An open request to epidemiologists: please stop querying self-reported sleep duration

A recent epidemiology study [1] contains the latest in a long series of U-shaped relationships [2–5] linking a health problem with self-reported sleep duration: the outcome of interest (in this case, visual impairment) worsens as self-reported sleep duration deviates from the referent value (typically 7 ± 1 h). In an era where we contemplate whether AHI is the right metric for sleep apnea [6], how to integrate new information on sleep electrophysiology into our legacy scoring systems (or whether positive airway pressure (PAP) is effective or even dangerous [7,8]), should we not also reasonably wonder how mechanistically useful or clinically actionable information can be contained in cross-sectional self-reported sleep duration? Sleep duration seems like a natural thing to ask about, and it would seem that adult patients could provide an approximation. An et al. cite Lockley et al. [9] as evidence of subjective-objective agreement to address the self-reporting limitation. However, that study (of blind adults) in fact noted “large and inconsistent differences with absolute sleep parameters,” and suggests utility mainly for tracking over time. More broadly, several concrete arguments exist that the answers patients/study participants provide to self-reported sleep duration queries contain so little information that they should be avoided as an isolated metric in epidemiology studies.

First, people with sleep complaints and sleep disorders may not accurately estimate their sleep on a single night of objective testing. Rather, people’s estimates depend heavily on extraneous factors including demographics [4], comorbidities including insomnia [10], the timing of the query [11], and the way the question is asked [12]. Conversely, individuals without sleep complaints, tend to over-estimate sleep duration, largely because they are amnestic for natural nightly awakenings. These observations can be interpreted as contaminating the responses people give to the seemingly simple query of: “how many hours do you sleep, on average?”

Second, there is a “differential diagnosis” regarding the response to the duration questions typical of epidemiology studies. Consider an individual who answers “5 h.” With respect to (unknown) objective sleep duration, this value could be: a) accurate reporting of consistent short sleep time, b) accurate reporting of short nocturnal sleep time, but without taking into account naps, c) accurate reporting of the average of highly fluctuating sleep times, d) under-estimation due to misperception insomnia, e) and/or under-estimation for reasons unrelated to insomnia. We can perform the same exercise of the differential diagnosis of long self-reported sleep duration. That the referent range of “normal” duration itself has a differential diagnosis completes the uncertainty trifecta; we have uncertainty and thus misclassification across the entire range of reported answers.

Third, duration is only one “axis” of sleep, and does not consider sleep quality. By their nature (with massive sample sizes), epidemiological studies do not have objective polysomnography (PSG) data on the participants. Thus, we know nothing about common (and often undiagnosed) forms of pathophysiology such as sleep disordered breathing and periodic limb movements of sleep – each of which impacts sleep quality.

Fourth, statistical control for other self-reported factors also contain uncertainties, Distinctions between self-reported and actual behaviors have been discussed elsewhere, such as medication compliance [13], diet [14], smoking [15], and alcohol [16]. These major categories of uncertainty call into question correlations measured by epidemiological studies based on self-reported sleep duration. Given these concerns, one wonders about the risks associated with continuing this trend of asking simple queries of duration and making epidemiological and even genomic inferences that may spill over into individual patient care discussions. Without objective data on duration and quality, we will continue to debate the meaning of self-reported duration; we will wonder if increasing duration will mitigate the repeatedly reported manifold risks of short sleep duration, and if we are internally consistent. We will also want to know if restricting sleep in longer-than-average sleepers will benefit them, and most concerning perhaps, the general population learning of such results might either become anxious about their sleep duration or, if they are savvy regarding study design, lose faith in this field of inquiry.

In an era when low-cost, scalable objective physiological monitoring is increasingly available with wearable devices, we hope to put the era behind us of the U-shaped curve in epidemiology based purely on self-reported sleep duration. Instead, we should focus our increasingly limited resources on what matters: a comprehensive phenotypic profile that pairs subjective measures of sleep with objective measures, over multiple nights. Only with this kind of high quality data will we have the potential of making useful mechanistic or actionable inferences from large studies.