Published on in Vol 5 (2025)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/69321, first published .
Public Discourse Toward Older Drivers in Japan Using Social Media Data From 2010 to 2022: Longitudinal Analysis

Public Discourse Toward Older Drivers in Japan Using Social Media Data From 2010 to 2022: Longitudinal Analysis

Public Discourse Toward Older Drivers in Japan Using Social Media Data From 2010 to 2022: Longitudinal Analysis

Authors of this article:

Akito Nakanishi1 Author Orcid Image ;   Masao Ichikawa2 Author Orcid Image ;   Yukie Sano3 Author Orcid Image

1Graduate School of Science and Technology, University of Tsukuba, Ibaraki, Japan

2Institute of Medicine, University of Tsukuba, Ibaraki, Japan

3Institute of Systems and Information Engineering, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, Ibaraki, Japan

Corresponding Author:

Yukie Sano, PhD


Background: As the global population ages, concerns about older drivers are intensifying. Although older drivers are not inherently more dangerous than other age groups, traditional surveys in Japan reveal persistent negative sentiments toward them. This discrepancy suggests the importance of analyzing discourse on social media, where public perceptions and societal attitudes toward older drivers are actively shaped.

Objective: This study aimed to quantify long-term public discourse on older drivers in Japan through Twitter (subsequently rebranded X), a leading social media platform. The specific objectives were to (1) examine the sentiments toward older drivers in tweets, (2) identify the textual contents and topics discussed in the tweets, and (3) analyze how sentiments correlate with various variables.

Methods: We collected Japanese tweets related to older drivers from 2010 to 2022. Each quarter, we (1) applied to the Japanese version of the Linguistic Inquiry and Word Count dictionary for sentiment analysis, (2) employed 2-layer nonnegative matrix factorization for dynamic topic modeling, and (3) applied correlation analyses to explore the relationships of sentiments with crash rates, data counts, and topics.

Results: We obtained 2,625,807 tweets from 1,052,976 unique users discussing older drivers. The number of tweets has steadily increased, with significant peaks in 2016, 2019, and 2021, coinciding with high-profile traffic crashes. Sentiment analysis revealed a predominance of negative emotions (n=383,520, 62.42%), anger (n=106,767, 17.38%), anxiety (n=114,234, 18.59%), and risk (n=357,311, 58.15%). Topic modeling identified 29 dynamic topics, including those related to driving licenses, crash events, self-driving technology, and traffic safety. The crash events topic, which increased by 0.28% per year, showed a strong correlation with negative emotion (r=0.76, P<.001) and risk (r=0.72, P<.001).

Conclusions: This 13-year study quantified public discourse on older drivers using Twitter data, revealing a paradoxical increase in negative sentiment and perceived risk, despite a decline in the actual crash rate among older drivers. These findings underscore the importance of reconsidering licensing policies, promoting self-driving systems, and fostering a more balanced understanding to mitigate undue prejudice and support continued safe mobility for older adults.

JMIR Infodemiology 2025;5:e69321

doi:10.2196/69321

Keywords



As the global population ages, concerns about older drivers are increasing. The share of the global population aged 70 years or older is projected to rise from 6.4% in 2023 to 11.7% in 2050 [1]. Japan is one of the most rapidly aging societies, with 23.6% of its population being 70 years or older in 2023. Concurrently, the proportion of driver’s license holders among older Japanese people is also increasing, reaching 16.6% for those aged 70 years or older in 2023 [2]. In response to their decline in physical and cognitive functions, as well as tragic traffic crashes involving them, the National Police Agency of Japan has incrementally tightened licensing policies for them in 1998, 2002, 2009, 2017, and 2022 [3].

Research on older drivers’ traffic safety frequently examines their risk of traffic crashes, neurological and physical impairments, and crash causes. A global meta-review [4] finds that older drivers have a higher risk of crashes and injuries even when controlling for travel time or distance, partly because they experience declines in vision, cognitive performance, reaction time, and physical ability, despite their strong adherence to traffic laws. This makes older drivers a higher-risk group in traffic than younger drivers, who are more likely to be involved in crashes due to law violations or careless behavior. In addition to traffic crashes, studies have also discussed the effects of driving cessation, such as increased depressive symptoms [5,6] and reduced quality of life [7]. In Japan, research on older drivers is a relatively new field [8] and covers a range of topics, including dangerous driving associated with dementia [9], qualitative surveys of older drivers and their families [8,10,11], and discussions on licensing policies [12,13].

Although many studies highlight the risks associated with older drivers, they are not significantly more dangerous than other age groups in Japan when considering both their at-fault crash rate and the harm they cause to others. As shown in Figure 1, the at-fault crash rate (per 100,000 licensed drivers) for car and motorcycle drivers aged 70 years or older was 384 in 2023, which was higher than for those in middle age groups (30‐59 y: 301 and 60‐69 y: 313), but lower than for those in younger age groups (16‐19 y: 1025 and 20‐29 y: 497) [14]. Furthermore, the rate among those aged 70 years or older substantially decreased in the past decades from 874 in 2010 to 384 in 2023. It is also noted that their at-fault fatal crashes tend to result in fewer fatalities for occupants of other vehicles compared with fatal crashes caused by drivers in other age groups [15]. These statistics demonstrate that the risk of traffic crashes and resultant injuries imposed by older drivers is not as high as perceived, as driving failure frequencies do not significantly differ between age groups [4].

Figure 1. Traffic crash rates per 100,000 licensed drivers by year, categorized by age group (16‐29 y, 30‐59 y, 60‐69 y, and 70 y and above).

Despite the decreasing crash rate, negative sentiments toward older drivers are still persistent in Japan. According to a survey by the National Police Agency, 85% (n=1698) of respondents (balanced by age and sex) perceive older drivers as dangerous, and 80% (n=1596) believe that licensing for older drivers should be revised [16]. Such sentiments are potentially influenced by media reports on older drivers and their crashes as well as stereotypes about older people. Recent studies suggested a negative tone toward older drivers in news reports [17], an increase in newspaper articles about them [18], and underreported of their crashes killing themselves [19]. However, discourse regarding older drivers in social media has not been explored despite the widespread use of various social media platforms.

Today, social media plays various roles in formal and informal communication at individual, organizational, and societal levels. The number of social media users escalates globally [20], with nearly 50% of Japan’s population using Twitter (now X), a leading text-based social media platform [21]. Social media not only facilitates access to a broad range of information but also serves as a platform for public discussion, allowing individuals to express their views, opinions, and sentiments. As the reliance on online health information also intensifies [22], research into health-related topics on Twitter has expanded to include discussions on food security [23], type 1 diabetes [24], and various aspects of COVID-19, such as mask-wearing [25], vaccines [26-28], and the pandemic [29]. Consequently, applying text mining techniques, frequently employed to quantify the textual content and sentiment of tweets in these fields, offers valuable insights to facilitate a direct interpretation of societal discussions around health-related issues.

Given social media’s impacts on public opinions and policy-making, gaining insights on older drivers and their crashes from social media—beyond conventional surveys [17,30]—is crucial for strategic planning of traffic safety. In this study, we aimed to (1) examine the sentiments toward older drivers in tweets, (2) identify the textual contents and topics discussed in the tweets, and (3) analyze how sentiments correlate with various variables. To achieve these goals, we conducted sentiment analysis, topic modeling, and correlation analysis, using over 2.6 million tweets on older drivers posted in the past 13 years.


Study Workflow

Figure 2 illustrates the workflow of this study, covering data collection and processing, sentiment analysis, topic modeling, and correlation analysis.

Figure 2. Diagram of the study workflow. Tweets are collected and preprocessed, followed by sentiment analysis, topic modeling, and correlation analysis. Each figure and table number corresponds to the main text. J-LIWC: Japanese version of the Linguistic Inquiry and Word Count Dictionary; NMF: nonnegative matrix factorization.

Data Collection and Processing

We collected tweets about older drivers posted between January 1, 2010 and December 31, 2022, using the Twitter application programming interface v2 [31] with Japanese keywords “(高齢 OR 老人) AND (運転 OR ドライバー)”, or literally, “(old age OR older people) AND (driving OR driver).” To ensure the relevance of collected tweets, we examined the contents in a random sample of 1000 tweets and determined that 967 tweets pertained to older drivers or their transportation issues.

As part of text preprocessing, we removed symbols, emojis [32], URLs, and hashtags (#...), and normalized characters by converting them to lowercase and half-width forms. Tweets were then tokenized into words using MeCab [33] with the mecab-ipadic dictionary, with stopwords removed (Multimedia Appendix 1) and verbs and adjectives lemmatized. In addition, we consolidated tweets from the same user within the same quarter by merging their tokenized word lists into single documents. This approach reduces individual biases, enables dynamic trend analysis by quarter, and improves topic modeling accuracy by processing aggregated and extended textual content [34]. After pooling tweets by user and quarter, we removed words with an occurrence rate of less than 0.1% per quarter, resulting in 9287 unique words. Hereafter, we refer to the collection of quarterly tweets aggregated by each user as a “document.”

Sentiment Analysis

To clarify the Twitter users’ feelings toward older drivers, we first conducted sentiment analysis using the Japanese version of the Linguistic Inquiry and Word Count Dictionary (J-LIWC) [35]. Sentiment analysis quantifies the sentiment expressed in documents, using techniques ranging from word-based and context-based methods to deep-learning approaches. We adopted a word-based method for its interpretability and consistency with topic modeling, specifically using J-LIWC, which is psychologically validated and reliable in both English and Japanese.

The J-LIWC includes 69 categories in 4 broad categories—“psychological processes” (37 categories), “linguistic processes” (14 categories), “punctuation” (12 categories), and “other grammar” (6 categories). For this study, we used the “psychological processes” of the J-LIWC, aligning with our research objectives. The “psychological processes” consists of 10 subcategories, namely “affective processes,” “negative emotions,” “social processes,” “cognitive processes,” “perceptual processes,” “biological processes,” “drives,” “relativity,” “personal concerns,” and “informal language,” and under these subcategories, there are several further subcategories, as illustrated in Table 1. Throughout this paper, we refer to the further subcategories of the J-LIWC as sentiments.

Table 1. Proportion of documents containing words corresponding to 10 sentiments in the Japanese version of the Linguistic Inquiry and Word Count Dictionary’s “affective processes,” “negative emotions,” and “drives” within 614,429 documents discussing older drivers posted from 2010 to 2022. In addition to proportions, results of linear regression (the formula Yt=β0+β1Xt+ϵt) are represented by β0,β1, R2, and P values.
Subcategory and sentimentTweets, n (%)β0β1R2P value
Affective processes
Positive emotions217,300 (35.37)31.00.110.154.004
Negative emotions383,520 (62.42)46.10.360.462<.001
Negative emotions
Anxiety114,234 (18.59)14.00.120.395<.001
Anger106,767 (17.38)11.00.170.686<.001
Sadness23,609 (3.84)1.700.050.230<.001
Drives
Affiliation105,984 (17.25)15.60.040.049.120
Achievement179,969 (29.29)21.70.180.416<.001
Power213,716 (34.78)28.10.150.362<.001
Reward105,067 (17.10)13.10.100.412<.001
Risk357,311 (58.15)42.80.330.384<.001

Topic Modeling

Next, we employed 2 layers of nonnegative matrix factorization (NMF) [36], a form of dynamic topic modeling. Topic modeling aims to extract latent textual data and generate topic distributions categorized as “static” (ignoring time) and “dynamic” (incorporating temporal variation) [37,38]. While many previous studies use latent Dirichlet allocation [39], it does not account for temporal trends, thus unsuitable for longitudinal studies. Several models have been developed to capture topic dynamics over time, such as dynamic topic model (DTM) [40], which is an extended version of latent Dirichlet allocation, 2 layers of NMF [35], and BERTopic [41]. DTM and 2 layers of NMF rely on a bag-of-words approach, whereas BERTopic uses word embeddings. This study opted for 2 layers of NMF as it provides a richer set of top words, higher coherence scores, faster performance than DTM, and, compared with BERTopic, offers objective evaluation metrics, enables topic distribution generation for each document, and eliminates the need for labor-intensive testing [36,42].

Two layers of NMF is a method for extracting dynamic topic distributions by applying a 2-stage NMF, an unsupervised dimensionality reduction technique that decomposes data into nonnegative factors. In the first stage, document-term matrix At for each time unit (quarter) t undergoes NMF, producing a “window topic,” which represents topic distributions within each time unit. In the second stage, all “window topic”–term matrix B (a combination of all At) are further decomposed via NMF, resulting in a “dynamic topic,” which incorporates temporal changes.

Following the previous study [36], we determined the optimal number of dynamic topics by testing a range between 25 and 90, selecting the one with the highest Topic Coherence using Word2Vec score [43]. TC-W2V score is defined as the mean pairwise cosine similarity between term vectors embedded using the word2vec model [44]. In this study, we employed the Japanese Social Media Corpus [45] as a large-scale word2vec model, trained on approximately 2 million Japanese words from social media and web sources (Multimedia Appendix 1).

Statistical Tests

In each section, we employ simple linear regression to analyze temporal trends in data counts, sentiments, and topics. The simple linear regression model is expressed as follows:

Yt= β0+β1Xt+ϵt

where Yt represents the dependent variable (eg, tweet count at quarter t) and Xt denotes the independent variable (eg, quarter t). The coefficient β1 indicates the rate of change, while the intercept β0 represents the baseline value when Xt is zero. The term ϵt accounts for random fluctuations not explained by the model. For this analysis, we use the actual data count, the proportion of documents containing a given sentiment, and the mean topic proportion per time unit.

In addition, we examine the correlation between sentiments and traffic crash rates, data counts, and topics using Pearson correlation analysis. The correlation coefficient (r) is computed as follows:

r=(XtX)(YtY)(XtX)2(YtY)2

where Xt (eg, negative emotions at quarter t) and Yt (eg, tweet count at quarter t) represent the paired values of the independent and dependent variables, respectively, and X- and Y- denote their mean values. A positive r indicates a positive correlation, while a negative r suggests an inverse relationship.

To quantify the proportion of variance explained, we also compute the coefficient of determination (R2) for regression models and the squared correlation coefficient (r2) for correlation analysis. Significance levels are calculated under the null hypotheses of β1=0 for linear regression and r=0 for correlation analysis.

Ethical Considerations

This study did not require institutional review board approval because it is an observational study that used only publicly accessible data and reported aggregate results with no personal identifiers. To maintain privacy and confidentiality, all personal identifiers were removed during data processing. Transparency was maintained throughout the study, with clear communication of its purpose, methods, and findings.


Summary of Tweet Counts and Users

Our dataset contained 2,625,807 tweets from 1,052,976 unique users, comprising original tweets (n=767,419, 29.25%), retweets (n=1,678,133, 63.91%), replies (n=171,781, 6.54%), and quoted tweets (n=8474, 0.32%). To analyze primary opinions or original tweets, we excluded retweets and eliminated referenced text from quoted tweets and replies. Figure 3 illustrates the quarterly counts of original tweets and unique users discussing older drivers over a 13-year period. There has been a consistent presence of tweets and users, with a moderate increase over time. This trend is supported by linear regression (tweet count: β0=−3011.09, β1=832.75, R2=0.256, P<.001; user count: β0=−1325.86, β1=534.90, R2=0.238, P<.001).

Figure 3. Tweet and user counts related to older drivers by quarter from 2010 to 2022. Black vertical dotted lines indicate the 3 major tweet peaks.

Since 2016, the number of tweets per quarter has exceeded 10,000, rising to over 20,000 per quarter from 2019 onward. The highest peak occurred in the second quarter of 2019 (approximately 162,000 tweets), driven by a crash on April 19, 2019, in Higashi-Ikebukuro, Tokyo, where an 87-year-old man mistakenly stepped on the accelerator, killing a 3-year-old child and her mother and injuring 9 others [46]. The second peak, in the fourth quarter of 2016 (approximately 67,000 tweets), coincided with a series of crashes in Tochigi (November 10, 2016), Tokyo (November 12, 2016), and Miyazaki (November 13, 2016), where older drivers struck and killed pedestrians, sparking active public debate about older drivers [47]. Another significant peak in the fourth quarter of 2021 (approximately 60,000 tweets) followed a crash in Osaka, where an 89-year-old man mistakenly stepped on the accelerator, resulting in pedestrian casualties [48]. Hereinafter, we plot black dotted lines to indicate the 3 major tweet peaks in each figure.

Sentiments Toward Older Drivers

After processing, our final dataset comprised 614,429 documents posted by 404,689 unique users. Table 1 presents the proportion of documents containing words associated with sentiments in the primary emotional subcategories (“affective processes” and “negative emotions”) as well as “drives.”

Across all quarters, negative emotions (n=383,520, 62.42%) were more prevalent than positive emotions (n=217,300, 35.37%) in “affective processes.” Within the subcategory of “negative emotions,” anxiety (n=114,234, 18.59%) and anger (n=106,767, 17.38%) were more frequent than sadness (n=23,609, 3.84%). In “drives,” risk (n=357,311, 58.15%) was the most frequently observed sentiment. Sentiments in other categories, none of which exceeded an average proportion of 50%, are presented in “S2. Sentiment analysis Details” in Multimedia Appendix 1. For instance, family (n=65,069, 10.59%) in “social processes,” insights (n=261,040, 42.48%) in “cognitive processes,” and health (n=80,967, 13.18%) in “biological processes” were the most predominant sentiments in their respective categories.

Temporal trends of sentiments are plotted in Figure 4, with the numeric results of linear regression presented in Table 1. Regarding “affective processes,” while positive emotions shows a slight increase (β1=0.11, P=.004), negative emotions increased more prominently by 0.36% per quarter (P<.001), rising from approximately 40% to 60%, with spikes aligning with the 3 previously discussed data peaks, as indicated by black dotted lines. All sentiments in “negative emotions” display an increasing trend, with anxiety (n=114,234, 18.59%, β1=0.12%, P<.001) and anger (n=106,767, 17.38%, β1=0.17%, P<.001) showing relatively strong upward trends compared with sadness (n=23,609, 3.84%, β1=0.05%, P<.001). Regarding “drives,” risk remained consistently above 40% and showed a noticeable increase over time (β1=0.33%, P<.001) relative to other sentiments. Additional analyses can be found in “S2. Sentiment Analysis Details” in Multimedia Appendix 1, indicating that insights (n=261,040, 42.48%, β1=0.25%, P<.001) and hear (n=87,281, 14.21%, β1=0.19%, P<.001) were relatively dominant and exhibited increasing trends.

Figure 4. The Japanese version of the Linguistic Inquiry and Word Count Dictionary quarterly trends in “affective processes,” “negative emotions,” and “drives.” Black vertical dotted lines indicate the 3 major tweet peaks as illustrated in Figure 3.

Topics About Older Drivers

Topic modeling with 2 layers of NMF identified 29 dynamic topics that achieved the highest coherence score of 0.22 in an experiment varying the number of topics between 25 and 90 (Multimedia Appendix 1).

We further selected 16 topics that constituted at least 3.5% of all documents on average, naming them based on the top-weighted words and representative original tweets for each topic. Finally, we selected 9 topics where the R2 value exceeded 0.1 in simple linear regression, represented by a red dashed line, as illustrated in Table 2. The details of all 16 topics, including typical tweets (in both Japanese and English), as well as the regression results for the 7 topics whose R² did not exceed 0.1, are provided in Multimedia Appendix 1.

Table 2. Description of 9 topics, including topic names, top 10 weighted words, proportions, and results of linear regression (the formula Yt=β0+β1Xt+ϵt).
Topic nameTop 10 weighted wordsTweets, n (%)β0β1R2P value
Topic 1:
License surrender
return, license, voluntary, revoke, system, forcefully, surrender, certificate, revocation, discount43,270 (7.04)4.370.080.278<.001
Topic 2:
Crash events
crash, cause, increase, traffic, report, death, decrease, frequent, occur, prevention35,396 (5.76)2.730.070.358<.001
Topic 5:
Traffic safety
traffic, safety, campaign, prevention, bicycle, drinking, nationwide, walking, public, child25,139 (4.09)6.13−0.050.123.010
Topic 6:
Ikebukuro incident
runaway, Ikebukuro, Tokyo, NHK, defendant, crash, bereaved, family, memoriala24,430 (3.98)−0.100.090.343<.001
Topic 7:
Self-driving technology
self-driving, social, Japan, technology, necessary, safety, soon, experiment, popularization, development24,235 (3.94)1.950.050.151.004
Topic 8:
Social issues
issue, social, consider, life, rural, solution, necessary, difficult, traffic, local23,935 (3.90)1.570.050.294<.001
Topic 9:
License renewal
renewal, license, course, take, exam, test, center, excellent, Emperor, Kobe23,525 (3.83)5.83−0.050.126.010
Topic 10:
Discussing senior driving
say, bad, hear, go, oneself, come, same, can, young, complain23,511 (3.83)1.300.070.402<.001
Topic 14:
Driving errors
brake, accelerator, step, mistake, wrong, pedal, crash, parking, MTb, operation21,741 (3.54)2.230.040.199<.001

aThe frequently mentioned name of the individual involved in this crash has been masked to consider ethical concerns.

bMT indicates Manual Transmission, which is one type of vehicle transmission system.

The most prevalent topics were License surrender (n=43,270, 7.04%) and Crash events (n=35,396, 5.76%), followed by Traffic safety, Ikebukuro incident, Self-driving technology, Social issues, License renewal, Discussing senior driving, and Driving errors. Other topics include Thoughts on older drivers, Prevalent older drivers, and News media (“S3. Topic Modeling Details” in Multimedia Appendix 1).

Figure 5 illustrates the temporal trends of topic proportions for 9 topics with a red dashed line, representing the simple linear regression result for each topic. The proportions of topics 1 (License surrender, β1=0.08, P<.001), 2 (Crash event, β1=0.07, P<.001), 6 (Ikebukuro incident, β1=0.09, P<.001), 8 (Social issues, β1=0.05, P<.001), 10 (Discussing senior driving, β1=0.07, P<.001) and 14 (Driving errors, β1=0.04, P<.001) show consistent increases over time. Notably, topic 1 peaked in 2016, when measures to promote license surrender in rural areas gained attention. Topics 2 and 8 correspond to tweet count peaks, while topic 6 coincides specifically with the fourth quarter of 2019 subsequent periods. Although less pronounced, topic 7 (Self-driving technology, β1=0.05, P=.004) also exhibits an increasing trend, whereas topics 5 (Traffic safety, β1=−0.05, P=.01) and 9 (License renewal, β1=−0.05, P=.001) are declining. Additional analysis can be found in “S3. Topic Modeling Details” in Multimedia Appendix 1). Notably, News media exhibits topic peaks in the fourth quarter of 2016, reflecting unique characteristics not captured by simple linear regression.

Figure 5. Temporal trends of 9 topic mean proportions, shown as solid blue lines, with red dashed lines representing regression lines. Gray shaded areas represent 95% CIs for the regression lines. Black vertical dotted lines indicate the 3 major tweet peaks as illustrated in Figure 3.

Correlation Analysis on Sentiments

Table 3 and Table 4 present the results of the correlation analysis for positive emotions and negative emotions in “affective processes,” risk in “drives” and anxiety, anger, and sadness in “negative emotions,” across 4 crash rates, 2 data counts, and 9 topics. While the correlation analysis for crash rates among individuals aged more than 70 years and 80 years is based on yearly data (n=13) due to the limited availability of statistical data [14], the other analyses are based on quarterly data (n=52).

Table 3. Results of the correlation analysis for positive emotions, negative emotions, and risk.
VariablesPositive emotionsNegative emotionsRisk
Rr2P valueRr2P valueRr2P value
Crash rate (n=13)
16‐29–0.7470.559.003–0.7500.563.003–0.7020.493.007
30‐59–0.7140.510.006–0.7840.615.002–0.7390.546.004
60‐69–0.6870.473.009–0.7990.638.001–0.7540.568.003
Over 70–0.7120.507.006–0.7870.619.001–0.7440.554.004
Data count (n=52)
Tweet count0.3170.101.020.6830.466<.0010.6770.459<.001
User count0.3110.097.030.6740.454<.0010.6710.450<.001
Topic (n=52)
Topic 10.0390.002.780.3960.157.0040.3640.132.008
Topic 20.1730.030.220.7610.579<.0010.7190.517<.001
Topic 5–0.0370.001.79–0.1730.030.220–0.1110.012.430
Topic 60.4250.180.0020.5460.298<.0010.5540.306<.001
Topic 70.0030.000.980.3510.123.0100.3810.146.005
Topic 80.2700.073.050.7140.510<.0010.6930.480<.001
Topic 90.2170.047.12–0.4960.246<.001–0.5260.277<.001
Topic 100.2640.069.060.4250.181.0020.3650.133.008
Topic 140.1770.031.210.430.185.0010.410.168.003
Table 4. Results of the correlation analysis for anxiety, anger, and sadness.
VariablesAnxietyAngerSadness
Rr2P valueRr2P valueRr2P value
Crash rate (n=13)
16‐29–0.7540.568.003–0.9380.881<.001–0.7420.550.004
30‐59–0.7950.631.001–0.9250.856<.001–0.7440.554.004
60‐69–0.8180.669<.001–0.9130.833<.001–0.7430.551.004
Over 70–0.7880.620.001–0.9200.847<.001–0.7350.540.004
Data count (n=52)
Tweet count0.4090.168.0030.5190.269<.0010.4420.196.001
User count0.4070.166.0030.5160.266<.0010.4380.192.001
Topic (n=52)
Topic 10.4190.175.0020.2970.088.0300.1580.025.260
Topic 20.5260.277<.0010.5480.300<.0010.3800.144.005
Topic 5–0.1210.015.390–0.1860.035.190–0.1320.017.350
Topic 60.3750.141.0060.6770.459<.0010.5120.262<.001
Topic 70.2910.085.0400.2350.055.0900.0000.000≥.99
Topic 80.4970.247<.0010.4850.235<.0010.3960.157.004
Topic 9–0.5140.264<.001–0.3860.149.005–0.0280.001.840
Topic 100.3850.148.0050.6820.465<.0010.2870.082.040
Topic 140.2460.061.0800.2580.066.0700.3660.134.008

For the crash rate of individuals aged between 60 and 69 years, negative emotions (r=−0.80, P=.001) and anxiety (r=−0.82, P<.001) exhibit strong negative correlations, similar to those observed for individuals aged more than 70 years. Reversely, anger exhibits the strongest negative correlation with the youngest age group (16-29 y). Regarding the tweet count, negative emotions (r=0.68, P<.001), risk (r=0.68, P<.001), and anger (r=0.52, P<.001) exhibit moderate to strong positive correlations, similar to the user count.

For topics, topics 2 (Crash events, r=0.76, P<.001), 6 (Ikebukuro incident, r=0.55, P<.001), 8 (Social issues, r=0.71, P<.001), and 9 (License renewal, r=−0.50, P<.001) are significantly correlated with negative emotions, similar to risk. Furthermore, while positive emotions do not show significant correlations with any topic, other sentiments are correlated with multiple topics: anxiety with topics 2, 8, and 9; anger with topics 2, 6, 8, and 10 (Discussing senior driving); and sadness with topic 6.


Principal Findings

In this study, we conducted a quantitative analysis of discourse regarding older drivers from 2010 to 2022 in Japan using Twitter. The number of tweets and users discussing older drivers has increased since 2016, with peaks observed in 2016, 2019, and 2021. Sentiment analysis revealed that negative emotions were more prevalent and increased over time compared with positive emotions. In addition, contexts related to anxiety, anger, and risk were prevalent and showed an upward trend. Topic modeling identified themes primarily related to driving licenses, crash events, personal perspectives, and traffic issues. Chronologically, despite decreasing trends in Traffic safety and License renewal, topics such as Crash events and License surrender are increasing. Finally, correlation analysis revealed that negative emotions were negatively correlated with crash rates among older drivers, positively correlated with tweet counts, and positively associated with topics such as Crash events and Ikebukuro incident.

This study quantified issues surrounding older drivers through Twitter, a leading social media platform, contributing to advancements in research on public health and ageism. The function of social media, such as posting and sharing tweets on any subject along with figures and URLs in Twitter, enables to prompt the swift dissemination of information [49] and derive population-level inferences. These have fostered new fields in public health sectors such as infoveillance [50], digital epidemiology [51], and digital disease detection [52]. They also led to an increase in health-related studies in Japan including hospitals [53], disease information [54], and eHealth literacy [55]. In addition to public health, the issue of older drivers has the aspects of ageism, which encompass our thoughts (stereotypes), emotions (prejudice), and behaviors (discrimination) toward others based on age [56]. Ageism toward older people is especially high [57] and can adversely affect physical and mental health, such as reduced cognitive function [58], shorter life expectancy [59], deteriorated mental health [60], and increased isolation [61], prompting initiatives to counteract it [62]. Research on ageism is seen across multiple domains including health research [63], mental health services [64], long-term care facilities [65,66], workplaces [67,68], and media representations [69-71], with recent Twitter-based approaches employing thematic [72-76], descriptive [75,76], computational modeling [77,78], and quantitative text analyses [79-82]. This study, therefore, addresses the issues surrounding older drivers from both public health and ageism perspectives, contributing insights toward mitigating these challenges.

An abrupt increase in tweet counts sporadically observed over time is likely attributable to high-profile crashes caused by older drivers. The highest peak in the second quarter of 2019 corresponding to topic 6 (Ikebukuro incident) coincided with a crash that occurred in Higashi-Ikebukuro, Tokyo, drawing significant media and online attention for several months due to widespread sadness and anger toward the man, leading to the highest recorded number of driver’s license surrenders in 2019 [2]. The peak in the fourth quarter of 2016, associated with topic 8 (Social issues) and topic 11 (News media) coincided with a series of crashes, and the peak in the fourth quarter of 2021, followed by a crash in Osaka, led to an increase in tweets related to topic 2 (Crash events) and topic 14 (Driving errors).

Regarding sentiments, all 3 peaks coincided with peaks in negative emotions and risk, with average proportions remaining somewhat higher after the peaks than before, while positive emotions were less relevant. Thus, the data peaks of tweets about older drivers are largely facilitated by contexts of negative emotions and risks, supported by the trends in their specific crashes aggregated by topics.

The detected topics help us understand how society forms opinions and perceptions toward older drivers in the long run. Among the 16 topics accounting for over 3.5%, we identified 7 increasing topics and 2 decreasing topics. Driving licenses are discussed in both topics 1 and 9, with public discourse increasingly focusing on license surrender (topic 1 is rising) rather than renewal (topic 9 is declining), suggesting a social shift toward discouraging older adults from driving and a potential impact on the number of license surrenders [2]. Topic 2 (Crash events), including tweets such as “It seems that crashes involving older drivers are occurring frequently” and “Let’s prevent crashes involving older drivers,” reflects general discussions about such events, often triggered by specific incidents, as seen in data peaks in 2016 and 2019. In contrast, topic 6 (Ikebukuro incident) includes a summary of the Ikebukuro incident and the perpetrator [46], showing a notable increase in 2019, and topic 14 (Driving errors) highlights that crashes caused by older drivers often occur due to physical or operational limitations, as shown in meta-review [4]. Over the data collection period, topic proportions shifted from topic 5 (Traffic safety; prominent until 2015) and topic 7 (Self-driving technology; high between 2013 and 2019) to topics 8 (Social issues) and 10 (Discussing senior driving), both of which have recently increased. These topics express concerns about older drivers from societal (topic 8) and family (topic 10) perspectives. Other recurring topics include Thoughts on older drivers, Prevalent older drivers, and News media, which consistently appear in this discourse. Notably, topic 11 (News media), which includes references to major media outlets, showed a high proportion between 2016 and 2017, suggesting a potential media overemphasis on older driver–related issues. (Example tweets for these and other topics not discussed in the main text are provided in “S3. Topic Modeling Details” in Multimedia Appendix 1).

Our analysis clarifies the sentiments in public discourse and their correlations, particularly negative emotion, anxiety, anger, and risk, which have shown a persistent and increasing trend over time. The proportion of documents expressing negative emotions rose from approximately 40% to 60%, accounting for an increase of 1.4% per year. This trend is strongly positively correlated with data count and with topics 2 (Crash events), 6 (Ikebukuro incident), and 8 (Social issues). In contrast, negative emotions are most significantly and strongly negatively correlated with the crash rates of older adults (aged 60-69 y and 70+ y; P=.001), although similar trends are also observed in other age groups. These suggest that negative perceptions toward older drivers arise from heightened attention to their crash events and recognition of the issue as a social problem. Similar to negative emotions, the sentiment of risk increased from 40% to 60% (1.3% per year), with peaks around 70% in 2016 and 2019. This indicates that public perception includes not only negativity toward older drivers but also a sense of danger regarding their driving. Within “negative emotions,” anxiety and anger, each accounting for approximately 20%, are more prevalent than sadness. Although they do not show sharp peaks, both have steadily increased over time. While their correlation patterns are similar to those of negative emotions, stronger correlations are observed; the negative correlation between anxiety and topic 9 (License renewal) suggests decreasing focus on continued driving, while the positive correlation between anger and topic 10 (Discussing senior driving) indicates emotional overrepresentation of older drivers in discourse. Positive emotions show no significant correlation with any of the variables. Sadness, however, is correlated with topic 6 (Ikebukuro incident), suggesting compassion for the mother and child who were victims in the incident. Overall, driving by older people is increasingly perceived as negative and risky, and this perception is influenced by strong negative emotions such as anxiety and anger, which amplify public discourse on the topic despite a declining trend in their actual crash rates.

Finally, we present our perspectives. First, negative stereotypes and prevailing public sentiment toward older drivers may influence licensing policies [3]. Our analysis reveals contrasting trends between topics related to license renewal and license surrender, suggesting a shift in public perception toward encouraging older individuals to stop driving. However, this shift tends to overlook the potential adverse effects of driving cessation, as noted in previous studies [5-7]. Although analysis of the relationship between public discourse and licensing policies remains limited, it is essential to reconsider current policies that may discourage older people from driving—such as overly strict licensing requirements [4,5] and campaigns promoting license surrender [2]—in favor of approaches that both reduce traffic crashes and ensure their safety.

Second, the disproportionate public attention given to traffic crashes involving older drivers is a significant concern. Although the anger directed at tragic crashes involving older drivers is understandable, many older individuals follow traffic laws and drive responsibly [4], therefore it is unfair that the entire older population faces negative consequences as a result. To address this, it is necessary to implement risk-based licensing policies that focus on high-risk individuals, as well as to promote autonomous driving technologies and infrastructure that support continued safe mobility for older adults.

Third, public debate following crashes involving older drivers often centers on specific high-profile incidents rather than on the overall frequency or statistical context of such crashes. This may be partly due to media tendencies to sensationalize older driver involvement, as seen in the temporary increase in topic 11 (News Media). Media use influences the formation and stability of public opinion clusters [83] and can contribute to heightened public anxiety during periods of uncertainty, such as pandemics [84]. Providing the public with a more accurate and balanced understanding of older drivers—including countering media-driven bias—is crucial to fostering constructive discourse. This may involve encouraging more responsible media reporting and promoting public access to objective information on older driver safety.

In summary, this study offers a long-term analytical approach that integrates sentiment and textual content, contributing to research on public discourse related to aging issues and, more broadly, public health. Our findings enhance the understanding of societal perceptions and policy implications surrounding older drivers in Japan and may inform future discussions on aging societies around the world.

Limitations

Our study had several limitations. First, while Twitter is one of the most widely used social media platforms in Japan [85], Twitter users do not necessarily represent the general population [86]. Furthermore, Twitter provides access only to tweets from public accounts, deterring extreme or controversial textual content to be tweeted or we might have missed such tweets because past tweets can be deleted. In addition, display algorithms were modified around 2011 due to a rapid increase in Twitter users, resulting in nonstationarities in the dataset. To mitigate such user biases, several approaches have been proposed, including evaluating users based on the authenticity and bias of their engagements [87], and accounting for population demographics and word ambiguity [88].

Second, while the keyword (“(高齢 OR 老人) AND (運転 OR ドライバー)” in Japanese) effectively targets discussions about older drivers, there is still room for refinement. “運転” (driving) and “ドライバー” (driver) are the words related to driving, representing specifically older driving by combing words “高齢” (denoting old age) or “老人” (denoting older people). Although adding “事故” (crash) was considered, this word could collect data unrelated to crashes like falls in daily life, without words related to driving. Other age-related words like “80歳のドライバー” (80-year-old drivers) or “祖父” (grandfather) specify too detailed ages or familial relations, which are less general and more restrictive, so we did not use them aligning with our research objectives. In addition, we found that a small portion (33 out of 1000) of the sampled tweets were not related to older drivers but represented older people in nondriving contexts such as riding the bus or being crash victims. Therefore, while the data from this study appears to be sufficiently valid, there is a need to refine the keyword or explore other methods to collect more accurate data on older drivers.

Third, our methods have room for improvement. Word-based approaches for sentiment analysis and topic modeling offer valid and intuitive interpretations for emerging themes such as older driver–related discourse, but they may not fully capture contextual nuances such as negations or compound expressions. Although the J-LIWC dictionary covered 48% (4458 out of 9287) of the words used in this study—indicating substantial coverage with room for refinement—our findings remain persuasive, especially when compared with the Japanese version of the Moral Foundations Dictionary [89], which covers only 3% (275 words) (“S3. Sentiment Analysis Details” in Multimedia Appendix 1). Furthermore, the coherence score for the dynamic topic model was 0.22, lower than the commonly accepted baseline of 0.36 [36]. This may be influenced by the linguistic characteristics of Japanese and the use of a Japanese word2vec model. Finally, as correlation analysis does not imply causation, more rigorous statistical testing is needed. Therefore, although our word-based method using reliable dictionaries offers high interpretability, future work should explore alternative methods—such as deep-learning–based sentiment analysis or BERTopic for topic modeling—and incorporate techniques like supervised machine learning or community detection to gain deeper insights [24,41,90,91].

Conclusions

Despite the actual decline in crash rates among older drivers, previous surveys in Japan reveal persistent negative sentiments toward them, suggesting the need to understand discourse on social media, which serves as a platform for public debate. To quantitatively assess public awareness, we collected and analyzed tweets related to older drivers. The results revealed an increase in the number of tweets from 2010 to 2022, with certain peaks aligning with heightened attention to crashes involving older drivers. Negative emotions and risk consistently displayed high and rising levels, primarily correlated with the topic of crash events. Furthermore, there are diverse topics related to drivers’ licenses, crash events, and traffic safety. These imply unfair public recognition toward older drivers, suggesting the need for reconsidering current license policies, promoting self-driving systems, and facilitating accurate and balanced understanding, in order to provide continued safe mobility for older adults.

Acknowledgments

This work was supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI grants 23K21516 and 23K28192.

Conflicts of Interest

None declared.

Multimedia Appendix 1

Supplementary details of the analysis.

DOCX File, 1264 KB

  1. United Nations. Population division world population prospects 2024. 2024. URL: https://population.un.org [Accessed 2025-04-23]
  2. National Police Agency. Driver’s license statistics: 2024. 2024. URL: https://www.npa.go.jp/publications/statistics/koutsuu/menkyo.html [Accessed 2024-07-16]
  3. National Police Agency. Report on measures against older drivers’ traffic crashes. 2020. URL: https://www.npa.go.jp/koutsuu/kikaku/koureiunten/menkyoseido-bunkakai/prevention/final_report.pdf [Accessed 2024-08-12]
  4. Duc-nghiem N, Hoang-tung N, Kojima A, Kubota H. Traffic safety of elderly road user: the global trend and the Japanese case. J JSCE, Ser D3. 2016;72(5):I. [CrossRef]
  5. Marottoli RA, de Leon CFM, Glass TA, et al. Driving cessation and increased depressive symptoms: prospective evidence from the New Haven EPESE. J American Geriatrics Society. Feb 1997;45(2):202-206. URL: https://agsjournals.onlinelibrary.wiley.com/toc/15325415/45/2 [CrossRef]
  6. Chihuri S, Mielenz TJ, DiMaggio CJ, et al. Driving cessation and health outcomes in older adults. J Am Geriatr Soc. Feb 2016;64(2):332-341. [CrossRef] [Medline]
  7. Farquhar M. Elderly people’s definitions of quality of life. Soc Sci Med. Nov 1995;41(10):1439-1446. [CrossRef] [Medline]
  8. Asahara K, Joichi Y, Tanimoto Y, Hikawa Y, Miyoshi Y. Literature review of research trends on older adult drivers in Japan from 2009 to 2019: analysis by text mining. J Yonago Med Assoc. 2020;71:18-26. URL: https://repository.lib.tottori-u.ac.jp/records/4315 [Accessed 2025-06-06]
  9. Nakano T, Ogura Y, Yamasaki H, Yamada M. Assessment method of elderly cognitive function while driving. IEEJ TransEIS. 2016;136(4):502-508. [CrossRef]
  10. Nakahira Y, Okuda M, Torii J. Awareness in care managers and welfare volunteers regarding driving by elderly people. Bull Ehime Pref Univ Health Sci. 2011;8(1):19-25. URL: https://cir.nii.ac.jp/crid/1520853834764493440 [Accessed 2025-06-06]
  11. Watanabe T, Fujikake K, Suzuki R. Survey on driving behavior of elderly drivers with dementia from the perspective of caregiving families. J Jpn Med J. 2009;4443:89-93. URL: https://cir.nii.ac.jp/crid/1522262178388046464 [Accessed 2025-06-06]
  12. Ichikawa M, Inada H, Nakahara S. Effect of a cognitive test at license renewal for older drivers on their crash risk in Japan. Inj Prev. Jun 2020;26(3):234-239. [CrossRef] [Medline]
  13. Ichikawa M, Inada H, Nakahara S. Increased traffic injuries among older unprotected road users following the introduction of an age-based cognitive test to the driver’s license renewal procedure in Japan. Accid Anal Prev. Mar 2020;136:105440. [CrossRef] [Medline]
  14. National Police Agency. Traffic accident statistics: Annual report 2023. 2023. URL: https://www.npa.go.jp/publications/statistics/koutsuu/toukeihyo_e.html [Accessed 2024-07-16]
  15. Ichikawa M, Nakahara S, Taniguchi A. Older drivers’ risks of at-fault motor vehicle collisions. Accid Anal Prev. Aug 2015;81:120-123. [CrossRef] [Medline]
  16. National Police Agency of Japan. Materials for the prevention measures for older drivers. 2019. URL: https://www.npa.go.jp/koutsuu/kikaku/koureiunten/menkyoseido-bunkakai/prevention/final_report.pdf [Accessed 2024-06-12]
  17. Hoshi M. Is driving by the elderly dangerous. Economic Prism. 2015. URL: https://cir.nii.ac.jp/crid/1520010380785444992 [Accessed 2024-07-17]
  18. Nemoto M, Taniguchi A. Qualitative analysis of newspaper of traffic safety for older drivers. 2020. Presented at: Proceedings of the 62nd Conference on Infrastructure Planning and Management of the Japan Society of Civil Engineers. URL: http://library.jsce.or.jp/jsce/open/00039/202011_no62/62-29-01.pdf
  19. Ichikawa M, Tanaka R, Nakanishi A, Sano Y. News coverage of older drivers' fatal car crashes: Is it over-represented? J Epidemiol. 2024;advpub:JE202402. [CrossRef]
  20. Kemp S. Digital 2024: Global overview report. 2024. URL: https://datareportal.com/reports/digital-2024-global-overview-report [Accessed 2024-08-15]
  21. World Population Review. Twitter users by country / X users by country 2024. 2024. URL: https://worldpopulationreview.com/country-rankings/twitter-users-by-country [Accessed 2024-08-15]
  22. Song H, Omori K, Kim J, et al. Trusting social media as a source of health information: online surveys comparing the United States, Korea, and Hong Kong. J Med Internet Res. Mar 14, 2016;18(3):e25. [CrossRef] [Medline]
  23. Molenaar A, Lukose D, Brennan L, Jenkins EL, McCaffrey TA. Using natural language processing to explore social media opinions on food security: sentiment analysis and topic modeling study. J Med Internet Res. Mar 21, 2024;26:e47826. [CrossRef] [Medline]
  24. Bedford-Petersen C, Weston SJ. Mapping individual differences on the internet: case study of the type 1 diabetes community. JMIR Diabetes. Oct 15, 2021;6(4):e30756. [CrossRef] [Medline]
  25. Xu WW, Tshimula JM, Dubé È, et al. Unmasking the Twitter discourses on masks during the COVID-19 pandemic: user cluster-based BERT topic modeling approach. JMIR Infodemiology. 2022;2(2):e41198. [CrossRef] [Medline]
  26. Zhang J, Wang Y, Shi M, Wang X. Factors driving the popularity and virality of COVID-19 vaccine discourse on Twitter: text mining and data visualization study. JMIR Public Health Surveill. Dec 3, 2021;7(12):e32814. [CrossRef] [Medline]
  27. Boon-Itt S, Skunkan Y. Public perception of the COVID-19 pandemic on Twitter: sentiment analysis and topic modeling study. JMIR Public Health Surveill. Nov 11, 2020;6(4):e21978. [CrossRef] [Medline]
  28. Kobayashi R, Takedomi Y, Nakayama Y, et al. Evolution of public opinion on COVID-19 vaccination in Japan: large-scale Twitter data analysis. J Med Internet Res. Dec 22, 2022;24(12):e41928. [CrossRef] [Medline]
  29. Xue J, Chen J, Chen C, Zheng C, Li S, Zhu T. Public discourse and sentiment during the COVID 19 pandemic: using latent Dirichlet allocation for topic modeling on Twitter. PLOS ONE. 2020;15(9):e0239441. [CrossRef] [Medline]
  30. Show 59th year research report: study on the actual conditions and awareness of elderly drivers. Japan Safe Driving Center. 1985. URL: https://www.jsdc.or.jp/Portals/0/pdf/library/research/s59_1.pdf [Accessed 2024-07-18]
  31. Twitter API. 2023. URL: https://developer.x.com/ [Accessed 2025-04-23]
  32. kyokomi. Emoji. 2024. URL: https://github.com/kyokomi/emoji [Accessed 2024-07-18]
  33. Sato T, Hashimoto T, Okumura M. Implementation of a word segmentation dictionary called mecab-ipadic-NEologd and study on how to use it effectively for information retrieval (in Japanese). Proceedings of the Twenty-three Annual Meeting of the Association for Natural Language Processing. 2017:NLP2017-NLPB6-1. URL: https://github.com/neologd/mecab-unidic-neologd
  34. Hong L, Davison BD. Empirical study of topic modeling in twitter. Presented at: Proceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; Jul 25-28, 2010:80-88; Washington DC District of Columbia. [CrossRef]
  35. Igarashi T, Okuda S, Sasahara K. Development of the Japanese version of the linguistic inquiry and word count dictionary 2015. Front Psychol. 2022;13:841534. [CrossRef] [Medline]
  36. Greene D, Cross JP. Exploring the political agenda of the European parliament using a dynamic topic modeling approach. Polit Anal. Jan 2017;25(1):77-94. [CrossRef]
  37. Blei DM. Probabilistic topic models. Commun ACM. Apr 2012;55(4):77-84. [CrossRef]
  38. Vayansky I, Kumar SAP. A review of topic modeling methods. Inf Syst. Dec 2020;94:101582. [CrossRef]
  39. Blei DM, Ng AY, Jordan MI. Latent Dirichlet allocation. J Mach Learn Res. 2003;3:993-1022. URL: https://www.jmlr.org/papers/volume3/blei03a/blei03a.pdf [Accessed 2025-06-06]
  40. Blei DM, Lafferty JD. Dynamic topic models. Presented at: Proceedings of the 23rd International Conference on Machine Learning; Jun 25-28, 2010:113-120; Pittsburgh, Pennsylvania. [CrossRef]
  41. Grootendorst MR. BERTopic: neural topic modeling with a class-based TF-IDF procedure. arXiv. Preprint posted online on 2022. [CrossRef]
  42. Egger R, Yu J. A topic modeling comparison between LDA, NMF, Top2Vec, and BERTopic to demystify Twitter posts. Front Sociol. 2022;7:886498. [CrossRef] [Medline]
  43. O’Callaghan D, Greene D, Carthy J, Cunningham P. An analysis of the coherence of descriptors in topic modeling. Expert Syst Appl. Aug 2015;42(13):5645-5657. [CrossRef]
  44. Tomas M, Chen K, Corrado G, Dean J. Efficient estimation of word representations in vector space. arXiv. Preprint posted online on 2013. [CrossRef]
  45. Sakaki T, Mizuki S, Gunji N. BERT pre-trained model trained on large-scale Japanese social media corpus. 2019. URL: https://github.com/hottolink/hottoSNS-bert [Accessed 2024-07-18]
  46. Asahi Newspaper Digital. Older driver incident in Ikebukuro: 87-year-old driver hits 10 people, killing two including a young girl. 2019. URL: https://www.asahi.com/articles/ASM4M46VCM4MUTIL014.html [Accessed 2024-08-13]
  47. Transport Ministry of Land, Infrastructure and Japan Tourism. Fukuoka prefecture road traffic environment safety promotion liaison committee 1st meeting: “Ansuiren seminar.”. 2016. URL: https://www.qsr.mlit.go.jp/fukkoku/site_files/file/iinkai_kaigi/dourokoutuu/161116shiryou.pdf [Accessed 2024-08-13]
  48. Asahi Newspaper Digital. 89-year-old suspect: “It’s time to stop driving” - accidents involving elderly drivers continue. 2021. URL: https://www.asahi.com/articles/ASPCK6KS1PCKPTIL041.html [Accessed 2024-08-13]
  49. Kwak H, Lee C, Park H, Moon S. What is Twitter, a social network or a news media? Presented at: Proceedings of the 19th International World Wide Web Conference; Apr 26-30, 2010:591-600; Raleigh North Carolina USA. [CrossRef]
  50. Eysenbach G. Infodemiology and infoveillance: framework for an emerging set of public health informatics methods to analyze search, communication and publication behavior on the internet. J Med Internet Res. Mar 27, 2009;11(1):e11. [CrossRef] [Medline]
  51. Salathé M, Bengtsson L, Bodnar TJ, et al. Digital epidemiology. PLOS Comput Biol. 2012;8(7):e1002616. [CrossRef] [Medline]
  52. Brownstein JS, Freifeld CC, Madoff LC. Digital disease detection--harnessing the web for public health surveillance. N Engl J Med. May 21, 2009;360(21):2153-2155. [CrossRef] [Medline]
  53. Sugawara Y, Murakami M, Narimatsu H. Use of social media by hospitals and clinics in Japan: descriptive study. JMIR Med Inform. Nov 27, 2020;8(11):e18666. [CrossRef] [Medline]
  54. Mitsutake S, Takahashi Y, Otsuki A, et al. Chronic diseases and sociodemographic characteristics associated with online health information seeking and using social networking sites: nationally representative cross-sectional survey in Japan. J Med Internet Res. Mar 2, 2023;25:e44741. [CrossRef] [Medline]
  55. Mitsutake S, Oka K, Okan O, et al. eHealth literacy and web-based health information-seeking behaviors on COVID-19 in Japan: internet-based mixed methods study. J Med Internet Res. Jul 11, 2024;26:e57842. [CrossRef] [Medline]
  56. World Health Organization. Global report on ageism. 2021. URL: https://www.who.int/publications/i/item/9789240016866 [Accessed 2024-07-18]
  57. Officer A, Thiyagarajan JA, Schneiders ML, Nash P, de la Fuente-Núñez V. Ageism, healthy life expectancy and population ageing: how are they related? Int J Environ Res Public Health. May 1, 2020;17(9):3159. [CrossRef] [Medline]
  58. Lamont RA, Swift HJ, Abrams D. A review and meta-analysis of age-based stereotype threat: negative stereotypes, not facts, do the damage. Psychol Aging. Mar 2015;30(1):180-193. [CrossRef] [Medline]
  59. Levy BR, Slade MD, Kunkel SR, Kasl SV. Longevity increased by positive self-perceptions of aging. J Pers Soc Psychol. 2002;83(2):261-270. [CrossRef]
  60. Wurm S, Benyamini Y. Optimism buffers the detrimental effect of negative self-perceptions of ageing on physical and mental health. Psychol Health. 2014;29(7):832-848. [CrossRef] [Medline]
  61. Wethington E, Pillemer K, Principi A. Research in social gerontology: social exclusion of aging adults. In: Riva P, Eck J, editors. Social Exclusion. Springer; 2016:177-195. [CrossRef] ISBN: 978-3-319-33033-4
  62. Officer A. A global campaign to combat ageism. Bull World Health Organ. Apr 1, 2018;96(4):295-296. [CrossRef] [Medline]
  63. Chang ES, Kannoth S, Levy S, Wang SY, Lee JE, Levy BR. Global reach of ageism on older persons’ health: a systematic review. PLOS ONE. 2020;15(1):e0220857. [CrossRef] [Medline]
  64. Bodner E, Yuval P, Wyman MF. Ageism in mental health assessment and treatment of older adults. In: Ayalon L, Tesch-Römer C, editors. Contemporary Perspectives on Ageism. Springer; 2018:241-262. [CrossRef] ISBN: 978-3-319-73820-8
  65. Lagacé M, Tanguay A, Lavallée ML, Laplante J, Robichaud S. The silent impact of ageist communication in long term care facilities: elders’ perspectives on quality of life and coping strategies. J Aging Stud. Aug 2012;26(3):335-342. [CrossRef]
  66. Band-Winterstein T. Health care provision for older persons: the interplay between ageism and elder neglect. J Appl Gerontol. Apr 2015;34(3):NP113-NP127. [CrossRef] [Medline]
  67. Bal AC, Reiss AEB, Rudolph CW, Baltes BB. Examining positive and negative perceptions of older workers: a meta-analysis. J Gerontol B Psychol Sci Soc Sci. Nov 2011;66(6):687-698. [CrossRef] [Medline]
  68. Kluge A, Krings F. Attitudes toward older workers and human resource practices. Swiss J Psychol. Mar 2008;67(1):61-64. [CrossRef]
  69. Bai X. Images of ageing in society: a literature review. Population Ageing. Sep 2014;7(3):231-253. [CrossRef]
  70. Loos E, Ivan L. Visual ageism in the media. In: Ayalon L, Tesch-Römer C, editors. Contemporary perspectives on ageism. Springer; 2018:163-176. [CrossRef] ISBN: 978-3-319-73820-8
  71. Sánchez-Román M, Autric-Tamayo G, Fernandez-Mayoralas G, et al. Social image of old age, gendered ageism and inclusive places: older people in the media. Int J Environ Res Public Health. Dec 18, 2022;19(24):17031. [CrossRef] [Medline]
  72. Bacsu JD, Fraser S, Chasteen AL, et al. Using Twitter to examine stigma against people with dementia during COVID-19: infodemiology study. JMIR Aging. Mar 31, 2022;5(1):e35677. [CrossRef] [Medline]
  73. Daneshvar H, Anderson S, Williams R, Mozaffar H. How can social media lead to co-production (co-delivery) of new services for the elderly population? A qualitative study. JMIR Hum Factors. Feb 12, 2018;5(1):e5. [CrossRef] [Medline]
  74. Jimenez-Sotomayor MR, Gomez-Moreno C, Soto-Perez-de-Celis E. Coronavirus, ageism, and Twitter: An evaluation of tweets about older adults and COVID-19. J Am Geriatr Soc. Aug 2020;68(8):1661-1665. [CrossRef] [Medline]
  75. Xi W, Xu W, Zhang X, Ayalon L. A thematic analysis of Weibo topics (Chinese Twitter hashtags) regarding older adults during the COVID-19 outbreak. J Gerontol B Psychol Sci Soc Sci. Aug 13, 2021;76(7):e306-e312. [CrossRef] [Medline]
  76. Litchman ML, Snider C, Edelman LS, Wawrzynski SE, Gee PM. Diabetes online community user perceptions of successful aging with diabetes: analysis of a #DSMA tweet chat. JMIR Aging. Jun 22, 2018;1(1):e10176. [CrossRef] [Medline]
  77. Fraser KC, Kiritchenko S, Nejadgholi I. Computational modeling of stereotype content in text. Front Artif Intell. 2022;5:826207. [CrossRef] [Medline]
  78. Fraser KC, Kiritchenko S, Nejadgholi I. Extracting age-related stereotypes from social media texts. Presented at: Proceedings of the 13th Language Resources and Evaluation Conference; Jun 21-23, 2022:3183-3194. URL: https://aclanthology.org/2022.lrec-1.341/ [Accessed 2025-06-06]
  79. Kong D, Chen A, Zhang J, et al. Public discourse and sentiment toward dementia on Chinese social media: machine learning analysis of Weibo posts. J Med Internet Res. Sep 2, 2022;24(9):e39805. [CrossRef] [Medline]
  80. Ng R, Indran N. Role-based framing of older adults linked to decreased ageism over 210 years: evidence from a 600-million-word historical corpus. Gerontologist. Apr 20, 2022;62(4):589-597. [CrossRef] [Medline]
  81. Ng R, Indran N. #ProtectOurElders: analysis of tweets about older Asian Americans and anti-Asian sentiments during the COVID-19 pandemic. J Med Internet Res. Mar 29, 2024;26:e45864. [CrossRef] [Medline]
  82. Stracqualursi L, Agati P. COVID-19 vaccines in Italian public opinion: Identifying key issues using Twitter and natural language processing. PLOS ONE. 2022;17(11):e0277394. [CrossRef] [Medline]
  83. Glogger I, Shehata A, Hopmann DN, Kruikemeier S. The world around us and the picture(s) in our heads: the effects of news media use on belief organization. Commun Monogr. 2023;90(2):159-180. [CrossRef] [Medline]
  84. Greenhawt M, Kimball S, DunnGalvin A, et al. Media influence on anxiety, health utility, and health beliefs early in the SARS-CoV-2 pandemic-a survey study. J Gen Intern Med. May 2021;36(5):1327-1337. [CrossRef] [Medline]
  85. Ministry of Internal Affairs and Communications. Digital information flow in Japan. 2023. URL: https://www.soumu.go.jp/main_content/000919817.pdf [Accessed 2024-11-25]
  86. Mellon J, Prosser C. Twitter and Facebook are not representative of the general population: political attitudes and demographics of British social media users. Research & Politics. Jul 2017;4(3). [CrossRef]
  87. Uppada SK, Manasa K, Vidhathri B, Harini R, Sivaselvan B. Novel approaches to fake news and fake account detection in OSNs: user social engagement and visual content centric model. Soc Netw Anal Min. 2022;12(1):52. [CrossRef] [Medline]
  88. Weeg C, Schwartz HA, Hill S, Merchant RM, Arango C, Ungar L. Using Twitter to measure public discussion of diseases: a case study. JMIR Public Health Surveill. Jun 26, 2015;1(1):e6. [CrossRef] [Medline]
  89. Matsuo A, Sasahara K, Taguchi Y, Karasawa M. Development and validation of the Japanese moral foundations dictionary. PLOS ONE. 2019;14(3):e0213343. [CrossRef] [Medline]
  90. Wu Q, Sano Y, Takayasu H, Havlin S, Takayasu M. Twitter communities are associated with changing user’s opinion towards COVID-19 vaccine in Japan. Sci Rep. Apr 5, 2025;15(1):11716. [CrossRef] [Medline]
  91. Sano Y, Hori A. 12-year observation of tweets about rubella in Japan: a retrospective infodemiology study. PLOS ONE. 2023;18(5):e0285101. [CrossRef] [Medline]


DTM: dynamic topic model
J-LIWC: Japanese version of the Linguistic Inquiry and Word Count Dictionary
NMF: nonnegative matrix factorization


Edited by Tim Mackey; submitted 26.11.24; peer-reviewed by Abizer Adenwala, Ryota Kobayashi, Sayan Mitra, Yuyan Jiang; final revised version received 12.04.25; accepted 22.04.25; published 16.06.25.

Copyright

© Akito Nakanishi, Masao Ichikawa, Yukie Sano. Originally published in JMIR Infodemiology (https://infodemiology.jmir.org), 16.6.2025.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Infodemiology, is properly cited. The complete bibliographic information, a link to the original publication on https://infodemiology.jmir.org/, as well as this copyright and license information must be included.