Witchcraft as a Conflict-Mitigating Institution: A Randomized Vignette Experiment in Monrovia, Liberia

Respondents complete a household survey containing an embedded randomized vignette experiment.

Each respondent evaluates three hypothetical third-person scenarios in which a stranger enters a victim's home while the victim is away, steals money, and is later identified. Because the thief did not know whose house it was, the victim's attributes (demographics) are orthogonal to the offense by construction.

After the victim identifies the thief, we randomize, within respondent and using the survey software at form load: (i) the wrongdoer's community reputation for consulting traditional healers (our observable measure of engagement with witchcraft and access to supernatural recourse, i.e., known vs. not known to consult healers when facing problems); and (ii) the confidant the victim turns to (no one; a social third party (a friend, community leader, or pastor/priest/imam); or a traditional healer). The amount stolen and both actors' demographic attributes are also randomized per scenario.

After each scenario, respondents rate the likelihood of a range of reactions by the victim, including confrontation (arguing, threatening, fighting) and worry about physical versus supernatural harm. The experiment tests whether belief in witchcraft mitigates interpersonal conflict, through deterrence (a victim refrains from confronting a wrongdoer believed to have supernatural recourse) and substitution (a victim who takes the grievance to a healer no longer needs worldly confrontation).

2026-07-08

2026-08-31

Two primary outcomes, both rated for the victim (Actor 1) on a 0 to 3 likelihood scale: Supernatural-fear differential: the difference between the victim's rated worry that the wrongdoer may harm them through spiritual/supernatural means and their worry about physical harm (spiritual-harm worry minus physical-harm worry). This is the primary, identifying outcome for the deterrence hypothesis (H1). Confrontation index: an equally weighted standardized index of the likelihood that the victim argues face-to-face with, threatens, and physically confronts or fights the wrongdoer. This is the primary outcome for the substitution hypothesis (H2) and a secondary (corroborating) outcome for the deterrence hypothesis (H1).

The confrontation index is an equally weighted average of its three component items (argue face-to-face, threaten, physically confront or fight), each standardized (z-scored) using the mean and standard deviation of the full analysis sample. Each component is rated on a 0 to 3 likelihood scale (very unlikely to very likely); "don't know" responses are set to missing, and the index is the average of the non-missing standardized components. The same index construction is used for both the deterrence and substitution tests.
The supernatural-fear differential is the respondent's spiritual-harm worry rating minus their physical-harm worry rating for the same scenario (each on the 0 to 3 scale), defined only when both component items are non-missing. A positive treatment effect on this differential (i.e., the wrongdoer's healer reputation raising spiritual-harm worry relative to physical-harm worry) is the signature of fear of supernatural retaliation and distinguishes deterrence from a general "dangerous person" reaction, which would raise both worries equally.

All rated for the victim (Actor 1) on the 0 to 3 likelihood scale unless noted:
1. Individual components of the confrontation index, reported separately: likelihood the victim argues face-to-face with, threatens, and physically confronts or fights the wrongdoer.
2. The two components of the supernatural-fear differential reported separately: worry about spiritual/supernatural harm, and worry about physical harm.
3. Victim avoidance: likelihood the victim avoids the wrongdoer (corroborating outcome for substitution).
4. Victim gives up: likelihood the victim gives up trying to recover the money. Used as a rival-detector for substitution. A positive treatment effect here would indicate resignation rather than the redirection of the grievance into supernatural channels, and is interpreted as a warning rather than as support for the hypothesis.
5. Exploratory, without a directional prediction: likelihood the victim reports the matter to the police, and likelihood the victim feels angry.
6. Confidant contrast: the effect of the victim consulting a healer relative to talking to no one (the healer-vs-social-third-party contrast is the primary substitution test; the healer-vs-no-one contrast is reported as secondary).
7. Exploratory (offender-side, non-confirmatory): the survey also elicits ratings of the wrongdoer's likely reactions when informed of the victim's response (e.g., returning the money, apologizing, worry about supernatural harm). These are collected but are not part of the confirmatory analysis and carry no pre-specified predictions; any analysis of them will be labeled exploratory.

The individual confrontation components and the two harm-worry items are the disaggregated pieces of the primary outcomes, reported to show which specific reactions drive any index effect. Victim avoidance and giving up on the money are additional victim reactions: avoidance corroborates a reduction in direct confrontation, while giving up distinguishes genuine substitution (the victim still pursues the grievance, by supernatural means) from mere resignation. Police reporting and anger are reported without a signed prediction, since the direction each would move under deterrence is theoretically ambiguous. All secondary outcomes are analyzed with the same specification as the primary outcomes (respondent and scenario-position fixed effects; standard errors clustered by respondent).

The study is a household survey of approximately 600 adults across thirty communities in Greater Monrovia, Liberia, containing an embedded within-respondent randomized vignette experiment. Each respondent evaluates three hypothetical scenarios involving two fictional people, a victim and a wrongdoer, in which the wrongdoer steals money from the victim's home and is later identified.
Two treatments are randomized within respondent. First, the wrongdoer's community reputation for consulting traditional healers (known vs. not known). This is our measure of engagement with witchcraft and access to supernatural recourse. It is varied so that both levels appear across each respondent's three scenarios. Second, the confidant the victim turns to after identifying the wrongdoer is blocked so that each respondent sees all three conditions exactly once: talking to no one, to a social third party (a friend, community leader, or pastor/priest/imam), and to a traditional healer. The order of the three scenarios is randomized. Additional scenario features, including the amount stolen and both actors' gender, age, education, employment, and religiosity, are randomized independently in each scenario. All randomization is performed by the survey software on the data-collection tablet and stored with each submission, so that assignment is fully recoverable.
The design tests two mechanisms by which belief in witchcraft may mitigate interpersonal conflict. Under deterrence, a victim refrains from confronting a wrongdoer believed to have supernatural recourse, for fear of supernatural retaliation; this is identified from the wrongdoer's healer-reputation treatment. Under substitution, a victim who takes the grievance to a traditional healer no longer needs worldly confrontation; this is identified from the confidant treatment, comparing the healer condition to a social third party who holds constant the act of confiding while lacking the supernatural channel.

Analysis is at the respondent-scenario level, with respondent and scenario-position fixed effects and standard errors clustered by respondent.

Not available

Randomization was performed by the data-collection software (ODK Collect) on the enumerator's tablet at the start of each interview, using the software's built-in pseudo-random number generator. All treatment assignments: the wrongdoer's healer reputation, the victim's confidant, the scenario order, the amount stolen, and both actors' demographic attributes, were drawn on the device and stored with the submission, so that realized assignments are fully recoverable for analysis. Within-respondent blocking (each respondent seeing all three confidant conditions and both healer-reputation levels) was implemented in the survey logic. Separately, the thirty study communities were selected in advance by probability-proportional-to-size sampling using a fixed random seed in statistical software.

The unit of randomization is the respondent-scenario: each respondent evaluates three scenarios, and treatments are assigned to each of a respondent's three scenarios. Both treatments of interest are randomized within respondent: the victim's confidant is blocked so that each respondent sees all three conditions (no one, social third party, healer) exactly once, and the wrongdoer's healer reputation is blocked so that both levels appear across a respondent's three scenarios. The amount stolen and the two actors' demographic attributes are randomized independently at the respondent-scenario level. There is no group- or cluster-level treatment assignment: all treatment variation occurs within respondent, across scenarios.
Separately, at the sampling stage, the thirty study communities were selected by probability-proportional-to-size sampling; this is a sampling procedure, not treatment assignment.

No

600 respondents

1800 = 600 * 3 scenarios

600 respondents