What ‘Works’ Means in 20 Randomized Trials
The subject here is the broad randomized evidence for savoring interventions, and the task is to understand what its pooled result actually estimates. The meta-analysis is valuable not because it turns many studies into one simple verdict, but because it shows both a positive average and the variation surrounding that average. Reading it carefully gives us a better judgment than either “savoring works” or “the studies are too different to matter.”
What is hidden inside “20 randomized trials”?
Chen and colleagues synthesized 20 independent randomized controlled trials, 4,805 participants, and 45 effect sizes. An effect size is a common numerical representation of a difference between an intervention and its comparator. One trial can contribute more than one effect size when it measures several outcomes or reports more than one relevant comparison. That is why the number of effect sizes is larger than the number of trials, and why treating all 45 as if they came from 45 unrelated experiments would exaggerate the amount of independent evidence.
This distinction between trials and effect sizes is easy to miss in a headline. Suppose one study measures positive affect, negative affect, and depression. Those measurements may illuminate different parts of the same participants’ experience, but they are still connected by the same recruitment process, intervention, setting, and randomization. They should not be counted as three independent tests of savoring. The three-level approach used by the review is designed to respect that connection rather than treating every reported outcome as independent evidence.
The review searched PubMed, PsycINFO, the Cochrane Library, CINAHL, MEDLINE, and Google Scholar from their beginnings through March 6, 2025. It looked for randomized trials in humans that explicitly targeted savoring and measured emotional or mental-health outcomes before and after the intervention. Eligible outcomes included positive affect, negative affect, well-being, happiness, depression, and anxiety. Studies had to provide enough information to calculate a standardized effect. Interventions centered primarily on gratitude, kindness, or best-possible-self visualization were excluded when the savoring component could not be separated.
That last rule is more than housekeeping. A synthesis of all exercises that make people feel good would answer a broader question about positive activities. This review asked a more specific question about deliberate engagement with positive experience. The boundary protects the meaning of the pooled estimate, even though some real exercises—such as a gratitude letter—can contain more than one psychological ingredient. It also means the review’s result should not be quietly relabeled as evidence for every intervention that involves optimism, appreciation, pleasant activity, or social connection.
The included trials were varied in ways that matter to interpretation. Their median intervention duration was 14 days, with an interquartile range from 7 to 28 days. Eleven studies delivered savoring individually, four used computer-based online programs, and five used smartphone applications. The synthesis included both active and passive comparators. Fourteen samples were classified as Western and six as Eastern. Although the eligibility criteria did not impose an age restriction, every included sample was adult; university students were the youngest participants. The evidence is therefore broad relative to a single trial, but it is not evidence from every stage of life or every health condition.
The duration statistic also changes the implied time horizon. A median of two weeks tells us that the average study was testing a short intervention, not a lifelong practice. It can support a question about near-term change and perhaps a limited follow-up, but it cannot by itself establish what happens after months of continued practice or after the exercises stop. Likewise, the counts for online and smartphone delivery show that digital formats were represented, not that digital savoring has been tested with the same depth or precision as every other format.
The review used a three-level random-effects model. The name sounds technical, but the reason is straightforward. The analysis recognizes three sources of variation: ordinary sampling error, differences among multiple effects reported by the same study, and differences between studies. It then uses robust CR2 variance estimation for hypothesis tests and confidence intervals. This prevents several outcomes from one trial from masquerading as several fully independent trials.
“Random-effects” matters because the review is not assuming that every trial estimates one fixed effect with only measurement noise around it. It allows the underlying effect to vary from setting to setting. A short writing exercise in students, a smartphone intervention in adults, and a structured activity for older people may all belong to the savoring family while producing different results. The model estimates an average across that family and acknowledges that the family has members with different conditions.
A useful analogy is a set of thermometers taken in different rooms. The reading from each thermometer contains measurement noise. Several thermometers in the same room share the room’s conditions, so their readings are related. Different rooms may have genuinely different temperatures. A model that accounts for all three facts is more credible than one that simply averages every reading as if it were independent. But even the best averaging method cannot make the rooms identical. Statistical care protects the synthesis from one error; it does not remove the underlying diversity of interventions and participants.
The model is therefore part of the substantive story. The pooled estimate is not a vote counted across identical experiments; it is a statistical center drawn from related but nonidentical evidence. Understanding the center requires looking at both its magnitude and the distribution that produced it.
A positive average with a wide context around it
The overall pooled effect was Hedges’ g = 0.51, with a 95% confidence interval from 0.26 to 0.77 and p < .001. Hedges’ g is a standardized mean difference corrected for small-sample bias. Standardization allows researchers to combine outcomes measured on different scales: a depression questionnaire, a positive-affect scale, and a well-being measure do not use the same units, but their intervention-versus-comparator differences can be expressed relative to variability in the sample.
That standardization is useful and limiting at the same time. A standardized effect can tell us that the average difference is not trivial relative to variation in the measured outcome, but it does not tell us how many points a person would gain on a particular questionnaire. To make that translation, we would need the instrument, its scale, the baseline distribution, and a judgment about what constitutes a meaningful change. The pooled statistic does not contain those details on its own.
The positive sign was arranged to mean benefit from savoring. For outcomes such as depression or anxiety, a reduction was sign-reversed so that it could be combined with increases in positive affect or well-being. The result is convenient for synthesis, but it also means the pooled number is an abstraction across unlike outcomes. It should be read as evidence that the intervention condition tended to do better than its comparator across the included measures—not as a promise of a particular number of points on any one questionnaire.
The confidence interval describes uncertainty around the estimated average. In this synthesis, the interval remains above zero, which is consistent with a positive overall effect. It does not describe the range of individual responses. Some participants may have improved more, some less, and some not at all. Nor does it say how long the change lasted. A confidence interval around a pooled average is not a personal forecast and is not a clinical response rate.
The p-value answers a narrow question about compatibility with a null effect under the model; it does not measure importance, certainty for an individual, or the probability that savoring will work for a particular reader. The magnitude and interval carry more decision-relevant information. Even then, they must be read alongside the comparator and the heterogeneity. A precise estimate of an average across many unlike situations would still not tell us which situation a given person occupies.
The review also grouped outcomes into three broad categories. Negative emotional disorders or symptoms had a descriptive estimate of g = 0.61, with a 95% confidence interval from 0.31 to 0.91. Negative emotional states had g = 0.33, with an interval from −0.01 to 0.68 and p = .056. Positive psychological states had g = 0.50, with a 95% confidence interval from 0.24 to 0.75. These figures may tempt a reader to conclude that savoring is best for symptoms, weaker for feelings, and equally strong for positive states. That conclusion goes beyond the analysis.
The omnibus test asking whether outcome category moderated the intervention effect was not significant: F(2, 7.55) = 1.72, p = .242. In practical terms, the differences among the three point estimates were not established as reliable differences. The negative-emotional-state estimate is also less precise than the others and its interval includes zero. The correct reading is that the meta-analysis found positive evidence overall, with category-specific patterns that are informative but not a proven ranking of what savoring changes best.
This distinction matters because outcomes are not interchangeable in everyday life. A short-term shift in negative affect is not the same as a change in how someone evaluates their life. A lower symptom score is not automatically a diagnosis-level improvement. Flourishing, life satisfaction, positive affect, anxiety, and depression capture related but distinct aspects of experience. Combining them can reveal a broad signal, while also making it harder to say exactly which form of change a person should expect.
There is a further reason to resist a single “well-being” label. Some outcomes are close to the intervention’s immediate target: a person may report more positive emotion or less negative affect after practicing. Other outcomes are broader evaluations formed over time, such as satisfaction with life or flourishing. The latter may require changes in relationships, health, purpose, or daily circumstances that a brief exercise cannot directly control. The pooled result does not tell us that every outcome is equally malleable or equally important to every decision.
The largest caution comes from heterogeneity. Total I² was 86.61%, indicating that most of the observed dispersion in effects was not simple sampling noise. The model attributed 72.77% to between-study variance and 13.84% to within-study variance across outcomes. Both components were statistically supported. The exact percentages are properties of this model, not a percentage of people who will respond. Their meaning is that effects differed substantially across the studies and, to a smaller extent, among outcomes within studies.
High heterogeneity does not mean that the intervention is ineffective. It means that the average is combining effects that are not all alike. A positive average can coexist with small effects in some studies and larger effects in others; the pooled number does not reveal the exact distribution of individual responses, but the heterogeneity statistic warns us not to imagine a uniform one. The appropriate response is not to discard the average automatically or to ignore the variation, but to ask which differences have been measured well enough to explain it.
What might create that variation? The trials differed in age and health status, the activity used to elicit or revisit positive experience, the amount of practice, delivery mode, comparator, cultural setting, and outcome instrument. The review tested several of these as moderators, but a moderator analysis can only examine the variables that were measured and represented enough to compare. A nonsignificant test does not prove that context is irrelevant; it may also reflect limited power, wide uncertainty, or categories that are too broad to capture the real mechanism.
The heterogeneity therefore changes the question a reader should ask. Instead of asking, “What is the effect of savoring?” as if there were one context-free quantity, ask, “What does this average say about this family of interventions across these kinds of studies, and how similar is the situation I care about?” The first question produces a headline. The second produces a usable judgment.
The meta-analysis did perform several checks for selective reporting. Its funnel plot showed no clear asymmetry, the robust Egger test was not significant, and selection-model tests did not detect significant bias. Trim-and-fill gave an adjusted estimate of g = 0.48, while selection models gave g = 0.60. A PET-PEESE adjustment produced a near-zero estimate, but the authors treated that result cautiously because the method can overcorrect toward zero in a small, highly heterogeneous synthesis. The checks make it less likely that selective publication alone explains the positive average; they do not erase the other limitations.
About 30% of the included studies were judged to have high risk of bias, and most outcomes relied on self-report questionnaires. Self-report is not worthless—emotions and well-being are partly subjective experiences—but it can be influenced by recall, social desirability, expectations, and the desire to report that participation was worthwhile. The review also could not fully exclude unpublished or grey literature, and most interventions were brief. These facts do not reverse the average result. They define the kind of result it is: a positive signal from randomized evidence, with unresolved questions about objective change, clinical generalization, and durability.
The review’s design also determines what it cannot say about harm. The included studies focused on emotional and mental-health outcomes, not on a comprehensive assessment of every possible adverse consequence. The supplied evidence therefore supports discussion of potential benefit and uncertainty, but not a claim that savoring is risk-free in every population or that positive-emotion practices are appropriate in every clinical circumstance. That is another reason to keep a well-being signal separate from a treatment recommendation.
The most responsible summary is consequently two-part. First, across 20 adult RCTs, savoring interventions were associated with a statistically significant, moderate standardized advantage over their comparators. Second, the high between-study variation means the pooled effect is a center of gravity, not a universal setting. To understand how strong the causal claim is in any one comparison, we must now examine the comparator itself—especially the difference between doing more than nothing and outperforming another program that also creates positive emotion.