What should you know about Data Analysis?
A sound data analysis starts with a specific question, checks the source and structure of the data, prepares the records, explores relevant comparisons and communicates a conclusion with its limitations. Good charts and algorithms cannot compensate for a question that was never clearly defined.
Data Analysis at a glance
| Concept | What to know |
|---|---|
| Define | State the decision, outcome, population and period. |
| Collect | Identify sources, units, fields and permissions. |
| Validate | Check duplicates, missing values, ranges and types. |
| Explore | Summarize distributions, trends and group differences. |
| Analyze | Use appropriate calculations or models; test alternatives. |
| Communicate | Show evidence, assumptions, uncertainty and next actions. |
1. Define the question and unit of analysis
An open-ended instruction such as “analyze sales” leaves too many possible directions. “Did weekly order value decrease after the pricing change?” defines an outcome, a comparison and a time frame. A follow-up question might ask whether any change differs across customer groups.
Decide what one row represents before counting or averaging anything. A table of individual orders measures transactions; a table of customers measures people. The number of records does not equal the number of unique customers when some customers order repeatedly.
Write down the intended decision and what evidence would change it. The introduction to data analysis explains descriptive, diagnostic, predictive and prescriptive questions in more depth.
Question: “Were October returns more frequent than September returns?” Define the numerator (returned orders), denominator (eligible orders), and date convention before calculating rates.
Check this: A higher number of returned orders does not necessarily mean a higher return rate if total orders also rose.
2. Understand where the data came from
Record the source, time zone, coverage, collection method and any known exclusions. A marketing export may exclude visitors who declined tracking; an operations report may omit canceled orders. Those omissions affect which claims the dataset can support.
Check data permissions and keep private information out of reports unless it is necessary and appropriate. Before combining tables, confirm that customer identifiers, dates and categories mean the same thing across systems.
Consider whether the data is observational or produced by a controlled experiment. Observational comparisons are useful, but differences between groups can arise from other variables besides the one being studied.
One source stores timestamps in UTC and another in local time. Comparing “daily” activity without conversion can move observations into the wrong day.
Check this: A clean spreadsheet may still be unsuitable for the question if the source systematically misses part of the population.
3. Clean and validate the dataset
Look for duplicate records, impossible values, inconsistent spelling, missing entries and data stored in the wrong type. Confirm whether negative numbers are valid refunds, errors or a special code. Decide which problems require correction, which deserve flags and which should remain as observed.
A reliable cleaning process preserves the raw data, documents changes and checks the result. Do not silently delete unusual rows simply because they distort an average; first learn whether those rows are legitimate observations.
Spreadsheet users can follow the Excel data-cleaning guide. For more systematic checks, the exploratory analysis workflow introduces missingness, variable types and outlier review.
If the values are “US”, “U.S.” and “United States”, a category summary will split them unless the labels are standardized.
Check this: Deleting missing records can introduce bias when the missingness is related to the outcome.
4. Explore distributions, segments and trends
Start with record counts and simple summaries, then inspect the distribution of numeric variables. A mean can conceal an uneven spread, while a time-series total can hide seasonal effects. Use histograms for numerical distributions, bar charts for categories and line charts for ordered time points.
Compare groups using consistent definitions. A segment with 20 orders and 5 returns has a 25% return rate; one with 100 orders and 10 returns has a 10% rate, even though the second has more returns in absolute terms.
Explore alternative explanations when a pattern appears. The chart selection guide helps match questions to visuals; the EDA guide covers a practical sequence of checks.
Orders: Week A = 50 returns from 1,000 orders (5%); Week B = 60 from 1,500 (4%). Returns increased in count but declined as a share of orders.
Check this: Do not compare rates based on different definitions of “eligible” records.
5. Test explanations without overstating them
Descriptive analysis asks what happened; diagnostic analysis investigates why it may have happened. Predictive models estimate outcomes from patterns in observed data, while prescriptive work compares possible actions. Each requires different evidence and should not be treated as proof of causality.
When estimating a relationship, review confounding variables, data leakage, sample size and whether the model performs on records it has not seen. A high in-sample fit may not translate to reliable future forecasts.
For a proposed business change, report both an estimated effect and uncertainty when the design allows it. A small difference may be noise; a statistically detectable difference may still be too small to matter operationally.
A rise in sales after a campaign does not show the campaign caused the rise if a holiday, price change or distribution expansion happened simultaneously.
Check this: Correlation is not causation, and predictive accuracy is not the same as a causal explanation.
6. Communicate a result someone can verify
A useful result states the question, the method, the figures and the limitations. Include the time period, denominator, units and any data exclusions. Keep the chart readable, and use a short explanation to tell readers what comparison it displays.
Make recommendations proportionate to the strength of the evidence. A descriptive dashboard can identify a change worth investigating; it may not justify a major decision without additional checking.
For presentation design, review common chart mistakes. A clear written finding is often more useful than an elaborate dashboard containing dozens of unrelated metrics.
“The October return rate was 4% (60 of 1,500 eligible orders), down from 5% in September (50 of 1,000). Definitions were consistent; reasons for the change need further investigation.”
Check this: Avoid vague claims such as “performance improved” when the actual measurement and comparison are not stated.
Frequently asked questions about Data Analysis
What are the main steps of data analysis?
Define a question, inspect the data source, clean and validate records, explore patterns, apply suitable analyses, and communicate findings with caveats.
How is data analysis different from data visualization?
Data analysis includes the full process of asking, preparing, calculating and interpreting. Data visualization is one way to examine or communicate the results.
Is data cleaning part of analysis?
Yes. Type errors, duplicates and missing values can change computed results, so validation and documented cleaning are part of credible analysis.
Does a correlation show cause and effect?
No. A relationship may reflect confounding, selection effects or other mechanisms. Causal conclusions require appropriate designs and assumptions.
What should a basic analysis report contain?
A clear question, data definition, time period, method, main numerical results, a suitable chart or table, limitations and an actionable next step.
The articles below explore individual ideas with more detail, examples and practical steps.
Go to the article library ↓