Loneliness, as psychologists have measured it since the 1970s, is not a count of anybody’s relationships. It is a reading a person takes of their own social world, and that reading can come out badly for someone with a marriage, a standing Tuesday and a name to call at midnight. Low mood is a different measurement again — a report on how a person feels in general rather than on how they feel about their connections. The two turn up together often enough that the interesting question became which of them turns up first.
A depression score with loneliness inside it
The usual instrument in this literature has a measurement problem sitting inside it. The standard 20-item questionnaire for depressive symptoms, the CES-D, includes an item asking whether the respondent felt lonely. Score it as written and loneliness is already part of the depression number, so a finding that one predicts the other is partly a finding that something predicts itself.
John Cacioppo, Louise Hawkley and Ronald Thisted dropped that item, summed the remaining 19, and went looking for the order.
Their data came from the Chicago Health, Aging, and Social Relations Study, which recruited people born between 1935 and 1952 and living in Cook County, Illinois, through a probability design that deliberately oversampled African American and Latino American households. The resulting sample was 229 people, aged 50 to 68 at the first visit, 52.4 percent women, and split roughly evenly between White, Black and Hispanic participants by design. Participants came to the University of Chicago for a full day of testing once a year, from 2002 to 2006, and were paid $126 a visit. Attrition averaged about 7 percent a year, and 163 of the original 229 were still in the study at the fifth visit.
Loneliness was measured on the revised UCLA scale, 20 questions summed to a score between 20 and 80; the average in the first year was 36. The average depression score in that first year, on the 19-item version, was 9.8. These were people answering questions in a laboratory, not patients in a clinic.
Both measures held their shape. The paper describes their year-to-year stability as moderate, with correlations above .6 between a person’s own scores in different years. Measured at the same moment, loneliness and low mood correlated between .57 and .68 in the paper’s own table — close enough to be worth worrying about, far enough apart to be worth separating. The deposited manuscript’s text gives a lower floor of .42, which the table shows is a four-year lag rather than a same-moment figure.
Then the order came out lopsided. Loneliness in one year predicted a rise in depressive symptoms the following year, with an unstandardized coefficient of 0.18 and a confidence interval running from 0.09 to 0.30. The path in the other direction, depressive symptoms predicting a later rise in loneliness, did not reach significance: 0.10, with an interval from −0.05 to 0.20 that includes zero, though most of it sits above zero.
The coefficient is not large. Our own arithmetic on the paper’s numbers: someone a standard deviation lonelier than average, about ten points on that 20-to-80 scale, would be predicted not quite two depression points higher a year later, holding the earlier year’s depression score constant, on a scale where the sample averaged 9.8. The direction is the finding here, not the size of it.
Five other explanations, tested one at a time
Something else could be driving both halves. So the team ran the analysis again five times, each time adding a candidate culprit that was allowed to change from year to year.
The first was actual social isolation, measured the standard way as four yes-or-no facts: whether someone was married or living with a partner, had close friends or relatives, belonged to a club or group, and attended church at least twice a month. Adding it changed almost nothing. The loneliness-to-mood path stayed at 0.18. Social connection scored this way predicted neither loneliness nor mood a year out, though the confidence intervals around those nulls are wide enough that a real effect cannot be ruled out. Then came neuroticism, stressful life events, perceived stress, and social support. The forward path survived all of them, thinning to 0.13 once social support was in the model.
The lopsided order had one exception, and the authors put it in themselves. In the model controlling for stressful life events, the reverse path did reach significance, at 0.13 with an interval from 0.01 to 0.27. They describe that effect as small and say it warrants replication. They trace it to the 28 people who, at the study’s start, reported having had, or having been told by a doctor they had, emotional, nervous or psychiatric problems; excluding those people dropped the reverse path to 0.03 and out of significance, while the forward path stayed at 0.19. Their reading is that the influence of low mood on later loneliness is greater in people carrying a diagnosis than in people without one, and they note the diagnosed subgroup was too small to model on its own.
Back in the first model, one background variable did have a reliable year-to-year effect of its own, and only one: age. Older participants sat on a slightly lower loneliness track, roughly half a point per additional year of age. The pattern between loneliness and mood held anyway.
The experiment that would settle it
None of this establishes cause. A design that watches the same people answer the same questions for five years can show which measurement moves first; it cannot show that moving one would move the other. The authors are more permissive about this than we are — they write that cross-lagged path analysis is widely used to infer causal associations, and that experimental evidence would confer greater confidence in the causal role of loneliness on depressive symptoms “and vice versa”, a hedge that points in both directions.
Clinical depression was never measured, and clinical patients were not specifically recruited, so the study speaks to symptom scores in a general population rather than to diagnosed illness. Both measures were self-report, filled in by the same person in the same survey packet. And 229 people from one American county in the early 2000s is a particular set of lives rather than a general law.
A 2020 conference abstract points the other way. Jillian Minahan and Karen Siedlecki reported a cross-lagged analysis of 1,560 healthy adults aged 18 to 95 from the Virginia Cognitive Aging Project, across three time points, in which depression predicted later loneliness more strongly than loneliness predicted later depression. It is an abstract rather than a full paper and it spans a much wider age range, so it is not a clean contradiction. It is a reason to hold the Chicago result loosely.
Nobody here is a clinician and nothing above is a diagnosis. Low mood or a sense of being unreachable that has lasted months is something to take to a therapist or a counselor.
The paper does correct a common piece of advice. The usual response to someone who says they feel lonely is to point at their calendar and their contacts, on the assumption that the feeling is a readout of the arrangements. In this sample the arrangements did not reliably predict either measure a year ahead, and the feeling predicted one of them. That is an argument for asking people how they feel about their connections rather than counting the connections and inferring the rest.
Whether changing the feeling would change what follows it is a question this design cannot reach. The study that could would have to alter loneliness deliberately and then follow mood for years, in people recruited for that purpose rather than asked once a year for five years about the lives they already had.
