Enhancing Societal Resilience Research through AI-Powered Survey Infrastructure (AI-RESIL)

How to

improve the resilience of individuals and societies in times of crises is a key

question for public health. Research into resilience as the maintenance or

quick recovery of mental health after major life events has so far primarily

been based on panel surveys. Yet, the disadvantages of established measurements

might bias results in several ways: (i) some life events are underreported due

to recall bias, and the time gap between reported life events and survey

participation is often unknown and varies individually, and (ii) life events

themselves can lead to incomplete or missing responses.

This

project will address both problems through an Artificial Intelligence

(AI)-powered survey infrastructure that combines survey and web tracking data

collected in the new GESIS Panel.dbd with open-source Large Language Models

(LLMs). First, we will detect the onset of major life events (e.g., health

events, divorce) from web browsing behavior using theory-informed machine

learning and link these measures with repeated responses to survey items from

resilience research. Second, we will handle missing data in rating scales and

free-text responses by building LLM-proxies that role-play a set of

“characters’’ and simulate their survey responses through fine-tuning and

in-context learning. Third, we will analyze resilient responses after major

life events and the predictive value of psychosocial resilience factors for

mental responses by comparing (i) survey-based measures used in resilience

research with (ii) the augmented measures obtained from our AI-powered

infrastructure. This comparison will assess the added value of our AI approach

and shed light on potential biases of findings in resilience research.