What is Conjoint Analysis?
Cutting-edge techniques rising in popularity in the social sciences
VETA’s Chief Science Officer, Teppei Yamamoto, has developed conjoint analysis into a causal inference methodology. The method has attracted major interest across many scientific fields as a way to estimate people’s preferences. His published papers have been cited more than 17,000 times (as of October 2025, according to Google Scholar) – an extraordinary number for a social science researcher in Japan, which shows the extent of this method’s influence on modern academic research.
Conjoint analysis is a statistical method for uncovering individuals’ preferences by constructing combinations of different “levels” for various “attributes” and analyzing which combinations of these levels they prefer.
For example, suppose we conduct a conjoint analysis focusing on how the attributes of immigrants shape support for their acceptance into the United States. In this case, things such as the immigrant’s ‘past travel history to the United States,’ ‘reason for application,’ and ‘country of origin’ are ‘attributes,’ while ‘Mexico’ and ‘Iraq’ would be possible ‘levels’ of the ‘country of origin’ attribute.
To perform such a conjoint analysis, the levels of the various attributes are randomly combined to create a set of hypothetical profiles. Respondents then compare and consider these profiles, choosing which ones they feel a preference for – which allows us to estimate the importance and influence of each attribute on the respondent’s choices.
Returning to the example of estimating how different attributes influence respondents’ attitudes towards immigrants, in the figure below you can see the profiles of ‘virtual immigrants’ 1 and 2, which randomly combine levels of their various attributes.
Respondents are presented with multiple tasks in which two or more ‘virtual profiles’ are presented and they must choose between them, as shown below.
In conjoint analysis, the content of the choices (profiles) presented to the respondents is determined randomly, which allows us to precisely estimate the effect of each attribute independently of other conditions. A key feature is that this allows the measurement of causal effects, showing us how specific factors cause changes in people’s preferences.
The average effect of each attribute level on preference is measured by an index called AMCE (Average Marginal Component Effect). AMCE indicates how much the probability of being selected changes when one level of an attribute is changed, assuming that all other conditions remain in an average state. In other words, it clearly shows how people’s judgments change depending on the presence or absence of a certain condition.
For example, in the case we introduced earlier of measuring support for immigration, it’s possible to estimate results such as ‘there is, on average, an X% point higher probability of an immigrant from Germany being accepted than an immigrant from Mexico’.
Moreover, specific combinations of levels of multiple attributes may affect people’s judgements. Such interactions between different attributes are measured by an index called ACIE (Average Component Interaction Effect). Returning to our immigration example, we might find, for instance, that migrants’ educational backgrounds are only of significance to their acceptance when they are engaged in specific careers or occupations.
These and other analytical techniques make it possible to understand, in a statistically supported manner, which factors are important to people’s decision-making when confronted with complex choices.
Conjoint analysis has several advantages compared to conventional surveys and other such methods.
① Understanding the causal effects of each factor and its combinations
In real-world choices, various factors are intertwined – which means that even if we find a correlation between a certain factor and a choice, it isn’t clear whether this is a true causal relationship. Conjoint analysis addresses this by randomly combining factors, so the independent impact of each one and its various combinations on decision-making can be measured.
For example, suppose that in the real world, immigrants from country B are more likely to be preferred than immigrants from country A. This may not be due to a negative preference for country A itself; it could simply be that immigrants from country A tend to have lower levels of education or language skills. By using conjoint analysis, it is possible to estimate the independent effects of each of these correlated factors and reveal the true causal effects.
② Evaluate the relative importance of multiple factors
Decisions we make in the real world are often a matter of trade-offs. For example, when choosing a rental property, properties that are very convenient for transportation will also have higher rents. Traditional surveys using one-on-one or multiple-choice formats are not very suitable for analyzing complex choices that include such conflicting factors – after all, everyone would like to live somewhere convenient and pay less rent, but everyone also has their own unique willingness to trade off between those factors.
In conjoint analysis, trade-offs between factors are directly used as part of the question format, making it possible to estimate which of the conflicting factors the respondent is considering and with what weight.
③ Simultaneous testing of many factors at low cost
If you want to measure the independent effects of the many factors involved in complex choices, you usually need to conduct randomized controlled trials (RCTs) such as A/B tests for each factor. Naturally, such surveys are costly.
In conjoint analysis, multiple attributes are incorporated into every profile, and each respondent is repeatedly asked to choose between sets of profiles. Therefore, a lot of information about multiple factors can be obtained at once. This means that conjoint analysis allows us to efficiently verify a large number of hypotheses simultaneously.
④ High reliability; design that mirrors real-world choices
It is often pointed out that the answers people give in surveys do not necessarily match their behaviour in the real world. For example, if you are asked a broad question about what kind of cuisine you like to eat for lunch, you will probably give answers that are unconsciously influenced by your mood at the time or what you have eaten recently. This can cause unexpected problems if the results obtained from such a survey are used for market research by a food company!
If conjoint analysis were used instead, in addition to the type of cuisine, factors such as price range, store reputation, and atmosphere would also be combined, and the question would be more specific; ‘If you were going to lunch, which of these stores would you choose?’ Respondents envisage this situation in the real world while answering the question, and studies have proven that the results are more consistent with actual behavioral data.
⑤ High data quality even over long surveys
Many widely used awareness surveys and aptitude tests include more than 100 question items, so it’s unsurprising that respondents tend to get tired and bored! Answering many questions in the same format reduces their attention and care, so the quality of the data decreases as the survey progresses.
Conjoint analysis also requires repeated answers to questions in the same format. However, sicne the tasks take the form of a selection task that imitates a real-world situation, respondents continue to answer without getting bored. Respondents often tell us that they felt like they were playing a game while answering these tasks, and in fact, past studies have demonstrated that the quality of the data hardly deteriorates even when respondents are asked to answer larger numbers of tasks.
⑥ Suppress ‘social desirability bias’ and strategic responses
Social desirability bias arises when respondents to a survey consciously or unconsciously change their answers to try to seem more ‘acceptable’ to the survey researcher. This is a common problem with traditional political or social science surveys, especially when they address sensitive content. In fields such as recruitment, meanwhile, there is a tendency for respondents to strategically give answers that make them appear to be ideal applicants from the company’s perspective.
Conjoint analysis has come to be regarded by researchers as a powerful method for suppressing this kind of bias. By concealing sensitive items among multiple factors, it is possible to divert respondents’ awareness bias and draw out their true feelings.