Statistics for Data Analysis

Statistics sounds intimidating, but a Data Analyst uses a surprisingly small, practical slice of it. Statistics is really the science of making sense of numbers when there is variability — and almost all real data varies.

The payoff of this topic is huge: you will stop reporting a single average as if it told the whole story, you will know whether a difference you see in a report is real or just luck, and you will finally understand why "correlation is not causation" is the most repeated warning in the field. Everything here stays at the level a working analyst needs — no proofs, no calculus.

Basics (data types, metrics, data quality) and comfortable arithmetic. No math background needed — you will compute everything in Excel or Python anyway.

Every dataset is a sample — a subset — of a bigger reality (a population). One month of orders is a sample of all months; 200 surveyed users are a sample of all users. Statistics answers two questions analysts ask constantly: "What is the truth, roughly?" (descriptive stats) and "Is this difference real or just luck?" (inference). Without this lens, a report that compares two averages can confidently state a false conclusion — because the difference was noise.

Mean = sum ÷ count (the average). Median = the middle value when sorted. Mode = the most frequent value.

The trick is skew: a few huge values (a mansion in house-price data) pull the mean up while the median stays undisturbed. Example: five salaries — 100, 110, 120, 130, 1,000. Mean = 292 (misleading), median = 120 (honest). Analysts report median for income, house prices, session lengths — anything with outliers. Click AVERAGE in Excel for a clean symmetric spread; click MEDIAN when big values can yank the story sideways.

Center tells you where the middle is; spread tells you how squashed or scattered the values are.

• Range = max − min (fragile — one outlier blows it up)
• Variance = average squared distance from the mean
• Standard deviation (SD) = sqrt(variance) — spread in the data's own units

Low SD = values huddle near the mean (consistent delivery times); high SD = values swing widely (unpredictable). For a normal distribution about 68% of values sit within ±1 SD and 95% within ±2 SD. Comparing two groups with the same average but very different SDs tells you which is stable and which is wild.

Percentiles mark the sorted data cut-offs: the 25th percentile (Q1), the 50th (median, Q2) and the 75th (Q3) split the data into quarters. The Interquartile Range (IQR) = Q3 − Q1 = the middle 50% band.

IQR is the professional's outlier detector: values below Q1 − 1.5×IQR or above Q3 + 1.5×IQR are classic outliers. Two practical uses: report "the middle 50% of orders were $25–$60" (a far truer picture than an average), and flag suspicious records before analysis.

A distribution is just a histogram of where the values fall. The famous normal distribution (bell curve) shows up when many small factors add up: heights, measurement errors, long-run averages. Its shape lets you estimate how common extreme values are.

Real business data is often skewed instead: revenue, order amounts and salaries pile up near the low end with a long right tail (right-skewed) — the same reason medians beat means. Recognizing skew immediately tells you which summary and which chart will be honest, and whether a log/normalized transform might improve your analysis.

Analyzing a whole population is often expensive or impossible, so analysts work on samples. A sample is only useful if it fairly represents the population — that is the crux of sampling bias.

Classic bias traps: surveying only app users for a product used on desktop and mobile; judging satisfaction only from complaints (angry people self-select); A/B tests that include only power users. Whenever you hear "based on N surveys", ask how were these people chosen? A biased sample produces confident garbage no matter how clean the analysis is.

Correlation (r, from -1 to +1) measures whether two variables move together. Ice-cream sales and drowning accidents correlate strongly in summer — one does not cause the other; a third force (warm weather) drives both.

Correlation ≠ causation because a relationship could be coincidence, due to a lurking variable, or even reversed. To prove cause you need experiments (A/B tests, controlled trials) or very strong domain logic. A good analyst reports associations honestly and labels causal claims only when the evidence truly supports them.

Outliers are extreme values that sit far outside the pack. They are not automatically errors — a whale account ordering 1,000 units is real, valuable data.

Analyst protocol: detect → investigate → decide. Detect with IQR or z-scores; investigate whether it's an error, a data-entry bug, or a genuine event; then decide — flag it, exclude it (documenting why), cap it, or keep it split for a separate analysis. Never silently delete; never let one absurd row pull your whole average.

When a report says a change is statistically significant, it means the observed difference is unlikely to be pure chance.

• Null hypothesis (H₀): nothing is happening — the new button converts the same as the old (difference = 0)
• Alternative hypothesis (H₁): a real effect exists
• Test statistic + p-value: a p-value is the probability of seeing results at least this extreme if H₀ were true. p < 0.05 is the common "unlikely to be luck" line

Two caveats that matter daily: a small p-value doesn't mean a large effect (statistical ≠ practical significance), and a large dataset will flag even trivial differences as significant. Analysts always report the size of the effect, not just its p-value. Confidence intervals show the range the true value likely sits in — "conversion lifted 0.4% (±0.3%)" — and are the more honest cousin of a single point estimate.

Use a real dataset (house prices, salaries, deliveries — anything with outliers).

1. Compute mean, median, mode and SD; decide which to report and why.
2. Find quartiles and IQR; flag every IQR-based outlier and classify each as error vs genuine.
3. Draw the histogram; describe skew and what it means for the average.
4. Pick two numeric columns; compute correlation in Excel (=CORREL); write two plausible non-causal explanations.
5. Split your data into two groups (e.g. regions); compare means + SDs and judge if the difference looks real or within noise.

Interviewers probe statistics basics because the cost of misuse is high: a manager who chases a "7% lift" that is actually noise spends budget on nothing. In daily work you will use medians and SDs for summaries, IQR to police data quality, distributions to choose charts and transforms, and significance/effect-size reasoning whenever someone compares two periods or two groups. Every data story you tell later rests on these numbers being honest.

Business problem: The delivery team claims the new last-mile partner made orders faster. The analyst must check this with statistics before the company signs a contract.

Dataset: any order-delivery log with delivery minutes and the partner flag (60–90+ orders across the two partners). Create one if needed.

Objective: an evidence-based verdict — is the new partner faster, similar, or within noise?