The Wrong Politics for the Right Neighborhood? Political Cues, Rental Housing, and Spatial Sorting

We conduct a between-listing correspondence experiment in Bogotá’s rental housing market during the 2026 Colombian presidential election. Each listing receives one WhatsApp inquiry from a fictitious applicant. The main treatment is political affiliation, randomized through the applicant’s WhatsApp profile picture, which signals support for the left-wing candidate, support for the right-wing candidate, or nothing (no photo, the control). We independently randomize the applicant’s name, accent (a WhatsApp voice note), a neighborhood-belonging cue, and gender; name and accent signal perceived class and social origin, separating political from class-based discrimination. The primary outcome is whether the landlord responds within seven days confirming that the unit is available. Secondary outcomes include the speed, quality, and length of the landlord’s reply, and the number of messages needed to secure a viewing.

Political affiliation conditions. In Wave 1 (before the first-round election on May 31, 2026), four conditions are active: the main right-wing establishment candidate (Candidate 1), a right-wing outsider candidate (Candidate 2), the main left-wing candidate (Candidate 3), and a no-photo control. These were the three candidates most likely to reach the runoff. Signaling both right-wing profiles ensures that one condition matches the eventual right-wing finalist regardless of the first-round outcome. Because two of the four conditions are right-wing, we assign probabilities that equalize the Left, Right, and Control cells in expectation: Candidate 3 receives probability 1/3, Candidates 1 and 2 each receive 1/6, and the no-photo control receives 1/3. In Wave 2 (June 1 to August 21, 2026), the three remaining conditions (the two finalists and the no-photo control) are each assigned with probability 1/3. The no-photo condition uses WhatsApp’s default blank profile picture. The runoff takes place on June 21, 2026, in the middle of Wave 2, creating two sub-periods: one between the first round and the runoff, and one after the runoff result is known.
Other treatment dimensions. Four additional attributes are randomized independently with equal probability (50/50): name class (high vs. low, signaled through first and last names), accent class (high vs. low, via a voice note), neighborhood belonging (prior connection vs. interest only), and applicant gender (female vs. male).
Validation of class signals. First and last names were selected and validated in a two-step procedure. In the first step, we used administrative data (the Gran Encuesta Integrada de Hogares, GEIH, and school enrollment records) to identify name combinations empirically associated with different socioeconomic strata in Colombia. Second, respondents in a dedicated pre-experiment survey in Bogotá rated the perceived social class of each full name; we retain only the names respondents most consistently classified into the intended category. Accents were validated in parallel: natural voice actors recorded otherwise identical rental inquiries in higher- and lower-class speech styles, and the same survey respondents rated the perceived social class of each recording. Only the recordings most consistently classified into the intended category are used in the experiment.
Outcome coding. The primary outcome is effective response: a binary indicator equal to 1 if the landlord directs the applicant toward the property (confirms availability, offers a visit, or provides follow-up contact) within seven days of the inquiry. Secondary outcomes are: any response received (a descriptive check); quality of response (coded by human coders on a structured rubric and cross-validated with an AI tool, with human-coder fixed effects); word count of the landlord’s reply; response time, measured both as elapsed hours from the applicant’s first message to the landlord’s first reply and as categorical thresholds set blind to treatment; and the number of exchanges required to obtain an effective response (tenant search burden).
Neighborhood political context. Each property is matched to 2026 presidential electoral returns at four spatial scales: locality (localidad), urban planning zone (UPZ), Google Maps neighborhood boundary, and nearest polling place (by Euclidean distance). At each scale, the local political orientation is defined as either a binary winner-takes-all classification (right vs. left, based on the second-round outcome) or a continuous vote-share gap between right-wing and left-wing candidates. We treat the full set of spatial scales and both orientation definitions as a pre-specified family, reporting estimates across all of them. The binary measure at the nearest polling place serves as the benchmark for power calculations.
Neighborhood amenity context. We define high-amenity neighborhoods as those where renters are willing to pay a high premium above what observable characteristics alone would predict. We measure this premium through neighborhood fixed effects from a hedonic housing price model estimated on the full universe of rental listings scraped prior to the experiment, controlling for property characteristics, time effects, access to employment and transport, and neighborhood quality indicators including socioeconomic stratum, proximity to parks, transit, commercial activity, schools, health services, cultural amenities, public safety, and environmental quality. Amenity status is defined both as a binary high-versus-low classification and as a continuous standardized index. We report heterogeneous treatment effects across this measure and alternative spatial scales, treating them as a pre-specified robustness family.
Pooling and wave structure. Our main analysis pools applications from both waves, with wave fixed effects in all specifications. We also estimate heterogeneity across three pre-specified electoral stages: (i) pre-first-round (May 4–30), (ii) between the first round and the runoff (June 1–20), and (iii) post-runoff (June 22–August 21).
Multiple hypothesis correction. Two outcome families are pre-specified: the set of main outcomes (effective response, quality, length, response time, and number of interactions to effective response) and the set of political signal × neighborhood interaction terms. We apply corrections for multiple testing within each family.

2026-05-04

2026-08-21

Effective response: a binary indicator equal to 1 if, within seven days, the landlord or agent directs the applicant toward the property by confirming availability, offering to schedule a visit, or providing contact information for follow-up.

1. Any response received (binary; expected near-ceiling, analyzed as a descriptive check).
2. Quality of response: effort and completeness of the landlord’s engagement, human-coded on a structured rubric.
3. Length of response: word count across all landlord messages in the thread.
4. Response time: elapsed hours from the applicant’s first message to the landlord’s first reply, plus a categorical version with thresholds set blind to treatment.
5. Number of interactions required to obtain an effective response.

We implement a between-listing correspondence design. Each rental listing receives exactly one WhatsApp application from a fictitious applicant whose profile picture is randomized across the political affiliation conditions; the applicant’s name, accent, neighborhood belonging, and gender are independently randomized. The sample is rental listings from the two major online housing platforms in Bogotá, restricted to estratos 3 through 6. Data collection spans the 2026 Colombian presidential first round (May 31) and runoff (June 21), letting us study whether political discrimination varies with electoral salience.

Listing selection and sampling. Rental listings are scraped daily from the two main online housing platforms in Bogotá, restricted to properties located in the city and classified in estratos 3 through 6. The contact pool is further restricted to listings posted within the previous seven days: fresh listings are less likely to be already rented, and inquiries about them are indistinguishable from ordinary tenant search, reducing the risk of detection. Each day, a batch of 60 listings is drawn at random from this pool, half from lower-estrato properties (estratos 3–4) and half from higher-estrato properties (estratos 5–6), and scheduled for contact. Before assignment, each batch is de-duplicated against every listing and phone number contacted earlier in the study, so no property or landlord ever receives more than one inquiry. Treatment assignment is blocked by estrato so that the political conditions are balanced within each group.
One-contact rule and detection minimization. Each listing is contacted exactly once, and no landlord is contacted twice: once a phone number has been contacted it is blacklisted for the remainder of the study, even if it posts new listings. Inquiry messages are written in multiple equivalent phrasings, and the order of the message components after the greeting and open inquiry (voice note, name sign-off, belonging cue) is randomized independently for each application. These measures preserve ecological validity while minimizing the risk that landlords or agents recognize the experiment.
Device and account structure. Political affiliation is signaled through the profile picture of the WhatsApp account used to send the inquiry. The experiment uses multiple WhatsApp sending accounts, each with a fixed profile picture corresponding to one political condition. At the time of application assignment, a device is drawn at random from the pool carrying the profile picture that corresponds to the assigned political condition.
Message structure. Each inquiry has five components: (i) a greeting appropriate to the time of day; (ii) an open-ended message describing the type of apartment sought and the general area of the listing; (iii) a voice note in the assigned accent; (iv) a message with the applicant’s name; and (v) a message expressing interest in the neighborhood. Applications assigned to the neighborhood-belonging condition signal a prior connection to the neighborhood in component (v); the others express interest only. Each component has multiple phrasings, so no two applications are identical in wording. Components (iii) to (v) are sent in randomized order.
Follow-up protocol. If no effective response is received within 24 hours of the initial inquiry, a single follow-up message is sent reaffirming interest without adding new signals. A closing message thanking the landlord and declining further interest is sent at 96 hours regardless of response status. No new applicant signals are introduced in follow-up messages. Landlord messages arriving after the closing message are still recorded and count toward the seven-day outcome window.
Data collection and coding. All WhatsApp interactions are downloaded and coded daily, before the next sending session begins. This ensures no conversation is lost if a device is suspended or replaced. Outcomes are coded from the full interaction thread. Response quality is coded by human coders blinded to the political treatment condition, using a pre-specified rubric; coding is cross-validated with an AI tool.
Wave transition. Wave 1 runs from May 4 to May 30, 2026 (four sending weeks). Wave 2 begins on June 1, 2026, after the first-round result is known. The eliminated candidate’s devices are retired and removed from the active pool. Wave 2 runs through August 21, 2026 (approximately eleven sending weeks), spanning the pre-runoff period, the month following the June 21 runoff, and the two weeks following the August 7 inauguration.
Estimation. We estimate linear probability models with the interaction of each political signal and the political lean of the neighborhood. The primary quantity of interest is whether the penalty a left-wing applicant faces is concentrated in right-leaning neighborhoods. The right signal plays a distinct role: a penalty on the left with little or none on the right indicates discrimination directed against the left, rather than a generic reaction to any political image. We also report average effects of each signal across the city. A separate pre-specified specification interacts each political signal with the neighborhood quality index instead of political lean, to examine whether the heterogeneity tracks political lean, amenity quality, or both. All specifications include controls for name class, accent class, neighborhood belonging, gender, wave fixed effects, and a platform indicator. Standard errors are robust; we also report results clustered at the neighborhood level and with spatial standard errors (Conley 1999) as robustness checks.
Pooling across waves. The cross-wave specification pools the two right-wing candidates into a single right category. A Wave 1 specification enters each candidate separately.
Heterogeneity. Beyond the neighborhood interaction, we report heterogeneity by counterparty type (owner vs. agency), property stratum, electoral stage, neighborhood belonging, and gender. We also run an exploratory Causal Forest (Wager and Athey 2018) to detect heterogeneity without pre-committing to subgroups.
Neighborhood-level estimates.
To describe where the political penalty is largest, we estimate shrunken neighborhood-level treatment effects with an empirical Bayes, partial-pooling model in the spirit of Kline, Rose, and Walters (2022). The model lets the political-signal effect vary across neighborhoods and pulls each estimate toward the value predicted by its political lean and other covariates. Neighborhoods with few listings are shrunk toward the model prediction rather than reported as noise. We estimate these effects at the nearest-polling-place level, with UPZ and locality (localidad) as coarser cross-checks, and map the resulting posterior estimates.
Planned follow-up (separate pre-analysis plan). After a washout period following the end of data collection, we may recontact a subset of landlords with a short survey that elicits their political preferences and sociodemographic characteristics, to explore heterogeneity in the political-signal effect along these dimensions. This follow-up is not part of the present analysis: it will be governed by a separate pre-analysis plan and may require separate IRB review. We record it here only to note that the possibility was anticipated ex ante.

Randomization is performed by computer at the time of daily application assignment. Daily batches are drawn half from lower-estrato properties (estratos 3–4) and half from higher-estrato properties (estratos 5–6), and assignment is stratified so that the political conditions are balanced within each group.
Political affiliation. In Wave 1 (pre-first-round), Candidate 3 (Left) is assigned with probability 1/3; Candidates 1 and 2 (Right) are each assigned with probability 1/6; the no-photo control is assigned with probability 1/3. This weighting equalizes the Left, Right, and Control cells in expectation, because two of the four conditions are right-wing. In Wave 2 (post-first-round), the two finalists and the no-photo control are each assigned with probability 1/3.
Other dimensions. Name class (high vs. low), accent class (high vs. low), neighborhood belonging (prior connection vs. interest only), and applicant gender (female vs. male) are each drawn with equal probability (50/50), independently of the political affiliation draw and of each other.

The rental listing. Each listing receives exactly one application, and all treatment dimensions are randomized at the listing level.

No

Between 2,000 and 5,000 rental listings.

Between 2,000 and 5,000 applications (one per listing, since each listing is contacted exactly once).

Wave 1 (between 700 and 1,000 listings): Left condition 1/3; each right-wing candidate 1/6; no-photo control 1/3.
Wave 2 (between 1,300 and 4,000 listings): each of the two finalist conditions 1/3; no-photo control 1/3.
Pooled across both waves: left cell approximately 1/3; right cell approximately 1/3; no-photo control approximately 1/3.

Power is assessed by Monte Carlo simulation calibrated to preliminary data: the baseline (no-photo control) effective-response probability is about 0.56, and 66.5 percent of geolocated preliminary listings fall in right-leaning neighborhoods, the share over which the primary estimand is identified. Because the sample is bounded by the rental market, we fix the sample size and report minimum detectable effects at 80 percent power and a 5 percent two-sided test, at the expected sample of about 3,000 listings and at 4,000 and 5,000 should the market allow a larger sample. At the expected 3,000 listings, the minimum detectable effect is 6.7 percentage points for the average left penalty and 7.9 percentage points for the primary quantity, the left-signal penalty in right-leaning neighborhoods. At 4,000 and 5,000 listings the primary minimum detectable effect falls to 7.0 and 5.9 percentage points (6.0 and 5.3 for the average penalty). The interaction—whether the left penalty in right-leaning neighborhoods differs from the one in left-leaning neighborhoods—has a minimum detectable effect of about 14 percentage points at 3,000 listings, falling to about 10 at 5,000.