Writing what the data can defend
Evidence, framing, bias and the craft of data journalism
00:00–00:30 · 0.5 min
Central argument
Evidence sets the ceiling for a claim.
Concept
The task is epistemic calibration, not elementary prose style.
Example
Students bring already published data stories.
Say / demonstrate
Connect to their submitted sites. Today’s standard is the strongest consequential statement the evidence warrants, not the loudest possible interpretation.
Ask
Which sentence in your own story would you least want an editor to challenge?
Defensible answers
Any specific sentence is useful; preserve the question for the homework.
Common misconception
A more cautious-sounding sentence is necessarily a more rigorous one. Precision can support forceful conclusions.
Transition
Start with two accurate summaries that create different impressions.
The numbers can be right.
The story can still be wrong.
One accurate account
+20%Recorded cases rose from 100 to 120.
Another accurate account
−20%Cases per resident fell from 10 to 8 per 1,000.
00:30–01:45 · 1.25 min
Central argument
Accuracy at the level of individual numbers does not guarantee an accurate overall impression.
Concept
Estimand choice: absolute burden and population-standardised risk answer different questions.
Example
100/10,000 × 1,000 = 10; 120/15,000 × 1,000 = 8. Counts grow 20%; the rate falls 20%.
Say / demonstrate
Ask students to defend both statements. Demand the reporting question before choosing the lead. Check that coverage, case definitions and exposure are comparable.
Ask
Which account would you lead with for hospital capacity? For an individual’s risk?
Defensible answers
Capacity can justify the count; resident risk may justify the rate, if population is the relevant exposure. Both may belong in the story.
Common misconception
Rates are always more truthful than counts, or one of these statements must be mathematically false.
Transition
Separate what the data shows from the explanation we attach to it.
Analysis is not argument
- Observation
- Average transaction value falls from ₹100 to ₹60.
- Interpretation
- More small payments may be entering the system.
- Overclaim
- Indians are abandoning cash for everyday purchases.
What evidence would connect payment records to cash substitution?
01:45–03:15 · 1.5 min
Central argument
Observation, mechanism and publishable claim are distinct.
Concept
A mean is a ratio; composition, user entry and payment substitution are different explanations.
Example
₹10bn/100m = ₹100; ₹9bn/150m = ₹60. No cash-use observations appear in the example.
Say / demonstrate
Ask students to list competing explanations: a change in merchant mix, user mix, ticket sizes or large transfers. Cash substitution needs evidence about the alternative channel and preferably the same people over time.
Ask
What would you request before writing the cash sentence?
Defensible answers
Representative payment diaries or household/merchant surveys, transaction composition, coverage over time, and evidence of substitution rather than additional transactions.
Common misconception
An observed fall in the mean identifies its cause. “May” can flag a hypothesis but does not establish it.
Transition
Rank different kinds of claims by what they require.
The verb carries the burden of proof
A measured change
A valid comparison
A joint pattern
A causal role, even if partial
Causal evidence, not co-movement
A justified forecast and mechanism
05:45–07:00 · 1.25 min
Central argument
Verbs encode commitments about relationships.
Concept
Calibration of causal language and the evidential status of hypotheses.
Example
Training and employment rise together; that supports neither “drove” nor a causal forecast by itself.
Say / demonstrate
Have students underline the verb in a draft. “Contributed to” is not a harmless descriptive fallback. Qualifiers should communicate uncertainty in an evidence-backed claim, not disguise missing support.
Ask
Can “could have caused” ever be useful?
Defensible answers
As an explicitly identified hypothesis, with a reason to consider it and alternatives; not as a substitute for causal evidence in the finding.
Common misconception
Every causal verb requires a randomised trial, or adding “may” fixes an unsupported causal claim. Credible non-experimental designs exist.
Transition
Put the proposed explanation through a counterfactual test.
Decisions behind every data story
- QuestionWhose problem becomes visible?
- DatasetWho collected it, and why?
- InclusionWhich people disappear?
- DenominatorRisk, scale or exposure?
- TimeframeWhich baseline is fair?
- ComparisonAgainst what benchmark?
- ConclusionWhich sentence survives the checks?
09:15–10:30 · 1.25 min
Central argument
Editorial perspective is unavoidable; arbitrary and concealed choices are avoidable.
Concept
Transparency, robustness and analytical degrees of freedom.
Example
An “employment” story can study registered formal employees or all working-age adults; neither file chooses the reporting question.
Say / demonstrate
Ask students to name one decision they made in their submitted story. The standard is whether a reasonable alternative changes the result, and whether the choice is defensible for the question.
Ask
Which choice could reverse your own headline?
Defensible answers
Plausible answers include sample exclusions, population denominator, baseline or subgroup weighting. Ask for a concrete sensitivity check.
Common misconception
Objectivity means having made no choices. Revealing a choice does not by itself justify it.
Transition
Test one of the most consequential choices numerically.
The timeframe is an editorial decision
2017–2025: index 100 → 114, a rise of 14%.
12:00–13:30 · 1.5 min
Central argument
Period selection needs an editorial rationale and robustness checks.
Concept
Base effects, structural breaks, seasonality, data revisions and selecting on outcomes.
Example
Index values: 100, 102, 106, 70, 85, 105, 110, 112, 114. Window changes yield +14%, +62.9% and −0.9%.
Say / demonstrate
Toggle the windows. A disruption year can be appropriate for measuring recovery but not necessarily long-run improvement. Establish the question and period before seeing the preferred answer. Keep the longer context visible and record the data release. Annual data cannot diagnose monthly seasonality.
Ask
When is starting at the trough legitimate?
Defensible answers
When recovery from that documented disruption is the question, with the baseline explained and the longer view available. It is misleading if chosen to imply ordinary growth.
Common misconception
Every unusual baseline is forbidden; annual trends rule out seasonal effects; revisions cannot change earlier comparisons.
Transition
Ask whose observations the time series contains.
Two correct numbers.
An invalid comparison.
Survey A · ₹6,000
- Concept
- Disposable income
- Unit
- Per person, per month
- Coverage
- Formal workers, urban sample
- Period / prices
- 2024, current rupees
- Design
- 500 respondents
Dataset B · ₹180,000
- Concept
- Consumption
- Unit
- Per household, per year
- Coverage
- Registered households, old district
- Period / prices
- 2021, constant 2011 rupees
- Design
- All registered records
15:00–16:30 · 1.5 min
Central argument
Arithmetic cannot repair incompatible constructs and populations.
Concept
Measurement validity, target-population alignment and comparability.
Example
The two displayed amounts are not alternative estimates of the same quantity.
Say / demonstrate
Ask what transformations are possible and which mismatches remain. Annual-to-monthly division is straightforward; household-to-person conversion needs household information and proper weights. Inflation adjustment needs an appropriate index. A registry census is not necessarily coverage of the entire population.
Ask
After dividing B by 12, are the figures comparable?
Defensible answers
No. The construct, household unit, coverage, geography, prices and years still differ. A precise conversion can leave an invalid comparison.
Common misconception
A common currency or date makes income and consumption identical; a census of registered units is a population census.
Transition
Now compare distributions with identical means.
Poorest 10% In Rich Countries Earn More Than India's Richest 10%: World Bank
NDTV Business · Edited by Prateek Shukla
Published and updated 12 October 2026, 08:45 IST
What’s the strongest claim this headline makes?
30:00–31:30 · 1.5 min
Central argument
First reconstruct the claim before evaluating it.
Concept
Headline semantics: population, measure, group comparison and attribution.
Example
The headline is accurately typeset from the article as retrieved on 20 October 2026; it is not a fabricated screenshot.
Say / demonstrate
Ask students what “earn,” “richest” and a comparison between groups would ordinarily imply. Do not reveal the answer key or subsequent calculations yet. Separate what the headline says from what a student assumes the dataset contains.
Ask
Is the headline about cutoffs, averages, typical people, or every member?
Defensible answers
Students may read it as a group-level or typical-person comparison. The wording is ambiguous; record those readings rather than declaring a single mathematical proposition.
Common misconception
Every reader interprets a group comparison identically; an article’s attribution means the institution wrote the headline.
Transition
Read the whole article before judging its evidence.
Read the article like an editor
31:30–34:30 · 3 min
Central argument
Distinguish reported statistics, source interpretation and policy recommendations.
Concept
Source criticism and provenance of each step in a published argument.
Example
Use the original article, not a slide summary or its automated quick-read panel.
Say / demonstrate
Allow silent reading. Ask students to label a sentence’s evidential role, not whether they agree with its politics. Notice where a descriptive distribution comparison becomes an explanation or a recommendation.
Ask
Which sentence requires a different kind of evidence than the numerical comparison?
Defensible answers
An explanation or policy recommendation requires evidence beyond the distribution table. Attribution makes an expert opinion visible but does not validate it.
Common misconception
The assignment is to find as many errors as possible; a quoted interpretation must be either an established fact or a lie.
Transition
Rebuild the source comparison from its underlying records.
Reconstruct the comparison
India · upper-decile entry
90% at or below the cutoff.
Upper 10% begins above it.
US · bottom-decile cutoff
Bottom 10% lies at or below it.
90% lies above it.
| Selected national row | Year · measure | P10 | P90 |
|---|---|---|---|
| India | 2022 · C | 3.40 | 9.98 |
| US | 2022 · I | 25.13 | 172.99 |
| Canada | 2022 · I | 31.12 | 128.89 |
| Germany | 2022 · I | 28.44 | 127.86 |
| Australia | 2020 · I | 28.08 | 128.89 |
2021 international $ / person / day. C = consumption; I = income. Rank strips are not welfare scales.
34:30–37:30 · 3 min
Central argument
A percentile threshold locates a boundary, not the average of the adjacent group.
Concept
Weighted quantiles, survey reference years and the distinction between population ranks and value scales.
Example
The displayed P10 and P90 values come from national quantile rows in the retained 2021-PPP extract. Both PPP vintages were inspected.
Say / demonstrate
Explain the shaded population shares. India’s upper decile begins above its P90; the US lower decile lies below its P10. The strips are percentile ranks, not density curves. The table documents all four comparators without pretending their survey years are contemporaneous.
Ask
Where are the values for people inside the shaded groups?
Defensible answers
The boundaries do not provide those values. Group means need the distribution above or below the cutoff; the source also supplies bin averages and population shares.
Common misconception
A 2026 dataset release means all observations were collected in 2026; the two shaded blocks have the same monetary width.
Transition
Test exactly what follows from ordering those boundaries.
A threshold is not a group average
India’s P90 = $9.98/day
90% of Indians are at or below this value. The richest 10% are above it.
US P10 = $25.13/day
10% of Americans are at or below this value. The remaining 90% are above it.
Population ranks, not income bars. Selected World Bank PIP rows: India and US, 2022; 2021 PPP $ per person per day. These cutoffs vary by year.
What the data shows
The cutoff for entering India’s richest 10% is lower than the upper boundary of America’s poorest 10%.
What it does not show
That every person in India’s richest 10% earns less than every person in America’s poorest 10%.
A percentile tells you where a group begins or ends, not what everyone in that group earns.
Approximately 2 minutes
Explain
P90 and P10 are population-rank boundaries. The strips show shares of people, not welfare amounts. India’s selected 2022 consumption P90 is $9.9768; the US selected 2022 income P10 is $25.1296, in 2021 PPP international dollars per person per day. These are selected survey rows, not universal or latest-year values. “Richest” here means ranked by the survey’s monetary welfare measure, not wealth.
A cutoff is not the average of the people above or below it. Comparing two cutoffs cannot by itself establish decile means, still less an ordering of every individual. The headline overstates what the cutoff comparison alone proves and uses earnings language for India’s consumption measure. Do not declare its numerical conclusion false: separate population-share-weighted calculations from the selected PIP percentile bins give India’s top-decile consumption mean $14.07 and the US bottom-decile income mean $15.24. That ordering persists, but it compares different welfare concepts.
Ask
Do we know what every person in either decile receives? No. We know the boundaries; group means require separate calculation, and neither statistic describes every member.
Same dollars. Different measurements.
India
Consumption per person, per day
United States
Income per person, per day
Consumption measures what households use or spend. Income measures resources received. They are related, but not identical.
₹80,000/month ≠ $9.98/day
₹80,000: an individual earnings claim in nominal rupees.
$9.98: a percentile cutoff for per-person household consumption in 2021 PPP dollars.
A direct comparison requires matching definitions, populations, time periods and units.
Before comparing two numbers, check what each one actually measures.
Approximately 2 minutes
Explain
In the selected 2022 rows India measures consumption and the US measures income, both as daily household welfare per person. Per-person allocation does not mean every household member personally earns that amount. Consumption includes valued use, not just cash purchases; income can include transfers and non-labour resources. Savings and borrowing can separate consumption from current income.
PPP adjusts for price-level differences using a common 2021 price basis; it is not today’s market exchange rate. It does not harmonise survey concepts, years, populations or coverage. Here the welfare reference year is 2022 for both countries, distinct from the PPP reference year and release date. Household surveys can miss non-household populations and underrepresent the upper tail; do not invent a correction. The matching rows reproduce the article’s rounded values, but its exact original export remains unconfirmed.
The article attributes the ₹80,000 personal-earnings claim to the Economic Survey without an edition or table. A search of the official 2025–26 report did not locate that claim; this does not prove it false. Request the edition, page/table, reference period, earnings definition, ranked population and statistic. Converting currencies or days to months cannot reconcile individual earnings with per-person household consumption.
Ask
Does putting both figures in PPP dollars make them the same measure? No: common purchasing-power units do not turn consumption into earnings.
Be bold about what you know.
Be precise about what you don’t.
29:30–30:00 · 0.5 min
Central argument
The goal is disciplined ambition rather than maximal or timid claims.
Concept
Calibrated confidence and the boundary between observation and interpretation.
Example
The next section applies the same tests to a published article.
Say / demonstrate
Close the lecture at 30 minutes. Do not preview the case-study critique. Invite students to interrogate the headline before seeing the reconstruction.
Ask
What is the difference between a weak story and a bounded claim?
Defensible answers
A bounded claim can be consequential and well supported. A weak story may lack evidence or public relevance, regardless of confidence in its tone.
Common misconception
Precision drains all narrative power.
Transition
Read the published headline on its own terms.