Data exploration

[sigmafelix]

Objective

• Understanding the present PEGS data: list of variables, descriptions, and variable types
  • Data-wise: Missingness, value ranges
  • Mechanism-wise: the connection between covariate groups
• Developing additional covariates that were not included in the existing data

[sigmafelix]

I will try to keep this page updated without including any sensitive information.

Temporal inconsistency

• Some records had the earlier date of completion of health & exposure survey (H&E) than that of the date of enrollment; the number is negligible (18 out of 18,462 unique participants). --> As of 11/20/2023, confirmed that this issue does not matter as there is no right sequence of surveys (i.e., between enrollment and health & exposure survey)
• For climate change/extreme weather events/natural hazard impact analysis, I will use the coordinates of the most recent dates.

Potential issue of temporal misalignment in covariates

• Since the date of coordinate records (gis_event_date) spans from 2002-10 to 2020-10, census-based covariates (i.e., EJI) need to be reexamined to check any temporal misalignment: will ping Farida about this

Analysis design from data availability

• Non-spatial analysis with N(participants) maximized
  • BKMR at the first stage
  • BKMR-CM (causal mediation)
  • MixSelect (Ferrari and Dunson 2020)
• Subpopulation analysis by regions/the dichotomization NC-nonNC states
• Sensitivity: leveraging exposome survey a and/or b

[sigmafelix]

11/20/2023

Mental health outcomes can be derived from internal exposome survey (EB). Although the GIS master table only records the date of completion for the exposome survey A, I will consider this date as the completion date of EB. Related to this, I am looking at fulfilling the basic assumptions in causality/association: no temporal antecedence of results to attributable factors (avoiding reverse causality). Should be there no such issues in the cleaned data, I could state associative/causal interpretations from any models I will employ. Regarding the temporal orders, I am listing some issues found in the pre-calculated data below:

EJI and SVI

• EJI data are mostly based on 2015-2019 ACS data and CDC PLACES datasets, both at census tract level.
• SVI data are calculated in 2010, 2014, 2016, and 2018, which are at relatively finer temporal resolution; data elements with the latest but earlier year than the year of the completion date will be used.

Hazards

• PM2.5 constituents were aggregated for 19-year (2000-2018).
• TRI-SEDC values do not include known report years; should figure it out
• CACES fused emission estimates also do not include years.

[sigmafelix]

There are 11 participants who have no survey completion date. These participants are excluded in further analysis. Now I am calculating

• [x] the total days of experiencing smoke regardless of density in 30-, 60-, 120-day, and 1-, 2-, 3-, 5- year windows.