What can machine learning teach us about habit formation? Evidence from exercise and hygiene
Abstract
We apply a machine learning technique to characterize habit formation in two large panel data sets with objective measures of 1) gym attendance (over 12 million observations) and 2) hospital handwashing (over 40 million observations). Our Predicting Context Sensitivity (PCS) approach identifies context variables that best predict behavior for each individual. This approach also creates a time series of overall predictability for each individual. These time series predictability values are used to trace a habit formation curve for each individual, operationalizing the time of habit formation as the asymptotic limit of when behavior becomes highly predictable. Contrary to the popular belief in a "magic number" of days to develop a habit, we find that it typically takes months to form the habit of going to the gym but weeks to develop the habit of handwashing in the hospital. Furthermore, we find that gymgoers who are more predictable are less responsive to an intervention designed to promote more gym attendance, consistent with past experiments showing that habit formation generates insensitivity to reward devaluation.
Additional Information
© 2023 the Author(s). Published by PNAS. This article is distributed under Creative Commons Attribution-NonCommercial-NoDerivatives License 4.0 (CC BY-NC-ND). We thank Anthony Kukavica, Predrag Pandiloski, and Mira Potter-Schwartz for excellent research assistance. We thank Hengchen Dai for her help accessing and interpreting the Proventix data. We thank the Sloan Foundation G2018 11259 (CFC), the Behavior Change for Good Initiative (whose contributions were funded by the Robert Wood Johnson Foundation, the AKO Foundation, J. Alexander, M. J. Leder, W. G. Lichtenstein, the Pershing Square Fund for Research on the Foundations of Human Behavior from Harvard University and by Roybal Center grants (P30AG034546 and 5P30AG034532) from the National Institute on Aging), the Linde Institute (X.L.), and the Chen Neuroscience Institute at Caltech (A.B., CFC) for the financial support. We also thank audiences at BCFG, TA-DAH, Behavioral Ops (Shanghai), Stanford GSB, HKU, AAAI, and our internal science team meetings, for helpful ideas. We thank 24 h Fitness for partnering with the Behavior Change for Good Initiative at the University of Pennsylvania to make this research possible, as well as Proventix. Portions of this paper were developed from the thesis of A.B. Author Contributions. A.B., H.H., and C.C. designed research; A.B., H.H., K.L.M., A.L.D., and C.C. performed research; A.B., H.H., X.L., and C.C. analyzed data; and A.B., H.H., K.L.M., A.L.D., and C.C. wrote the paper. Data, Materials, and Software Availability. The data analyzed in this paper were provided by 24 h Fitness and Proventix. We have their legal permission to share the deidentified data. The data and code to replicate the analyses are available at https://osf.io/m8gdp/ (26). The authors declare no competing interest.Attached Files
Published - pnas.2216115120.pdf
Supplemental Material - pnas.2216115120.sapp.pdf
Files
Name | Size | Download all |
---|---|---|
md5:8d5d09c5b9d9bb9bedcab0d7983f129c
|
756.6 kB | Preview Download |
md5:7e7d69dfa8918ed181322ff1ca2be3a3
|
3.1 MB | Preview Download |
Additional details
- PMCID
- PMC10151500
- Eprint ID
- 121704
- Resolver ID
- CaltechAUTHORS:20230602-252014000.59
- Alfred P. Sloan Foundation
- G-2018-11259
- Tianqiao and Chrissy Chen Institute for Neuroscience
- Robert Wood Johnson Foundation
- AKO Foundation
- Pershing Square Fund for Research on the Foundations of Human Behavior
- NIH
- P30AG034546
- NIH
- 5P30AG034532
- Linde Institute of Economic and Management Science
- Created
-
2023-06-05Created from EPrint's datestamp field
- Updated
-
2023-06-05Created from EPrint's last_modified field
- Caltech groups
- Tianqiao and Chrissy Chen Institute for Neuroscience