Project Overview

Overview

In January 2016, Dr. Tim Slack at Louisiana State University was awarded an RFP-V grant of $1,929,832 to lead the GoMRI project entitled Understanding Resilience Attributes for Children, Youth, and Communities in the Wake of the Deepwater Horizon Oil Spill (RCYC). The project consisted of 1 collaborative institution (National Center for Disaster Preparedness, Columbia University, PI Beedasy) and approximately 8 research team members (including students).   The overarching aim of the RCYC project was to assess the public health impacts of the 2010 Deepwater Horizon oil spill (DHOS) in the Gulf of Mexico (Theme 5 in GoMRI RFP-V), with special emphasis on the impacts of the disaster on children and their families over time. The project was a three-pronged multi-method effort combining quantitative data from a household-level survey, qualitative data from focus groups, and the analysis of social media data (i.e., Twitter) following the disaster.   The broad set of research questions addressed by the RCYC study include: Q1: What are the impacts of disaster-related disruption on children and families exposed to the 2010 DHOS, both in terms of physical and mental health effects as well as social consequences, such as increased risk behaviors and decreased economic opportunities? Q2: What is the relationship of primary and secondary stressors on these outcomes? Q3: What attributes of children and families are related to greater resilience to negative disaster-related impacts? Conversely, what attributes of children and families are related to greater vulnerability to negative impacts? Q3a: How does resilience/vulnerability vary across key sociodemographic groups, economic/occupational types (e.g., fishers and oil/gas workers), and families with different levels of social capital (e.g., social network structures and trust) or attachment to the social or natural environment? Q3b: What role do online social networks play in facilitating resilience? Q4: What sorts of issues are children confronting as a result of the oil spill and what sorts of measures do children, families, and community stakeholders see as being needed in response? Q5: How do all of the above change over time?   The ultimate goals of this research project are to: 1) Assess the public health and social impacts of the DH oil spill with a special focus on children and their families; 2) Identify attributes of children and families associated with resilience to negative disaster impacts and, conversely, attributes of children and families associated with vulnerability to negative disaster impacts; 3) Build a variety of datasets that allow for the scientific analysis of these issues; 4) Train graduate students in disaster resilience research to help build the next generation of scholars dedicated to these issues; and 5) Make the information generated from this project actionable with the aim of helping facilitate disaster resilience and mitigate vulnerability. 

 

Research Highlights

As of December 31, 2019, this project’s research resulted in 21 scientific presentations and 4 datasetsbeing submitted to the GoMRI Information and Data Cooperative (GRIIDC), which are/will be made available to the public. The project also engaged 7 Master’s level and 4 PhD level students over its award period. Of special note, GoMRI Scholar Kathryn Keating was selected as a 2018 National Academy of Sciences Gulf Research Program Policy Science Fellow. Her one-year fellowship was with the RESTORE Council in New Orleans, LA. Significant outcomes of this project’s research according to GoMRI Research Theme Five are highlighted below.  

Theme Five Accomplishments

 

RCYC Longitudinal Household Survey

The RCYC household-level surveys leveraged prior research efforts by personnel at the National Center for Disaster Preparedness (NCDP) (Abramson et al. 2010; Abramson et al. 2013). A 2010 random-digit dial survey of coastal Louisiana and Mississippi revealed significant worry about the health impacts of the DHOS and motivated more in-depth study. Subsequently, in 2012, NCDP researchers used a multi-stage sampling design to select communities, census blocks, and households with children to build a dataset concerning the impacts of the DHOS in Louisiana, Mississippi, Alabama, and Florida. An impact index was calculated to identify spill-affected communities using three sources of data: (1) individual claims data from the Gulf Coast Claims Facility (zip code), (2) business claims data from the Gulf Coast Claims Facility (zip code), and (3) aggregated coastline oiling data from National Oceanic and Atmospheric Administration’s Shoreline Cleanup and Assessment Technique (latitude/longitude). Z-scores were calculated for each of the three variables by zip code and then summed to create a standardized index where higher values indicated more impacted areas. Within highly impacted areas, a two- stage cluster sampling design was utilized to randomly select census blocks, and within these blocks to randomly select households with children. Households identified through this process were then surveyed about oil spill exposure, health status, and related topics.1

 

In 2014, researchers from the NCDP returned to the DHOS affected areas in South Louisiana and conducted a face-to-face household survey (N=717). Since the initial surveys conducted in 2012 were anonymous, the research team revisited the previously interviewed addresses and collected identifiable information to populate a cohort database going forward. From each household, one adult age 18 years or older who was the parent or caregiver of a child in the household provided information for themselves, the focal child, and characteristics of their household. In cases where there was more than one child in the household, the child with most recent birthday was identified. Adult respondents were selected based on their self-reported ability to be the caregiver best able to answer questions about the focal child’s health. Of those surveyed, 91% (N=655) agreed to be followed-up for subsequent waves of data collection. The RCYC study then conducted follow-up interviews with the same adults, and centered on the same focal  children, in 2016 and 2018.2 Out of the 655 respondents who had agreed to be followed up from 2014 survey, approximately 74% were re-interviewed in 2016 (N=482) and (N=481) in 2018. Reasons for attrition included inability to relocate respondents, refusals, mortality, and incarceration. The survey instrument covered topics such as direct and indirect oil spill exposure, physical and mental health status, perceptions of recovery, demographic data, and a range of characteristics theoretically linked to social vulnerability and resilience.

 

The three waves of longitudinal cohort data available as a result of this effort are unique in human-subjects disaster research. Most research in this area is cross-sectional, allowing a snapshot of characteristics among a sample of the population of interest at a specific point in time. The RCYC survey data allow for understanding within unit (family household) change over time which provide the opportunity to evaluate trajectories of change (e.g., are people doing better over time, about the same, or becoming worse off?). Additionally, this is one of the few studies that have focused on the effects of oils spills on the long-term health of children. For example, in one paper that is currently under review, we consider the effect of DHOS exposure on health trajectories of children, an especially vulnerable population subgroup. Results from latent linear growth curve models show that DHOS exposure via physical contact and job/income loss both negatively influenced initial levels of child health. However, the effects of physical exposure dissipated over time while the effects of job/income loss persisted. This pattern holds for both general child health and the number of recent physical health problems children have experienced. These findings help to bridge the literature on disaster impacts and resilience/vulnerability, with the literature on socioeconomic status (SES) as a fundamental cause of health outcomes over the life course.

 

RCYC Focus Groups

Another innovative aspect of the RCYC project was to draw a subsample of qualitative focus group respondents from those participating in the quantitative household-level survey. Six focus groups were conducted in the following Louisiana parishes: Vermilion, Terrebonne, Jefferson/Lafourche, Plaquemines, Orleans, and St. Tammany. In November of 2017, focus group participants were drawn from a larger sample of 484 survey participants. Focus group participants were purposively selected with the intent of capturing people in each community with a range of DHOS experiences. Recruitment for the focus groups was planned with a target of approximately ten participants per community, with a total of 46 individuals ultimately participating across the six focus groups. Each focus group was facilitated by a team of three researchers: a lead facilitator, note-taker, and timekeeper. The sessions were held on weekday evenings in centrally located public venues in each community (e.g., a public library). All focus groups were audio recorded, and sessions lasted 90 minutes on average. The lead facilitator guided the conversation using a series of pre-formulated questions, probing for details to responses and redirecting side conversations as necessary.

 

Focus group audio recordings were transcribed and the qualitative data from the narrative transcripts were then coded by three researchers. An initial focus group codebook was developed using the RCYC survey instrument as a guide. The survey included modules such as “child health,” “adult health,” and “economic impacts,” among others. These modules were based on theoretical linkages to oil spill exposure suggested by the extant literature. As such, the initial codebook included codes such as: “responsible party”, “personal health”, “child health”, “financial loss”, and “career change.” With this codebook as a starting point, grounded theory methods were utilized to analyze the narrative text. To prepare the text for analysis, transcripts were unitized by a lead coder who identified meaningful conceptual breaks within the text. Unitization ensures that all coders are working with the same units of analysis within the narrative transcripts and is especially useful for conversational transcripts without clear breaks or instances where multiple people are speaking. This method also facilitates assessment of interrater reliability and the identification of passages subject to disagreements in coder interpretation.

 

In stage one of the data preparation, three researchers worked independently to code the passages from the narrative transcripts using variables in the codebook while simultaneously employing blind, open coding techniques to create new codes for emergent themes. In stage two, an iterative process of discussion around revisions to the initial codebook was used as new information emerged. New codes drawn from emergent themes included: “peer information”, “regulation”, “food insecurity”, “water issues”, “worker safety”, “environmental damage”, “emotional response”, “community dynamics”, “family dynamics”, “leaving community”, and “protective factor.” All transcripts were then coded using the revised codebook. Across the three researchers working to code the focus group data, inter-rater reliability for codes in this analysis was 99.2%, calculated using a simple percent-agreement method. In rare cases of discrepancies in coder interpretation, the group discussed the passage in an effort to reach inter-coder agreement, and when necessary, undertook a revision to the codebook. All differences in initial coder agreement were resolved via this iterative process.

 

By capturing descriptions of family experiences with the DHOS in people’s own words, the focus group data provide greater context and depth among this subsample than is allowed from the (more generalizable) household survey data alone. For example, in one paper that we currently have under review, the focus group data is used to yield insights into the strategies individuals and families used to cope with financial loss in the years following the DHOS. Key findings include: 1) long-term economic impacts persist, but are nuanced and differ across place; 2) for families living in multi-stressed environments, concerns about the DHOS spill over into other aspects of social functioning and become enmeshed with perceptions of other environmental stressors, and 3) economic exposure following the DHOS has impacted families differently based on social position and community social structure.

 

RCYC Social Media Analysis

A final accomplishment of the RCYC project was building a dataset on Twitter use following the DHOS. There is tremendous demand for real-time information in a disaster context. Online networks and emergent virtual communities powered by information and communication technologies contribute to addressing these information needs. In the aftermath of the DHOS, interactions on the Twitter platform shaped the online conversation through the use of hashtags such as #OilSpill, and the sharing of information about the clean-up and response efforts, including the use of dispersants and environmental and health impacts.

 

A set of historical tweets was identified based on a search strategy built on a set of keywords. The tweets were filtered by date, to include data posted for four months from the day the rig exploded on April 20, 2010 until August 20, 2010 (i.e., after the well was brought to a static state). The Twitter data was obtained through the service provider Texifter from Twitter’ GNIP power track. The initial search was conducted through the now defunct Sifter, Texifter’s search application. A total of 876,298 tweets was acquired from Texifter.

 

The dataset of tweets was hosted on Texifter’s collaborative text analytics platform DiscoverText. After a preliminary examination of more than two hundred tweets, it was found that not all postings were relevant to our study and further cleaning was required. For example, there were mentions of other spills, such as the 2010 Great Barrier Reef oil spill in Australia, while other tweets contained unrelated content that did not offer anything specific or useful about the DHOS. The dataset was cleaned by developing and applying data classifiers through machine learning. Small sets of randomly sampled tweets were manually coded, to build and train a machine classifier. The classifier was used to estimate the relevance of each uncoded data unit or tweet in relation to the oil spill. The machine learning API at the core of the classification process in DiscoverText, Uclassifier is a multinomial naïve Bayesian classifier with hybrid complementary naïve Bayes, class normalization and special smoothing and the result is the probability of an item belonging to a certain class.

 

Data manipulation started with the de-duplication of the tweets. The elimination of all duplicates or retweets of the same content limits the selection of identical tweets during the random sampling for labeling and classification processes. A machine classifier was trained to identify patterns of relevance to the oil spill in the entire corpus of tweets through an iterative group coding exercise for machine-learning. This process was comprised of several rounds of group coding, consensus-based adjudication, validation, and retraining of the classifier. Randomly selected sets of 100-tweets were coded by 3 to 5 coders, using the binary code-set ‘relevant’ and ’not relevant.’ The relevance of each tweet to the spill was determined by direct or indirect references to the DH spill in the tweet. Users often referenced the oil spill explicitly using popular hashtags or words directly related to the spill. Nonetheless, this was not always the case as tweets also referred to the oil spill implicitly or without using the common hashtags (e.g., when users expressed their personal opinions or their political views about the general consequences, and the community and institutional responses).