The other swings between 500 and 12,000. You would plan stock very differently. Three simple measures of spread:

• Range: biggest minus smallest. Quick, but one odd day ruins it.
• Interquartile range: the width of the middle half of days. Hard to fool.
• Standard deviation: the typical distance from the mean. Most libraries print it.

In pandas, df["amount"].describe() prints all of these in one go. The first week of learning Python for data science gets you to that line.

Imagine you ask 200 shoppers at one Westlands supermarket what they pay for mobile data. Does the answer describe Kenya? No. It describes shoppers at that supermarket on that day.

A sample speaks for a population only if it was chosen fairly. Three questions to ask every time:

1. Who could not have ended up in my sample? (People without smartphones, people who shop elsewhere.)
2. Who chose to answer? (The angry and the keen answer more than the content.)
3. Is the sample big enough for the difference I care about? A bigger sample shrinks random error. It does not fix a biased one.

The same warning applies to your own data. A table of M-Pesa transactions holds only the people who used M-Pesa. Before you generalise, write down who is missing.

For example, imagine a savings service with a USSD menu. Version A asks for the amount first. Version B asks for the phone number first. You send each version to 1,000 made-up users.

• For example, version A: 1,000 users, 300 finished a deposit, a rate of 30.0%.
• For example, version B: 1,000 users, 345 finished a deposit, a rate of 34.5%.

B is 4.5 percentage points ahead. Is that real, or luck? The standard error of the difference is about 2.1 points (work: the square root of 0.30 x 0.70 / 1,000 plus 0.345 x 0.655 / 1,000). The gap divided by the standard error is about 2.2. A gap above 2 is unlikely to be luck alone, so luck alone would produce it about 3 times in 100.

Put it in plain words for the product owner: "B probably finishes more deposits. The true gain is likely between 0.4 and 8.6 points. Ship B, and keep watching."

Rules that stop false wins

• Decide the sample size and the metric before you start.
• Do not stop the test the day it looks good.
• Split users at random, not by day, or Friday paydays will fool you.

The result then has to reach a person who decides. Data storytelling shows how to write that one page.

This one surprises people. For example, imagine a bank flags suspicious transfers. Made-up numbers: 1 in 100 transfers is fraud. The alert catches 90% of fraud. It also wrongly flags 5% of honest transfers.

Take 10,000 transfers.

• Fraud: 100. The alert catches 90.
• For example, honest transfers: 9,900. The alert wrongly flags 495 of them.
• Total alerts: 90 + 495 = 585.

For example, of those 585 alerts, only 90 are fraud, about 15 in every 100. An alert that "catches 90% of fraud" is wrong about five times in six when it fires.

That is Bayes' rule without the formula. Start from how rare the thing is, then update with the evidence. Two lessons follow. For rare events, accuracy figures mislead. And an alert team needs a cheap way to clear false alarms, or it will drown. The fraud topic gets its own build in our guide to mobile money fraud detection with machine learning.

1. Redo each example above on paper, then in a spreadsheet, then in Python.
2. Change one number and predict what moves.
3. Use free lessons from Khan Academy for gaps, not as a starting wall.
4. Apply each idea to a dataset from our six-month data science roadmap.

No maths degree is needed for any of this, and no school can promise a job from it. It is the thinking that lets you say "I am not sure yet" in a meeting, and mean it. The Data Science to AI Engineering programme starts on 11 January 2027 and costs KES 140,000.