Pricing a Resignation by the Departures It Causes

When a nurse resigns, the hospital pays more than the cost of hiring her replacement. Until the post is filled, her shifts are spread across the colleagues who stay. Heavier workload is among the best-documented drivers of burnout, and burnout is among the best-documented drivers of the next resignation. Current attrition-prediction models ignore this chain. They score each employee's individual risk of leaving from a static table of attributes. They are judged on classification accuracy, mostly on one synthetic dataset. None of them asks which departures would be most expensive once the follow-on departures they trigger are counted. We propose to estimate that quantity directly. We call it the cascade-adjusted cost of a departure: the replacement cost of the person who leaves, plus the expected replacement cost of the extra departures their absence causes among teammates. We then test whether ranking employees by this cost, rather than by individual risk, directs a fixed retention budget toward larger savings.

The model is deliberately small and mechanistic. A departure can raise teammates' chance of leaving through two channels. The first is workload: the person's duties are redistributed for as long as the vacancy stays open. The second is social: people leave because someone they know left. These two channels imply different remedies. Fast backfilling, for example with temporary staff, addresses the first. Retention conversations address the second. They also make a sharp, testable prediction that separates them: only the workload channel should depend on how long the vacancy stays unfilled. Single-crystal X-ray diffraction enters as an evaluation discipline, not as a metaphor. Crystallographers fit a few physically meaningful parameters to indirect measurements. They report a standard uncertainty on every parameter, as in a cell length written as 21.874(7) Å. They judge the fit by residual indices, the summary numbers that measure how far the model's predictions sit from the observed data. We adopt the same practice. The field needs it: the attrition-ML literature reports the same algorithm on the same dataset with results that cannot be reconciled.

Core Claim

The economic harm of attrition is concentrated in departures that open long vacancies in overloaded teams. A two-channel hazard model can identify those departures before they happen. Ranking employees by the cost such a model predicts should capture substantially more realized attrition cost under a fixed intervention budget than three alternatives: ranking by individual risk, ranking by individual risk multiplied by replacement cost, or ranking by a strong network turnover predictor that does not separate the channels.

Grounding in Prior Work

Two strands of survey research motivate the workload channel. Wong et al. study 502 Hong Kong public-hospital nurses. They find that burnout mediates the path from job demands to turnover intention, and that a unit increase in work overload raises burnout by 0.590. They also find that pay satisfaction does not weaken this pathway. Lee et al. use a multilevel model of 558 real-estate agents nested in 73 offices and find burnout positively predicts turnover intention (coefficient 0.568). Both studies are cross-sectional and measure intention rather than actual departure, so neither can show the dynamic feedback we model. That feedback is a departure that raises overload, which in turn produces the next departure. Fessehatsion adds a useful warning: collegial support had no detectable effect among Eritrean teachers. The social channel may therefore be context-dependent, which is one reason we estimate the two channels separately rather than assume both exist.

The social channel is motivated by Teng et al.. In a longitudinal dataset of about 3,000 employees at a high-tech firm, they report that 91% of resignations followed earlier departures by managers or peers. Their contagion-aware neural network (CEHNN) reaches 0.914 AUC-ROC. AUC-ROC is the probability that the model ranks a randomly chosen leaver above a randomly chosen stayer. This is the strongest existing method on our question, and we reuse its central insight that departures are not independent events. We differ in three ways. Their model predicts who will leave, while ours attributes induced departures to a specific earlier departure. Theirs measures closeness through interaction counts, while ours adds vacancy-driven workload as an explicit, separately estimated channel. Theirs is scored on accuracy, while ours is scored on cost.

The individual-risk tabular line constrains our design more than it motivates it. A Comparison of Machine Learning Approaches for Predicting Employee Attrition reports logistic regression at 85.01% AUC-ROC on the synthetic IBM HR dataset. Retention is All You Need reports XGBoost at 89.12% accuracy and uses explanation-based "what-if" simulations. The transformer study reports a tabular transformer at 0.7452 AUC-ROC and XGBoost at only 0.6542 AUC-ROC on nominally the same 1,470 records. Its authors concede that the transformer memorizes small datasets and is sensitive to hyperparameters. All three treat employees as independent observations. None puts replacement costs into its decision rule, although Zhu et al. document nurse replacement costs of up to $88,000. The disagreement between these papers is the concrete case for refinement-style reporting.

From the crystallographic side, refinement reports such as Liu et al. supply the reporting norm we borrow. These reports give every refined parameter with its uncertainty and judge the fit by residual indices ($R_{gt}(F)=0.0557$, $wR(F^2)=0.1325$). We do not claim that organizations behave like lattices. The "vacancy redistributes stress" analogy suggested by the research file is suggestive, but it adds nothing testable beyond the hazard model below, so we leave it aside. We also leave aside the observation in the phase-retrieval tutorial that phase retrieval is formally a single-layer quadratic network. It does not bear on this decision problem.

Research Program

Our model is a discrete-time hazard model. A hazard is the probability that an employee leaves in a given period, given that they are still employed at its start. For employee $i$ in period $t$:

$$\lambda_i(t) = \sigma\big(\beta^\top x_i(t) + \gamma\, \Delta W_i(t) + \delta\, S_i(t)\big),$$

The symbols are as follows. $\sigma$ is the logistic function, which maps any number to a probability. $x_i(t)$ is the usual set of individual covariates: tenure, overtime, pay, and survey satisfaction where available. $\Delta W_i(t)$ is $i$'s excess workload from unfilled vacancies on their team, computed from rosters as uncovered shifts divided by remaining headcount. $S_i(t)$ is a decayed count of recent departures among $i$'s close colleagues. The coefficient $\gamma$ measures the workload channel and $\delta$ the social channel.

The cascade-adjusted cost of $j$'s departure is

$$C_j = c_j + \sum_{k \neq j} c_k \big[P(k \text{ leaves} \mid j \text{ leaves}) - P(k \text{ leaves} \mid j \text{ stays})\big],$$

Here $c$ is the replacement cost by role, and the probabilities are computed by simulating the fitted model forward over a fixed horizon. The simulation uses the organization's empirical distribution of backfill delays.

Fitting follows refinement practice. Every coefficient is reported with a bootstrap standard uncertainty. A held-out set of whole teams, never seen during fitting, gives a residual index: the normalized gap between predicted and observed departures per team per period. We keep a data-to-parameter ratio comparable to what crystallographers require before trusting a refinement, which is the reason for the small model rather than a deep one. Before using real data, we check that the model's parameters can actually be recovered, the way crystallographers test a refinement procedure on structures with known answers. We attach simulated team structures and simulated cascades with known $\gamma$ and $\delta$ to the IBM HR covariates, then confirm that the fitting procedure returns those values within their reported uncertainties.

Real data must carry team membership, rosters or shift records, departure dates, and the date each vacancy was filled. No such dataset is public. Teng et al.'s data is proprietary, and the IBM HR set has no team structure. The main empirical work therefore depends on a data partnership with a hospital system, where rosters are routinely kept and replacement costs are documented. A secondary partner from another sector, ideally with weaker interpersonal ties, would test how far the results carry over.

Baselines

The do-nothing reference allocates a retention budget of $k$ interventions uniformly at random. This sets the floor for realized cost captured. A second naive reference ranks employees by replacement cost alone, which asks whether expensive people are simply expensive to lose.

The individual-risk baselines are logistic regression and XGBoost on the same covariates $x_i$. On IBM HR they reach 85.01% AUC-ROC for logistic regression (Comparison) and 89.12% accuracy for XGBoost (Retention is All You Need). These figures use different metrics and are not directly comparable. The same XGBoost scored 0.6542 AUC-ROC in the transformer study, so we rerun all baselines ourselves rather than quote them. The cost-sensitive version multiplies predicted risk by $c_i$. This is the competitor a practical reviewer would propose first.

The strongest existing method is CEHNN, with 0.914 AUC-ROC reported by Teng et al. on their proprietary data. We reimplement it on the partner data with the same interaction features. We give it the same advantage as our model: its risk multiplied by $c_i$. We also give it a simulation-based cascade cost obtained by deleting node $j$ and re-scoring neighbors. That makes the comparison about the two-channel structure, not about which method gets to simulate.

Every method gets the same data, the same temporally forward splits (train on earlier years, test on later ones), the same team-held-out validation, and a matched hyperparameter search budget.

Two ablations isolate the key ingredient. One sets $\gamma=0$, removing workload. The other sets $\delta=0$, removing the social channel. The proposal beats the field if it captures more realized cost in the top $k$ than cost-weighted CEHNN across a range of budgets. Realized cost here means each targeted person's own replacement cost if they left, plus the replacement costs of teammate departures in the following window. A useful margin would be on the order of a tenth of total cost captured, with bootstrap intervals that exclude zero. It must also beat both ablations. If the $\gamma=0$ ablation performs just as well, the workload claim fails.

Why not just…?

Why not just make XGBoost cost-sensitive? This is the real threat. If induced departures are a small share of total attrition cost, weighting individual risk by replacement cost captures nearly everything, and our model adds complexity for nothing. We do not argue this away. We measure it. The first analysis on partner data estimates what fraction of realized cost comes from departures that follow a teammate's vacancy within the window. If that fraction is small, we report the negative result and stop.

Why not just use CEHNN or a graph neural network? Such models already know that departures cluster. They can predict that a team is at risk, but they do not say whether the risk comes from shifts nobody is covering or from friends leaving. That distinction decides whether money should go to agency staff or to retention conversations. Deleting a node and re-scoring the network gives a rough cascade estimate, and we include it as a baseline. The decisive backfill-delay test below can only be run on a model that separates the two channels.

Why not just raise pay for at-risk staff? Wong et al. found that pay satisfaction did not weaken the overload–burnout–turnover pathway. Lee et al.'s negative salary coefficient (−0.236) is smaller than the burnout coefficient (0.568) in the same model. We include pay as a covariate and report its fitted effect alongside $\gamma$. If pay dominates, cost-sensitive targeting with raises is the answer.

Why bring crystallography in at all? Only as a validation standard, and we would drop it if a lighter standard served equally well. The case for keeping it is the record in the file itself. The same algorithm on the same 1,470 records is reported at 0.6542 AUC-ROC in one paper and with competitive accuracy in another, and none of these figures carries an uncertainty. Reporting every parameter with its uncertainty, judging fit by held-out residuals, and testing parameter recovery on simulated data with known answers are what make the channel estimates trustworthy enough to act on.

Decisive Checks

The sharpest test uses natural variation in backfill delay. Hold fixed the departure itself and the departed person's ties to the team. If the workload channel is real, teammates' hazard should rise with the number of days the vacancy stays open. Social contagion predicts no such dependence once the departure has happened. A flat relationship between delay and teammate hazard falsifies the workload half of the claim. A rising one, robust across sites and controlled for unit-level staffing shortages that might cause both long delays and departures, supports it and directly prices faster backfilling.

The second check is parameter recovery on the semi-synthetic IBM HR cascades. It shows that $\gamma$ and $\delta$ can be separated at realistic team sizes. If they cannot, the real-data estimates are uninterpretable. The third check is the forward-in-time cost-capture comparison against cost-weighted CEHNN described above. Together these show whether the cascade-adjusted cost of a departure is a quantity that can be measured, and whether it is worth measuring.