What can hypothetical choices tell us about unobservable behaviours? Nicolas KRUCIEN, Verity WATSON, Mandy RYAN Health Economics Research Unit University of Aberdeen 10th World Congress in Health Economics Dublin, 13/07-16/07/14 1/26
What can hypothetical choices tell us about unobservable behaviours?
Nicolas KRUCIEN, Verity WATSON, Mandy RYAN
Health Economics Research Unit University of Aberdeen
10th World Congress in Health Economics
Dublin, 13/07-16/07/14
1/26
Context • Growing evidence that participants to a DCE do not process the
information as suggested by the RUM model – ANA: Some attributes are excluded from the utilities
computation and comparison – RRM: Utilities comparison would mainly based on the
direct comparison of some attributes • Accounting for these behavioural phenomenon can lead to
qualitatively different DCE results • Models have been developed to account for these departures
from the standard RUM model, but weak theoretical foundations
2
Objective
• To improve modelling of preferences in a DCE by allowing respondents to use more flexible information processing strategies [IPS] (‘as if’ approach).
• Two main propositions: 1) Limited Attention instead of classical non-attendance 2) Co-existence of several IPS within the same choice
3
Method: Proposition 1
• Limited attention – Inspired by the Rational Inattention (RI) theory (Sims, 2003)
• Main components of the RI framework: – Agents have limited amount of cognitive resources they can
allocate to the decision problem – They have to decide which pieces of information (attributes)
is worth looking at (Comparison of costs and benefits) – The final choice (2nd stage) is conditional upon information
trade-offs made at the 1st stage – Agents still act rationally (Optimisation under constraint)
4
Method: Proposition 1
• “Impossible” to specify a cost function without additional information on attributes' valuation by the respondents
• Respondents simplify the information processing by ‘guessing’ about attributes’ values
5
Method: Proposition 1 • Different ‘guessing’ processes are compared: • Look at one alternative and use the observed values to infer
those of the remaining alternative – Probability of similarity: {0%, 10%, 20%}
• Look at one alternative and make ‘simple’ guesses about values of the other alternative – Actual values are replaced by expected values based on the
assumption that levels are equally likely • Look at one alternative and make ‘intelligent’ guesses about
the values of the other alternative – Actual values are replaced by expected values that took into
account the probability of each level to occur so far
6
Method: Proposition 2 • Random Regret Minimisation (RRM) (Chorus et al, 2008)
– Close to RUM in terms of GoF but still leads to important differences (predicted probabilities)
– Agents’ decision making is based on attributes comparison (Low order) vs. alternatives comparison (High order)
7
Attribute Alt [A] Alt [B] Alt [C]Cost (in £) 10 5 5
Time (in min) 30 20 30
U(A) U(B) U(C)
BR(C)
BR(C)
Method: Proposition 2 • Choice of RUM/RRM as an IPS is made for each attribute
separately depending on several factors (importance; format; uniqueness)
• 16 ‘hybrid’ models are estimated and compared to identify which attribute is best accounted for by either RUM/RRM
8
RUM RRM
DCE
• Secondary data analysis • Patients' preferences for
the role of the pharmacist in management of drug therapy • 4 attributes:
– Travel + Waiting time to GP [3 levels] – Travel + Waiting time to pharmacy [3 levels] – Chance of receiving best treatment [3 levels] – Cost [4 levels]
• 16 tasks including 3 alternatives {A, B, SQ} • 204 respondents (3,265 observations)
9
Results • Choice proportions • Limited Attention
Remark: Best ANA model “Pharma time” omited: LL= -2811.1
10
Utility RegretLOOK [A] INFER [B] (Rate: 0%) -2800.6 -2804.9LOOK [A] INFER [B] (Rate: 10%) -2805.3 -2809.6STANDARD -2811.1 -2813.1LOOK [A] INFER [B] (Rate: 20%) -2821.9 -2825.4LOOK [A] ASSUME DYNAMIC [B] -2834.6 -2836.5
Information processingLog-Likelihood
Table. Comparison of the different 'guessing' processes (Top 5)
A B SQ31.3% 12.0% 56.7%
Results • Hybrid IPS (+ “Look [A] Infer [B] at 0%”)
11
GP time Pharma time Chance Cost LogLikR R U R -2793.6R U U R -2793.7U R U R -2795.1U U U R -2795.2R U R R -2797.2
Table. Comparison of the different hybrid RUM-RRM models (Top 5)
Results
12
Beta 0.7184 * 0.4807 * 0.6058 * 0.5988 *SE 0.0528 0.0372 0.0522 0.0541
Beta -0.0139 * -0.0087 * -0.0133 * 0.0090 *SE 0.0014 0.0009 0.0014 0.0009
Beta -0.0001 -0.0003 0.0070 * -0.0049 *SE 0.0020 0.0013 0.0024 0.0016
Beta 0.6836 * 0.4684 * 0.6864 * 0.6824 *SE 0.0363 0.0249 0.0388 0.0388
Beta -0.0624 * -0.0399 * -0.0722 * 0.0513 *SE 0.0047 0.0030 0.0052 0.0036
LogLik -2811.1 -2813.1 -2800.6 -2793.6
Pharma time
Chance
Cost
AttributeTable. Comparison of the different discrete choice models
(1): Regret={GP time; Pharma time; Cost} - Utility={Chance}
RUM RRM RUM - LA HYBRID - LA (1)
ASC_SQ
GP time
Conclusion • Evidence that participants to a DCE make some assumptions
about the content of the alternatives – Is ANA generated by the researcher?
• In line with regret theory (Loomes & Sugden, 1982), respondents’ preferences are partially based on the anticipated performance of a considered option + other alternative(s) – Why is it true for only some attributes?
• Application of RI theory to DCM choice modelling is an interesting research avenue (raising several methodological issues) – Use of eye-tracking technology to collect data about respondents'’
“information- trade-offs among the attributes
13
Thank you for your attention
14