Experimental Design Details
The experiment comes right after the first questionnaire, which asks the respondent’s basic demographic information (see the attachment “Questionnaire.pdf” for the first questionnaire). While the survey is conducted in-person, the experimental part is applied on a tablet which the interviewer hands to the respondent answers the questions privately so that the interviewers cannot see the woman’s responses in the experiment. The interviewees can ask questions to the interviewers while answering the questions.
In the experiment, each woman is randomly assigned to one of ten treatments which consist of different hypothetical partner choice scenarios (see the attachment “ExperimentTreatments.xlsx”, sheet “Experiment_english”, for the sets of scenarios and “Questionnaire.pdf” for the exact instructions). While the treatments 1-9 ask about women’s own preference, the treatment 10 asks about what women think their friends prefer. We will discuss the treatment 10 later.
In treatments 1-9, each treatment consists of twelve scenarios and each scenario consists of two hypothetical male profiles. Each profile has four of the following attributes that differ by the treatment, but the salary is included in all treatments: Affection, Attractiveness, Compatibility, jealousy with controlling behaviour, Economic instability, Education, Faithfulness, Frustration, Hardworkingness, Respectfulness, Salary, Trustworthiness, jealousy without controlling behaviour. The degrees of four attributes vary across the twelve scenarios and the salary of the profile depends on women’s household income asked at the beginning of the experiment; specifically, the salary a multiple of women’s household income with the multiplication factor being one of the following: 0.5, 0.6, 0.7, 0.8, 0.9, 0.95, 1, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2. For each profile, a woman is asked to give a rating with a scale of 1 to 7 by his desirability as her partner.
For each woman, the order of the twelve scenarios shown to her and which profile appears on which side of the screen (left or right) are randomized. Also, the position of the four attributes of the profile (first, second, third, or fourth) is randomized at the beginning of the experiment for each woman but fixed afterward; for example, one woman may see salary in the first, frustration in the second, affection in the third, and education in the fourth of the attribute list while another woman may see education in the first, frustration in the second, affection in the third, and salary in the fourth of the attribute list even if they are assigned to the same treatment.
To estimate the preference for partner attributes, we follow Wiswall and Zafar (2018). Specifically, we follow the specification presented in section V.B. and the estimation procedure in section V.D. In the experiment, each woman i observes J=12 pairs of profiles, with a vector of K=4 partner attributes X. In each of the J pairs, they are asked to assign a point value between 1 and 7 for each of the profiles, denoted pj1 and pj2, where 1 and 2 indicate each profile, and j indicates the particular pair being seen among all J comparisons.
We use this data to estimate equation (6) of Wiswall and Zafar (2018) for each woman i. Note that this model is estimated for each woman using 12*2=24 unique observations, hence the parameter vector beta_i is indexed by i. For the reasons laid out in Wiswall and Zafar (2018), we estimate the model using a least absolute deviations (LAD) estimator.
For inference, we follow the same block bootstrap inference procedure as Wiswall and Zafar (2018) section V.D.
To make the interpretation of the parameter estimates easier, we compute the willingness-to-pay (WTP) using equation (8) of Wiswall and Zafar (2018) section V.F. We are interested in the median WTP for each partner attribute and especially the violent attributes. We also look at the distribution of the WTP.
We exclude from our sample women who have the same ratings for all the 24 profiles (e.g., give ratings 1 for all profiles) or women who give a different rating only to a single profile because we cannot estimate beta_i for those women.
Our main interest is in documenting median WTP across the entire population of surveyed women, representative of all women aged 18-55 living in urban areas of Chile. However, we will additionally document heterogeneity in median WTP calculations as laid out previously, stratifying on the following characteristics:
• Whether the woman has suffered domestic violence in the past, and the type of violence suffered
• Whether the woman has children
• The woman’s beliefs about support networks and state responses in cases of DV
• Her beliefs over the chances that she will suffer DV in the next 12m asked in an elicited beliefs module which comes later in the survey.
• Her income (grouped), her education level, her age and religion
We will correct for multiple inference in the heterogeneity analyses using a False Discovery Rate correction described by Anderson (2008).
Within the experiment we conduct two checks to examine the reliability of survey experiment results. The first consists of repeating one experiment (a group of 12 profiles), where instead of asking women about their choices (as in experiments 1-9) we ask them the decisions they would advise for a friend (experiment 10). This experiment (experiment 10) has exactly the same profiles as experiment 1, with the only difference being that women are asked about rating the partners for her friend, instead of for herself. We will compare WTP and parameter estimates from experiment 1 (self) to experiment 10 (friend), to consider whether women’s responses depend on the framing of the question, and how they consider partnership choices for themselves, compared to what they think is best for a close third party. Specifically, we will estimate WTP for experiment 10 exactly as following calculations laid out above, however we will present this separately to examine how women consider partnership choices for themselves versus friends. We additionally include an incentivised component in the experiment, asking women what they believe an average person in the survey population will respond to a specific profile set which they have observed, with a small monetary incentive for responses closest to the true observed mean ratings. We will use these ratings to determine whether respondents correctly perceive social norms relating to violence, and whether individuals who believe that violent traits are less important are more willing to accept violent relationships.
Anderson, Michael L. 2008. “Multiple Inference and Gender Differences in the Effects of Early Intervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects.” Journal of the American Statistical Association 103 (484): 1481–95.
Wiswall, Matthew, and Basit Zafar. 2018. “Preference for the Workplace, Investment in Human Capital, and Gender.” The Quarterly Journal of Economics 133 (1): 457–507. https://doi.org/10.1093/qje/qjx035.