Consistency of Risk Preference Measures: An Artefactual Field Experiment from Rural China

ABSTRACT

A variety of measures have been developed to elicit individual risk preferences. How these measures perform in the field, in particular in developing countries with non-student subjects, is still an open question. We implement an artefactual field experiment in rural China to investigate (i) consistency across incentivised experimental risk measures, (ii) consistency in risk preferences elicitation between non-incentivised survey measures and incentivised experiments, and (iii) possible explanations for risk preference inconsistency across measures. We find that inconsistent risk preferences across survey and experimental measures may be explained by ambiguity preferences. In the survey, subjects may mix risk and ambiguity preferences.

Acknowledgements

We acknowledge the financial support of the China Scholarship Council, the North-South Centre of ETH Zurich, and Southwestern University of Finance and Economics. We thank James Andreoni, Michael Bennet, Michael Kuhn, Peter Martinsson, and Jintao Xu for their comments and suggestions. Complementary funding was provided by the Alexander von Humboldt-Foundation. This research was partly conducted while at ETH Zurich, Chair of Environmental Economics and Policy. Experimental instructions, data, and codes are available upon request. All errors and omissions are our own responsibility.

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplemental Material

Supplementary Material is available at https://doi.org/10.1080/00220388.2017.1336542

Notes

1. See, for instance, studies by Barham, Chavas, Fitz, Salas, and Schechter (2014), Cardenas and Carpenter (2008), Dohmen et al. (2011), Hill (2009), Liu (2013), Liu and Huang (2013), Tanaka, Camerer, and Nguyen (2010), and Ward and Singh (2015).

2. See, for instance, studies by Blais and Weber (2006), Hill (2009), Dohmen et al. (2011), or Vieider et al. (2015) for measures relying on simple survey questions; and Anderson and Mellor (2009), Ding et al. (2010), Dohmen et al. (2011), or Vieider et al. (2015) for measures relying on hypothetical gambles, lotteries, and investments.

3. See for instance, studies by Crosetto and Filippin (2013), Deck et al. (2008), Eckel and Grossman (2002), Gneezy and Potters (1997), He, Martinsson, and Sutter (2012), Holt and Laury (2002), or Lejuez et al. (2002) for measures based on experiments.

4. A promising option in overcoming some of these challenges is the use of tablets. However, difficulties could still be encountered if participants are computer illiterate, if there are restrictions in importing and transporting these technologies, and if there is lack of electricity or battery chargers. All these challenges as well as insurance needs can imply additional financial and time costs.

5. In the pre-test, we mainly tested whether farmers could understand the questionnaire and the experimental instructions. In the first pre-test, we used a parallel gains/loss AH game following the original design by Andreoni and Harbaugh (2010). The AH game on gain worked well. However, due to its unusual design that asks subjects to choose their least preferred option, the AH game on loss confused subjects. The laboratory experiment conducted by Andreoni and Harbaugh (2010) has reported a similar problem. We decided to drop the loss part and only keep the gain part of the AH experiment in the second pre-test. After receiving positive feedback, we started the final experiments four days after the second pre-test.

6. $1 ≈ CNY 6.

7. In total, 34 subjects did not participate in the study.

8. The experimental instructions are available on request.

9. The proportion of subjects that did not answer the control questions correctly is very small (about 1.7%). This also implies that the scope for selection bias is very small, and that the inconsistency across the different elicitation measures might not be the result of poor understanding of the tasks.

10. In this study, we scale up the values of the outcomes in the original Holt and Laury (2002) study to make the maximal outcome equal to that in the Andreoni and Harbaugh (2010) experiment, so that the two experiments are comparable. All scaled-up values of outcomes are adjusted to the nearest multiples of 0.5 for the purpose of easing implementation.

11. A parallel gains/loss HL task is conducted but the task over losses is not in the scope of this study and is not reported. Therefore, each lottery over gains has a 1/20 chance of being selected.

12. Hardeweg et al. (2013) find that about 40 per cent of subjects choose the midpoint and 25 per cent of subjects choose the two extreme points. Charness and Viceisza (2016) find a high proportion (27%) at one extreme point.

13. The CRRA coefficient r estimated in the AH experiment is allocated into intervals of CRRA coefficient r shown in Table 2 for the purpose of comparing the AH experiment and the HL experiment.

14. Except for land size, the coefficients of these control variables are not significant at conventional statistical levels.

15. We also control for individual socio-demographic characteristics (age, gender, education, working off-farm, household size, income, and land size), village fixed effects, and cluster the standard errors at the village level. The effects of gender and household size are significant at the 5 per cent statistical level in both models. The effects of age and education are weakly significant (at the 10% statistical level), while the coefficients of the other variables are not significant.

Additional information

Funding

This work was supported by the China Scholarship Council; Alexander von Humboldt-Foundation; North-South Centre of ETH Zurich;