Choice-Based Conjoint

Choice-Based Conjoint Tasks, Part-Worths and Health Preference Trade-Offs

Choice-based conjoint (CBC) is the marketing research name for conjoint studies built on repeated choices between described alternatives, and health economists usually call the same task a discrete choice experiment. Readers of health preference research meet both names and need to know whether they describe one method or two. The term comes from marketing research and sits inside conjoint analysis, and CBC outputs include relative importance, willingness to pay, maximum acceptable risk and simulated choice shares. The full modelling workflow for health studies is described under discrete choice experiments.

One task, several names

Conjoint analysis is a family of stated preference methods that infer the value of attributes from responses to profiles combining them. The ISPOR checklist for conjoint analysis in health counts discrete choice experiments as one member of that family, alongside other multi-attribute formats, and asks whether the elicitation format was rating, ranking or choice. Sawtooth Software defines choice-based conjoint by the task, in which respondents choose concepts from sets of concepts instead of rating or ranking them.

The ISPOR task force on statistical analysis notes that discrete choice experiments in market research are often called choice-based conjoint studies, and that marketing studies mainly aim to predict choices, whereas outcomes research focuses on preferences for changes in attributes and levels. Germany's Institute for Quality and Efficiency in Health Care (IQWiG) used the same framing in a 2014 pilot, describing a discrete choice experiment as the choice-based variation of conjoint analysis. Not every choice modeller accepts the shared label. Louviere, Flynn and Carson published a 2010 paper titled "Discrete choice experiments are not conjoint analysis". A study is therefore best judged by its task format and statistical model, not by its name.

From marketing research to health preference studies

The choice format joined conjoint questionnaires to the discrete choice models of econometrics. Sawtooth's technical paper credits a 1983 article by Louviere and Woodworth with providing a theoretical foundation for choice-based conjoint, and lists three features of their approach, namely choice sets built from complex experimental designs, an optional "none of these" alternative, and aggregate estimation by multinomial logit rather than estimation for each respondent. The paper dates large-scale commercial use to the 1990s, after the first CBC software release in 1993, with hierarchical Bayes estimation of individual part-worths following in the mid-1990s.

The ISPOR statistical report traces the economic basis to McFadden, who showed that conditional logit is consistent with random utility theory, and dates a rapid increase in health applications to the middle of the 1990s. The 2013 ISPOR design report states that conjoint methods, particularly discrete choice experiments, had become the most frequently applied stated preference approach in health care. A review of 301 health discrete choice experiments published from 2013 to 2017 found that 47 reported using Sawtooth's software for the experimental design.

Building the choice tasks

A CBC study describes each alternative by a fixed list of attributes, each set at one of a few levels, and asks respondents to pick one alternative from each choice set. The ISPOR checklist asks whether an opt-out or status quo alternative should be included and whether the experimental design was justified and its properties evaluated.

Sawtooth's paper contrasts fixed orthogonal designs, which often use a single questionnaire version and sometimes split respondents into blocks, with randomised designs in which each respondent sees a unique, carefully chosen set of tasks. Its software generates 300 versions by default, seeks level balance and orthogonality, and the vendor reports that any efficiency loss against strictly orthogonal designs is usually in the range of 5 to 10%. Minimal overlap of levels within a task is optimal for main effects, but the default procedure, Balanced Overlap, allows some overlap because this may improve the precision of interaction effects at some cost to main effects. Health studies have moved towards efficient designs that use prior information about the coefficients, and the 2013 to 2017 review reports more use of D-efficient designs.

Estimating part-worths from choices

CBC data are analysed with the same random utility models as any discrete choice experiment. This section covers only what the worked example needs and the estimator most tied to CBC software. Conditional logit gives the following.

$$P_{nj} = \frac{\exp(V_{nj})}{\sum_{m \in C_n} \exp(V_{nm})}, \qquad V_{nj} = \sum_{k=1}^{K} \beta_k x_{njk}$$

where $P_{nj}$ is the probability that respondent $n$ chooses alternative $j$, $C_n$ is the set of alternatives shown, $V_{nj}$ is the systematic utility of alternative $j$, $x_{njk}$ is the level of attribute $k$ in that alternative, $\beta_k$ is the preference weight (part-worth) for attribute $k$ and $K$ is the number of attributes. The logit form follows from assuming independent, identically distributed extreme value errors, which implies independence of irrelevant alternatives.

Conditional logit returns one set of mean weights for the sample. Hierarchical Bayes, built into CBC software, instead estimates weights for each respondent and borrows from the sample distribution only as far as that person's choices are insufficient. The ISPOR report describes conditional logit for each person's choices at the lower level and, at the upper level,

$$\beta_n \sim N(b, W)$$

where $\beta_n$ is the vector of preference weights for respondent $n$, $b$ is their population mean and $W$ is their variance-covariance matrix, estimated iteratively with a Gibbs sampler. Sawtooth reports that most of its CBC customers use hierarchical Bayes for the market simulation models they deliver. Mixed logit and latent class models are covered on the discrete choice experiment page.

What the estimates are used for

Part-worths have no natural unit, so results are reported as differences or ratios. The ISPOR statistical report lists money equivalence (willingness to pay), risk equivalence (maximum acceptable risk) and time equivalence among the trade-offs a choice study can estimate.

The report measures an attribute's importance as the difference between the weights of its best and worst levels, over the range of levels in the experiment. Dividing each difference by the sum across attributes gives a share.

$$RI_k = \frac{R_k}{\sum_{h=1}^{K} R_h}$$

where $RI_k$ is the relative importance of attribute $k$ and $R_k$ is the difference between its highest and lowest level weights. The marginal rate of substitution between two attributes is the ratio of their weights; it becomes willingness to pay when the denominator is cost and maximum acceptable risk when it is a risk.

$$WTP_k = -\frac{\beta_k}{\beta_c}, \qquad MAR = -\frac{\beta_b \, \Delta_b}{\beta_r}$$

where $WTP_k$ is the money given up for a one-unit gain in attribute $k$, $\beta_c$ is the (negative) weight on cost, $MAR$ is the largest risk increase accepted for a benefit gain of $\Delta_b$ units, $\beta_b$ is the weight per unit of benefit and $\beta_r$ is the (negative) weight per unit of risk.

Marketing CBC also feeds a market simulator, which applies the estimated weights to scenarios never shown to respondents and reports shares of choice. Sawtooth notes that simulators built on aggregate logit inherit the independence of irrelevant alternatives property, so a new product takes share from the others in proportion to their shares, so adding a near-copy of an existing product can give it an unrealistic gain. Simulators built on individual or latent class estimates are less exposed.

Worked example of an illustrative treatment choice study

The figures are hypothetical and chosen for easy arithmetic. People with a chronic inflammatory condition choose between two treatments described by four attributes, and a conditional logit model with linear coding returns these coefficients.

1. Relative importance. The ranges are 0.040 × 40, 0.25 × 4, 0.60 and 0.012 × 50, summing to 1.60 + 1.00 + 0.60 + 0.60 = 3.80. For remission the calculation is as follows.

$$RI_{remission} = \frac{1.60}{3.80} \approx 0.421$$

where $RI_{remission}$ is remission's share of the total utility range. Infection risk takes 26.3% and administration and cost 15.8% each.

2. Maximum acceptable risk. For a remission gain of 20 percentage points, with the signed coefficients shown below.

$$MAR = -\frac{0.040 \times 20}{-0.25} = \frac{0.80}{0.25} = 3.2$$

where $MAR$ is the rise in one-year infection risk, in percentage points, that exactly offsets the gain. On average, respondents would accept up to 3.2 points of extra risk for that benefit.

3. Willingness to pay. For the tablet rather than the injection the calculation is as follows.

$$WTP_{tablet} = -\frac{0.60}{-0.012} = 50$$

where $WTP_{tablet}$ is the monthly payment, in pounds, that respondents would make to take a daily tablet instead of a weekly injection. A 10-point remission gain (utility 0.40) is worth 0.40 divided by 0.012, about £33.33 a month.

4. A simulator scenario and the near-copy problem. Treatment A offers 80% remission, 3% risk, an injection and a £25 cost. Treatment B offers 60% remission, 1% risk, a tablet and a £10 cost; £10 was not a tested level, but linear coding allows values inside the tested range. The systematic utilities are as follows.

$$V_A = 0.040 \times 80 - 0.25 \times 3 - 0.012 \times 25 = 3.20 - 0.75 - 0.30 = 2.15$$

where $V_A$ is the systematic utility of treatment A, with the injection as the reference level. For treatment B the utility is as follows.

$$V_B = 0.040 \times 60 - 0.25 \times 1 + 0.60 - 0.012 \times 10 = 2.40 - 0.25 + 0.60 - 0.12 = 2.63$$

where $V_B$ is the systematic utility of treatment B. Exponentiating gives about 8.585 for A and 13.874 for B, so A's share in a two-product scenario is as follows.

$$S_A = \frac{8.585}{8.585 + 13.874} \approx 0.382$$

where $S_A$ is the simulated share of choice for A. Adding a second tablet identical to B under another brand, the aggregate logit simulator gives A 8.585 divided by 36.333, about 23.6%, and each tablet 38.2%. People who preferred A have no reason to switch to a copy of B, so A's share should stay near 38%; the drop is the near-copy problem. A share choosing neither treatment would need an estimated opt-out constant, which this design lacks.

5. Range dependence. Had infection risk run from 1% to 10%, its range would be 0.25 × 9 = 2.25 and the total 5.05, so its importance would rise to 44.6% and remission's would fall to 31.7%, with every coefficient unchanged. Willingness to pay and maximum acceptable risk would not move, because they rest on coefficients alone.

Use in health technology assessment and benefit-risk decisions

In the two examples below, CBC results serve as weights for combining outcomes and as evidence on how much risk patients accept for a benefit. Both uses require the sample, attributes and levels to match the decision.

IQWiG's 2014 pilot (working paper GA10-03, whose English summary is a translation of the authoritative German text) tested whether a choice-based conjoint study could weight the outcome-specific results of its efficiency frontier method in chronic hepatitis C. Patients and professionals chose between two hypothetical treatments described by seven attributes, analysed by logistic regression; sustained virological response received the largest weight (coefficient 0.804). The institute judged weighting by this method feasible for patients, found that the linearity of weights could be tested, and noted unresolved methodological issues. It treats conjoint analysis as a method of multi-criteria decision analysis within health technology assessment; such weights combine outcome results and are not the health state utilities used for QALYs.

The US Food and Drug Administration's non-binding guidance on patient preference information for devices, issued in March 2026 and superseding its 2016 guidance, says a discrete choice experiment may be optimal when trade-offs among several attributes are needed. It names preference weights, maximum acceptable risk and minimal acceptable benefit as parameters that can support a device's benefit-risk assessment.

Where choice-based conjoint results mislead

Several misreadings arise when outputs are treated as if they were independent of the design. The worked example shows two of them. Relative importance depends on the level ranges, so shares from studies with different ranges cannot be compared, and aggregate logit simulators misallocate share when similar products are added.

The ISPOR statistical report observes that hierarchical Bayes does not estimate a sample-wide mean directly and cannot test significance for any individual, so individual part-worths are not precise personal measurements. Sawtooth's paper acknowledges that choices are an inefficient way to elicit preferences, since each answer carries less information than ratings of every alternative. Simulated shares describe stated choices, not uptake, and willingness to pay is only as credible as the cost attribute. Because one study may carry any of the three names, reports should state the task format, design method and estimator.

Sources

• Bridges JFP, Hauber AB, Marshall D, Lloyd A, Prosser LA, Regier DA, Johnson FR, Mauskopf J. Conjoint analysis applications in health: a checklist. A report of the ISPOR Good Research Practices for Conjoint Analysis Task Force. Value in Health. 2011;14(4):403-413. https://doi.org/10.1016/j.jval.2010.11.013
• Food and Drug Administration. Incorporating Voluntary Patient Preference Information over the Total Product Life Cycle: Guidance for Industry, Food and Drug Administration Staff, and Other Interested Parties. FDA; 30 March 2026. https://www.fda.gov/media/92593/download
• Hauber AB, González JM, Groothuis-Oudshoorn CGM, Prior T, Marshall DA, Cunningham C, IJzerman MJ, Bridges JFP. Statistical methods for the analysis of discrete choice experiments: a report of the ISPOR Conjoint Analysis Good Research Practices Task Force. Value in Health. 2016;19(4):300-315. https://doi.org/10.1016/j.jval.2016.04.004
• Institute for Quality and Efficiency in Health Care (IQWiG). Choice-based Conjoint Analysis: pilot project to identify, weight, and prioritize multiple attributes in the indication "hepatitis C". Executive summary of working paper GA10-03, version 1.1 (English translation). Cologne: IQWiG; 2014. https://www.iqwig.de/download/ga10-03_executive-summary-of-working-paper-11_conjoint-analysis.pdf
• Johnson FR, Lancsar E, Marshall D, Kilambi V, Mühlbacher A, Regier DA, Bresnahan BW, Kanninen B, Bridges JFP. Constructing experimental designs for discrete-choice experiments: report of the ISPOR Conjoint Analysis Experimental Design Good Research Practices Task Force. Value in Health. 2013;16(1):3-13. https://doi.org/10.1016/j.jval.2012.08.2223
• Louviere JJ, Flynn TN, Carson RT. Discrete choice experiments are not conjoint analysis. Journal of Choice Modelling. 2010;3(3):57-72. https://doi.org/10.1016/S1755-5345(13)70014-9
• Louviere JJ, Woodworth G. Design and analysis of simulated consumer choice or allocation experiments: an approach based on aggregate data. Journal of Marketing Research. 1983;20(4):350-367. https://doi.org/10.2307/3151440
• Sawtooth Software. The CBC System for Choice-Based Conjoint Analysis. Version 9. Technical Paper Series. Orem, UT: Sawtooth Software; 2017. https://content.sawtoothsoftware.com/assets/0891a76f-93d3-4838-a38d-8ac0a2cda519
• Soekhai V, de Bekker-Grob EW, Ellis AR, Vass CM. Discrete choice experiments in health economics: past, present and future. PharmacoEconomics. 2019;37(2):201-226. https://doi.org/10.1007/s40273-018-0734-2