Men's Perspectives on Female Employment in India: A Discrete Choice Experiment

NA

2025-07-15

2025-09-15

Impact of the four job attributes ((1) gender composition of the workplace; (2) safety of commuting options; (3) nature of work; (4) expected salary) on the probability of choosing a specific job

NA

NA

We conduct a discrete choice experiment (DCE) to elicit husbands' preferences on different types of hypothetical job offers to their wives. DCEs are a survey method used to measure the relative importance of different characteristics that respondents weigh against each other when making certain choices (here: accepting wives' employment). DCEs have benefits over other stated preference techniques (e.g., ranking or rating exercises) because they mimic real-world choice scenarios while reducing cognitive complexity for respondents.

The DCE will be constructed as an unlabeled design with two alternatives per choice set. Specifically, we present ten different choice sets, asking respondents to choose between two hypothetical job types (Job A vs. Job B) for their wife and also offering an opt-out option, i.e., refusing both Job A and Job B. In each choice set, job descriptions vary along the following four attributes and levels:

1. Gender composition of the workplace:
- predominantly female
- predominantly male
- mixed

2. Safety of commuting options:
- walking to the workplace
- public transport to the workplace
- long-distance/overnight travel

3. Nature of work
- agriculture
- construction
- office-based (e.g., nurse, teacher, engineer…)
- service- base (e.g., house help, cleaning, sales, factory…)

4. Expected salary
- higher than husband's own income
- equal to husband's own income
- lower than husband's own income

The selection of specific combination of job profiles in the ten choice sets was implemented in the software Ngene. The experimental design that will be used was determined as a D-efficient design based on weak priors for the main effects of the above attributes (without interactions). Statistical efficiency was measured by the D-optimality criterion (D-error), the most widely used metric in this regard. D-optimal or D-efficient designs minimise the determinant of the asymptotic variance-covariance matrix, ensuring minimum variation around the parameter estimates.

Main analysis:
We will firrst employ a conditional logit model (McFadden et al., 1973) for repeated choices to analyse the attributes’ main effects on respondents’ preferences for different job types. Assuming the random utility component to follow an independently and identically distributed
type 1 extreme-value distribution, the conditional logit allows us to estimate the probability of choosing Job A as a function of its attribute levels and the attribute levels of Job B (and vice versa). In addition to estimating the conditional logit to analyse respondents’ choices, we will consider employing a mixed (random parameter) logit, given that it relaxes the independence of irrelevant alternatives (IIA) assumption and removes limitations with regard to unobserved preference heterogeneity. The influence of the four attributes on respondents’ probability of selecting Job A or Job B will then be interpreted using odds ratios. We will further assess predicted probabilities for job selection on each attribute level. Based on the weak priors employed in the experimental design, we expect the following directions for the respective attributes:
1. Respondents will be more likely to opt for a job with predominantly female colleagues, followed by mixed workspaces, with predominantly male workspaces being the least preferred option.
2. Respondents will prefer travel by foot over travel by public transport over overnight travel.
3. Respondents will reveal a preference for service- and office-based jobs over agricultural jobs. Construction work will be the least preferred option.
4. Respondents will have a higher preference for women earning less than their own income (in view of status threats), followed by equal income levels and then more than their own income.

In addition to the main effects, we will assess whether certain individual-level characteristics predict differences in the importance of each attribute for respondents’ choices. To this end, we will extend the conditional logit model by adding individual-level predictors of a potential
heterogeneity in attribute-based prioritization preferences. This hybrid conditional logit model will allow us to assess systematic heterogeneity in preferences arising from observed socioeconomic and other individual-level characteristics by interacting the latter with a certain
attribute level.

For each respondent, the sequence of the ten choice sets will be randomised to account for possible ordering effects, possible learning effects, and respondent fatigue.

Individuals

No

NA

2250 male respondents

Power calculations based on the procedure proposed by Bekker-Grob et al. (2015) indicate that, with 80% statistical power and an alpha level of 5%, and priors ranging between 0.1 and 0.2, we will be able to detect the main parameter effects with a sample of N = 650. The target sample site exceeds this number and would therefore also account for possibly smaller priors.