When Roommates Misreport: Envy-Free Rent That Survives Self-Report Bias

Fair rent-splitting tools ask each housemate to state, in dollars, how much each room is worth to them. The tool then assigns rooms and sets prices so that nobody would rather swap: no one prefers another person's room at that person's price. This property is called envy-freeness. The guarantee holds only with respect to the numbers people typed in. Survey research has shown for decades that such numbers are not honest-but-noisy draws around the truth. They carry structured distortions. People anchor on defaults, compress their answers toward the middle of the scale, and shade answers toward how they would like to appear. This proposal asks a narrow question: which kinds of reporting distortion actually break envy-freeness in rent division, and can we build a pricing rule that spends its safety margin on those kinds and no others?

The central observation is simple. Envy between housemates depends only on the differences between how much a person values one room versus another. So any bias that shifts all of one person's numbers up or down by the same amount cannot cause envy. A bias that shrinks a person's stated differences can, and it does so in a predictable direction. It endangers exactly those people who were given a room they like less, in exchange for a discount sized to their understated preference. That points to a specific design: estimate the shape of reporting bias from an experiment where true values are known, then protect envy-freeness against that shape rather than against generic noise.

Where the gap is

The deployed rule on Spliddit.org is the maximin rule of Which is the Fairest (Rent Division) of Them All?. It first picks the room assignment that maximizes total stated value. It then solves a linear program for the envy-free prices that make the worst-off housemate as well off as possible. That paper motivates this proposal in two ways. It shows on 13,277 real instances that the choice among envy-free prices matters to people. It also takes the reported values as accurate.

Robust Rent Division is the closest prior work, and it constrains what counts as new here. Peters, Procaccia and Zhu show that maximin leaves housemates on a knife-edge, where a small change in true value creates envy. They propose two fixes. Lexislack lexicographically maximizes the sorted pairwise envy margins, which gives the allocation that stays envy-free over the largest ℓ1 ball around the reports. A second, sample-based method maximizes the probability of envy-freeness, or minimizes expected envy, under a noise distribution. In their tests that distribution is Uniform, Normal, or "Biased Normal" synthetic noise added to 1,000 Spliddit instances. Their uncertainty model is generic and chosen by the analyst, and they do not ask what human reporting error actually looks like.

That is the gap this proposal fills. It is not "add noise to rent division," which Peters et al. already did. It is to measure the structure of real reporting error, prove which parts of that structure envy-freeness is sensitive to, and match the robustness set to it.

Fair Rent Division on a Budget supplies further motivation. It notes that asking for exact room values plus a budget raises the cognitive load of reporting, and so the error.

The self-report literature in the file comes from health and psychology, far from markets, so its role is to motivate rather than to supply methods. Studies of dietary logs (Yuan et al.) needed objective biomarkers before self-reports could be calibrated. Measurement-invariance work (Yue et al.) and nurse self-assessment studies (work readiness survey) document directional, image-driven inflation rather than zero-mean error. Taken together, these papers set the methodological lesson used here: bias can only be estimated where ground truth exists. That is why the proposal is built around an induced-value experiment rather than raw Spliddit logs, which contain no truth.

Core Claim

We model a housemate's report for room $k$ as $\hat v_{ik} = a_i + \alpha_i v_{ik} + \epsilon_{ik}$. Here $v_{ik}$ is the true value, $a_i$ is a person-level offset (anchoring on the equal split, or on the rent total that the interface forces the reports to sum to), $\alpha_i$ is a person-level scale (below 1 means compression toward the middle), and $\epsilon_{ik}$ is residual noise.

Under this model, the true envy margin of housemate $i$ toward housemate $j$ is

$$s_{ij} = \frac{\hat v_{i}(r_i) - \hat v_{i}(r_j) - (\epsilon_{i r_i} - \epsilon_{i r_j})}{\alpha_i} - (p_i - p_j),$$

where $r_i$ is $i$'s room and $p_i$ is $i$'s price. The offset $a_i$ has dropped out entirely.

When $\alpha_i < 1$, the true preference gap is larger than the reported one. A housemate who received their reportedly better room gains margin. A housemate compensated with a discount for a reportedly worse room loses margin, because the discount was computed from an understated gap.

Lexislack's ℓ1 ball spends slack equally on all pairs. Our claim is that a rule whose uncertainty set is aligned with the measured $(\alpha, \epsilon)$ structure will achieve a higher envy-free rate under true values, at equal or better worst-off utility. The rule recenters on the debiased estimate and allocates extra slack to the discounted-room pairs that compression makes fragile. A second, falsifiable claim is that offset biases, which may be the most visible distortion in reports, are harmless. Any method that spends robustness on them wastes it.

Research Program

The work has three parts.

Analysis. We will formalize the invariance and sensitivity results above for the general affine-plus-noise report model. This includes deriving which envy constraints become binding under compression versus expansion. We will also check whether the Spliddit constraint that reports sum to the rent couples $a_i$ with $\alpha_i$ in a way that matters.

Measurement. We will run an induced-value elicitation study. Participants in groups of three to five are shown apartment rooms described by attributes. Each attribute carries a hidden payoff that defines their true value, and payment depends on the realized true utility of their assigned room and price. Participants then report through a Spliddit-style interface.

Two manipulations separate the bias mechanisms. A default-anchor condition pre-fills the equal split, against a blank form. A repeated-elicitation condition separates stable person-level scale $\alpha_i$ from trial-to-trial noise $\epsilon_{ik}$, which speaks to the self-report literature's open problem of conflating random recall error with systematic distortion. We will fit the hierarchical affine model and report the population distribution of $\alpha_i$ and $a_i$.

Mechanism. We will build a bias-aware rule with two parts. It first estimates the debiased center from the fitted population model. It then runs the expected-envy-minimization program of Peters et al. over samples drawn from the fitted structured model, with correlated, person-level scale draws, rather than independent per-entry noise. We reuse their integer-programming formulation and their sample-complexity reasoning, under which roughly $O(n^2 \log n)$ samples suffice. That reasoning depends on the complexity of the allocation class, not the noise shape, so it should carry over. The new parts are the noise model and the recentering.

Baselines

The naive reference point is the maximin rule run directly on reports, as deployed on Spliddit and described in Which is the Fairest (Rent Division) of Them All?. Its reported advantage is in disparity: it improved the gap between most- and least-satisfied housemates by "9% of the total rent" over earlier methods, and users rated it more than a full point higher on a 1–5 scale. Any robust rule must not give back much of that disparity advantage. Beating maximin means a higher true-value envy-free rate with disparity, as a share of rent, no more than a small tolerance worse.

The strongest existing competitors are the two rules from Robust Rent Division. The file reports only qualitatively that, at noise level 0.05, maximin's envy-free probability "drops sharply while robust rules maintain much higher" rates. No exact envy-free-rate figure was found on Lacuna for Lexislack or the distributional rules, so we will reimplement both. Lexislack is run on the same reports, with no tuning. The distributional expected-envy rule is run with each of Uniform, Normal and Biased Normal noise, with the noise scale tuned on a held-out portion of the induced-value data. That gives it the same calibration data our method gets, so the comparison is about the shape of the noise model and not about access to data. All rules receive the same instances, the same number of noise samples, and the same solver time limit.

Debias-then-maximin is a plug-in baseline. It rescales each report by the fitted $\alpha$ and runs maximin, isolating the value of correcting the center without adding robustness.

Two ablations isolate the key ingredient. The first uses our recentered rule with an isotropic uncertainty set of matched total variance, which tests whether the anisotropic, compression-aligned structure matters. The second uses our structured set without recentering.

Beating the strongest baseline means a statistically significant gain in true-value envy-free rate and expected true envy over the best-tuned Peters et al. rule, on held-out induced-value groups, at no worse worst-off true utility.

Why not just…?

Why not just use Lexislack, since its ℓ1 ball already covers small biases of any shape? It covers them only up to a fixed radius, and it spends that radius equally on all pairs, including pairs our analysis says compression makes safer. If measured bias is mostly idiosyncratic noise with $\alpha_i$ near 1, Lexislack should match our rule, and we will report that outcome as a finding that structure does not matter. The isotropic-set ablation and the head-to-head on true values settle this directly. This is a real threat, and the measurement study decides it.

Why not just run Peters et al.'s distributional method with their Biased Normal noise? Their noise is a synthetic model chosen without reference to human data. Our hypothesis is that real error is correlated within a person, through a shared scale, in a way that independent per-entry noise cannot represent even when its variance is tuned. Giving their method the same calibration data and a tuned scale makes the comparison fair. If it ties, then noise shape is irrelevant and only variance matters, which would be a useful negative result.

Why not just debias the reports and run ordinary maximin? That corrects the center but puts the allocation back on maximin's knife-edge, now sitting on an estimated value that has its own error. Peters et al.'s brittleness result suggests this will fail at modest error levels. The debias-then-maximin baseline and the no-recentering ablation separate the two contributions. If debiasing alone captures most of the gain, the proposal reduces to "calibrate, then use any robust rule," and we will say so.

Proposed Experiments

The primary test uses held-out induced-value groups. For each rule we compute, under true values, the envy-free rate, expected envy, worst-off utility, and disparity as a share of rent, which is the Gal et al. metric. The decisive check is the interaction predicted by the analysis. Under the fitted compression model, maximin's and Lexislack's true-envy violations should concentrate among housemates who received discounted, less-preferred rooms, and our rule should remove that concentration. If violations are spread evenly across pairs, the core mechanism is wrong.

The second test transfers the fitted bias distribution onto the 1,000-instance Spliddit sample used in Robust Rent Division. Treating the stated profiles as true values, we generate biased reports from the fitted model and rerun all rules. This checks whether the gains hold on real preference structures rather than only on our lab-designed ones. The default-anchor manipulation adds a falsification test: it should move $a_i$ substantially while leaving envy outcomes unchanged, confirming the invariance prediction.

Finally, as a limited probe of the strategic-misreporting question that both Peters et al. and the budget paper leave open, we will compute best-response gains for a single misreporting housemate under each rule on the induced-value profiles. The aim is to see whether recentering on a population bias model creates new manipulation opportunities, without claiming incentive compatibility.