To analyze the effects of horticultural therapy and activities for older adults with dementia through a meta-analysis, 29 articles that met the criteria for meta-analysis were selected as the subjects out of total 147 research articles published over 20 years from 2003 to 2022. The criteria for inclusion in the meta-analysis were empirical studies that have been conducted in groups, allowing for pre- and postcomparisons between control and experimental groups, using experimental or quasi-experimental design methods, and were quantitative studies clearly presenting the statistical figures (mean, standard deviation, p-value) for calculating effect sizes (Cho and Kim, 2015). The data was collected by searching journals and theses/dissertations related to horticultural therapy and horticultural activities with the keyword ‘horticulture for older adults with dementia’ on the RISS (Research Information Sharing Service) website provided by the Korea Education and Research Information Service.

Out of the 147 articles searched, duplicate articles, articles unrelated to horticultural therapy/activities, articles not targeting older adults with dementia or mixed with other age groups, non-experimental studies, and articles that integrated other programs with horticultural therapy were excluded. Lastly, articles where meta-analysis could not be applied (those not clearly providing statistical figures (mean, standard deviation, p-values)) were also excluded, ultimately leaving 29 articles.

The findings of the 29 selected articles were coded using the MS Excel program. During the coding process, data for the experimental group, such as the pre- and post-test number of participants (N), mean (M), and standard deviation (SD), were entered according to specific variables (purpose of therapeutic intervention, size of the participating group, frequency per week, total number of sessions, and duration per session). Articles on the effects of horticultural therapy often use multiple measures to verify effects depending on the researcher’s interests. This results in multiple effect sizes in a single study; and independence cannot be guaranteed if these effect sizes are used as they are, and this may affect the overall effect size in the meta-analysis compared to when a single effect size is used. In this case, one of the calculated effect sizes must be randomly selected, or the average of the effect sizes must be calculated (Kim, 2006). Thus, since a single effect size cannot be calculated by taking the average when there are several dependent variables in the article, individual effect sizes were calculated. Thus, there were 102 cases of effect sizes calculated from total 29 articles selected. The coded results were analyzed using the Comprehensive Meta-Analysis (CMA) version 4 program.

When entering the values into CMA4, the direction of the effect of the CMA program option was set as positive for the positive scale indicating that increased measures lead to a positive effect, and the direction of the effect of the CMA program option was set as negative for the negative scale indicating that decreased measures lead to a positive effect, thereby ensuring consistency in the direction of effect sizes.

To determine the differences in the effectiveness of horticultural therapy for older adults with dementia, a meta-analysis was conducted on the effect sizes of 29 articles. Considering the variations in research methods, samples, and intervention methods among the articles, the analysis was conducted using a random-effects model. A total of 102 cases of effect sizes were calculated from 29 articles selected based on the established criteria. The overall average effect size for horticultural therapy targeting older adults with dementia was ES = 0.522 as shown in Table 1. Since the 95% confidence interval did not contain zero, the average effect size appeared to be significant. According to Cohen’s criteria (1988) for interpreting average effect sizes, an effect size below 0.3 is considered small, 0.5 is medium, and 0.8 or higher is large (Hwang, 2018). The overall average effect size was 0.522, which was medium. This finding is consistent with the result of meta-analysis for evaluating the appropriateness of horticultural therapy research by Jang (2010), which reported a medium effect size of 0.52 for horticultural therapy targeting dementia patients.

The results of analyzing the effect sizes for different therapeutic interventions (psychological-emotional, social, cognitive, and physical domains) of horticultural therapy programs are as follows (Table 2). The physical domain was measured through assessment of activities of daily living, grip strength tests, pinch tests, and tapping tests, and this domain showed the largest effect size, as a result of analyzing 41 cases in 16 articles. The cognitive domain was measured through the Korean Mini-Mental State Examination (K-MMSE), dementia severity rating, and neuropsychological tests, and this domain showed the effect size of 0.567 with significance, as a result of analyzing 17 cases in 13 articles (p = .001). The social domain was measured through interpersonal relationship scales and communication tools, and this domain showed the effect size of 0.541 but with no significance, as a result of analyzing 3 cases in 3 articles (p = .640). The psychological-emotional domain was measured through the Korean Geriatric Depression Scale Short Form (KGDS-SF), the Korean Geriatric Depression Tool, and self-esteem scales, and this domain showed the smallest effect size of 0.464, as a result of analyzing 41 cases in 22 articles (p = .031). The overall average effect size of all domains of the dependent variables had no significance (p = .974). The effect size of the physical domain was the largest at 0.583 (p < .001). This aligns with research showing that horticultural therapy applied to older adults had the physical effect of reducing stress hormones and improved muscle movement through outdoor gardening therapy (Lee, 2008). It is also consistent with the result that the physical domain showed the largest effect size as a result of a meta-analysis on the application of horticultural therapy to older adults with dementia in South Korea (Kang and Kang, 2021). This suggests that horticultural therapy is effective in recovering and maintaining physical functions of older adults with dementia.

The group sizes of the participants were classified into four groups: under 10 participants, 11–20 participants, 21–30 participants, and 31 participants or more. Table 3 shows the results of analyzing the average effect size by group size.

The effect size for the group with under 10 participants was significant at 0.371 (p = .001). The 95% confidence interval for the overall effect size was between .145 and .597. For the group with 11–20 participants, the effect size was not significant at 0.382 (p = .130), and the 95% confidence interval for the overall effect size was between -.113 and .878. The group with 21–30 participants had the effect size of 1.272, which was significant (p < .001).

Comparing the results of analyzing the effect sizes of horticultural therapy on older adults with dementia based on the group size of participants, it was found that the effect size was large when the participating group had 21–30 participants. This is consistent with the result that studies with larger groups showed larger effect sizes than those with smaller groups participating in horticultural therapy for older adults with dementia (Park, 2023).

The number of sessions per week for horticultural therapy targeting older adults with dementia was classified into once a week, twice a week, 3 times a week, and other, after which the effect sizes were compared. The results of the analysis are shown in Table 4.

The effect size of conducting horticultural therapy once a week on older adults with dementia was significant at 0.555 (p < .001). The 95% confidence interval for the overall effect size was between .299 and .810. The effect size of conducting the program twice a week was significant at 0.539 (p = .009), and the 95% confidence interval was between .134 and .943. The effect size of conducting the program 3 times a week was not significant at 0.384 (p = .580). The effect size of the ‘other’ category was also not significant at 0.342 (p = .217).

As a result of analyzing the average effect sizes based on the number of sessions per week, it was found that conducting the program once a week was more effective with a larger effect size. This is consistent with the results of the meta-analysis on art therapy programs suggesting that conducting the program once a week was adequate for older adults (Rim, 2013).

The total number of sessions in horticultural therapy for older adults with dementia was classified based on the criteria used in studies by Choi (2020) and Park (2023), into three groups: under 10 sessions, 11–20 sessions, and 21 sessions or more. As shown in Table 5, when 21 or more sessions of the program were conducted, the effect size was the largest at 0.603 (p < .001). The effect size for 11–20 sessions was not significant at 0.509 (p = .130), and the effect size for under 10 sessions was significant at 0.325 (p = .001).

As a result of analyzing the effect sizes based on the total number of sessions, it was found that conducting 21 sessions or more was more effective than conducting 11–20 sessions or under 10 sessions. This aligns with the result that more sessions of horticultural therapy for older adults with dementia lead to larger effect sizes (Park, 2023), and the result that conducting more than 12 sessions of art therapy for older adults with dementia shows larger effect size than conducting under 12 sessions (Jung, 2024).

The duration per session of horticultural therapy for older adults with dementia was classified into three groups: less than 60 minutes, 60–90 minutes, and 90 minutes or more. The results of comparing the effect sizes are shown in Table 6. As a result of examining the effect sizes based on the duration per session, it was found that the effect size of the program with a duration of 90–120 minutes per session was the largest at 0.637, excluding programs not providing the duration (p = .040). Moreover, the effect size of the program with a duration of less than 60 minutes was 0.515, but there was no significance (p = .092), and the effect size of the program with a duration of 61–90 minutes was 0.467 (p < .001). However, the result of Meta-ANOVA showed that there was no significance between groups (p = .919), making simple comparison between groups impossible. Considering the physical strength and attention span of older adults with dementia, the larger effect size for programs with a duration of 60 minutes or more indicates the need for further research.

Publication bias refers to the problem that arises when collected articles do not fully encompass the area of interest or topic, indicating the error of either publishing or not publishing the study results depending on the attributes or direction of the results (Higgins and Green, 2011). This occurs when published research fails to represent all research findings, and it may also include biases related to language in addition to the issue of unpublished articles (Oh, 2002). As a result of a funnel plot analysis conducted to examine the overall distribution of individual effect sizes in this study, it was visually identified that the effect sizes of the articles slightly deviated from the symmetry formed by the central line (average effect size) (Fig. 1). This suggests the presence of publication bias in the study results.

To further assess publication bias, the trim and fill method by Duval and Tweedie (2000) was employed. This method is known to enable unbiased estimation of effect sizes by estimating and filling up unreported effect sizes when the distribution of effect sizes is asymmetrical in the funnel plot (Lee, 2014). As a result, 5 articles were added, and the effect size after correction was 0.367, showing a difference from 0.522 before correction (Table 7). The statistical analysis results also reveal that the effect sizes are asymmetrical. These results confirmed the risk of publication bias in this study. Nonetheless, both the calculated average effect sizes and the corrected average effect sizes fall within the range of medium effect size compared to the interpretation criteria of Cohen (1988), indicating that there is no problem in interpretation. However, the research findings must be accepted in consideration of the possibility that the effect size may have been overestimated.