A meta-analysis of the effectiveness of gratitude interventions on well-being across cultures

Significance

Most people aspire to lead happy lives, and gratitude interventions such as journaling or writing gratitude letters are popular strategies to enhance happiness worldwide. However, their effectiveness across cultures remains unclear. We synthesized data from 145 studies spanning 28 countries and found that gratitude interventions result in small increases in well-being. Notably, the effectiveness of gratitude interventions varied significantly between countries. Also, methodological factors influenced their effectiveness: Interventions were more effective when positive emotions were measured as a well-being outcome, multiple types of gratitude interventions were combined, or randomized controlled trials were employed. These findings highlight the importance of exploring cultural influences and optimizing intervention designs to maximize their impact on well-being.

Abstract

Gratitude practice is a popular strategy for promoting happiness worldwide. However, is it equally effective for people from different cultural backgrounds? To answer this question, we conducted a preregistered meta-analysis on the effectiveness of gratitude interventions on well-being across cultures. Using data from 145 papers, 163 samples, 727 effect sizes, and 24,804 participants from 28 countries, we found that gratitude interventions led to small overall increases in well-being, Hedges’ g = 0.19, 95% CI [0.15, 0.22]. Moderation analyses revealed significant between-country differences in the effects of gratitude interventions. However, we found no significant evidence of moderators explaining such cross-cultural variability, highlighting the need for more primary research in this domain. Subsequent analyses indicated that three methodological characteristics moderated the effects: well-being outcomes, intervention types, and randomization. The effectiveness of gratitude interventions was greater when well-being outcomes were measured as positive affect, when multiple types of gratitude interventions were combined, and when randomized controlled trials were employed. Overall, the results appeared to be robust against publication bias and the presence of influential cases. These findings contribute to a deeper understanding of the effectiveness of gratitude interventions on well-being across cultural contexts and methodological characteristics.

Leading a happy life is considered one of the most important goals (1) and ideals (2) to most people around the world. People strive to move closer to such a goal or ideal and use various strategies to promote their happiness. Gratitude practice is one of the most popular happiness-promoting strategies recommended by both laypeople and the mainstream media (3). Gratitude practices involve activities such as keeping a gratitude journal and writing gratitude letters. Although these practices have roots in many of the world’s religious, philosophical, and cultural traditions (4), empirical research on the effectiveness of gratitude in promoting happiness originates from Emmons and McCullough (5) and Seligman et al. (6). In these seminal studies, the authors implemented gratitude interventions aimed at focusing on the positive aspects of life or on writing a letter to and then visiting someone to whom one felt grateful, and found that the interventions generally improved well-being. Since then, numerous studies on gratitude interventions have been published.

As research on gratitude interventions has surged, several meta-analytic reviews have been published over the past decade (7–11). These meta-analyses have indicated positive effects of gratitude interventions on well-being. For example, one meta-analysis included five studies comparing gratitude interventions with measurement-only control conditions and 20 studies comparing them with alternative-activity conditions (8). Results indicated that gratitude interventions had a small effect on life satisfaction and depression (Cohen’s d = 0.31 in the measurement-only comparison and d = 0.17 in the alternative-activity comparison). Another meta-analysis analyzed 38 studies and found weighted Cohen’s ds of the gratitude intervention relative to the neutral condition at the postintervention as follows: 0.17 for life satisfaction (k = 19), 0.18 for positive affect (k = 19), 0.25 for happiness (k = 9), 0.30 for well-being (k = 3), 0.13 for depression (k = 9), and 0.05 for negative affect (k = 15) (9).

Despite these important previous meta-analytic efforts, they had several limitations. First, no work has examined cultural variations in the effectiveness of gratitude interventions. This is a theoretically important research gap, given that understanding cultural differences could shed light on the mechanisms responsible for the effectiveness of gratitude interventions. In fact, prior cross-cultural studies have demonstrated that Anglo-Americans tend to benefit more from gratitude interventions than Asians or Asian Americans (e.g., refs. 12–15). While these studies suggest potential national, cultural, and religious differences in the effectiveness of gratitude interventions, these differences have not been investigated meta-analytically. Furthermore, these studies only compared Americans with Asians or Asian Americans. Thus, it remains unclear whether cultural differences beyond the East-West dichotomy emerge and what accounts for those differences.

Second, previous meta-analyses took a traditional analytic approach. For instance, they assumed independence between effect sizes. To avoid interdependence among effect sizes, some meta-analyses aggregated effect sizes per sample when there were multiple effect sizes for outcomes or multiple comparisons between conditions within the same sample (7, 8, 10). Alternatively, other meta-analyses conducted separate meta-analyses for each outcome and each comparison between conditions (9, 11). However, these traditional approaches may lead to information loss and reduced statistical power due to smaller number of effect sizes, which increases the risk of false negatives (16). Moreover, when effect sizes within a sample are heterogenous, averaging them may obscure the examination of the heterogeneity inherent within the sample. Consequently, previous meta-analyses may not be sufficiently precise to draw firm conclusions about the magnitude and heterogeneity of the effectiveness of gratitude interventions.

Third, previous meta-analyses, especially recent ones, were not sufficiently comprehensive to investigate the effectiveness of gratitude interventions. Ref. 11 focused solely on children in schools, ref. 7 only included research on certain outcome variables (i.e., depression and anxiety), and ref. 10 only included certain intervention types (i.e., expressed gratitude interventions). This restricted scope may result in a small set of studies and limit the generalizability of their conclusions. Last, more than 60 new studies have been published since the most recent comprehensive meta-analyses (8, 9). Thus, an update is imperative to provide a more complete picture of the current state of the literature.

Present Meta-Analysis.

In this meta-analysis, we addressed the limitations of the previous meta-analyses and examined the effectiveness of gratitude interventions on well-being across cultures. In this paper, we use “well-being” as an umbrella term that refers to an individual’s overall psychological, emotional, and social functioning and encompasses all the outcomes of gratitude interventions included in this meta-analysis. Specifically, we include life satisfaction, positive affect, negative affect, happiness, composite well-being, and depression as outcomes, as these are the most frequently examined variables in the existing literature. Life satisfaction, positive affect, negative affect, happiness, and composite well-being all assess subjective evaluations of one’s life and emotional experiences. Life satisfaction reflects the cognitive component, while positive and negative affects capture the emotional components (17). Happiness and composite well-being are more global constructs, incorporating both cognitive and emotional aspects (18). Depression is typically assessed through depressive symptoms and includes not only emotional components but also cognitive, behavioral, and physical components (19).

Our main goal was to investigate the moderating effects of culture on gratitude interventions, with country serving as a proxy for culture. First, as research on gratitude intervention has been predominantly conducted in the United States, we compared the effects between the United States and other countries. Second, we compared the effects of gratitude interventions across countries. Third, we examined whether country-level cultural variables moderate the effects. Moreover, we conducted moderator analyses on study, sample, and measurement characteristics. For moderation analyses, we initially tested each moderator separately and then tested statistically significant moderators together.

In addition to the preregistered analyses, we further performed a model-selection analysis to identify the model that best explains the intervention effects and the relative importance of the respective moderators. For explanatory purposes, we conducted a series of mini- meta-analyses that break down effect sizes by well-being outcomes, types of interventions, and types of control conditions to identify any differential findings.

Furthermore, this study sought to provide the most comprehensive, precise, and transparent meta-analysis to date on the effectiveness of gratitude interventions on well-being. Unlike previous meta-analyses, we utilized more advanced meta-analytic techniques to obtain more precise parameter estimates. Specifically, we conducted a three-level random-effects model that included all effect sizes from the same sample and modeled three sources of variance at different levels: sampling variance (Level 1), within-sample variance (Level 2), and between-sample variance (Level 3). We also encompassed all groups, intervention types, outcome measures related to well-being and thoroughly analyzed publication bias. Additionally, we updated the evidence by including all studies recorded before May 12th, 2024, and preregistered our analysis plans. By doing so, this study aimed to contribute to advancing knowledge and guiding future research on gratitude interventions.

Results

Descriptive Statistics of the Included Studies.

This meta-analysis included a total of 145 papers, 163 samples, 727 effect sizes, and 24,804 participants. Out of 163 samples, 138 samples reported multiple effect sizes, with an average of 4.46 effect sizes per sample (range: 1 to 32). All studies were published between 2003 and 2024 and conducted in 28 countries. Across samples, the average age was 29.5 y (SD = 13.4), and the average percentage of females was 65.3% (SD = 20.0).

Outliers and Influential Effect Sizes.

Next, we examined outliers and influential cases in the effect sizes, as they may impact the precision and robustness of the meta-analysis. Following the preregistered guidelines (20), effect sizes were classified as outliers or influential if they a) exhibited a considerable absolute difference in fit values, Cook’s distance, hat value, or covariance ratio, and b) their externally standardized residuals were greater than 1.96 or less than −1.96. Thirteen out of 727 effect sizes were identified as influential. Removing these 13 effect sizes yielded similar pattern of findings, including a similar overall effect size, Hedges’ g = 0.17, SE = 0.02, 95% CI [0.14, 0.20], P < 0.001 (See the SI Appendix for the full results without influential cases; SI Appendix, Figs. S1–S3 and Table S1). As the findings were robust to the presence of influential cases, we decided to retain them in the subsequent analyses.

Overall Effect Size and Heterogeneity.

The overall effect size for the difference in well-being between gratitude and control conditions was Hedges’ g = 0.19, SE = 0.02, 95% CI [0.15, 0.22], P < 0.001 (Fig. 1). This indicates that gratitude interventions had a small positive overall effect on well-being compared to control conditions. However, the effect sizes were highly heterogenous, τ² = 0.03, σ² = 0.02, Q(726) = 1,377.80, P < 0.001, I² at Level 2 (the amount of within-sample heterogeneity) = 21.9%, I² at Level 3 (the amount of between-sample heterogeneity) = 35.4%, 95% Prediction Interval (PI) = [−0.21, 0.59]. The significant Cochrane’s Q test indicates the presence of heterogeneity, and the total I² statistic across both within- and between-sample levels (57.2%) suggests moderate heterogeneity (21). The 95% PI includes 0, suggesting that the effect of the intervention in future studies may vary across studies and contexts. Taken together, the observed amount of effect size variability exceeds what would be expected by sampling error alone, suggesting the need for moderator analyses.

Fig. 1.

Orchard plot showing the overall effect of gratitude interventions on well-being based on 727 effect sizes from 145 papers and 163 samples. Blue dots represent individual effects, with their size indicating the precision of each effect (larger dots reflect higher precision). The black circle in the middle represents the overall effect size, with the thick whiskers around it showing 95% CI. k represents the number of effect sizes, and the number in parentheses indicates the number of samples.

Publication Bias Analyses.

Next, to assess the possibility of publication bias, we initially conducted Precision-Effect Test and Precision-Effect estimate with SE (PET-PEESE) and Three-Parameter Selection Model (3PSM) analyses. Both the slopes in the PET (b = 0.41, SE = 0.23, P = 0.074) and PEESE (b = 0.81, SE = 0.47, P = 0.087) regression models were marginally significant, which may raise concerns about publication bias and/or small-sample bias. However, even after controlling for potential bias, the overall effect size remained statistically significant, Hedges’ g = 0.15, SE = 0.03, 95% CI [0.10, 0.21], P < 0.001. Next, in the 3PSM analysis, we assumed that studies showing statistically significant (P < 0.05) and positive effects of gratitude interventions on well-being would be more likely to be published than studies showing nonsignificant (P ≥ 0.05) or negative results. However, there was no credible evidence of such bias: In fact, results showed a reversed pattern, where studies reporting nonsignificant or negative results were 1.75 times more likely to be published than those reporting significant and positive results. After adjusting for this reversed selection bias, Hedges’ g increased from 0.15 (unadjusted effect size) to 0.20 (adjusted effect size), with the adjusted effect size showing a better fit, ${\chi }^2\left(1\right) = 11.80$, P < 0.001.

We further tested for publication bias using techniques that account for the multilevel and dependent nature of the data. First, visual inspection of the contour-enhanced funnel plot suggested asymmetry in the plot (Fig. 2A). This was further supported by Egger’s MLMA test, which revealed a significant association between observed effect sizes and their SE (b = 0.41, P = 0.037), indicating potential small-sample bias. Next, we used a significance funnel plot and sensitivity analysis. As shown in Fig. 2B, the significance funnel plot displayed that there was a difference between the overall effect size across all samples (g = 0.19, 95% CI [0.15, 0.22], P < 0.001) and the “worst-case” effect size with only nonaffirmative samples (g = 0.07, 95% CI [0.04, 0.10], P < 0.001). However, affirmative samples would have needed to be 8.2 times more likely to be published than nonaffirmative samples to attenuate the overall effect size (g = 0.19) to the worst-case effect size (g = 0.07). Indeed, the sensitivity analysis demonstrated that even when assuming extreme publication bias (η = 200; the ratio of publishing affirmative over nonaffirmative findings), both the bias-adjusted effect size and 95% CI remained above zero (SI Appendix, Fig. S4). Taken together, although some evidence of small-sample bias and selection bias was found, we conclude that the effects of gratitude interventions on well-being appear to be robust against these biases.

Fig. 2.

Funnel plots. (A) Contour-enhanced funnel plot. Hedges’ gs are plotted against the SE. Dots represent individual studies. Shading in the triangular regions indicates significance. (B) Significance funnel plot. Black diamond represents the robust clustered estimate in all studies. Gray diamond represents the robust clustered estimate in nonaffirmative studies only. Studies lying on the diagonal line have P = 0.05.

Moderator Analyses: Cultural Influences.

We conducted a series of multilevel meta-regression analyses to explore sources of heterogeneity. We examined nine continuous and nine categorical moderators in total. Because multiple comparisons were tested in the moderator analyses, we adjusted the significance level to P = 0.003 (P = 0.05/18) using the Bonferroni correction. For categorical variables, a significant P value in the Omnibus test indicates that there was a statistical difference between the categories of the moderator, and the estimates for each category of the moderator represent the overall effect size (i.e., mean g) within the category.

The primary purpose of the present meta-analysis was to test whether the effect of gratitude interventions on well-being would vary across cultures. Thus, we first performed three preregistered meta-regression analyses to examine whether culture would moderate the overall effect. Table 1 presents the results for cultural moderator analyses. First, we ran a multilevel meta-regression analysis using the US vs. non-US variable as a moderator. The results revealed that the US variable did not significantly moderate the effects of gratitude interventions on well-being, F(1, 161) = 7.51, P = 0.007 (which does not meet the Bonferroni-corrected threshold of P = 0.003). This suggests that the effects of gratitude interventions on well-being in the United States do not differ from their effects in other countries.

Table 1.

Results of cultural moderator analyses for the effectiveness of gratitude interventions on well-being

Moderator	k ES	Omnibus test	B or mean g	95% CI	SE	P
US	727	F(1, 161) = 7.51				0.007
US	366		0.14	[0.10, 0.18]	0.02	<0.001
Non-US	361		0.24	[0.18, 0.30]	0.03	<0.001
Country	694	F(17, 135) = 207.21				<0.001
Australia	44		0.22	[0.02, 0.41]	0.10	0.028
Belgium	2		0.40	[0.18, 0.62]	0.11	<0.001
Canada	18		0.19	[0.02, 0.36]	0.08	0.024
China	35		0.36	[0.03, 0.69]	0.17	0.034
France	4		0.09	[−0.22, 0.41]	0.16	0.556
Germany	11		0.26	[0.13, 0.38]	0.06	<0.001
Hong Kong	16		0.60	[0.34, 0.87]	0.14	<0.001
India	25		0.18	[−0.01, 0.37]	0.10	0.065
Ireland	32		0.20	[0.09, 0.31]	0.06	<0.001
Japan	18		0.04	[−0.12, 0.21]	0.08	0.626
Malaysia	7		0.43	[0.17, 0.70]	0.13	0.002
Netherlands	6		0.23	[−0.06, 0.51]	0.14	0.116
Poland	9		0.19	[0.06, 0.31]	0.06	0.004
South Korea	14		0.34	[0.04, 0.64]	0.15	0.027
Spain	12		0.49	[0.12, 0.86]	0.19	0.010
Switzerland	48		0.13	[0.07, 0.18]	0.03	<0.001
United Kingdom	27		0.19	[−0.03, 0.41]	0.11	0.083
United States	366		0.14	[0.10, 0.18]	0.02	<0.001
Country predictors
Individualism	727	F(1, 161) = 8.01	−0.06	[−0.11, −0.02]	0.02	0.005
GDP per capita	723	F(1, 160) = 0.95	−0.02	[−0.06, 0.02]	0.02	0.331
Tightness	643	F(1, 146) = 0.50	0.06	[−0.11, 0.23]	0.09	0.481
Relational mobility	575	F(1, 139) = 6.22	−0.06	[−0.11, −0.01]	0.02	0.014
Residential mobility	727	F(1, 161) = 4.69	−0.04	[−0.08, −0.00]	0.02	0.032
Religiosity	727	F(1, 161) = 4.30	−0.05	[−0.09, −0.00]	0.02	0.040
Dominant religion	723	F(4, 157) = 2.78				0.029
Christian	603		0.16	[0.12, 0.19]	0.02	<0.001
Hindu	25		0.18	[−0.00, 0.36]	0.09	0.053
Muslim	7		0.43	[0.18, 0.69]	0.13	<0.001
Pluralistic	19		0.33	[0.13, 0.53]	0.10	0.002
Unaffiliated	69		0.37	[0.18, 0.57]	0.10	<0.001

Note: k_ES = the number of effect sizes. B for continuous variables represents unstandardized regression coefficients. B for categorical variables represents the overall estimates of effect sizes for each category of the moderator. CI = confidence interval. SE = standard error. Ten countries for the country moderator and Jewish for the dominant religion were omitted from the analysis because the cluster-robust omnibus Wald test could not be performed.

Next, we conducted another multilevel meta-regression analysis using individual countries as moderators. When all 28 countries were included in the analysis, the cluster-robust omnibus Wald test could not be performed partly because the effect sizes for some countries were derived from a single sample with the same sample size. We thus excluded those 10 countries (Brazil, Israel, Italy, Kenya, New Zealand, Philippines, Romania, Singapore, South Africa, and Taiwan) from the analysis. The results indicated that the effects of gratitude interventions on well-being did vary across countries (Fig. 3). No significant effects of gratitude intervention were found in France, India, Japan, the Netherlands, and the United Kingdom. However, gratitude interventions were significantly effective in promoting well-being in the other countries. The point estimate for the effect size for gratitude interventions was greater than zero in every country.

Fig. 3.

Country differences in the effects of gratitude interventions on well-being. A meta-regression analysis examining the impact of gratitude interventions on well-being across 18 countries. Ten other countries were excluded due to uncalculated SE. Each dot represents an individual effect, with its size indicating the precision of the effect (larger dots reflect higher precision). The black circles represent the overall effect size for each country, with the thick whiskers showing 95% CI. k represents the number of effect sizes, and the number in parentheses indicates the number of samples for each country.

Last, given the significant difference in the effects of gratitude interventions across countries, we examined whether seven country-level predictors (individualism, GDP per capita, tightness, relational mobility, residential mobility, religiosity, and dominant religion) moderated the effect. None of the seven predictors were found to have a moderating effect on well-being.

Moderator Analyses: Study, Sample, and Measurement Characteristics.

Next, we conducted moderator analyses for study, sample and measurement characteristics (Table 2). We did not find the moderating effects of age, gender composition, research designs, types of control conditions, dose (i.e., the number of sessions assigned), or publication status. This suggests that the effects of gratitude interventions on well-being were not significantly influenced by the mean age of participants, the proportion of females in the sample, research designs, types of control conditions, or the number of sessions that individuals completed for the interventions. Although the types of control conditions were not statistically significant, it is worth noting that effect sizes for gratitude interventions were not significantly different from those for other positively valenced activities. This suggests that gratitude interventions may not be more effective than other positive interventions. Additionally, the lack of significant differences between published and nonpublished studies further supports the limited evidence of publication bias in this meta-analysis.

Table 2.

Results of study, sample, and measurement moderator analyses for the effectiveness of gratitude interventions on well-being

Moderator	k ES	Omnibus test	B or mean g	95% CI	SE	P
Publication status		F(1, 161) = 0				0.999
Published	573		0.19	[0.15, 0.23]	0.02	<0.001
Unpublished	154		0.19	[0.09, 0.29]	0.05	<0.001
Age	596	F(1, 135) = 0.76	0.02	[−0.03, 0.07]	0.02	0.384
Gender composition	712	F(1, 157) = 1.53	−0.03	[−0.09, 0.02]	0.03	0.218
Dose	716	F(1, 159) = 0.14	−0.01	[−0.04, 0.03]	0.02	0.709
Outcomes	727	F(5, 157) = 5.05				<0.001
Positive affect	175		0.27	[0.22, 0.33]	0.03	<0.001
Negative affect	169		0.12	[0.07, 0.17]	0.03	<0.001
Life satisfaction	146		0.18	[0.12, 0.24]	0.03	<0.001
Depressive symptoms	111		0.15	[0.09, 0.22]	0.03	<0.001
Happiness	99		0.17	[0.12, 0.22]	0.03	<0.001
Composite well-being	27		0.28	[0.14, 0.42]	0.07	<0.001
Intervention types	727	F(6, 156) = 3.92				0.001
Counting blessings	356		0.17	[0.12, 0.22]	0.03	<0.001
Remembering gratitude	51		0.25	[0.16, 0.33]	0.04	<0.001
Expressing gratitude	181		0.15	[0.11, 0.20]	0.02	<0.001
Counterfactual	4		0.36	[0.19, 0.53]	0.09	<0.001
Psychoeducational	27		0.18	[0.07, 0.30]	0.06	0.003
Mixed	83		0.26	[0.19, 0.34]	0.04	<0.001
Other	25		0.15	[0.09, 0.20]	0.03	<0.001
Study designs	727	F(5, 157) = 2.98				0.014
Post only	99		0.24	[0.17, 0.31]	0.03	<0.001
Follow-up only	5		0.20	[0.10, 0.29]	0.05	<0.001
Pre–Post	376		0.18	[0.14, 0.23]	0.02	<0.001
Pre-Follow-up 1	173		0.14	[0.09, 0.19]	0.03	<0.001
Pre-Follow-up 2	50		0.17	[0.10, 0.23]	0.03	<0.001
Pre-Follow-up 3	24		0.11	[0.05, 0.18]	0.03	<0.001
Control types	727	F(3, 159) = 3.71				0.013
Measurement only	145		0.26	[0.17, 0.34]	0.04	<0.001
Neutral	551		0.16	[0.13, 0.20]	0.02	<0.001
Negatively valenced	20		0.45	[0.18, 0.72]	0.14	0.001
Positively valenced	11		−0.15	[−0.55, 0.25]	0.20	0.446
Randomization types	727	F(1, 161) = 16.01				<0.001
RCT	662		0.20	[0.17, 0.24]	0.02	<0.001
Quasi-experimental	65		0.05	[−0.01, 0.12]	0.03	0.120

Note: k_ES = the number of effect sizes. B for continuous variables represents unstandardized regression coefficients. B for categorical variables represents the overall estimates of effect sizes for each category of the moderator. CI = confidence interval. SE = standard error. GDP per capita was log-transformed and Dose was square-root transformed.

In contrast, outcome measures, intervention types, and randomization types significantly moderated the effects of gratitude interventions. Specifically, although gratitude interventions significantly enhanced all types of well-being outcomes—ranging from g = 0.12 (95% CI [0.07, 0.17]) for Negative Affect to g = 0.28 (95% CI [0.14, 0.42]) for Composite Well-being—the effect was strongest when the outcome measures were Positive Affect or Composite Well-being. Similarly, all types of gratitude interventions significantly promoted well-being, ranging from g = 0.15 (95% CI [0.11, 0.20] for Expressing gratitude to g = 0.36 (95% CI [0.19, 0.53]) for Counterfactual gratitude. However, the effects were stronger when multiple gratitude interventions were combined or when counterfactual gratitude interventions were implemented, although the effect of counterfactual gratitude interventions should be interpreted with caution given the small number of effect sizes. Finally, gratitude interventions were more effective in individual-level randomized controlled trials than in quasi-experimental designs, b = 0.15, 95% CI [0.08, 0.23], t(161) = 4.00, P < 0.001.

At a reviewer’s request, we also explored the potential moderating role of trait gratitude. The sample’s level of trait gratitude was inversely related to the effects of gratitude interventions on well-being, b = −0.18, 95% CI [−0.28, −0.07], SE = 0.05, t(37) = −3.38, P = 0.002.

Moderator Analyses: Simultaneous Model.

Next, we conducted a meta-regression analysis by simultaneously including statistically significant moderators (i.e., outcome, intervention type, randomization). The results indicated that the analysis including all significant moderators was also statistically significant, F(12, 150) = 5.01, P < 0.001. However, it was difficult to interpret specific findings because the three moderating variables each had six categories, seven categories, and two categories, respectively. Thus, we went on to examine the best model explaining the gratitude intervention effects and the relative importance of multiple moderators on gratitude interventions by using model-selection analysis with these three significant moderators. We examined the main effects of each moderator, and thus the model-selection analysis considered a total of eight (=2³) models ranging from a model with no moderators to one including all three moderators. Model selection was based on the Akaike Information Criterion, and the relative importance of each moderator was estimated as the sum of Akaike weights for all models that included each moderator. A cut-off value for the relative importance was set at 0.8. The analysis showed that the best model among the eight models included all three moderators, and the relative importance of all three moderators was greater than the cut-off value of 0.8, together suggesting that all moderators substantially explained heterogeneity observed in effect sizes within and between studies (See SI Appendix, Fig. S5 for the plot showing the relative importance of the moderators).

Mini Meta-Analyses.

Next, we conducted a total of 17 mini meta-analyses by breaking down effect sizes by well-being outcomes (six categories), types of interventions (seven categories), and types of control conditions (four categories). In these mini meta-analyses, we performed moderation analyses using the same moderators as in the main meta-analysis. The reason for conducting these mini meta-analyses on these three variables is that they have been identified as key factors explaining the effects of gratitude interventions on well-being in previous systematic reviews and meta-analyses (7, 8, 9, 22). We found that the patterns of the mini meta-analyses were largely similar to those of the main meta-analysis. Notably, consistent with the main analysis, the US variable and the seven country-level variables did not moderate the effects, except for dominant religion for happiness outcomes, residential mobility for depressive symptoms, and individualism and tightness for composite well-being. Between-country differences were observed when the well-being outcome was depressive symptoms, when the intervention types were counting blessings, expressing gratitude, or mixed interventions, and when the types of control conditions were measurement-only or neutral conditions. The complete results regarding well-being outcomes (SI Appendix, Table S2), types of interventions (SI Appendix, Table S3), and types of control conditions (SI Appendix, Table S4) can be found in the SI Appendix.

Discussion

The present meta-analysis synthesized the results from previous investigations into the effectiveness of gratitude interventions on well-being and examined the moderating effect of culture on these interventions. It also conducted additional moderation analyses on study, sample, and measurement characteristics, along with a model-selection analysis and 17 mini meta-analyses. With data extracted from 145 papers, 163 samples, 727 effect sizes, and 24,804 participants, we found that gratitude interventions had a positive effect on well-being, Hedges’ g = 0.19, 95% CI [0.15, 0.22], P < 0.001. Subsequent outlier analyses and publication bias analyses indicated that this effect was robust against influential cases and small-sample/selection biases, respectively, and that the interventions were especially effective in promoting positive affect and composite well-being. These findings are consistent with previous meta-analyses (7–11) as well as with the results of three preregistered studies of gratitude interventions (23–25). However, our meta-analysis extends past meta-analyses in several ways. Unlike previous meta-analyses, our meta-analysis employed more advanced methodologies that account for nonindependent effect sizes. This allowed us to analyze 727 effect sizes—significantly more than the 14 (11) to 166 (9) effect sizes used in previous meta-analyses—within a single model. As a result, we were able to derive more precise and generalizable overall effect size estimates. In addition, we were able to test various moderating variables that had not been adequately investigated due to limited number of effect sizes. Furthermore, some previous meta-analyses focused on specific populations (11), specific outcomes (7), or specific types of gratitude intervention (10). Our meta-analysis covers all populations, outcomes, and types of interventions, providing a more comprehensive overview of the literature, while also offering detailed information on each specific population, outcome, and type.

What does Hedges’ g of 0.19 (r = 0.095) observed in this meta-analysis mean? According to a study (26) that analyzed 708 correlations derived from 87 meta-analyses in social and personality psychology, rs of 0.10, 0.20, and 0.30 can be considered small, typical, and relatively large, respectively. Thus, the effect size reported here can be interpreted as small. However, a Hedges’ g of 0.19, or r = 0.095, is comparable to the effect size of the relationship between alcohol consumption during pregnancy and preterm birth (r = 0.09) and the effect size of taking antihistamines for a runny nose [r = 0.11, ref. 27]. Given that doctors recommend avoiding alcohol during pregnancy and prescribe antihistamines for runny noses based on these findings, the effect of gratitude interventions on well-being—although it may appear small—merits serious consideration. Furthermore, considering that gratitude interventions are simple, low-cost activities; that the outcome variables (e.g., life satisfaction) are relatively resistant to change; and that most participants in this meta-analysis came from nonclinical populations where ceiling effects are likely to occur (28), this effect size is particularly noteworthy. Additionally, many researchers argue that seemingly small effects can lead to significant consequences over time and at scale (29, 30). Therefore, when practiced regularly in daily life, the effects of gratitude interventions could be nontrivial.

The main contribution of this meta-analysis was to examine cultural differences in the effectiveness of gratitude interventions on well-being. We found significant between-country differences in the effects of gratitude interventions on well-being; gratitude interventions significantly increased well-being in many countries, including the United States, Switzerland, Spain, South Korea, Poland, Malaysia, Ireland, Hong Kong, Germany, China, Belgium, and Australia. Conversely, no significant effects were observed in France, India, Japan, the Netherlands, and the United Kingdom. However, the seven preregistered country-level variables we examined (individualism, GDP per capita, tightness, relational mobility, residential mobility, religiosity, and dominant religion) did not significantly moderate these effects at P values below our Bonferroni-corrected critical values.

How can we explain these findings? One possibility is that, although cultural differences in the effectiveness of gratitude interventions may exist, the country-level cultural variables examined in this meta-analysis might not sufficiently account for the variability in intervention effects across cultures. We selected these seven variables based on individual studies (12–15) demonstrating cultural differences in the effectiveness of gratitude interventions between Western and East Asian cultures, but no significant moderating effects were found. However, there may be other country-level predictors that are theoretically relevant to gratitude interventions and better explain the cultural differences. For example, previous research has shown substantial variations across cultures in ideal affect (desired emotional states; 31) and thinking styles (analytic vs. holistic; 32). Thus, it would be fruitful to examine additional country-level predictors in future meta-analytic efforts. Relatedly, we used country as a proxy for culture. However, there may be other proxies for culture that could explain the differential effects of gratitude interventions across cultures. Although information regarding cultural factors other than country was rarely reported in most of the included studies, it would be beneficial to explore cultural differences using other proxies, such as individuals’ religion, ethnicity, and language.

Another possibility is that the effects of gratitude interventions may not differ significantly across cultures. This would suggest that gratitude interventions are generally positive and effective regardless of cultural contexts, highlighting their potential for broad applicability across diverse cultural settings. If that is the case, how can we explain the country-level differences observed in the meta-analysis? These between-country differences may be attributed to factors other than cultural differences. Indeed, we found substantial heterogeneity in sample and measurement characteristics across studies conducted in different countries. For instance, in the cases of France and the Netherlands, the number of effect sizes was relatively low, with only 4 and 6, respectively. Additionally, most studies (for example, refs. 33–36) used well-being outcomes such as depressive symptoms and negative affect, which tend to have smaller effect sizes than other well-being outcomes such as positive affect. These sample and measurement characteristics may explain why the effects of gratitude interventions were not significant in these countries. On the other hand, Hong Kong showed relatively large effects of gratitude interventions on well-being. This may be partly because a couple of studies (37, 38) asked participants in the control conditions to engage in activities that elicited negative emotions (e.g., counting misfortunes). Such considerable heterogeneity observed in studies conducted across different countries may have contributed to the between-country differences and obscured the explanatory power of country-level variables. Therefore, to examine cultural differences more closely, future research should investigate country differences and country-level predictors by implementing standardized gratitude interventions across countries in terms of sample and measurement characteristics.

If gratitude interventions are equally effective across cultures, how then can the cultural differences reported in previous individual studies be explained? One plausible explanation lies in the importance of culturally or religiously tailored gratitude interventions. For instance, prior research has indicated that Asians tend to experience not only positive emotions but also negative emotions (e.g., indebtedness, guilt), when expressing gratitude toward someone rather than something (14, 15). Using our meta-analytic dataset, we further investigated the moderating effects of different types of gratitude interventions among Asians only. Although most types of gratitude interventions generally enhanced well-being, expressing gratitude to others (e.g., writing a gratitude letter) did not yield a statistically significant effect, Hedges’ g = 0.14, 95% CI [−0.03, 0.31], SE = 0.08, P = 0.106). This suggests that, for Asians, gratitude interventions that do not concurrently elicit negative emotions may be more effective. Additional support for culturally tailored approaches comes from a study conducted in Malaysia, where participants were instructed to associate their blessings with Allah (39). This Islamic-based gratitude intervention led to greater happiness compared to a secular-based gratitude intervention or a control condition. Similarly, in Hong Kong, researchers integrated meditation-like prompts into gratitude interventions, enhancing participants’ subjective well-being (37). Collectively, these findings may have implications for how to implement gratitude interventions across cultures: Culturally aligned interventions may be more effective than a one-size-fits-all approach.

The additional moderation analyses regarding study, sample, and measurement characteristics revealed interesting patterns. First, the mean age and gender composition of the sample did not significantly moderate the effects of gratitude interventions. However, it should be noted that this refers to the mean age and gender composition of the sample, not the age and gender of individuals. Since research has shown that individuals tend to respond differently to gratitude experiences based on their age (40) and gender (41), caution should be taken when generalizing these findings to individuals. Second, consistent with previous meta-analyses (7, 8, 10), the number of sessions participants engaged in for gratitude interventions appears to be less critical. It may be that voluntarily participating for an optimal duration is more important for enhancing well-being than simply engaging in a long-term intervention set by researchers (42). Third, the effects of gratitude interventions were greater when the well-being outcome was positive affect. This aligns with previous studies suggesting that gratitude functions to enhance positive emotions more than alleviate negative ones (for example, ref. 43). Therefore, to maximize their effects, gratitude interventions should be aimed at elevating positive affect rather than reducing negative affect. Fourth, the effectiveness of gratitude interventions was also significantly moderated by the type of interventions. Specifically, the effects were greater when multiple gratitude interventions were implemented rather than a single intervention. This may be because when individuals cannot choose the type of gratitude intervention (as in random assignment), they are more likely to find an activity that suits them when multiple options are available (42). Fifth, the effect size was larger for individual-level randomized controlled trials (RCTs) compared to quasi-experimental designs. This could be due to the higher internal validity of RCTs, which better control for potential confounding variables and biases, thus providing more reliable estimates of the intervention’s true effect. Last, trait gratitude negatively moderated the effect of gratitude interventions on well-being, indicating that samples with lower levels of trait gratitude gained more benefits from the interventions. This finding aligns with individual studies showing that gratitude interventions were most effective in improving well-being of individuals lower in trait gratitude (44, 45). However, again it is important to note that this refers to the trait gratitude of the samples, not of individuals. Taken together, these moderation analyses inform us about when and how the effects of gratitude interventions can be maximized.

Although the present meta-analysis has many strengths, several limitations warrant discussion. First, our inclusion criteria restricted our review to well-being outcomes that have been extensively studied in the literature. Although it allowed for a more thorough examination of the well-studied outcomes, it also means that other important aspects of well-being such as psychological well-being and anxiety were not included in our analysis. As more studies accumulate in these well-being outcomes, future meta-analyses could benefit from incorporating a broader range of well-being outcomes to provide a more comprehensive understanding of gratitude interventions. Second, our search was limited to studies written in English, which could introduce a potential monolanguage bias (46). This may limit the generalizability of our findings, as it is likely that relevant research published in other languages may yield different findings. To address this issue, future meta-analyses should test the robustness of our findings by including studies written in other languages. Third, in this meta-analysis, the search strategy was developed by the authors; however, to effectively identify potentially eligible studies, the search strategy should ideally be developed by an experienced librarian and peer-reviewed by a second librarian. Finally, it is important to note that the findings are limited to the 28 countries included in this meta-analysis. As such, it remains unclear whether our findings generalize to other cultural contexts. We encourage researchers to explore the effects of gratitude interventions in more diverse cultural contexts, as this would contribute to a richer understanding of how gratitude interventions may function across different cultural contexts.

Despite these limitations, our meta-analysis provides the most up-to-date, comprehensive, and advanced findings on the effectiveness of gratitude interventions on well-being. It also offers a thorough meta-analytic investigation of cultural differences in gratitude interventions. We hope this meta-analysis will help better understand the current state of gratitude interventions and highlight fruitful avenues for future research.

Materials and Methods

Transparency and Openness.

We preregistered the analysis plans for the meta-analysis on the Open Science Framework (https://osf.io/s4a5m). We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines in presenting the meta-analysis (47; See SI Appendix, Table S5 for the PRISMA checklist).

We initially preregistered our analysis plans on July 3rd, 2023, and updated them on August 27th, 2024. The updates included the following changes to the initial preregistered plan:

• 1.Inclusion criteria: We decided to include not only published papers but also unpublished papers, as well as not only studies using individually randomized trials (randomized controlled trials) but also quasi-experimental designs (e.g., group-level randomization).
• 2.Meta-analytic approach: We opted for a correlated and hierarchical effects (CHE) model under robust variance estimation (RVE; ref. 48) as our baseline model, instead of a cross-classified multilevel model, because the CHE model better reflects the interdependent nature of effect sizes in the meta-analysis.
• 3.Detecting influential cases: We explicitly stated in the updated preregistration how we will detect influential cases and analyze the data both with and without these cases.
• 4.Moderator analysis: We stated that we will include publication status and randomization as moderators, given that our updated data include both published and unpublished papers, and both individually randomized trials and quasi-experimental designs. We also included religiosity as a moderator, as it may influence the effectiveness of gratitude interventions on well-being.
• 5.Moderator analysis: We specified that we will first conduct separate meta-regression analyses for each moderator to assess the heterogeneity explained by each. We stated that we will then perform a multiple meta-regression analysis entering the statistically significant moderators from the univariate analyses simultaneously.
• 6.Publication bias analyses: In addition to the PET-PEESE method and the 3PSM, we stated that we will run additional publication bias analyses including Egger’s multilevel meta-analysis (MLMA) test, a significance funnel plot, and sensitivity analysis to address the dependent nature of the data.

Literature Search and Inclusion Criteria.

We conducted the initial literature search between September 2015 and March 2016. First, we searched American Psychological Association (APA) PsycINFO using “gratitude” as the search term to find as many potentially relevant papers as possible. In addition, we reviewed the reference sections of meta-analytic articles on gratitude or positive psychology interventions. The inclusion criteria at this stage were as follows: a) papers investigated the effects of a gratitude intervention as the independent variable, b) papers assessed life satisfaction, positive affect, negative affect, happiness, depressive symptoms, or a composite measure of the aforementioned variables (e.g., the mean of standardized scores of life satisfaction, positive affect, and negative affect) as outcome variables, c) papers randomized individual subjects (i.e., randomized controlled trials) and included a comparison condition, d) papers were published in a peer-reviewed journal, e) papers provided sufficient statistical information to calculate standardized effect sizes, and f) papers were written in English. Based on these inclusion criteria, we manually screened the papers and included 33 papers in our initial dataset (See SI Appendix, Fig. S6 for the PRISMA flowchart of the initial literature search).

We resumed our literature search between April and September 2023, again using APA PsycINFO and reviewing the reference sections of review articles on gratitude and positive psychology interventions. We last updated the search on May 12th, 2024, with two major changes to our inclusion criteria. First, we decided to include not only published papers but also unpublished papers to minimize potential publication bias. Second, we decided to include quasi-experimental designs, which involve group-level random assignment, in addition to individual-level randomized controlled trials. Additionally, we identified studies that recruited participants from multiple countries. Since the main purpose of this meta-analysis was to examine differences in the effects of gratitude interventions across countries, we decided to exclude the multicountry studies if they did not present data separately by country or if the authors did not respond to requests for country-specific data. Our final inclusion criteria were a) papers that investigated the effects of a gratitude intervention as the independent variable and included a comparison condition, b) papers that assessed life satisfaction, positive affect, negative affect, happiness, depressive symptoms, or composite well-being as outcome measures, c) papers that used randomized controlled trials or quasi-experimental designs, d) papers that provided sufficient statistical information, and e) papers written in English.

We expanded the electronic databases to include APA PsycINFO, Web of Science Core Collection, PubMed, and ProQuest Dissertations & Theses Global, in addition to reviewing the reference sections of relevant review articles. We applied the following search terms: gratitude AND (intervention* OR exercise* OR practic* OR training OR express* OR experiment* OR effect* OR strateg* OR promot* OR random* OR manipulat* OR treatment OR influence) AND (“life satisfaction” OR “positive affect” OR “positive emotion*” OR “negative affect” OR “negative emotion*” OR happiness OR happy OR well-being OR wellbeing OR depress*). This search strategy resulted in 4,092 papers. After removing duplicates, 2,042 papers remained for further screening.

We initially conducted the screening process using ASReview (49), an open-source tool that leverages active learning to help researchers perform systematic literature reviews more efficiently and transparently. To train the active learning model, we created a “prior knowledge” set by labeling one paper (included in the initial dataset in March 2016) as relevant to gratitude interventions and another paper as irrelevant. Following training, ASReview ordered the remaining papers from most to least relevant and presented each paper’s title and abstract one by one. The first author then marked them as relevant or irrelevant based on the titles and abstracts.

The rules for stopping the screening process followed the guidelines of refs. 50 and 51. All three conditions must be met to stop the screening:

• 1.The estimated number of papers related to gratitude intervention was determined, and screening continued until that estimated number was identified. The estimation method was as follows: 1) A subsample of 10% of the total papers (N) from the literature search was created (i.e., 2,042*0.1 = 204). 2) Within this subsample, the number of papers relevant to gratitude interventions (r) and those irrelevant (i) were manually identified by the first author, and the proportion of relevant papers [r/(r + i)] was calculated (i.e., r = 36, i = 168, 36/204 = 0.176. 3) This proportion was then multiplied by the total number of papers: N*[r/(r + i)] = 2,042*0.176 = 359. 4) The resulting number was then multiplied by 0.95: 0.95*N*[r/(r + i)] = 0.95*359 = 341. Thus, screening continued until 341 papers were identified as relevant to our search.
• 2.Screening continued until all 33 papers identified as relevant to gratitude interventions in the initial literature search were found.
• 3.Screening continued until 50 consecutive papers were identified as irrelevant to gratitude interventions.

As a result of the screening process using ASReview, 1,176 of the 2,042 papers were reviewed, and 351 were identified as relevant to our search. After obtaining the full texts, two researchers independently assessed whether the papers met the inclusion criteria. Any disagreements were resolved through discussion. If sufficient information was unavailable, we directly requested it from the authors. We contacted 51 first or corresponding authors, and 49% (25 out of 51) provided their data. Ultimately, a total of 145 papers were included in our final dataset (See Fig. 4 for the PRISMA flowchart of the last literature search. The full list of individual studies included in the present meta-analysis can be found in the SI Appendix).

Fig. 4.

PRISMA flowchart of the final literature search.

Data Extraction.

At least two independent researchers extracted study/sample, and measurement characteristics as potential moderators. For study/sample characteristics, we coded the following: publication status (published study vs. gray literature), country (the nationality of participants), age (mean age of participants in years in the total sample), gender composition (percentage of females in the total sample), and trait gratitude (mean level of trait gratitude across conditions). For measurement characteristics, we coded the following: dose, intervention types, outcome measures, research designs, control condition types, and randomization. For dose, we counted the number of sessions participants engaged in for gratitude or control conditions (e.g., daily for 7 d = 7, once a week for 3 wk = 3). For intervention types, we coded the following categories: a) counting blessings, b) remembering gratitude, c) expressing gratitude (e.g., gratitude letter/visit), d) counterfactual gratitude (i.e., imagining that a positive event might never have happened and reflecting on why it was surprising to be part of life; ref. 52), e) psychoeducational intervention, f) mixed, and g) other. We coded for the following outcome measures: a) life satisfaction, b) positive affect, c) negative affect, d) happiness, e) depressive symptoms, and f) composite well-being. Regarding research designs, the following categories were coded: a) post only, b) follow-up only, c) pre–post, d) pre-follow-up 1, e) pre-follow-up 2, and f) pre-follow-up 3. Types of control conditions included a) measurement-only, b) neutrally matched activity (e.g., listing daily activities), c) negatively valenced activity (activities eliciting negative emotions, e.g., counting misfortunes), and d) positively valenced activity (activities eliciting positive emotions, e.g., imagining the best possible self). For randomization, we coded a) randomized controlled trials and b) quasi-experimental designs.

For the main analyses, we gathered seven country-level predictors that may be linked to the effects of interventions, including individualism (53), GDP per capita (54), tightness (55), relational mobility (56), residential mobility (57), dominant religion (58), and religiosity (59).

Analytic Strategies.

Effect size and variance calculations.

We computed Hedges’ g as our effect size measure to examine the standardized mean difference in well-being between gratitude intervention and control conditions. To compute effect sizes and variances, we used means, SD (or SE), and sample sizes per condition, or other statistics (e.g., t or F statistics), along with the formulas to combine effect sizes from different research designs (i.e., posttest only with control designs and pretest-posttest-control group designs; see SI Appendix for the formulas used to compute effect sizes and variances) (60, 61). We coded the effect sizes so that higher numbers would indicate greater intervention effects, and thus reverse-coded Hedges’ gs for negative affect and depressive symptoms. Of note, for more precise estimation, gratitude conditions were compared to measurement-only or neutrally matched conditions rather than positively valenced or negatively valenced conditions when multiple control conditions were present in a single study.

Heterogeneity indices.

To assess heterogeneity, we calculated between-sample variance (τ²), within-sample variance (σ²), Cochrane’s Q, I² statistic, and 95% PI. A significant Q value indicates the overall presence of heterogeneity. The I² statistic represents the proportion of the variance in the effect size due to within-sample (Level 2) and between-sample (Level 3) variances. I² statistic is interpreted as low (25%), moderate (50%), and substantial (75%) heterogeneity (21). A 95% PI estimates the range where future studies are likely to fall based on current evidence. If the interval does not include zero, effects are likely beneficial across contexts in future research; if it includes zero, effects remain uncertain.

Meta-analytic approach.

We conducted a three-level random-effects model, specifically employing a CHE model under RVE (48). This approach was chosen because effect sizes were dependent in several ways: Multiple effect sizes were derived from the same sample (hierarchical effects), and multiple outcome measures or repeated observations of the same scale over time were included within the same sample (correlated effects). The CHE model offered several advantages. First, instead of averaging effects for each sample, it allowed us to include all effect sizes from the same sample, thereby increasing precision and statistical power. Second, it quantified three sources of variance at different levels: sampling variance (Level 1), variance between effect sizes within the same sample (Level 2), and variance between samples (Level 3). We assumed a constant sampling correlation (ρ) of 0.30 between pairs of effect sizes from the same sample. However, varying ρ in sensitivity analyses did not alter the results (See SI Appendix, Table S6 for the sensitivity analyses with varied ρ). We used the restricted maximum likelihood to estimate variance components and obtained cluster-robust variance estimates along with a small-sample adjustment.

Publication bias analyses.

We examined publication bias following recommendations by ref. 62. We employed the PET-PEESE method and the 3PSM, as these methods have been shown to be more reliable than others (62, 63). The PET-PEESE method combines two regression-based techniques: PET, which predicts effect sizes from their SE, and PEESE, which uses squared SE as predictors (64). A statistically significant slope in either regression model suggests the potential presence of small-sample bias. Both PET and PEESE also provide estimates of the true underlying effect size after accounting for such bias. If the PET estimate is not statistically significant, it is used as the final PET-PEESE estimate. Conversely, if the PET estimate is statistically significant, the PEESE estimate is used instead (64).

Similarly, 3PSM detects selection bias—the tendency to publish only significant results—by modeling three parameters: an effect size parameter, a heterogeneity parameter, and a weight parameter (i.e., the probability that studies are published depending on whether their results are statistically significant; 65, 66). 3PSM assumes that studies reporting statistically significant (P < 0.05) and positive effects are more likely to be published than those reporting nonsignificant (P ≥ 0.05) or negative effects. It then estimates the relative likelihood of publication for studies with nonsignificant or negative results compared to those with significant and positive results. Based on these estimated parameters, 3PSM provides an adjusted effect size that corrects for selection bias.

In addition, since these two tests do not account for dependencies among effect sizes, we implemented three more advanced techniques that consider these dependencies. First, alongside the contour-enhanced funnel plot (67), we used the Egger’s MLMA test (68). The contour-enhanced funnel plot displays effect sizes on the X-axis and SE on the Y-axis, with contour lines indicating different levels of statistical significance. If small-sample bias exists, the funnel plot appears asymmetrical. Egger’s MLMA test statistically assesses this asymmetry by regressing each effect size on its SE; a significant slope indicates asymmetry in the funnel plot. Second, we used a significance funnel plot (69) to visualize the extent to which effect sizes for nonaffirmative samples (i.e., negative and nonsignificant effect sizes) are smaller than those for all samples. We refer to the “worst-case” effect size as the point estimate obtained when conducting a meta-analysis using only the nonaffirmative samples. Finally, using Mathur and VanderWeele’s sensitivity analysis (69), we calculated bias-adjusted effect sizes by varying the degree of η from 1 to 200, where η represents the ratio of publishing affirmative over nonaffirmative findings.

Moderator analyses.

We conducted a series of multilevel meta-regression analyses to examine whether cultures, study/sample, and measurement characteristics would moderate the effectiveness of gratitude interventions on well-being. Continuous moderators were mean-centered, and categorical moderators with more than two categories were dummy-coded. We estimated separate meta-regression models for each moderator and then conducted a meta-regression analysis that simultaneously included the moderators identified as significant in the separate individual models. The significance of the moderating effect was determined using an omnibus F test. For categorical moderators, we used a no-intercept specification so that the coefficients represented the estimated effect for each category. If the omnibus F test was statistically significant, we further examined whether each category within a moderator was statistically significant. Due to the large number of moderators tested, we adjusted the significance level using the Bonferroni correction. We conducted the meta-analysis in R using the following R packages: metafor (version 4.6-0) (70), weightr (version 2.0.2) (71), PublicationBias (version 2.4.0) (72), and glmulti (version 1.0.8) (73).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Anonymized excel file and R files data have been deposited in Meta-analysis of gratitude interventions across cultures (https://osf.io/g5hp6) (74).