Development work is full of things we care about but cannot observe directly: a child's reading ability, a woman's empowerment, a household's food insecurity. We only ever see responses — answers to items, ticks on a scale.

The obvious approach — add up correct answers, or count 'yes' responses — treats every item as equal and every score gap as the same size. But a hard item is not worth the same as an easy one, and the jump from 4 to 5 correct is not the same as 8 to 9.

Classical Test Theory is the world of the total score: number correct, percentage right, the count of 'yes' responses on a scale. It is what nearly every report and dashboard already uses.

Cronbach's alpha estimates reliability from a single sitting by asking how consistently the items hang together. Rules of thumb put 'acceptable' around 0.70 and up — but the number is widely misread.

An item's CTT 'difficulty' is just the proportion who got it right (the p-value). Give the same item to a high-ability school and a struggling one and that proportion changes — so the item looks 'easier' or 'harder' depending on who sat it.

A person's CTT ability is their score on this particular test. Put them on an easier form and the score climbs; a harder form and it falls. The person's standing is tangled up with the difficulty of the form they happened to take.

CTT reports a single standard error of measurement for everyone. But in reality a test measures a middling student far more precisely than it measures a top scorer (who finds every item easy) or a very weak one (who finds every item hard).

The central move of IRT is deceptively simple: place each person's trait (θ) and each item's difficulty on one shared continuum. A person is somewhere on the line; so is every item.

By convention θ is scaled to have mean 0 and standard deviation 1 in the reference group, usually running from about −3 to +3. Negative means lower ability / less of the trait; positive means more.

IRT never says a person will get an item right. It gives the probability of a correct (or endorsing) response, as a function of where the person and the item sit on the scale.

That probability follows a smooth logistic function of θ. Far below the item, P is near 0; far above, near 1; in between it climbs through an S-shape. This curve is the heart of IRT.

For a food-insecurity or empowerment scale there is no 'right' answer — the curve gives the probability of endorsing the item ('yes, we worried about food'). Severe items are endorsed only by households high on the insecurity continuum.

Everything that follows is detail on that one S-curve: where it sits (difficulty, b), how steep it is (discrimination, a), whether it has a floor (guessing, c), how much information it carries, and whether it behaves the same for everyone (DIF).

A valid ICC is monotonically increasing: more of the trait never lowers the chance of a correct response. If a fitted curve dips or wiggles, something is wrong — a mis-keyed item, a trick question, or a broken assumption.

Given a person's pattern of right and wrong answers and the fitted ICCs, the software finds the θ that makes that pattern most likely. That θ — not the raw count — is the person's IRT score.

Two children each get 6 of 12 right. One passed the six easy items; the other passed six hard ones and missed easy items. CTT calls them equal. IRT, reading which items, places them at different θ.

Because b is expressed in θ units, you can ask: is this person above or below this item? A child at θ = 0.5 is comfortably above an easy item (b = −1) but below a hard one (b = +1).

On a food-insecurity scale, b is better read as severity: how much insecurity it takes before a household endorses the item. 'We worried about food' has a low b; 'a household member went a whole day without eating' has a high b.

b is empirical: it is set by how people actually respond, not by how hard the item looks. A question that seems advanced may be easy if everyone was taught it; a 'simple' item may be hard if the wording confuses people.

Near its difficulty, a high-a item changes a person's success probability a lot for a small change in θ. That sensitivity is exactly what lets it tell two nearby people apart — it carries more information (next section).

A low-a (flat) item gives almost the same success probability to people across a wide range of θ. Low-trait and high-trait people answer it similarly — so it adds little to distinguishing them.

If an item shows negative discrimination, higher-trait people do worse on it than lower-trait people — the ICC slopes downward. That should never happen for a sound item.

On an empowerment scale, a high-a item is one whose endorsement cleanly separates more- from less-empowered women near its severity. A low-a item — one answered similarly regardless of empowerment — is dead weight.

An extremely steep item measures superbly — but only in a razor-thin band of θ. A test built only of very steep items can measure one narrow region brilliantly and everywhere else poorly.

On a 4-option multiple-choice item, even a child who knows nothing has roughly a 1-in-4 chance of being right. So the ICC should not fall to zero at low θ — it should flatten out at a lower asymptote.

If you fit a model with no guessing parameter to multiple-choice data, the floor created by guessing gets misread as the item being 'easier' than it is — biasing b and distorting low-ability scores.

The 1PL / Rasch model fixes a to be the same for every item, so items differ only in difficulty (b). The ICCs are parallel S-curves — same shape, shifted left or right.

The 2PL frees a, so each item has its own slope as well as its own difficulty. ICCs can now be steep or flat, crossing one another. It fits more datasets but needs more respondents.

The 3PL adds c, the guessing asymptote — built for multiple-choice ability tests where lucky guesses are real. It is the most flexible of the three but the hungriest for data and the trickiest to estimate.

The test information function is just the sum of the item information functions. Where many items pile up, the test measures precisely; where items are sparse, it measures poorly.

Precision and information are two sides of one coin: the standard error of measurement at any θ is 1 / √information. High information → small standard error; low information → large error.

If a programme classifies children as 'at grade level' or not, the decision happens at one θ — the cut-score. That is exactly where you want maximum information and minimum error.

Because items and people share a scale, a computer-adaptive test can pick each next item to be maximally informative at the test-taker's current θ estimate — honing in fast.

Sample-free items, test-free scores, a clean common metric — these gifts depend on the model actually fitting the data. Validation is the work of checking that it does.

Fit asks whether the observed responses match what the fitted ICCs predict — overall and item by item. A badly fitting item's empirical curve departs from its modelled S-curve.

If girls genuinely have higher reading ability than boys, they will score higher — that is a real trait difference, not DIF. DIF is when girls and boys at the same ability still differ on a specific item.

Large learning assessments — India's NAS, citizen-led ASER, and cross-national studies — lean on IRT to place children on a single proficiency scale and to compare across grades, states and years.

To track learning over years you must change the questions (security, age-appropriateness) without changing what the score means. Equating via anchor items keeps the ruler fixed while the items rotate.

The Food Insecurity Experience Scale (FIES) — eight yes/no experience items — is modelled with a Rasch/1PL approach so that severity is comparable across countries and languages, underpinning SDG indicator 2.1.2.

Women's empowerment indices and agency scales combine items on mobility, decision-making and asset control. IRT checks whether they measure one coherent trait, ranks items by severity, and flags items that work differently across regions or castes.

Asset-based wealth indices (the kind behind NFHS/DHS wealth quintiles) ask whether a household owns particular assets. IRT and related latent-trait methods place households on a wealth continuum from the pattern of ownership.

A θ of 1.2 means nothing to a programme officer or a parent. Part of doing IRT well is translating the scale back into language people act on — proficiency bands, 'can read a paragraph', percentile, or a clear cut-score.