What the Mental-Health Evidence Actually Shows
The subject here is the observed relationship between online upward social comparison and psychological functioning. The key question is what the pooled evidence can establish, not whether a single headline number sounds alarming. Reading the synthesis carefully means separating the average association from an individual forecast, comparing the outcomes on a common scale without hiding their differences, and treating the study design as part of the result.
The Frontiers meta-analysis is useful precisely because it does more than report that “comparison is bad.” It combines multiple outcomes, retains dependence among estimates from the same studies, tests whether the pattern differs by context, and states where the evidence remains weak. Its conclusion supports concern about a broad pattern, especially for social-evaluative emotions, while leaving causal direction and individual variation unresolved.
What the pooled number means
The synthesis brought together 94 effect sizes from 54 independent samples, representing 36,583 participants. An effect size here is a correlation between upward social comparison and a psychological outcome. The researchers used a three-level random-effects model because one study can contribute several related estimates. A study might measure more than one outcome, or more than one comparison-related association, and those estimates are not fully independent. Treating them as if they came from unrelated studies would make the evidence look more precise than it is.
A random-effects model also changes the meaning of the headline result. It estimates an average across a distribution of true effects rather than assuming that every study is observing one fixed relationship. The pooled estimate is therefore a centre of gravity: individual effects may be smaller, larger, absent, or even point in another direction. In the omnibus model, positively valenced outcomes were reverse-coded so that a positive number consistently represented poorer functioning. The resulting average correlation between online upward comparison and psychological maladjustment was 0.330, with a 95% confidence interval from 0.289 to 0.370.
That is evidence of a reliable average association in the published dataset. It is not a 0.330 probability that a person will become distressed after looking at someone better off, and it is not a percentage of psychological harm caused by social media. Correlation does not tell us how much the comparison changed the outcome, whether the outcome changed the comparison, or whether a third factor influenced both. The estimate is meaningful because it is based on a large synthesis and its interval excludes zero; it is limited because its design and component studies do not turn it into a causal or personal prediction.
The outcome domains show why a single average can conceal important differences. In the cross-domain comparison model, the largest average estimate was for social-evaluative negative emotions, at 0.438, followed by anxiety at 0.382, depression at 0.306, lower wellbeing at 0.268, and lower self-esteem at 0.263. The overall differences among outcome categories were statistically significant. Social-evaluative negative emotions were not significantly different from anxiety, but their average association was significantly larger than the estimates for depression, wellbeing, and self-esteem.
The ordering is informative, but it should not be read as a proven ladder of effects. The categories describe different kinds of outcomes. Envy, body dissatisfaction, loneliness, and fear of negative evaluation stay close to a specific social comparison: they concern a target, a relationship, or anticipated judgment. Depression and anxiety are broader symptom domains. Wellbeing and self-esteem summarize wider evaluations of life and self. A stronger association with a proximal emotion may mean that the outcome is more tightly coupled to the comparison episode; it does not show that the emotion causes the broader condition.
The sign of the number also depends on the model. In the cross-domain model, wellbeing and self-esteem were reverse-coded because higher scores ordinarily indicate better functioning. Reversing them lets the reader compare “more maladaptive functioning” in the same positive direction as anxiety or depression. In separate outcome-specific analyses, the original directions were preserved: upward comparison was positively associated with anxiety and depression, and negatively associated with wellbeing and self-esteem. Those models reported estimates of 0.380 for anxiety, 0.268 for depression, −0.310 for wellbeing, −0.251 for self-esteem, and 0.458 for social-evaluative negative emotions.
The small differences between the cross-domain and outcome-specific figures are not a contradiction. They come from different parameterizations and analytic subsets. The cross-domain model is designed to compare categories after aligning their direction; the separate models describe each outcome in its original direction. A careful reader should ask which model a number came from before ranking outcomes or repeating it as if all estimates were interchangeable.
The broad pattern is therefore more precise than the slogan. Online upward comparison is associated, on average, with higher anxiety, depression, and social-evaluative negative emotions and with lower wellbeing and self-esteem. The association appears across multiple domains rather than being confined to one feeling. At the same time, the domain differences suggest that the comparison episode may be most closely tied to specific social-evaluative reactions. That is consistent with the mechanism discussed earlier, but it remains an interpretation of differing average associations rather than a tested causal sequence.
The domain ordering should not be mistaken for a ranking of how serious the outcomes are. A correlation is a standardized summary of co-variation, not a common unit of suffering. The measures used for envy, anxiety, depression, wellbeing, and self-esteem differ in wording, time frame, and sensitivity. A larger average association for one domain says that the measured variables moved together more strongly in these studies; it does not by itself say that the outcome is more clinically important, more damaging to every person, or easier to change.
The multilevel method matters for this interpretation. Suppose one study reports an association with envy and another with body dissatisfaction, while both also report wellbeing. Those estimates contain information about distinct outcomes, but the estimates from the same sample share participants, recruitment conditions, and measurement choices. Retaining them can preserve information that would be lost by averaging everything into one number, but it also requires the model to represent their dependence. The three-level approach is an attempt to do both. It does not remove the underlying heterogeneity; it prevents the analysis from pretending that repeated measures from one study are as independent as entirely separate samples.
The confidence interval supplies another useful discipline. The interval around 0.330 describes uncertainty about the pooled average under the model and data, not the range of responses that must contain every individual. Individual experiences can fall outside it, and a future study in a different context can produce a different estimate. The random-effects framing makes that variation part of the claim. A strong average pattern and substantial variation are not competing interpretations; they are both features of the evidence.
This diagram is an evidence pipeline, not a causal path. The studies are combined, their outcome directions are made comparable for one analysis, and the resulting pattern is then interpreted in light of moderator tests and design limitations. Nothing in the arrows means that one measured outcome necessarily produces the next.
Where the conclusion narrows
The researchers tested whether age group, cultural background, or data-collection year changed the average association within the outcome-specific models. None of these moderators reached statistical significance. The available data therefore did not detect a systematic difference between the coded groups or across the observed years. That result is narrower than saying that age, culture, and history do not matter.
There are several reasons to keep that distinction visible. Age was simplified into adolescent versus adult samples, even though developmental pressures can change within both groups. Cultural background was coded as a broad Western–Eastern distinction, which can hide variation in values, social norms, class, platform use, and self-construal. Data-collection year was treated as a continuous measure across a period of rapid platform change, but a single slope may not capture shifts in interface, algorithm, or user practice. Uneven subgroup sizes also limit the power of moderation tests. A non-significant result can mean that the available evidence did not resolve a difference, not that no difference exists.
The exploratory design comparisons make the same point. The average reverse-coded association was similar in cross-sectional and longitudinal or intensive-longitudinal samples, and the direct comparison was not significant. That does not establish that study design is irrelevant. The longitudinal and experimental groups were comparatively small and uneven, while cross-sectional studies supplied most of the evidence. A test with limited representation of the designs most capable of clarifying direction cannot settle the causal question.
Study quality adds another boundary. Measurement of comparison and outcomes was generally supported by validated instruments, but sample selection was a major concern. More than half of the samples—53.7%—were rated as having unclear selection risk, and 9.3% as high risk. Many participants came from convenience sources such as university populations or online user groups. Such samples can reveal a real relationship while still failing to represent the full range of ages, cultures, socioeconomic positions, and platform practices. The problem is not that convenience samples are automatically wrong; it is that their average may not travel unchanged to a different population.
The review also included only published studies. Egger’s tests did not detect small-study effects in the overall dataset or the individual outcome categories, but the authors explicitly caution that this does not prove publication bias is absent. Tests for asymmetry have limited power and detect only some forms of selective reporting. A synthesis can therefore be methodologically careful and still describe a literature shaped by which studies reached publication and which did not.
Most importantly, much of the underlying evidence is cross-sectional. A survey can find that people who report more upward comparison also report more distress, but it cannot by itself establish the order of events. Distress may encourage more comparison, comparison may contribute to distress, or both may be influenced by another factor. A reciprocal pattern is plausible: someone who feels excluded may search for social information, encounter more threatening standards, and then feel even more inadequate. The average association remains relevant under that possibility, but its interpretation changes from “comparison causes harm” to “comparison and poorer functioning are linked in a way that requires directional testing.”
The practical reading rule is consequently simple but demanding: treat 0.330 as evidence of an average pattern in a heterogeneous published literature, not as a personal forecast or causal estimate. Use the outcome-specific pattern to ask what kind of experience is most closely related to comparison. Read the null moderator findings as unresolved context rather than proof of uniformity. And treat the study-quality information as part of the conclusion, not as a footnote added after the conclusion has already been made.
This calibrated reading does not weaken the finding. It makes the finding usable. The synthesis supports taking online upward comparison seriously as a psychologically specific process rather than hiding behind broad measures of screen time. It also tells us what stronger evidence would need to add: more balanced longitudinal and experimental studies, finer cultural and developmental measures, better representation beyond convenience samples, and designs that record the comparison, the immediate appraisal, the subsequent behaviour, and the later psychological outcome.
That last requirement points to a different question. The mental-health synthesis observes naturally occurring comparison and its correlates. What happens when comparison is deliberately supplied as feedback about a defined behaviour, with a control condition and a measurement point after the intervention? Randomization can improve the claim about what the tested intervention changed, even though it cannot automatically isolate comparison from every accompanying technique. The next chapter turns to that narrower behavioural question.