Development work runs on numbers — targets, indicators, budgets, surveys, dashboards. Data literacy is the ability to read, question, interpret and communicate data, and to use it to make better decisions. It is not statistics for its own sake; it is judgement.

Every dataset embeds choices: what to measure, what category to use, whom to ask, what to ignore. Those choices carry power. A data-literate practitioner reads the choices, not just the digits.

Most development M&E lives in structured data, but a growing share — call-centre logs, photos, social media, satellite imagery — is unstructured and needs different tools.

These are among the largest demographic and health surveys in the world. Their size is what lets them speak reliably about districts — but not about your single panchayat.

Every scheme produces administrative data as a by-product of delivery: MGNREGA muster rolls, school enrolment (UDISE+), health records (HMIS), ration transactions, immunisation registers.

Mobile-phone records, satellite night-lights, transaction logs and remote sensing increasingly supplement official statistics — useful where surveys are slow or coverage is thin.

Most things we care about — poverty, empowerment, health, learning — are concepts, not numbers. Measurement is the bridge from an abstract concept to an observable indicator.

A proxy stands in for something hard to measure. Household assets proxy for wealth; night-light intensity proxies for economic activity; mid-upper-arm circumference proxies for acute malnutrition.

Indices like the Human Development Index or the Multidimensional Poverty Index (MPI) combine several indicators into a single score for easy comparison.

NITI Aayog's National MPI bundles 12 indicators across health, education and standard of living — nutrition, child mortality, schooling, cooking fuel, sanitation, housing, assets and more.

Before any fancy analysis, describe the data. Two questions answer most of it: what is typical (central tendency) and how much do values vary (dispersion).

A percentile is the value below which a given share of cases fall. The 25th percentile (Q1) has a quarter of households below it. Wealth quintiles — five 20% bands — are how NFHS and NSS routinely report inequality.

Many natural measurements (height, birth weight, measurement error) follow a roughly normal distribution — symmetric, bell-shaped, defined by its mean and standard deviation (SD).

A raw count without its denominator is almost meaningless. '500 dropouts' could be a crisis or a triumph depending on whether the base is 600 or 60,000.

Visualisation is not decoration. A well-made chart reveals patterns the table hides — trends, gaps, outliers, relationships — and lets a busy decision-maker grasp them in seconds.

Edward Tufte's principle: maximise the data-ink ratio. Every gradient, shadow, 3-D effect and clip-art icon competes with the data for attention — and usually wins.

Instead of cramming ten states into one tangled line chart, draw ten tiny identical charts side by side — small multiples. The eye compares shapes effortlessly when scale and layout are shared.

The correlation coefficient r runs from −1 (perfect negative) through 0 (none) to +1 (perfect positive).

Two variables can move together for several reasons, only one of which is 'A causes B'.

Ice-cream sales correlate with drowning deaths. Ice cream does not cause drowning — summer heat drives both. The confounder is the real story.

Test enough pairs of unrelated variables and some will correlate strongly by sheer chance. A tight correlation is a clue, never a proof.

Anscombe's quartet is four datasets with identical means, variances and correlation (r = 0.82) — yet utterly different shapes: one linear, one curved, one a single outlier driving everything.

A district with higher average income may have higher literacy on average — that does not mean the richer individuals within it are the literate ones. What holds for districts need not hold for people.

Simpson's paradox: a relationship visible in the whole dataset can flip when you break it into subgroups. A scheme can look worse overall yet be better in every region — if the regions differ in size and baseline.

When you pick the worst-performing districts and re-measure them later, they usually look better — even with no intervention. Extreme values contain extra luck that does not repeat. This is regression to the mean.

A well-chosen sample of a few thousand can describe a population of millions — the principle behind NFHS, NSS and every opinion poll. The magic is not size; it is representativeness.

A counter-intuitive truth: for a large population, accuracy depends on the absolute sample size, not the share of the population sampled. 1,500 people describe a state and a country about equally well.

When some groups are deliberately over-sampled (to study them reliably) or respond less, surveys apply weights so each respondent represents the right number of real people.

Analysts often spend the majority of a project just preparing data — finding errors, reconciling formats, handling gaps. Glamorous analysis sits on a large, unglamorous foundation of cleaning.

Metadata is data about your data: what each variable means, its units, allowed values, how and when it was collected, and what you changed.

A survey figure of 42% is shorthand for 'about 42%, give or take'. The confidence interval — say 39–45% — is the honest version. A point estimate without its range overstates certainty.

Even a 99%-accurate test for a rare condition produces mostly false positives — because the healthy vastly outnumber the sick. Ignoring the underlying rate is one of the commonest reasoning errors.

If coverage rises from 40% to 50%, that is a 10 percentage-point increase — but a 25 percent increase. Mixing the two is a classic way to exaggerate or hide change.

'Cases doubled!' sounds alarming — but a rise from 2 to 4 is a doubling of almost nothing. Relative change without the absolute base is designed to impress, not inform.

If you slice the data twenty ways, roughly one slice will show a 'significant' result by chance. Reporting only that slice — p-hacking — manufactures false findings.

In development data, rows are people — often poor, often without power over how their information is used. Data ethics is not paperwork; it is respect made operational.

Removing names is not enough. A combination of village + age + caste + occupation can re-identify one person — especially in small areas where few people share those traits.

The DPDP Act, 2023 is India's first comprehensive data-protection law. It sets duties for anyone handling personal digital data — including NGOs and researchers.

An average hides the people behind it. A programme can report good overall numbers while failing Dalit, Adivasi, disabled, or women beneficiaries. Disaggregation is how inequity becomes visible.

Communities are often extracted from — surveyed repeatedly, with the knowledge and value flowing to outside institutions while respondents see nothing back.