Replication in behavioural science fails mostly for accounting reasons. The second study costs nearly as much as the first, and almost no funding line is written to pay for it. What follows is how that arithmetic shapes whether a finding survives, who pays for the second try, and what would make repeating a study cheap enough to be routine.

The Case That Would Not Die

A behavioural finding usually enters the field as a preprint: a tidy effect, a striking graph, a sample of a few dozen students. It gets shared widely, cited early, and folded into lecture slides before any journal has accepted it. The original lab often reports a second successful experiment in the same paper, which makes the result look sturdier than a single run.

Then independent labs try to repeat it. Some report effects in the same direction but smaller. Some report nothing. The split rarely tracks method quality in an obvious way. It tracks resources. Groups with standing participant pools and salaried staff can afford a high-powered attempt; groups running on a single small grant cannot, and their underpowered failures get read as noise rather than evidence.

This site has argued that funding panels weigh a null result before the outcome is known, and the pattern holds here. A replication is judged on whether it will produce something publishable, not on whether it will produce something true.

Who Pays for the Second Try

Replication grants exist, but they are small and scarce relative to the original work they check. A typical replication award covers a fraction of what the first study cost, and deadlines are annual. A lab that misses the window waits a year, by which time the postdoc who would have run it has moved on.

Funders reward novelty. Review criteria in most national and charitable schemes ask what is new, what is innovative, what has not been done. Confirming someone else's result scores badly on every one of those questions. A researcher who spends a grant cycle on replication has a thinner CV than one who spends it on a new claim.

Career credit compounds the problem. A null finding is citable in principle and nearly invisible in practice. Hiring committees count first-author papers, and a well-run failure is still a failure in that count. The rational move for an early-career researcher is to avoid the second try entirely.

Consider the view from a mid-career lab head deciding where to place a talented doctoral student. Assigning that student to replicate a contested effect is a service to the field, but it is also a bet against the student's own job prospects. The student will graduate with a thesis chapter that few outside a narrow circle will cite, while peers who chased a novel effect will have a headline result to trade on. The lab head knows this, and the knowledge shapes the assignment even when no one says it aloud.

Infrastructure Costs Nobody Budgets

Participant recruitment is the bottleneck. A high-powered replication of a small effect can need several times the original sample, recruited from a population that has already seen the task. Online panels help, but attention checks and bot filtering eat into usable data, and the per-participant cost has risen across most platforms. For a medium-sized online study, the fee per completed response often lands in the low single digits of dollars, and that figure can double once you account for participants who fail attention checks and must be replaced.

Lab time competes with new projects. A behavioural lab has finite hours in a week and finite research assistants. Every session spent repeating a 2014 result is a session not spent on the study that might win the next grant. The opportunity cost is real and it is calculated, quietly, in every lab meeting.

Data sharing adds legal overhead. Ethics approvals, consent forms and data transfer agreements were drafted for the original study, not for a second lab in another country. Statisticians are scarce and expensive, and the analysis plan for a replication is often more demanding than the original because the question is narrower.

Equipment is a smaller but stubborn line. Reaction-time studies need reliable hardware and software licences; a replication that switches from a lab keyboard to an online browser introduces timing variability that can swamp the effect under test. The original lab may not have documented its exact setup, and chasing those details consumes weeks before a single participant is run.

How a Verdict Actually Forms

Meta-analyses pool the attempts and weight them by sample size. That weighting is defensible and it has a known bias: small original studies that happened to find a large effect get outweighed by larger replications, but the pooled estimate still drifts toward whatever got published first, because published first is what exists.

Publication bias tilts the record. Journals have historically preferred positive results, so the file drawer holds the failures. Correcting for that bias statistically is possible and contested; the corrections move the estimate, and reviewers disagree about which correction is right, which delays the verdict further.

Pre-registration changes the record slowly. Registering a protocol before data collection removes some degrees of freedom and makes an undisclosed analysis visible. It does not fund the study, and it does not speed up the years it takes for a body of pooled evidence to settle. Consensus lags the evidence by roughly a decade in most behavioural literatures.

Two labs settling a measurement question between them is a different problem from a field settling a behavioural claim, as a related piece on cold atoms settling a quasiparticle mass debate illustrates. Physics can sometimes buy its way to closure with one decisive instrument. Behavioural science mostly cannot.

Even when a verdict does form, it rarely travels back to the people who used the original finding. A textbook that cited the 2014 effect will not issue a correction; a popular book that built a chapter on it will keep selling. The lag between a meta-analytic conclusion and its removal from teaching materials can run another five to ten years, which means a generation of students learns the effect and only later, if ever, learns that it failed to replicate.

What Would Make Replication Cheaper

Shared participant pools across institutions would cut the largest line item. A consortium that maintains a vetted panel and allocates sessions to member labs turns recruitment from a per-study scramble into a standing service. Some university networks already do a version of this for online studies; the model has not spread to the labs that need it most.

Standing replication funds with fast review would help more than large annual calls. A rolling pot with a short application, reviewed in weeks rather than months, matches the pace at which a suspicious result appears. The objection is real: fast review means less scrutiny, and a badly designed replication can do as much damage as a badly designed original. The counter is that a slow fund mostly funds nothing.

Journals that publish null findings give the work a citation. Templates that cut setup time, from shared consent language to standard analysis scripts, lower the fixed cost of starting. None of these fixes the incentive at the top, where novelty is still what gets funded.

Another route is to treat replication as a service that labs sell to each other rather than as a project each lab must fund alone. A lab that has already built the task and the recruitment pipeline can run a repeat for a fraction of the original cost, and it can charge that fraction to the lab that wants the check. The arrangement is rare today because no standard contract exists for it, but the underlying economics are straightforward.

Three Moves for Working Researchers

Budget the replication into the original grant. A line item for an independent repeat, named in the application, is harder to cut later than a request made after the result is contested.

Name the lab that will try to repeat the study before data collection begins. A named collaborator with a protocol in hand turns an intention into a commitment that survives the end of the grant cycle.

Publish the protocol and register it publicly. A registered protocol is citable, and it gives whoever repeats the work a fixed target instead of a moving one.

Treat a null result as a citable output. Put it on the CV, list it in the annual report, and cite it in the next grant application as evidence of rigour rather than as a gap.

Ask funders for a standing replication line. Programme officers hear from applicants, and a request repeated across enough applications becomes a line in the next call. A related piece on spectrograph throughput setting the cost of a survey night makes the same point about instruments: the price of a second look is set long before anyone decides to take it.