No CrossRef data available.
Article contents
The challenges of forecasting resilience
Published online by Cambridge University Press: 02 September 2015
Abstract
Developing prospective models of resilience using the translational and transdiagnostic framework proposed in the target article is a challenging endeavor and will require large-scale data sets with dense intraindividual temporal sampling and innovative analytic methods.
Information
- Type
- Open Peer Commentary
- Information
- Copyright
- Copyright © Cambridge University Press 2015
References
Allen, N. E., Sudlow, C., Peakman, T., Collins, R., Dal-Ré, R., Ioannidis, J. P. & Franco, E. L. (2014) UK Biobank data: Come and get it. Science Translational Medicine
6:224ed4. doi: 10.1126/scitranslmed.3008601.Google Scholar
Bollen, J., Mao, H. & Zeng, X. (2011) Twitter mood predicts the stock market. Journal of Computational Science
2(1):1–8.Google Scholar
Butler, E. A. & Randall, A. K. (2013) Emotional coregulation in close relationships. Emotion Review
5(2):202–10.Google Scholar
Coan, J. A., Schaefer, H. S. & Davidson, R. J. (2006) Lending a hand: Social regulation of the neural response to threat. Psychological Science
17(12):1032–39.Google Scholar
Fowler, J. H. & Christakis, N. A. (2008) Dynamic spread of happiness in a large social network: Longitudinal analysis over 20 years in the Framingham Heart Study. BMJ
337:a2338. doi: 10.1136/bmj.a2338.Google Scholar
Hastie, T., Tibshirani, R. & Friedman, J. (2009) The elements of statistical learning: Data mining, inference, and prediction, second edition. Springer.Google Scholar
House, J. S., Landis, K. R. & Umberson, D. (1988) Social relationships and health. Science
241(4865):540–45.Google Scholar
Kanner, A. D., Coyne, J. C., Schaefer, C. & Lazarus, R. S. (1981) Comparison of two modes of stress measurement: Daily hassles and uplifts versus major life events. Journal of Behavioral Medicine
4(1):1–39.Google Scholar
Kessler, R. C., Chiu, W. T., Demler, O., Merikangas, K. R. & Walters, E. E. (2005) Prevalence, severity, and comorbidity of 12-month DSM-IV disorders in the National Comorbidity Survey Replication. Archives of General Psychiatry
62(6):617–27.Google Scholar
Kessler, R. C., Sonnega, A., Bromet, E., Hughes, M. & Nelson, C. B. (1995) Posttraumatic stress disorder in the National Comorbidity Survey. Archives of General Psychiatry
52(12):1048–60.Google Scholar
Killingsworth, M. A. & Gilbert, D. T. (2010) A wandering mind is an unhappy mind. Science
330(6006):932. doi: 10.1126/science.1192439.Google Scholar
King, G. & Zeng, L. (2001) Logistic regression in rare events data. Political Analysis
9(2):137–63.Google Scholar
Lindquist, M. A. & McKeague, I. W. (2009) Logistic regression with Brownian-like predictors. Journal of the American Statistical Association
104(488)1575–85.Google Scholar
Master, S. L., Eisenberger, N. I., Taylor, S. E., Naliboff, B. D., Shirinyan, D. & Lieberman, M. D. (2009) A picture's worth: Partner photographs reduce experimentally induced pain. Psychological Science
20(11):1316–18. doi: 10.1111/j.1467-9280.2009.02444.x.Google Scholar
Rutledge, R. B., Skandali, N., Dayan, P. & Dolan, R. J. (2014) A computational and neural model of momentary subjective well-being. Proceedings of the National Academy of Sciences
111(33):12252–57.Google Scholar
Schapire, R. E. (1990) The strength of weak learnability. Machine Learning
5(2):197–227.Google Scholar
Schilbach, L., Timmermans, B., Reddy, V., Costall, A., Bente, G., Schlicht, T. & Vogeley, K. (2013) Toward a second-person neuroscience. Behavioral and Brain Sciences
36(4):393–414. doi: 10.1017/S0140525X12000660.CrossRefGoogle Scholar
Uchino, B. N., Cacioppo, J. T. & Kiecolt-Glaser, J. K. (1996) The relationship between social support and physiological processes: A review with emphasis on underlying mechanisms and implications for health. Psychological Bulletin
119(3):488–531.Google Scholar
Wager, T. D., Davidson, M. L., Hughes, B. L., Lindquist, M. A. & Ochsner, K. N. (2008) Prefrontal-subcortical pathways mediating successful emotion regulation. Neuron
59(6):1037–50. doi: 10.1016/j.neuron.2008.09.006.Google Scholar
Weiss, G. M. (2004) Mining with rarity: A unifying framework. ACM SIGKDD Explorations
6(1):7–19.Google Scholar
Zaki, J. & Williams, W. C. (2013) Interpersonal emotion regulation. Emotion
13(5):803–10. doi: 10.1037/a0033839.Google Scholar
Kalisch et al. present a thought-provoking translational and transdiagnostic framework for studying resilience. In this commentary, we apply their theoretical framework toward prospective prediction of resilient responses to negative life events. Prospective prediction is employed in many domains that depend on accurately forecasting a future state. For example, investors develop models to predict the future value of companies and markets, and epidemiologists develop models to predict the spread of disease. In the area of resilience, a well-formulated model should be able to both forecast the trajectory of an individual's resistance and recovery and generalize across forms of psychopathology and contexts. Such models could transform the study of mental health, but it is not clear how close we are to developing them.
Here, we describe three conceptual challenges for applying Kalisch et al.'s model of resilience in a forecasting framework: (1) resilience is a process unfolding over time, not an outcome that can be measured at a discrete time point; (2) cognitive processes alone are unlikely to predict resilience accurately; and (3) low base rates pose a challenge to predictive accuracy. To help overcome these challenges, we will need studies with large, diverse samples and dense intraindividual temporal sampling.
1. Defining resilient outcomes
Kalisch et al. define resilience as the empirically observed absence of lasting mental health problems following adversity and propose that it can be operationalized as the change in mental health symptoms before and after an adverse event, with a slope of zero indicating a resilient outcome. But at which time points should such a slope be measured? As time passes after a stressful event, the likelihood of returning to a baseline measure becomes greater, increasing the apparent “resilience” independent of any characteristics of the individual. Alternatively, we could estimate the functional form, or shape, of symptom severity as it unfolds across time.
As resilience is likely a dynamic process reflecting multiple mechanisms operating on different timescales, modeling the temporal trajectory may be particularly informative about which mechanisms are involved. This endeavor will require dense sampling of intraindividual data across time and the application of emerging statistical techniques for modeling trajectories, such as functional data analysis (Lindquist & McKeague Reference Lindquist and McKeague2009).
2. Multiple resilient processes
Kalisch et al. adopt a predominantly cognitive view of resilience, proposing a fundamental role for positive appraisal style, which comprises three distinct intrapersonal processes: (1) the initial appraisal, (2) subsequent reappraisal, and (3) inhibiting alternative interfering appraisals. We agree that appraisal and reappraisal are critical (Wager et al. Reference Wager, Davidson, Hughes, Lindquist and Ochsner2008); however, to develop accurate, generalizable models of resilience, we will likely need to incorporate a broader set of mechanisms, including interpersonal ones. Social support can attenuate negative affective responses (Coan et al. Reference Coan, Schaefer and Davidson2006; Master et al. Reference Master, Eisenberger, Taylor, Naliboff, Shirinyan and Lieberman2009) and has been associated with positive long-term health benefits (House et al. Reference House, Landis and Umberson1988; Uchino et al. Reference Uchino, Cacioppo and Kiecolt-Glaser1996). These processes are likely not fully describable in terms of intrapersonal appraisals, but rather will require models of bidirectional, interpersonal feedback loops (Butler & Randall Reference Butler and Randall2013; Schilbach et al. Reference Schilbach, Timmermans, Reddy, Costall, Bente, Schlicht and Vogeley2013; Zaki & Williams Reference Zaki and Williams2013). For example, our feelings of happiness appear to be directly influenced by our peers and can propagate dynamically through our social network over time (Fowler & Christakis Reference Fowler and Christakis2008).
Therefore, as we move toward prospective models of resilience, it will be important to incorporate both intra- and interpersonal processes. Ensemble algorithms from statistical learning offer a promising approach to integrate multiple mechanisms into a single model (Hastie et al. Reference Hastie, Tibshirani and Friedman2009; Schapire Reference Schapire1990).
3. The base rate problem
One of the challenges of selecting training data for a predictive model is dealing with a very high base rate of resilient outcomes and an extraordinarily low base rate of significant negative life events on a daily basis. To make this more concrete, based on the lifetime prevalence of depression (Kessler et al. Reference Kessler, Chiu, Demler, Merikangas and Walters2005), the probability of an individual not being depressed on a given day is roughly 99.99%. Using Bayes' rule to combine this high base rate of not being depressed with a low frequency of significant traumatic life events (0.002%; Kessler et al. Reference Kessler, Sonnega, Bromet, Hughes and Nelson1995) reveals a very low conditional probability that an individual will not be depressed given an adverse life event (less than 5%). Therefore, in the general population, resilience defined as a null change across time is actually the standard response, and it will be difficult to identify when true resistance to and recovery from adversity occurs (King & Zeng Reference King and Zeng2001; Weiss Reference Weiss2004).
To account for these statistical issues, trauma researchers typically have focused on examining resilience to shared traumatic events such as the collapse of the World Trade Center. Such an approach will not be sufficient to develop predictive models of resilience, however, as these events are sampled a posteriori rather than prospectively, precluding baseline assessment, and it remains unclear how well mechanisms generalize beyond this experience. Therefore, to increase the predictive power of such forecasting models, it will be necessary to collect large-scale data sets and find a way to increase the frequency of events to train the model (Kanner et al. Reference Kanner, Coyne, Schaefer and Lazarus1981). Concerted, nationally funded efforts such as the 500,000-person UK Biobank project will help (Allen et al. Reference Allen, Sudlow, Peakman, Collins, Dal-Ré, Ioannidis and Franco2014), and new avenues to large-scale data collection are continually developing with the rapid proliferation of social media, mobile sensing, and cloud computing. For example, using experience sampling of mood from mobile devices, researchers recently collected more than 500,000 samples from approximately 30,000 people (Killingsworth & Gilbert Reference Killingsworth and Gilbert2010; Rutledge et al. Reference Rutledge, Skandali, Dayan and Dolan2014). Furthermore, general public mood can be assessed by mining Twitter feeds, and these metrics appear to modestly predict other global metrics such as the Dow-Jones Industrial Average (Bollen et al. Reference Bollen, Mao and Zeng2011).
4. Conclusion
Though the challenges we have raised in developing predictive models of resilience are substantial, they are inherent to many other problems (e.g., predicting the stock market, forecasting weather, etc.) and are by no means insurmountable. Resilience research can learn from other fields outside of psychology and neuroscience, which have addressed parallel problems with predicting complex and rare events. Billions of dollars are poured into financial markets, and the most powerful supercomputers in the world are continually running simulations to improve our weather forecasts. Why should improving our mental health by predicting resilience be any less important?