Immersion. Because neither the slow looking nor immersive mindset instructions, nor their combination, resulted in differential levels of immersion compared to the condition with no specific viewing instructions, we conducted a series of exploratory analyses on participants in the eight gallery conditions to better understand the relations between virtual gallery experiences and flourishing. These exploratory analyses focused on participants in the eight virtual gallery conditions.
Immersion scores from each of the four gallery visits were used as indicator variables in a confirmatory factor analysis to obtain an overall immersion value for each participant. These analyses were completed using the lavaan, lmtest, and sensemakr packages [94,96,97]. The confirmatory factor analysis fit the data well (χ2 = 1165.68, df = 6, pFigure 4 and Supplementary Table S12).Like the pre-registered analyses, pre-test flourishing scores were the most consistent and strongest predictor of post-test flourishing. There were, however, several aspects of flourishing that were predicted by immersion during gallery visits. Higher levels of immersion were associated with greater increases in engagement (b = 0.13, p < 0.001, f2 = 0.03), learning (b = 0.11, p = 0.001, f2 = 0.02), meaning (b = 0.04, p = 0.05, f2 = 0.01), respect (b = 0.05, p = 0.03, f2 = 0.01), anxiety (b = 0.05, p = 0.03, f2 = 0.01), stress (b = 0.05, p = 0.05, f2 = 0.01), and autonomy satisfaction (b = 0.09, p = 0.003, f2 = 0.02); however, these associations were small in magnitude. One potential reason for the counterintuitive relation between increases in anxiety and stress and increases in immersion may be that immersive engagement is cognitively demanding, which may result in heightened anxiety and stress. Additionally, there was a significant interaction for overall competency—people low in competency at pre-test showed greater gains when they were less immersed during their gallery visits (see Figure 7, panel F).
Qualitative analysis. In our second set of exploratory analyses, we analyzed the open-ended descriptions of participants’ experiences provided at the end of each of their visits to the virtual gallery.
To summarize the open response data from each participant’s gallery visit, we employed topic modeling [98]. Topic modeling provides a relatively interpretable, bottom-up summary of a corpus in a functionally similar way to the use of exploratory factor analysis in the analysis of survey data. Formally, topic models are a form of latent variable model estimated from document-level word co-occurrences that describes a corpus of texts as a mixture of a discrete set of latent topics (e.g., Document 1 belongs mostly to Topic A, but also has a little bit of Topic B, and so-on). Topics, in turn are described as a probability distribution over the words in the corpus vocabulary, such that words that frequently co-occur with one another (at the document level) will be grouped together in the same topic (e.g., in the latent Topic A, the words “couch” and “sofa” might occur with a high probability, while words such as “bankruptcy” and “lobster” might occur with a low probability).Text preprocessing was performed in part using the quanteda package [99]. Preprocessing proceeded as follows (see analysis code for further implementation details). We pooled all of participants’ responses across the multiple sessions into a single corpus. Texts were converted to lowercase, and punctuation was removed. We then removed responses with fewer than five words and removed words from the vocabulary if occurring in less than five texts Words below this threshold were first stemmed so that low-frequency terms (e.g., “activity” and “activities”) could be retained as a single stemmed term (e.g., “activ*”) if the word stem occurred with sufficient frequency. A custom set of 117 stopwords were also removed. After pruning the vocabulary, we then identified and joined high frequency multi-word phrases (e.g., “still life”). This produced a final corpus of 3050 text responses consisting of a vocabulary of 5957 words and phrases.With this corpus, we then estimated a series of correlated topic models (using the stm package [100]) with varying numbers of topics (ranging from 10 to 30). After manually inspecting model quality across these different models (e.g., topic coherence and redundancy), we settled on a final model consisting of 14 topics used for the following analyses. Note that the topic model construction process was performed entirely independently of, and prior to, the inferential analyses reported below.After the final set of topics were identified, these topics were then grouped into three categories: topics focused on participant descriptions of the art present during their visit (8 topics), topics focused on participants’ emotional states and feelings during the visit (4 topics), and topics related to navigation or qualities of the virtual gallery (2 topics). Based on these groupings, we were able to calculate the proportion of each response that belonged to these three categories. Across all responses, 56.52% (SD = 23.72%) of responses were art focused, 29.59% (SD = 20.57%) were feeling focused, and 13.89% (16.50%) were gallery or navigation focused. For the purposes of these analyses, we only focus on the proportions belonging to the art and feelings topic groupings.
Due to the dependency inherent in these topic groupings, analyses for the art and feelings topics were run separately. To obtain overall topic scores, separate confirmatory factor analyses using topic proportions for the four gallery visits as indicators were run to obtain overall latent topic scores for each participant. For each of the weekly gallery visits, the total proportion of each participant’s open-ended response that corresponded to art topics and feeling topics were used as the indicators for the factor analysis. The models fit the data well (Art: χ2 = 274.54, df = 6, p = 2 = 261.25, df = 6, p = Figure 5 and Figure 6 and Supplementary Tables S13 and S14).Although the two topic categories did not consistently predict flourishing outcomes, each were associated with two flourishing qualities. People whose open-ended responses featured greater descriptions of the artworks experienced greater decreases in negative emotions (b = −0.61, p = 0.01, f2 = 0.01) and sense of community (b = −0.56, p = 0.03, f2 = 0.01) and increases in overall thriving (b = 0.44, p = 0.04, f2 = 0.01). Additionally, the interaction between pre-test flourishing and art topic descriptions was significant for anxiety—people who were experiencing greater anxiety at pre-test showed greater reductions at post-test when their experience descriptions featured higher proportions of art-related topical content (see Figure 7, panel A).Conversely, people whose open-ended responses featured greater descriptions of their emotional state and feelings experienced greater increases in negative emotion (b = 0.80, p = 0.001, f2 = 0.02), depression (b = 0.49, p = 0.02, f2 = 0.01), and stress (b = 0.65, p = 0.006, f2 = 0.01) and decreases in overall thriving (b = −0.65, p = 0.002, f2 = 0.02) and positive feelings (b = −0.52, p = 0.01, f2 = 0.01). Like all prior analyses, all effects were small in magnitude. Additionally, four interactions between feeling topics and pre-test flourishing were significant. People who were lower in overall thriving, competency, and self-worth at pre-test showed greater gains at post-test when their experience descriptions featured lower proportions of feeling topics (see Figure 7, panels B, C, and E). People who were higher in depression at pre-test showed greater reduction at post-test when their experience descriptions featured lower levels of feeling-related topical content (see Figure 7, panel D).
留言 (0)