DESCRIPTIVE VS. INFERENTIAL STATISTICS: THE FULL BREAKDOWN (FOR PEOPLE WHO DON’T WANT A MATH LECTURE)
Any time you’ve been in a meeting, and seen someone say, “the data shows x,” you’ve probably witnessed either one of two distinct operations — and both are very easy to get wrong — especially for the more seasoned analyst.
Let’s fix that. No jargon, scary formulas, or opaque charts defining just what descriptive and inferential statistics are, whether to use them and where they sneakily fall down.
The One-Sentence Rundown of Everything
ü Descriptive statistics: describe what you currently know (in your current data set).
ü Inferential statistics: helps you to make an educated guess about data you have never seen before based on small-data samples.
That’s it. The rest of what is contained in this article is detail to the above concept.
Part 1: Descriptive Statistics — “Here’s What’s in Front of Me”
What if you had the exam results of each and every student in one classroom — 30 students. Desired to summarize the performance of the class. This is descriptive statistics: describing a dataset that you have all data for.
The Core Tools
Measures of Central Tendency (the “typical” value):
· Mean — the average; sum all the numbers and divide by the count.
· Median — The number in the middle of all the numbers listed when they are arranged in order. Useful when a few large or small numbers (such as one billionaire among a group of regular folks) would skew the mean.
· Mode — the most frequently occurring value
Measures of Spread (how spread out the data is):
· Range = highest value — lowest value
· Standard deviation & variance — the average distance values are from the mean.
· Interquartile range (IQR) — Spread of the “middle 50%” of data, use to identify outliers
Visual Tools — Bar charts, histograms, pie charts, box plots, scatter plots. — These are not bells and whistles — a histogram can show you the distribution is skewed in just a few seconds, which you would take several paragraphs to describe in text.
Where You’ll Use This for Example:
· Summing the month’s sales to put in a report
· Reporting off the average customer satisfaction score
· If you’re creating a dashboard, e.g. Showing current user counts, average time to complete, churn rates
· Calculating survey data results (“60 percent agreed that…” etc.)
Relevant Software:
Excel/Google Sheets, Python (pandas, matplotlib, seaborn), Tableau, Power BI, SQL
The beauty of descriptive stats: Nothing gets past your dataset. If you know all the 30 students’ scores, then the average IS the average. Not at all uncertain, not in any way probable.
Part 2: Inferential Statistics — “Here’s My Best Guess about what I can’t see”
Now imagine a different scenario. Not all of the people in a country are accessible — there are 30 million people and not 30 students. Rather, you poll 1,000 people and wish to make an estimation of how the entire country feels.
This is inferential statistics; making a claim from a sample and stating how sure you are about the claim.
The Core Tools
Hypothesis Testing — A methodical approach to the question of “Is this difference real or it’s random noise?” You hypothesize (or set up a null hypothesis) that nothing interesting is happening and then look for evidence provided by your data that you can reject your null hypothesis.
p-value — is a number that approximately answers the question: “How much evidence is there that nothing is happening but for chance to give rise to data this extreme?” A small p-value (typically less than 0.05) is used to reject the null hypothesis, but it does not imply that the null hypothesis is false 95% of the time. A small p value (usually less than 0.05) is evidence against the null hypothesis, but it is not evidence that the null hypothesis is wrong 95% of the time.
Instead of a guess, you provide a range — that’s called a confidence interval. “42% plus or minus 3%” is a far more complete statement of the truth than a simple 42%, because it conveys some uncertainty.
Regression Analysis — Making predictions about the relationship between variables (such as amount of revenue that is expected to change for a $1 increase in ad spending) and extending that relationship beyond your data set.
t-tests, ANOVA, chi-square tests — Different flavors of hypothesis tests for comparing groups, means or categorical relationships.
Get Ivone Atieno’s stories in your inbox
Join Medium for free to get updates from this writer.
Some of the places you’ll find yourself using this.
ü Polling and forecasting by political parties
ü Before rolling out a website redesign to all users, it is best to test it with A/B testing. A/B testing a website redesign prior to its full implementation is a good step to take.
ü Drug tests to check if a drug is effective
ü To predict the demand of next quarter based on a proportion of past transactions
ü When collecting some data and you need to comment on a larger population that you haven’t measured completely
Common Software
Python (SciPy, stats models, scikit-learn), R, SPSS, SAS.
The one thing that is apparent in inferential statistics is the uncertainty. You can never be certain — You work with probabilities, margins of error and confidence levels.
Descriptive vs. Inferential: Side-by-Side
The Mistakes Even Experienced Analysts Make
This is where it really counts, as when you get this wrong, you’re not just looking bad, you’re making real choices off of real flawed information.
1. Reporting a sample statistic, as if it were a population fact.
You ask 200 customers about the quality of the service they receive, and 70% say they are satisfied, and you report that 70% of our customers are satisfied (no ifs about it). The statistic for the 70% is just a sample statistic. A confidence interval or margin of error is not included, and you are unwittingly giving a guess away as fact.
Fix: If it came from a sample, state it and provide a range and NOT a fixed number.
2. Taking a small p-value as evidence
The p value < 0.05 is NOT an indication that “there is a 95% chance that this effect is real.” If there was no effect there, these kinds of data would occur by chance 5% of the time. That’s a subtle distinction — and one of the most common statistical blunders — and it is easy to confuse these two.
Fix: Accept a low p-value as evidence but not proof. Use it in conjunction with effect size and context — not a yes/no decision.
3. Solving problems by drawing inferences about cause and effect from descriptive charts
When a marketing campaign was introduced, and sales increased immediately, a chart that displays this is a description of what has occurred. Other factors might have contributed to the spike, so it’s not evidence that the campaign created that rise. Descriptive statistics reveal patterns, but don’t prove if the pattern is meaningful or causal.
Fix: You must have an actual inferential test, preferably a controlled experiment (such as an A/B test) not just “look at this trend line.”
4. Cherry-picking the sample
One of the sneakier errors is running an A/B test but stopping it when results look good (“p-hacking”) since it looks like valid inferential statistics (p-values and everything!) and is actually in violation of the assumptions required in order to make the p-value meaningful in the first place.
Fix: Pre-determine the number of samples and length of test to be used and adhere to them.
6. Excluding some of the data in favor of others
Ten reviews, all the same, that state a product is “amazing” will sound like a lot, but it’s really just a single strand of straw. Ten reviews, all the same, that say a product is “amazing” is compelling, but it’s only a single strand of straw compared to a population of ten. Descriptive statistics based on a “tiny sample” are assumed to be inferential conclusions about all the users, but with a tiny sample you cannot claim any confidence level in any inferences.
Fix: Question any general statement with the question “how many data points is this actually based on?”
7. Overlooking the significance being statistically significant does not mean practically significant.
Even the most insignificant of differences can be called “statistically significant” given a large enough sample. The new site design may only be gaining 0.02% clicks but with a low p-value — this is statistically significant but not practically. A new website design could be gaining 0.02% clicks with a low Ph-value — technically significant, not really significant at all.
Fix: Always ask 2 questions: Is it statistically significant? AND What is its actual impact on the business size?
Bringing It Together
It’s not a case of descriptive versus inferential statistics; they are partners in crime, performing different tasks:
v Descriptive statistics can be used to understand and communicate the data that you have in front of you.
v Inferential statistics enable you to make responsible claims about what you have learned from that data with honest statements of the level of confidence that you have.
The mistake isn’t using one or the other. The mistake is using the language of one while actually doing the other — describing a sample and talking about it like it’s the whole population, or eyeballing a chart and calling it proof.
The next time you come across stats in a report, a headline or a slide show, ask yourself one question: “Is this information describing what we know, or just what we speculate about?” In fact, it is the only one question to help you avoid statistical misinterpretation — and it’s the difference between an analyst who runs numbers and one who actually understands them.
