Explaining Online Health
Information Seeking

Applying the Planned Risk Information Seeking Model to Behavioral Data

Marko Bachl & Elena Link

Freie Universität Berlin | Johannes Gutenberg-Universität Mainz

2025-06-14

Background & Motivation

• Online health information-seeking established informational practice; search engines common starting point
• Models like the Planned Risk Information Seeking Model (PRISM) well established to explain self-reported health information seeking intentions (Kahlor, 2010; Link et al., 2021; Ou & Ho, 2022)
• Studies using passive observations of information seeking behaviors limited to few demographic predictors (Bach & Wenz, 2020; Bachl et al., 2024)

Research Question

Are the PRISM predictors of information seeking intentions transferable to information seeking behaviors?

Credit: Brett Jordan on Unsplash

PRISM predictors

We considered the following direkt positive paths to online information seeking:

• Attitudes towards online information seeking
• Perceived seeking control
• Negative affective risk responses
• Seeking-related subjective norms
• Risk perception: Severity and susceptibility of illness
• Current and desired levels of knowledge

(Kahlor, 2010; Link et al., 2021; Ou & Ho, 2022)

Methods

Methods: Survey & Tracking

• Stratified sample (age, gender, education) by commercial market research institute
• Predictors of online health information-seeking following established PRISM conventions
  • Attitudes towards online information seeking, Perceived seeking control, Negative affective risk responses, Seeking-related subjective norms: Multi-item latent scales, CFA
  • Risk perceptions (severity and susceptibility of illness), current and desired level of knowledge: Single items
• Three months of passively recorded browsing behavior (desktop & mobile) at URL level
• 728 participants after data cleaning

Credit: mk. s on Unsplash

Methods: O-HISB

• Extraction of 170k search queries from search engine visit URLs
• LLM-based zero-shot classification (Gilardi et al., 2023; Heseltine & Clemm von Hohenberg, 2024; Hoof et al., 2025) using the open weights reasoning model DeepSeek-R1-Distill-Llama-70B (DeepSeek-AI et al., 2025) deployed on a secure server operated by KISSKI.

Credit: Enzo Tommasi on Unsplash

Methods: O-HISB

Prompt

Your task is to assist in the classification of search queries for a research project on health communication.

You will receive a search query that was submitted to a search engine like Google. Your task is to classify the search query as either health-related or not health-related.

We define health according to the WHO: Health is a state of complete physical, mental and social well-being and not merely the absence of disease or infirmity.

Searching for medical professionals (for example, doctors, nurses) and health-related institutions (for example, health insurance, hospitals, nursing homes) are also considered health-related, even if no further context is given for the purpose of the search.

Searching for health-related issues of animals specifically is not considered health-related.

If there is insufficient content to make a reliable classification, categorize the query as not health-related.

Methods: O-HISB

• Extraction of search queries from search engine visits
• LLM-based zero-shot classification (Gilardi et al., 2023; Heseltine & Clemm von Hohenberg, 2024) using the open weights reasoning model DeepSeek-R1-Distill-Llama-70B (DeepSeek-AI et al., 2025) deployed on a secure server operated by KISSKI.
• 5 iterations for each search query; classification as health-related only if 5 out of 5.
• Two measures:
  • Binary indicator: O-HISB = 5 or more health-related searches
  • Count indicator: Number of health-related searches

Credit: Enzo Tommasi on Unsplash

Sample distribution of O-HISB

Results

Predicting O-HISB (binary): Tjur’s \(R^2\) = 0.02

Predicting O-HISB (count): Nagelk. \(R^2\) = 0.04

Summary & speculation

• Only few and weak effects of PRISM predictors on behavior

• Intention-behavior gap? (limitation: seeking intention not measured)
• Researcher-participant gap? (limitation: participants’ understanding of “health” not measured, annotation not yet systematically validated)
• Self-report-observation gap? (limitation: self-reported behavior not measured)

Credit: Tim Mossholder on Unsplash

Conclusion

Mismatch has important theoretical and methodological implications — systematic research needed

Comments are welcome - thank you!

Marko Bachl & Elena Link

marko.bachl@fu-berlin.de

elena.link@uni-mainz.de

Download presentation

Sample (n = 728)

Variable	Summary
Mean age (SD)	49.92 (11.81)
gender [Weiblich], %	48.2
gender [Männlich], %	51.6
gender [Nicht-binär], %	0.0
gender [Agender], %	0.1
gender [Anderes und zwar:], %	0.0

Variable	Summary
education [Kein Abschluss], %	0.3
education [Volksschul-/Hauptschulabschluss], %	13.6
education [Mittlere Reife / Realschulabschluss], %	35.3
education [Abitur / Fachabitur], %	20.5
education [Berufsakademie], %	6.3
education [Universitäts-/Hochschulabschluss], %	22.7
education [Promotion], %	1.4
ohis_binary [Yes], %	48.6
Mean ohis_count (SD)	14.85 (33.25)

References

Bach, R. L., & Wenz, A. (2020). Studying health-related internet and mobile device use using web logs and smartphone records. PLOS ONE, 15(6), e0234663. https://doi.org/gmmdrj

Bachl, M., Link, E., Mangold, F., & Stier, S. (2024). Search engine use for health-related purposes: Behavioral data on online health information-seeking in Germany. Health Communication, 39(8), 1651–1664. https://doi.org/gtg7gv

DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., … Zhang, Z. (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. https://arxiv.org/abs/2501.12948v1

Gilardi, F., Alizadeh, M., & Kubli, M. (2023). ChatGPT outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30), e2305016120. https://doi.org/gsqx5m

Heseltine, M., & Clemm von Hohenberg, B. (2024). Large language models as a substitute for human experts in annotating political text. Research & Politics, 11(1), 20531680241236239. https://doi.org/gtkhqr

Hoof, M. van, Trilling, D., Meppelink, C., Möller, M., Judith, & Loecherbach, F. (2025). Googling politics? Comparing five computational methods to identify political and news-related searches from web browser histories. Communication Methods and Measures, 19(1), 63–89. https://doi.org/gt45n9

Kahlor, L. (2010). PRISM: A planned risk information seeking model. Health Communication, 25(4), 345–356. https://doi.org/fg6c8h

Link, E., Baumann, E., & Klimmt, C. (2021). Explaining online information seeking behaviors in people with different health statuses: German representative cross-sectional survey. Journal of Medical Internet Research, 23(12), e25963. https://doi.org/gnrmz3

Ou, M., & Ho, S. S. (2022). A meta-analysis of factors related to health information seeking: An integration from six theoretical frameworks. Communication Research, 49(4), 567–593. https://doi.org/gq23kt