Choice models traditionally assume that individuals act as utility maximisers, considering all information at their disposal in a compensatory fashion. However, cognitive limitations and choice complexity may lead individuals to adopt different choice heuristics and not fully attend all the information available. Further, when confronted with virtual choices in a survey, task engagement may also play an important role. We explore the relationship between cognitive effort and information attendance in decision-making processes, considering latent constructs associated with thinking styles (i.e. intuitive vs rational) and survey engagement. A novel residential-location stated-choice survey, incorporating the use of interactive aids, allowed us to achieve a measure of cognitive effort based on self-reported indicators and response time. Defining our benchmark as a choice involving a full information search, we also produced information attendance indicators based on recordings of the choice paths followed by the respondents and their stated non-attended pieces of information. With these data, we estimated a Multiple Indicator Multiple Cause (MIMIC) model showing that greater engagement leads to more information attended, while intuitive decision-makers exhibit lower information attendance in complex scenarios. Additionally, individuals with higher educational levels tend to engage in more analytical thinking, resulting in more information search, while intuitive thinking increases with age. Our findings provide a framework for understanding how cognitive effort, information attendance, engagement and thinking styles shape decision-making processes, particularly when individuals deviate from compensatory behaviour. These insights are valuable for a better understanding of the heuristic choice process and for the design of improved stated-choice surveys. Finally, we offer a few recommendations on how to address the cognitive effort and information attendance trade-off in designing stated choice surveys.
Mixed logit models are widely used to recover individual-level preferences for secondary analysis, yet both first-stage sampling uncertainty and variability in conditional distributions are often overlooked. This technical note illustrates the implications of ignoring these sources of uncertainty and provides reproducible R code, compatible with the apollo package, to better approximate the empirical sampling distribution and improve the reliability of second-stage inference.
Measuring the role of time in subjective value is essential to understanding human decision making. Subjective value for delayed prospects can be elicited through binary choices between options or through price equivalents, yet these methods systematically diverge, producing preference reversals in which people select smaller-sooner options in choice but assign higher bids to larger-later options. Rather than reflecting genuinely unstable preferences, we show these reversals arise from a cognitive process specific to pricing: dynamic anchoring on the payoff of the prospect being evaluated. We develop a price accumulation model that incorporates this anchoring and adjustment process, and fit it to both pricing and choice data using a deep learning approach to likelihood-free parameter estimation. Across three studies involving a large representative pricing sample (N = 300), a smaller within-participant gains experiment, and within-participant losses experiment, we show that the same underlying discount rate governs both choice and pricing, and that preference reversals can be explained by the anchoring component present in pricing but absent in choice. Extending this to delayed losses, we find that apparent loss aversion in pricing is similarly driven by stronger anchoring rather than by changes in discounting, while loss aversion in choice reflects shallower discounting rates for losses relative to gains. These findings reconcile choice-price inconsistencies under a common theory of value and suggest that pricing-based approaches may offer a practical pathway for reducing impulsive decision-making.
The rank-ordered logit (ROL) model is frequently used to aggregate rankings from multiple individuals into a single ranking that corresponds to the unobserved utility derived from the ranked alternatives. When these individuals exhibit heterogeneous preferences, they can be grouped to reduce bias. However, the common practice of fitting ROL models separately for each group may overlook shared patterns across groups, resulting in a loss of information. We introduce the sparse fused rank-ordered logit (SFROL) model, an extension of the ROL model, that allows joint learning from heterogeneous ranking data, where information from different groups is utilised to achieve better model performance. In this framework, the observed rankings are modelled as a function of covariates pertaining to the ranked alternatives. By imposing penalties on the likelihood function, we allow for both the sharing of information on the corresponding coefficients across groups and the shrinkage of the coefficients to zero to improve model interpretability, parameter estimation and prediction. Simulation studies generally indicate superior performance of the proposed method compared to existing approaches across a wide variety of scenarios. The use and interpretation of the method are illustrated on a sweet potato consumer preference application. An R package containing the proposed methodology can be found at https://CRAN.R-project.org/package=SFPL.
Estimating average marginal effects (AME) from logit models is a common approach for assessing attribute importance in conjoint analysis, but it rests on strong parametric and behavioral assumptions. Recent work has formalized conjoint analysis within Rubin’s potential outcomes framework and introduced the Average Marginal Component Effect (AMCE) as a nonparametric causal estimand, but AMCE relies on additional identification assumptions. The Conditional Randomization Test (CRT) does not require these assumptions for testing whether an attribute has any causal effect and can also be used to diagnose violations of key AMCE assumptions. This research note presents the first comprehensive empirical comparison of AME, AMCE, and CRT using four public datasets that vary in task complexity and attribute count, with the aim of providing practical guidance for applied researchers across conjoint designs of differing complexity. We find that designs with many attributes and many tasks are more prone to violating AMCE assumptions, whereas simpler designs tend to be more robust, making AMCE a preferred causal estimand in such settings. When AMCE assumptions fail, AME estimates, when consistent with CRT results, provide a more credible basis for inference than AMCE. When AME is not statistically significant but CRT suggests an effect, interaction terms can be explored within the logit model to capture preference heterogeneity and reconcile the discrepancy. CRT can therefore inform both assumption checking and parametric specification, although it should be used with caution because Lasso-based CRT implementations can occasionally yield counterintuitive outcomes in applied settings.
