Mental health-related conversations on social media and crisis episodes: a time-series regression analysis

Anna Kolliakou, Ioannis Bakolis, David Chandran, Leon Derczynski, Nomi Werbeloff, David Osborn, Kalina Bontcheva, Robert StewartScientific 2020 · DOI

TL;DR

This study reveals a significant temporal association between the volume of mental health-related conversations on Twitter and daily mental health crisis episodes in London. The findings suggest that spikes in online discussions about depression and schizophrenia can precede and, in some cases, coincide with increased demand for mental healthcare services, offering a potential early warning system.

What Problem Does This Paper Address?

The paper addresses the growing concern about the impact of public discourse on mental health, particularly in the age of social media. While traditional media's influence on suicide and self-harm has been studied, the dynamic and pervasive nature of social media presents a new challenge. Researchers are keen to understand if the fluctuations in online conversations about mental health can serve as a proxy for population-level distress and predict surges in demand for mental healthcare services. This is crucial for developing proactive public health interventions and optimizing resource allocation in mental healthcare systems.

Methods

The study employed a time-series regression analysis to investigate the relationship between daily Twitter post volumes and daily mental health crisis episodes. They conducted both primary and replicated analyses using data from two London mental healthcare providers. Key methodological aspects include:

  • Time-Series Analysis: This statistical approach is used to analyze sequential data points collected over time, allowing researchers to identify patterns, trends, and correlations between variables that evolve chronologically.
  • Lagged Regression: To explore temporal relationships beyond immediate correlation, the researchers used lagged analyses. This means they examined how Twitter activity on one day might be associated with crisis episodes on subsequent days (up to 7 days later).
  • Confounder Adjustment: The models were adjusted for several potential confounding factors, including autocorrelation (the correlation of a time series with its own past values), calendar year, daily temperature, and bed occupancy levels, to isolate the specific effect of Twitter conversations.
  • Above-Median Analysis: To quantify the impact, they dichotomized tweet volumes into "above-median" and "below-median" days and assessed the percentage increase in crisis episodes on higher-volume days.

Datasets

The study utilized two primary datasets:

  • Twitter Data: A random 10% sample of all tweets published between January 2010 and December 2014, archived by the University of Sheffield. This dataset was filtered for keywords related to depression, schizophrenia, and mental health stigma.
  • Mental Healthcare Provider Data: Anonymized clinical data from the electronic health records (EHRs) of two London mental healthcare providers:
  • South London and Maudsley NHS Foundation Trust (SLAM): Data from January 2010 to December 2014, covering inpatient, home treatment team, and crisis house referrals.
  • Camden and Islington NHS Foundation Trust (C&I): Data from the same period, also including incident admissions to inpatient, home treatment team, and crisis house services.

Key Insights and Ideas

  • Social Media as a Public Mental Health Barometer: The study strongly suggests that Twitter conversations about mental health are not just noise but can reflect and potentially influence real-world mental health distress, leading to increased demand for services.
  • Temporal Dynamics are Crucial: The relationship isn't static. Initial positive associations (more tweets, more crises) on the same day or the next day evolve, becoming negative by day 4, and then potentially positive again by day 7. This complex pattern hints at both immediate precipitating effects and longer-term risk associations.
  • Replicability Across Services: The findings were consistently observed in two independent London mental healthcare services, increasing confidence in their generalizability within this urban context.
  • Schizophrenia Tweets Show Stronger Association: Days with above-median schizophrenia-related Twitter posts were associated with a notable 15% increase in crisis episodes, highlighting the potential sensitivity of this population to online discourse.

Key Results

  • Higher volumes of depression and schizophrenia-related tweets were associated with higher numbers of same-day crisis episodes in both SLAM and C&I.
  • After adjusting for confounders, a 7-day lagged analysis revealed a curvilinear pattern: positive associations on Day 1, negative associations by Day 4, and a return to positive associations by Day 7.
  • On days with above-median schizophrenia-related Twitter posts, SLAM crisis episodes were 15% higher (p<0.001).
  • Similar increases were observed for supportive depression (9% higher, p<0.001) and supportive schizophrenia (10% higher, p<0.001) tweet days.

Limitations and Future Work

  • Measurement Error: The authors acknowledge that both Twitter posts and crisis episodes are proxies. Twitter activity may not perfectly capture exposure to mental health content, and crisis episodes are just one indicator of population mental health. This suggests the observed effects might be underestimates.
  • Causality: While associations are strong, direct causality is difficult to establish. It's unclear if Twitter content directly influences individuals or if underlying events driving both tweets and crises are the true cause. The impact of social media exposure on those experiencing crises is also unknown.
  • Specificity: The study did not investigate whether specific types of tweets (e.g., stigmatizing vs. supportive) or specific diagnoses were linked to particular crisis episode types.
  • Geo-location: The use of general Twitter data, rather than geo-located tweets from specific areas, limits the ability to pinpoint local impacts.

Future Work: The authors suggest developing monitoring systems to identify high-risk days and communicate this information to mental health services for proactive support. Further research is needed to explore the causal pathways, investigate the impact of specific tweet content, and utilize geo-located data for more localized insights. They also emphasize the need for mixed qualitative and quantitative approaches to understand these complex relationships and develop effective interventions.

Summary updated

Abstract

We aimed to investigate whether daily fluctuations in mental health-relevant Twitter posts are associated with daily fluctuations in mental health crisis episodes. We conducted a primary and replicated time-series analysis of retrospectively collected data from Twitter and two London mental healthcare providers. Daily numbers of 'crisis episodes' were defined as incident inpatient, home treatment team and crisis house referrals between 2010 and 2014. Higher volumes of depression and schizophrenia tweets were associated with higher numbers of same-day crisis episodes for both sites. After adjusting for temporal trends, seven-day lagged analyses showed significant positive associations on day 1, changing to negative associations by day 4 and reverting to positive associations by day 7. There was a 15% increase in crisis episodes on days with above-median schizophrenia-related Twitter posts. A temporal association was thus found between Twitter-wide mental health-related social media content and crisis episodes in mental healthcare replicated across two services. Seven-day associations are consistent with both precipitating and longer-term risk associations. Sizes of effects were large enough to have potential local and national relevance and further research is needed to evaluate how services might better anticipate times of higher risk and identify the most vulnerable groups.
BibTeX
@inproceedings{kolliakou2020mental,
  author = {Anna Kolliakou and Ioannis Bakolis and David Chandran and Leon Derczynski and Nomi Werbeloff and David Osborn and Kalina Bontcheva and Robert Stewart},
  title = {{Mental health-related conversations on social media and crisis episodes: a time-series regression analysis}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-57835-9},
}
Figure