Single-Item Measures

What it is

Most published psychology and epidemiology measures things with batteries: ten items for depressive symptoms, nine for a sense of life worth living, dozens for personality. A battery buys two things. It lets researchers report whether the items hang together, and it spreads the risk that any one wording is misread. A single item buys neither, and it is used anyway, constantly, for a reason that has nothing to do with rigour: questionnaire space. A postal survey going to fifty-five thousand people has a hard limit, every item added costs response rate, and a cohort designed in the 1980s to study diet and cancer was not going to hand nine of its items to a wellbeing construct.

So single items are not a mark of carelessness. They are a budget decision, and the interesting question is what a finding built on one answer licenses.

Three limits are structural rather than incidental. The first is that there is no level, so there is no dose. A yes-or-no item sorts a population into two groups and nothing more; a three-option item adds one intermediate group. Neither gives you a gradient, and without a gradient a central piece of evidence for a causal reading, that more of the exposure goes with more of the outcome, is unavailable in principle.

The second is that there is no object. A question of the form "do you have X?" records whether people say yes. It does not record what their X is. This sounds pedantic until the construct in question is one where the object is the whole argument. If one person's answer rests on a grandchild, another's on a research career and a third's on a church, a single item treats all three as the same exposure, and any claim about which kinds of X are good for you is simply not in the data.

The third is the uncertain middle, and it is the one researchers get caught by. When a three-option item offers yes, uncertain or no, the uncertain group can be large. When a later study simplifies the same question to yes or no, those people have to go somewhere, and nobody knows where. That is misclassification in the strict sense, and it biases towards finding nothing rather than towards finding something, which is a useful thing to know when a study comes back null.

Against all of that sits a genuine defence, and it should be made properly. A single item inside a well-designed prospective cohort is a very different object from a single item in an online survey. If the response rate is 94.6 per cent, the outcome is a death certificate coded to ICD-10 rather than a self-report, the model carries twenty covariates, the analysis is repeated with the first two years of deaths excluded, and the authors test for effect modification across nine strata and find none, then the weak link is the exposure and everything around it is strong. The honest verdict is specific: findings like that establish that the answers separate people. They do not establish what the answers were about. Those are different claims, and the gap between them is where popular writing does its damage, because a framework with no instrument at all can borrow the findings of a measured construct indefinitely without ever being exposed to a test.

In effect

The research case: the Ohsaki item, and 36.4 per cent in the middle

Toshimasa Sone and colleagues' 2008 cohort is the clearest worked example available. The exposure, verbatim from the paper: "Ikigai was assessed through the subject's response to the question, 'Do you have ikigai in your life?' The subjects were asked to choose one of three answers: 'yes,' 'uncertain,' or 'no.'" One question, three options, asked once in late 1994, to 43,391 adults aged 40 to 79, who were then followed across 269,989 person-years during which 3,048 of them died. The distribution is itself a finding: 59.0 per cent yes, 36.4 per cent uncertain, 4.6 per cent no. More than a third of a very large sample would not commit either way. That single item carries a fully adjusted all-cause mortality hazard ratio of 1.5, with a confidence interval from 1.3 to 1.7, and a stroke figure of 2.1. It also carries nothing whatever about what any of those 25,596 people who said yes were thinking of.

Where the researchers say it themselves: Okuzono and colleagues, 2022

The best statement of the problem in this literature was written by researchers about their own data. Sakurako S. Okuzono and colleagues used a binary version, "Do you have Ikigai?", yes or no, in the Japan Gerontological Evaluation Study, and made it the first of nine stated limitations: "First, our Ikigai measure was a single binary item that asked if respondents had Ikigai or not. Sone et al assessed whether participants had Ikigai by asking 'Do you have Ikigai in your life?' with three response options (yes/uncertain/no). In their analytic sample (n = 43,391), 36.4% indicated they were uncertain, suggesting our binary measure is prone to misclassification. Moreover, we could not assess varying levels of Ikigai and its sources, which can result in heterogeneity in the impacts of Ikigai on health and wellbeing." All three structural limits in one paragraph, quantified from somebody else's sample, and their own call for future work asks for "information on kind of activities from which people drive Ikigai", which is the object the item cannot hold. A third cohort, Kozo Tanno and colleagues' 73,272 participants in 2009, reports only that the questionnaire included "a question about ikigai"; the full text is paywalled, so even the wording is unknown and must not be assumed to match.

Commerce and product design: the one-question metric

Product and customer-research teams run on single items by design, and the trade-off is identical. The Net Promoter question, "how likely are you to recommend us", is one item on an eleven-point scale, chosen because it is short enough to put at the end of a transaction and get an answer. An app store rating is one item. A post-support "was this helpful?" thumb is a binary item with a very large uncertain middle that has nowhere to go, which is why helpfulness rates look implausibly high: the ambivalent close the dialog. None of this makes the numbers useless, and a one-question metric with a 60 per cent response rate often beats a twelve-question survey with a 4 per cent one. What it means is that a company watching a single-item score climb knows that something about the answers moved and does not know what customers were answering about, so the explanation offered in the quarterly review is almost always supplied from outside the data. The research literature's discipline, stating the limitation in the paper and quantifying it, is the thing worth borrowing.

What it does not say

This page is not a verdict on single-item measures as a class. A substantial psychometric literature compares single-item and multi-item instruments and reports that for some constructs, notably global self-rated health, one item performs respectably. This site has not read that literature, so nothing here should be taken as endorsing or dismissing single items in general, in either direction.

It does not say the Ohsaki or JACC findings are wrong. They are large, carefully adjusted observational findings, and the exposure measure is their weakest component rather than a disqualification.

It does not say a null from a single-item study means no effect. Okuzono and colleagues' mortality result, 0.85 with an interval from 0.60 to 1.19, is underpowered rather than contradictory: their own fifth limitation states that their point estimate was similar to many prior studies while their confidence interval was wider and included the null, with only 6 per cent of the sample dying inside the three-year window. Misclassification from collapsing an uncertain group pushes in the same direction, towards finding nothing.

And it does not license the reverse inference that any framework claiming to describe the same construct inherits these findings. A four-circle career diagram asks four questions and requires them to intersect in a paid role. It has never been tested, no instrument operationalises it, and it shares nothing with the measured item except an English or Japanese label.

Who developed it

No single originator. This is long-standing survey and epidemiological practice rather than a named construct, and the cases described here belong to the researchers who used them: Toshimasa Sone and colleagues from 1994, Kozo Tanno and colleagues from 1988 to 1990, and Sakurako S. Okuzono and colleagues from 2013.

---