key: cord-0929660-lizugmgg
authors: Sugg, Margaret M.; Michael, Kurt; Hajnos, Sarah N.; Greene, Shannon; Runkle, Jennifer D.
title: Understanding the concurrent risk of mental health and dangerous wildfire events in the COVID-19 pandemic
date: 2021-09-21
journal: Sci Total Environ
DOI: 10.1016/j.scitotenv.2021.150391
sha: 7a01a592e6211062c1c3f4edb61081d270567af9
doc_id: 929660
cord_uid: lizugmgg

Little research has examined the mental health risks of concurrent disasters. For example, disasters like wildfires have been shown to have a strong association with psychological symptoms—the 2020 U.S. Western wildfire season was the worst on record and occurred while the country was still navigating the COVID-19 pandemic. We implemented two quasi-experimental analyses, an interrupted time series analysis, and a difference-in-difference analysis to evaluate the impacts of wildfires and COVID-19 on mental health crisis help-seeking patterns. Both methods showed no statistical association between exposure to wildfires and the seeking of mental health support during the COVID-19 pandemic. Results highlighted that 2020 wildfires were not associated with an acute increase in crisis texts for youth in the two months after the events, likely due to an already elevated text volume in response to the COVID-19 pandemic from March 2020 throughout the fall wildfire season (Aug to Oct 2020). Future research is needed outside of the context of the pandemic to understand the effects of extreme and concurrent climatic events on adolescent mental health, and targeted interventions are required to ensure youth and adolescents are receiving adequate support during these types of crisis events.

insight into the temporal and geographic variation of help-seeking behaviors and crisis mental health service support needs (Sugg et al. 2019a , Sugg et al. 2019b , Thompson et al. 2018) . CTL users are matched with a trained crisis counselor who engages with the user via text message. At the end of each conversation, CTL users are invited to an optional survey that collects information on age, demographics, and other relevant identifiers (e.g., LGBTQ+ (Pisani et al. 2019 ). In addition, the service employs a complex algorithm to label text conversations with a crisis response tag (e.g., suicidal thoughts, self-harm, depression) and conversations are triaged based on whether or not a texter expressing suicidal thoughts or self-harm (characterized as imminent risk).

We included the following crisis response binary variables as mental health proxies in our analysis: suicidal thoughts, self-harm, anxiety and stress, depression, relationship issues, abuse (emotional, physical, and sexual), substance abuse, bereavement, and isolation. We also included an outcome measure for all crisis text events as an overall measure of crisis volume for a given geographic area. CTL-data is limited to the area-code spatial scale, however, nearly 3 out of 4 participants are adolescents and likely reside in the same location as their phone number.

CTL data was restricted to participants who completed post-conversation surveys that included age, gender, and racial information.

Wildfire locations were identified using the geospatial multiagency coordination (GeoMac)'s National Interagency Fire centers as geographic information system (GIS) shapefiles of fire perimeter extent. Fire impact areas were restricted to the three largest fires that impacted large population centers the 1.) LNU Lightning complex fire in Napa 

Descriptive statistics were performed to examine the mean daily CTL volume and corresponding standard deviation. Paired tests were used to determine whether mean daily CTL volume for each outcome differed by intervention period (alpha=0.05).

Interrupted time series (ITS) allows for the comparison across time (pre/post design) within a single population (i.e., each individual serves as their own control) and is considered one of the stronger non-randomized experimental designs (Turner et al., 2020) . By utilizing a single population, ITS avoids selection bias and unmeasured confounders (i.e., underlying trends that change slowly over time, secular changes), which are significant limitations among studies that utilize between-group differences (Bernal et al., 2018 , Bernal et al., 2017 . In our study, the preintervention period was defined as January 01, 2020, to August 16, 2020 (227 days) and the postintervention period was defined as August 17 to October 15, 2020 (60-days) following each wildfire event: a) LNU Lightning Complex; b) SCU Lightning Complex; and c) Bobcat. The CTL-crisis response outcomes of each wildfire event were examined for the 1) impact group, the 2) control group, and 3) at the state level for California and Oregon for the 60-day postintervention period.

An interrupted time-series design captured the immediate impact of crisis text patterns using autoregressive integrated moving average models (ARIMA) to analyze repeated measures of daily CTL volume and address time-series autocorrelation. The ARIMA (p,d,q) parameters were specified using autocorrelation and partial autocorrelation across lag periods. The final ARIMA model (0,1,1) was confirmed using plot residuals, lag plots, histograms of the residuals, the Ljung-Box statistic (p-value>0.05), and PACF/ACF values of lags and residuals. All diagnostics confirmed no correlation and a relatively normal distribution among residuals. The final ARIMA models were constructed for all CTL conversations in wildfire impact areas and provided the forecasted expected text volume compared to the observed text volume (alpha=0.10). Expected conversion rates for the impact period were forecasted using Hyndmans J o u r n a l P r e -p r o o f and Khankara's ARIMA algorithm R (Hyndman & Khandakar 2008). A sensitivity analysis was computed for crisis response outcomes at the state level for California and Oregon for 60-days (August 17 to October 15) to ensure other spillover wildfire effects (e.g., smoke exposure) did not impact the larger geographical region.

One disadvantage of the use of interventional ARIMA in ITS is the inability to exclude some time-varying confounding events (e.g., wildfire and COVID-19) occurring at the same time (Bernal et al., 2018 , Li et al, 2021 . Thus, in an effort to minimize the potential confounding of co-occurring events, we also included control (i.e., counterfactual) locations where wildfire did not occur. ARIMA (0,1,1,) were also computed on locations not impacted by wildfires. This comparative ITS design is typically not done in other ITS analysis, as few as one-fourth of ITS studies include some form of control (Turner et al. 2020) . Comparative ARIMA models were constructed for the three control regions (Table 1, Figure 1 ), and forecasted results were compared to observed values (alpha=0.10).

Difference-in-difference (DID) is a separate quasi-experimental research study design for analyses of observational data using two-time points by differencing the change in the exposed group (locations impacted by wildfire) minus the change in the unexposed group (similar locations not affected by wildfire). One difference examines the impact of living in an area code with an impacted wildfire (exposed) compared to an area code with no wildfire impacts (control). A second difference examined the geographical comparator during the timeframe of the wildfire event (control) compared to the time frame outside the wildfire event (exposed). For this analysis, the CTL data was restricted to area codes that were exposed and unexposed (control). Unlike our ARIMA analysis, the use of DID allows for the inclusion of the control group(s) (i.e., counterfactual) and the control of additional variables (e.g., age, repeated measures) (Li et al., 2021) . In our study, generalized estimating equations were constructed for the following CTL outcomes 1.) Stress/Anxiety 2.) Isolation 3.) Depression 4.) Suicidal

Thoughts 5.) Self-Harm 6.) Substance Abuse and 7.) Abuse for each fire. GEE models accounted for repeated measures, using the actor I.D, using an autoregressive correlation J o u r n a l P r e -p r o o f structure (i.e., the unique identifier for each texter). After fitting the regression equation for each CTL outcome, we adjusted for the age of the CTL-texter in all final models, which showed no significant changes.

A final sensitivity analysis was computed for the Kincade fire (October 23, 2019) for area codes (707, 530) impacted by the wildfire to assess if wildfire had an impact on mental health prior to the pandemic period using the ARIMA (0,1,1) model. As control locations were not identified for these locations, no other analysis was conducted for this time period.

The volume of texts received by CTL increased during the pandemic period (starting in March 2020) and remained elevated throughout the study period (October 2020, Figure 2 ). The number of CTL texts for any reason during the exposure period were 14.76, 9.65, and 8.33 daily texts for the SNU lightning complex, the Bobcat, and the LNU lightning complex fire, respectively. Comparatively, the number of CTL texts for any reason during the non-exposure period (outside of the fire period) in the exact locations were 12.90, 8.01, 6.97 daily texts for the SNU lightning complex, the Bobcat, and the LNU lightning complex fire, respectively. Selfharm, stress/anxiety, isolation, relationship, and abuse were significantly higher in the wildfire time period (p-value < 0.10) than the non-wildfire time period for area codes impacted by the three fires. In contrast, significant declines in suicidal thoughts and bullying texts occurred for impacted area codes during the post-wildfire exposure period compared to the pre-wildfire period ( Table 2) .

The ARIMA (0,1,1) was the best fit model to examine the pre-/post-change in daily CTL texts for wildfire impacted areas. The daily crisis texts were forecasted for the impacted area codes for August 16 to October 15, 2020, for all CTL texts (Figure 3 ). Forecasted results were compared to observed values and were not significant across the pre-and post-intervention time period, except for select dates (Figure 2 ). Using the ARIMA (0,1,1), forecast results were replicated for the state level (California and Oregon) and also found not to be significant (alpha = J o u r n a l P r e -p r o o f 0.10) (Supplementary Figure 2) . The 2020 wildfire control regions (Table 1 ) also showed preversus post trends that were not statistically significant for all CTL texts, similar to locations impacted by wildfires (Figure 4) . Exceptions included the dates of the control location for the SCU lightning fire that were significantly higher from 09/05-09/07 and 09/21.

DID Results (During Fires) Figure 5 shows the odds ratio and 95% confidence interval for the difference and difference estimator (location exposure*time exposure), the time of exposure, and location exposure. There were no significant differences in outcomes after adjusting for age, gender, or race in the final models.

A final sensitivity analysis using ARIMA was computed for the Kincade fire (October 23, 2019) for area codes (707, 530) impacted by the wildfires to assess if this event was negatively associated with mental health prior to the pandemic. Results were not statistically significant, with the exception of two days (10/07/2019, 09/27/2019) across the 60-day time window (Supplemental Figure 3) .

The objective of this study was to examine the mental health impacts of concurrent disasters, the 2020 wildfires, and the COVID-19 pandemic on adolescent and young adult mental health in the Western U.S. Overall, we found no association between wildfire exposure and excess crisis text for mental health support among youth and adolescents. Our results are surprising as the 2020 wildfire season was one of the worst on record (Higuera and Abatzoglou 2021) and occurred amidst a global pandemic, which has upended the lives of many youth and adolescents. These results suggest that the mental health impacts from the unprecedented COVID-19 pandemic overshadowed the psychophysical implications of the 2020 wildfires, as the mental health impacts of COVID-19 have been substantial with significant increases in some crisis response outcomes (e.g., stress/anxiety, substance abuse, isolation, abuse, bereavement) among adolescents ). showing a dose-response relationship (Maguen et al. 2012 , Norris et al. 2002 , Goldman et al. 2014 ). These differences in research designs, exposure classifications, minimal consideration of resilience models, as well as the underlying COVID-19 pandemic may, in part, explain the differences in our results.

We did note a decrease in outcomes like depression and suicidal thoughts for exposed locations during the wildfire event. These decreases could be due to coming together during the community cohesion phase of disaster response (Townshend et al., 2015) . The Federal Emergency Management Agency (FEMA) and Substance Abuse and Mental Health Services Administration (SAMHSA)'s widely used model highlights six phases of a single disaster, with high emotions in the impact (phase two), heroic (phase three), and honeymoon phases (phase four), which is characterized by the availability of disaster assistance, optimism, and altruism.

Wildfire responses in our analysis may have been limited to these phases of emotion rather than the later phases that are characterized by disillusionment and abandonment (phase five) (DeWolf 2000). Our cohort of CTL users may follow similar patterns as the SAMSHA model. However, the SAMSHA model was constructed for a single disaster rather than multi-disaster events, and new frameworks are needed to understand the mental health patterns of concurrent and cascading disasters.

Most likely, the COVID-19 pandemic was the main driver of elevated crisis events during 2020, which resulted in the increased prevalence of texts for isolation, stress/anxiety, and J o u r n a l P r e -p r o o f Journal Pre-proof depression after March 2020 . The 2020 wildfires were not associated with an acute increase in crisis texts for youth in the two months after the events, likely due to an already elevated text volume in response to the COVID-19 pandemic. This result contrasts with other studies which have found elevated crisis text use six weeks after other natural hazards like the impact of Hurricane Florence on crisis text response in youth (2018), which occurred prior to the COVID-19 pandemic . Adolescents and young adults were likely already experiencing psychosocial stressors before this wildfire event. Therefore, the concurrent disaster did not result in elevated rates beyond what they were currently reporting. For instance, typical CTL users may have already been engaged in CTL texts prior to the wildfire event and continued their use of the service during these wildfire events. Alternatively, adolescents and youth may have demonstrated resilience during the wildfire 2020 event using community cohesion and coping strategies from the COVID-19 pandemic and applying them to the 2020 wildfire event (Masten and Motti-Stefanidi 2020) . Despite our null findings, CTL and other adolescent mental health services should focus on outreach during these critical disaster periods, as other studies have noted increases in adverse mental health conditions 6-, 12-, and 18-months post-wildfire events.

Our work had several notable strengths. First, quasi-experimental research designs have been underutilized in the wildfire and mental health literature. One advantage to these designs is the ability to address the potential for uncontrolled confounding by adding a control group (DID analysis) or limiting between-group differences (ITS analysis). Designs that do not account for uncontrolled confounding and the use of more traditional methods (i.e., spatial regressions) have indicated significant associations between natural hazards events and health consequences, despite quasi-experimental methods applied to the same scenario showing little to no association (Grabich et al., 2015) . Secondly, our study utilized a large, longitudinal data set of crisis texts which contrast with common cross-sectional surveys found in disaster-mental health literature that often fail to examine temporal trends and the delayed onset of psychosocial symptoms.

Additionally, unlike other ITS studies that focus on the region impacted (e.g., Pridemore et al. 2008 ) and fail to show comparative or counterfactual results, we further replicated our ITS J o u r n a l P r e -p r o o f analysis for wildfire exposed area codes (impact group) and non-wildfire exposed area codes (control group). We included multiple directly impacted locations (e.g., state-level and area-code level) in our ITS analysis to account for difficult-to-measure factors, such as smoke (e.g., common across the state), and the inclusion of representative control groups to ensure no trends were observed in these locations. Lastly, our work utilized two robust causal analytic techniques common to quasi-experimental research designs, ITS and DID, to ensure our null results are robust.

A major limitation to our study was the inability to directly distinguish between COVID-19 impacts and 2020 wildfire impacts during our exposure time period. However, we employed a quasi-experimental research design that included a pre-intervention period and a postintervention time period that also encompassed the start of the pandemic. At this time, methodologies are limited for assessing the risk of concurrent events like climate disasters and global pandemics. Our results would also benefit from the adjustment of confounding variables due to unmeasured population attributes, like underlying COVID-19 rates or mental health services use. However, due to the unique spatial scale of our data (area code), these data were not available. Moreover, we do not expect the distribution of these variables to change markedly within the two geographical areas. Hence our quasi-experimental approach, in which each texter serves as their own control, has likely accounted for any unmeasured confounding due to these temporal factors (Wing et al., 2018 , Herrington et al. 2021 ). Yet these approaches, like DID analyses, can be underpowered to detect small differences, and could also have contributed to our null findings (Wing et al., 2018) Additionally, the aggregate level exposure to wildfire events estimated in our study at the area code level did not necessarily equate to ‗actual' personal wildfire exposure, which is a strong predictor of mental health conditions (Goldmann and Galea, 2014) . Lastly, in our analysis, the post-intervention period was defined over a short window of time as the wildfire dates corresponded to a 60-day post-intervention period. However, previous studies have found a high prevalence of mental health and addiction as long as 24-months after the wildfire (Reifels et al. 2015, Moosavi et al.) . We choose this smaller time window due to numerous large events during September -December 2020, including the U.S. presidential election, social justice issues (i.e., civil unrest from the murders of George Floyd and Breonna Taylor), and continued spikes in COVID-19 outbreaks. Future work should extend the analysis to a longer time interval to J o u r n a l P r e -p r o o f examine the delayed mental health effects in this vulnerable age group once the community cohesion or ‗honeymoon' phase has subsided and disillusionment has set in. We also recommend that future research utilizes similar research frameworks on other mental health data sets to further explore the association between wildfires and mental health, especially in the context of compounding disasters.

This study was the first to examine the concurrent effects of mental health during the 2020 wildfire season and the ongoing COVID-19 pandemic. Our study leveraged a real-time anonymized data set of national crisis events in the U.S. from Crisis Text Line that allowed us to examine pre-and post-event crisis response to both the 2020 wildfires and COVID-19 pandemic.

We implemented a quasi-experimental framework to investigate the trends in crisis events using both a difference in difference and interrupted time series analysis. Our results showed no statistically significant increases in CTL-crisis events during the 2020 wildfire season, suggesting that the 2020 COVID-19 pandemic was the main driver of crisis in adolescents and young adults during the study period. Future research and advanced methodologies are needed to disentangle the complex effects of concurrent and cascading disasters. *Control locations were selected based on the absence of wildfire during the exposure period and similar racial, female to male ratio, age proportions, and when possible, employment rates.

J o u r n a l P r e -p r o o f 

Psychological impact of disasters on children: review of assessment and interventions

Depression and prolonged grief in the wake of disasters

Multisystem resilience for children and youth in disaster: Reflections in the context of COVID-19

Psychiatric morbidity following a natural disaster: An Australian bushfire. Social Psychiatry and Psychiatric Epidemiology

Population health effects of air quality changes due to forest fires in British Columbia in 2003 : Estimates from physician-visit billing data

Mental health effects in primary care patients 18 months after a major wildfire in Fort McMurray: risk increased by social demographic issues, clinical antecedents, and degree of fire exposure

60,000 disaster victims speak: Part I. An empirical review of the empirical literature

Protecting user privacy and rights in academic data-sharing partnerships: principles from a pilot program at crisis text line

Quasi-experimental evaluation of text-based crisis patterns in youth following Hurricane Florence in the Carolinas

Crisis support-seeking behavior and temperature in the United States: Is there an association in young adults and adolescents

Crisis text patterns in youth following the release of 13 Reasons Why Season 2 and celebrity suicides: A case study of

Adolescents in crisis: A geographic exploration of help-seeking behavior using data from Crisis Text Line. Social science & medicine

Social cohesion and resilience across communities that have experienced a disaster

Design characteristics and statistical methods used in interrupted time series studies evaluating public health interventions: a review

Designing difference in difference studies: best practices for public health policy research. Annual review of public health

Bushfire impact on youth

Active Rescue and Imminent Risk refer to more severe forms of suicidal thoughts

Risk: A CTL texter with suicidal thoughts and ideation with a plan to end life within 2-days

Active Rescue: CTL is unable to de-escalate the situation resulting in the initiation of active rescue

We thank our external funders for supporting this work. This project was supported by (grant SRG-0-160-19) from the American Foundation for Suicide Prevention, the CAREER award (grant #2044839) from the National Science Foundation, and the Appalachian State University's RIEEE concert award program. The content is solely the authors' responsibility and does not necessarily represent the official views of the American Foundation for Suicide Prevention, Crisis Text Line, or the National Science Foundation.