Look at NLInteraction and trial run

[sigmafelix]

NLInteraction--pub and NLInteraction--package is another mixture analysis method listed in PRIME publications. It relies on natural splines to fit nonlinear response functions of an exposure at certain quantiles of another exposures. The function runs in quite reasonable amount of time as it is based on less intensive matrix operations than BKMR.

---

Analysis side tasks

• [x] Test run without climate-related exposures
• [ ] With climate/extreme weather event exposures (i.e., #days of all/medium/severe smoke exposure; flood risk and actual flood damage)
• [ ] Examine the response functions with SVI/EJI

Method side tasks

• [ ] If it could be done in an acceptable amount of time, try replacing natural splines to monotone regression splines to impose strong assumption on the monotonicity in the increasing dose-response relationship
• [ ] Convert it to the faster implementation (e.g., Turing or Nimble)

[sigmafelix]

External exposome survey B questions comprehensively record mental health conditions in the past and the present. A methodological challenge is that the survey responses are categorical/ordinal variables. For example, the question on being nervous has six responses: missing - never - almost never - sometimes - fairly often - very often. After excluding missings (comprising 2-3% of responses), we have five ordinal categories. The motivating examples in the original NLInteraction paper were using continuous outcomes, which give limited implications for ordinal outcomes. I will find literature on treating ordinal variables in mixture models.

[sigmafelix]

The total number of days of all smoke intensity and rates of unemployment, bad mental health, minority, and elderly population at census tract level were considered to examine their two-way interactions in NLinteraction. Covariates such as individual traits and earthdata elements were also controlled. For the test run, I chose "in a past month, felt unable to cope with" as an outcome, which records the responses in five-point Likert scale. Even though this is a "no-brainer" fit of model just to find the model works at the real data, the result suggests that the other results from other model designs are potentially promising. The effect of smoke exposure was independent to minority rates, whereas the flipped patterns were found from unemployment and not good mental health status. One problem is that the 30-day smoke exposure is mostly zeros but the self-assessed mental health outcomes are all covering just in one month period from the date of survey completion.