Parallel forms and item banks: why a good online test has many versions

If every candidate who takes an online test sees the same questions, those questions will not stay secret for long. Parallel forms and item banks are how test developers keep unsupervised online testing secure without making it unfair, and the evidence behind them is worth understanding before you choose a test.

The security problem with unproctored online testing

Most employment testing now happens online, on the candidate's own device, without a supervisor in the room. That brings real benefits in speed, cost and candidate convenience, but also obvious risks: candidates can share questions, collude or have someone else sit the test. A widely cited panel review of unproctored internet testing in employment settings concluded that a test giving every candidate an identical fixed set of items is especially exposed to compromise, and discussed safeguards such as drawing items from large pools and verifying scores under supervised conditions.^[1]

The International Test Commission's guidelines on computer-based and internet-delivered testing set out good practice for this environment, including the security of test content and the use of supervised follow-up testing where stakes are higher.^[2] The ITC's later guidelines on test security extend this to the whole life cycle of a test, from development to retirement of compromised items.^[3] Our article on whether candidates can cheat psychometric tests with AI covers a newer side of the same problem.

Parallel forms and item banks

The traditional answer is parallel forms: two or more versions of a test built to the same blueprint, measuring the same construct with different items. Strictly parallel forms have equal means, equal spread of scores, equal reliability and equal correlations with other measures, so that it makes no difference to a candidate which form they receive. In practice, developers aim for forms that are equivalent within close limits and then demonstrate it with data.

An item bank takes the idea further. Instead of a handful of fixed forms, the developer calibrates a large pool of items, usually with item response theory, so that every item's difficulty and discrimination are known on a common scale. Tests can then be assembled from the bank in several ways: many pre-built equivalent forms; linear-on-the-fly testing, where each candidate receives a fresh form built to the same specification; or adaptive testing, where items are selected one at a time to suit the candidate's ability.^[1] In each case, the share of candidates who see any individual item falls sharply, which limits the value of leaked content.

Equivalence and equating

Giving candidates different versions only works if their scores mean the same thing. That is a measurement question, not just a writing one. Even carefully matched forms will differ slightly in difficulty, and a candidate should not be advantaged or penalised by the luck of the draw.

Equating is the statistical process that adjusts for these differences so that scores on different forms are interchangeable. Kolen and Brennan's standard text sets out the main data-collection designs, such as giving different forms to randomly equivalent groups, or linking forms through a set of common anchor items, and the methods used to place scores on a single scale.^[4] With an IRT-calibrated bank, equating is effectively built in, because all items share one scale, but the calibration itself must be checked, and items monitored over time for drift.

This matters for norms too. A percentile or band is only meaningful if it applies regardless of which form a candidate took, so the norms for a multi-form test should be built on equated scores. For an employer, the questions to ask are straightforward: how many forms or how large an item bank is there, how was equivalence shown, and how are scores equated?

Randomised item sets: benefits and limits

Randomisation is not a free lunch. Simply shuffling item order on a fixed test does little to stop content sharing. Drawing items at random from a pool without regard to their difficulty can produce forms that are noticeably easier or harder for different candidates, which is unfair. Good randomisation is constrained: items are drawn to match a blueprint for content and statistical properties, and the resulting forms are checked for equivalence.^[1,4]

There is encouraging evidence from practice. A study of a large operational programme of unproctored cognitive ability testing, which used supervised verification testing, reported item exposure and test overlap within acceptable levels, and found that, when unusual score patterns were examined, any fall from unproctored to supervised scores was negligible.^[5] That is evidence that good design reduces the risk, not that the risk disappears.

What this means for employers

A well-built online test should have many versions behind the screen, with evidence that they are equivalent and equated. That protects the integrity of results for everyone, including honest candidates. It also supports retesting: with parallel forms, a candidate can re-sit or be verified under supervision without simply recalling answers. Whatever the delivery method, interpret scores in bands alongside interviews and other evidence, rather than treating any single sitting as decisive. If you use online ability tests, ask your supplier how their item bank is built and maintained.