Data Journalism Intern

Most data journalism time is spent cleaning rather than analysing, and a story built on uncleaned data is a story waiting to be corrected.

This task takes you through the full process using freely available official data and free spreadsheet software.

Download one public dataset from an official source such as a government statistics portal, a ministry's open data page, a census release, or a regulator's published figures.

Choose one with at least several hundred rows and multiple columns, so there is something real to work with.

Before touching anything, document the dataset.

Record who published it, when, what period it covers, how it was collected, and what each column means according to the publisher's own documentation.

Now clean it, keeping the original file untouched and working on a copy.

Identify and handle missing values, deciding and recording whether to exclude or retain them.

Standardise inconsistent entries such as varying spellings of the same place or category.

Find and resolve duplicates.

Correct data types so numbers behave as numbers and dates as dates.

Flag outliers and investigate whether each is an error or a real extreme.

Keep a cleaning log recording every change, why you made it, and how many rows it affected.

This log is what makes your work reproducible.

Then analyse.

Calculate totals, averages, changes over time, and comparisons between categories.

Identify at least three findings that are genuinely newsworthy rather than merely true.

Interrogate the data before trusting it.

Check whether the totals in the dataset match any total the publisher states elsewhere, since discrepancies often reveal that the file is incomplete or covers a different scope than the title suggests.

Check whether every category that should appear actually does, as missing categories are easy to overlook and fatal to a conclusion.

Then separate the interesting from the merely true.

A finding is newsworthy when it is unexpected, when it affects a significant number of people, or when it contradicts a widely held belief.

For each of your three findings, state which of these applies, and discard anything that is simply a number nobody would dispute or care about.

Submit the original file, the cleaned file, the cleaning log, the data documentation, the integrity checks, and a findings note with the figures supporting each and your reasoning for its newsworthiness.

Charts make data understandable, but the same figures can be presented honestly or misleadingly depending on choices that often go unexamined.

This task uses the dataset you cleaned previously and any free charting tool.

First, study the principles.

Research which chart types suit which comparisons: change over time, composition of a whole, comparison between categories, distribution, and relationship between two variables.

Write a short reference note for yourself on when each is appropriate.

Now build five charts from your cleaned data, each answering a different question and each using a different chart type.

For every chart, write down the question it answers before you build it, so the chart serves the question rather than the other way around.

Apply accuracy rules strictly.

Start value axes at zero wherever truncating would exaggerate differences, and where you do not start at zero, state why clearly.

Label both axes with units.

Credit the data source on every chart.

Use colour to carry meaning rather than decoration, and ensure the chart still reads if printed in greyscale.

Then build one deliberately misleading version of any chart, using a truncated axis, a distorting scale, or selective date range.

Beside it, place the honest version.

Write a note explaining each chart type choice, and a separate analysis of the misleading version explaining exactly what the distortion does to a reader's impression.

Then test each chart on someone unfamiliar with the data.

Show it without explanation and ask what they think it says.

Where their reading differs from what the data actually shows, the chart has failed, and the fix is usually a clearer title, better labelling, or a different chart type rather than more decoration.

Write a title for every chart that states the finding rather than describing the contents.

A title reading Rainfall by district tells a reader nothing, whereas one stating which districts recorded the sharpest decline does the work the chart exists for.

Finally, strip each chart of anything that carries no information, including unnecessary gridlines, decorative shading, and three dimensional effects, which distort proportions and never aid comprehension.

Submit the five charts with their finding based titles, the misleading pair, the chart type reference, the reader test results, and both notes.

Data journalism fails when it presents numbers without meaning.

The craft lies in finding the human consequence inside a dataset and leading with that while keeping the analysis rigorous.

This task builds on your earlier cleaning and charting work and needs no new material.

Take your strongest finding from the cleaned dataset.

Before writing, test it properly.

Check whether the pattern holds across the full period or only a convenient slice.

Check whether it holds across categories or is driven by one unusual entry.

Check whether a simple explanation such as population change or a definitional shift accounts for it.

Write down the result of each test, because a finding that fails these checks must be abandoned rather than published.

Now research the context.

Find published reports, official statements, or academic work that explain why this pattern might exist.

Identify what experts in this field have said about it.

Write a news story of 800 to 900 words.

Lead with what the data means for people rather than with the number itself.

Introduce the figure early but explain it in plain language.

Include your strongest two charts positioned where they support the argument.

Quote at least two expert or official sources from published material, properly attributed.

State the limitations of the data honestly within the story rather than burying them.

Be disciplined about the language of causation.

Data shows that two things moved together, which is not the same as one causing the other, yet the words because of, led to, and resulted in slip into data stories constantly.

Review every sentence in your draft and replace any causal claim your evidence does not support with accurate wording such as coincided with or is associated with.

Then handle the numbers for readability.

Round to the precision that actually matters, convert raw figures into rates or proportions where that aids comparison, and give at least one number a concrete comparison a reader can picture.

Finally, write a short note identifying the single sentence in your story most likely to be challenged and stating exactly what evidence you would produce to defend it.

Submit the story with charts embedded, your verification test results, the causation review, the defence note, and a complete source list with links.

Serious data journalism publishes its method alongside its findings, so readers can judge the work and other journalists can reproduce it.

Writing this document also forces you to confront the weaknesses in your own analysis.

This task completes your data project.

Write a full methodology note covering your entire process.

Begin with the data source, including who published it, when, what it covers, and the exact link.

Record how you obtained it and in what format.

Document every cleaning decision from your log and explain the reasoning behind each.

Where you excluded rows, state how many and why.

Where you standardised entries, give examples of what was merged.

Set out your analytical method precisely enough that someone else could repeat it: the calculations performed, the comparisons made, and the tools used.

Then write the limitations section honestly.

State what the data does not capture, such as unrecorded cases or excluded populations.

Identify where definitions changed over the period and how that affects comparison.

Note the assumptions you made and how a different reasonable assumption would change your conclusion.

State plainly what your story does not prove, since the gap between correlation and cause is where most data stories go wrong.

Close with a reproducibility section listing every file, the order to use them in, and the steps to follow.

Write it for a sceptical reader rather than a sympathetic one.

Assume the person reading wants to find the flaw in your work, and answer their objections before they raise them.

This is a different exercise from explaining your process to a colleague, and it produces a far more useful document.

Then add a section on the decisions that could reasonably have gone another way.

At several points in any analysis there is a defensible alternative choice, such as a different time period, a different grouping, or a different treatment of missing values.

Name these points and state how a different choice would have altered your result.

Finally, include a correction commitment stating how you would handle it if an error were found after publication, and invite readers with relevant expertise to identify problems with the method.

Submit a methodology note of about 900 words written for publication alongside your story, together with the alternative decisions section and the file inventory.