Skip to content

Commit

Permalink
Merge branch 'master' of https://github.com/yhoriuchi/projoint
Browse files Browse the repository at this point in the history
  • Loading branch information
aaronrkaufman committed Aug 16, 2023
2 parents 39dc704 + 29535db commit 516fbdd
Showing 1 changed file with 3 additions and 5 deletions.
8 changes: 3 additions & 5 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -13,8 +13,6 @@ Some notes:
* The current version assumes that the outcome variable is a binary forced choice.

* This package is still under construction. Forthcoming features include the following:
+ Add an article explaining how to set up a Qualtrics survey
+ Add a function for subgroup comparisons.
+ Allow researchers to use weights for features and for respondents.
+ Allow researchers to use other outcome variables, such as rating.

Expand All @@ -35,11 +33,11 @@ There are some excellent R packages for conjoint analysis, including [cjoint](ht

* The [cjoint](https://cran.r-project.org/web/packages/cjoint/) and [cregg](https://thomasleeper.com/cregg/) packages assume that the unit of analysis is each profile even when researchers design a binary choice experiment. We define more straightforward *choice-level* MMs and AMCEs and allow users to estimate them ([Clayton et al., working paper](https://gking.harvard.edu/conjointE)).

* The profile-level MMs ([cregg](https://thomasleeper.com/cregg/)) are always attenuated toward 0.5 if a conjoint design includes "ties" for the attribute of interest ([Gander 2021](https://doi.org/10.1017/pan.2021.41)). Consider, for example, that the attribute of your interest is a candidate's party, which has two levels, {Republican, Democrat}. If these levels are randomly assigned to each profile, 50% of profile-pairs include ties. Therefore, even if strongly partisan respondents only care about a candidate's party and always prefer one party to another, they have to choose one of the two tied profiles randomly. Therefore, the (profile-level) marginal mean of Republican for Republican supporters and Democrat for Democrat supporters is 0.75 (= 1 * 0.5 + 0.5 * 0.5). We remove these ties to calculate MMs.
* The profile-level MMs ([cregg](https://thomasleeper.com/cregg/)) are always attenuated toward 0.5 if a conjoint design includes "ties" for the attribute of interest ([Gander 2021](https://doi.org/10.1017/pan.2021.41)). Consider, for example, that the attribute of your interest is a candidate's party, which has two levels, {Republican, Democrat}. If these levels are randomly assigned to each profile, 50% of profile pairs include ties. Therefore, even if strongly partisan respondents only care about a candidate's party and always prefer one party to another, they have to choose one of the two tied profiles randomly. Therefore, the (profile-level) marginal mean of Republican for Republican supporters and Democrat for Democrat supporters is 0.75 (= 1 * 0.5 + 0.5 * 0.5). We remove these ties to calculate MMs.

* The profile-level AMCEs ([cjoint](https://cran.r-project.org/web/packages/cjoint/)) may produce counter-intuitive estimates ([Abramson, Kocak, Magazinni, and Strezhnev, working paper](https://osf.io/preprints/socarxiv/xjre9/)). For example, suppose that the attribute of your interest is a candidate's race, which has three levels, {white, Black, and Asian American}. Specifically, you are interested in whether white respondents prefer a Black or Asian American candidate. In the case of head-to-head comparison, suppose that 75% of white respondents choose a Black candidate rather than an Asian American candidate. (This is an example of choice-level marginal mean.) But the profile-level AMCE of choosing a Black candidate using an Asian American candidate may be larger or smaller than 75%, surprisingly. This is because the AMCE's calculation includes *indirect* comparisons of other combinations. Under a set of assumptions, profile-level AMCEs can still be researchers' substantive quantities of interest ([Bansak, Hainmueller, Hopkins, and Yamamoto, forthcoming](https://doi.org/10.1017/pan.2022.16)) but they should not be the only quantities of interest for applied researchers. Therefore, we provide an option to estimate alternative and more intuitive quantities of interest, choice-level MMs and AMCEs.
* The profile-level AMCEs ([cjoint](https://cran.r-project.org/web/packages/cjoint/)) may produce counter-intuitive estimates ([Abramson, Kocak, Magazinni, and Strezhnev, working paper](https://osf.io/preprints/socarxiv/xjre9/)). For example, suppose that the attribute of your interest is a candidate's race, which has three levels, {white, Black, and Asian American}. Specifically, you are interested in whether white respondents prefer a Black or Asian American candidate. In the case of head-to-head comparison, suppose that 75% of white respondents choose a Black candidate rather than an Asian American candidate. (This is an example of the choice-level marginal mean.) But the profile-level AMCE of choosing a Black candidate using an Asian American candidate as the baseline may be larger or smaller than 75%, surprisingly. This is because the AMCE's calculation includes *indirect* comparisons of other combinations. Under a set of assumptions, profile-level AMCEs can still be researchers' substantive quantities of interest ([Bansak, Hainmueller, Hopkins, and Yamamoto, forthcoming](https://doi.org/10.1017/pan.2022.16)) but they should not be the only quantities of interest for applied researchers. Therefore, we provide an option to estimate alternative and more intuitive quantities of interest, choice-level MMs and AMCEs.

* Researchers often assume that each profile within a profile-pair needs to be independent. But this is not a critical assumption for conjoint analysis. In fact, researchers should consider a variety of research questions with cross-attribute constraints, such as a profile pair with one incumbent and one challenger, a profile pair with the sum of values being one (e.g., the expected chance of winning a seat between two candidates). Our framework and this software allow researchers to examine a variety of questions in more flexible and innovative manners.
* Researchers often assume that each profile within a profile pair needs to be independent. But this is not a critical assumption for conjoint analysis. In fact, researchers should consider a variety of research questions with cross-profile (note: not cross-attribute) constraints, such as a profile pair with one incumbent and one challenger, a profile pair with the sum of values being one (e.g., the expected chance of winning a seat between two candidates), a profile pair in which one column is fixed (e.g., to show the status quo or as the reference). Our framework and this software allow researchers to examine a variety of questions in more flexible and innovative manners.

* Last but not least, our package provides the easiest possible way to re-label and re-order attributes and levels.

Expand Down

0 comments on commit 516fbdd

Please sign in to comment.