Attrition Models That Know Pay Cannot Fix Burnout

Companies increasingly ask machine-learning models two questions about each employee: how likely is this person to quit, and what would change that? The second question is usually answered by "what-if" analysis. You take a trained classifier, raise the person's salary or cut their overtime in the input, and read off the new predicted quit probability. One decision-support system on the standard IBM HR benchmark reports such a simulated intervention cutting one employee's attrition probability "from 84% to 16.3%" (Retention is All You Need). Organizational psychology gives a different answer. In a survey of 502 Hong Kong nurses, job demands drove turnover intention through burnout, and satisfaction with pay did not weaken that burnout pathway (Wong et al.). Burnout here means emotional exhaustion and withdrawal. If the psychologists are right, a flat classifier that treats salary as an independent dial will recommend raises to exhausted employees whom raises will not keep. Those are the people whose departure is most expensive: up to $88,000 to replace a nurse, and from $88,000 to over $1 million for a physician vacancy (Zhu et al.).

We propose to build attrition models whose structure encodes the burnout pathway, so that pay can affect predicted risk only through the channels where psychology says it works. We then choose the models' settings by the money the resulting retention policy saves or wastes, not by accuracy or AUC. The test bed is a semi-synthetic workforce whose true causal structure we control, calibrated to the published nurse findings. On it we can measure the cost of wrong advice directly. That cannot be done on observational HR data, where nobody knows what a raise would actually have done.

Core Claim

The claim is that structure, not predictive accuracy, is what decides whether an attrition model gives sound retention advice. Two models can have nearly identical AUC yet recommend very different interventions. A model forced to route salary effects around the burnout pathway will spend a fixed retention budget better than an unconstrained model with matching held-out accuracy, whenever the true world resembles the one Wong et al. describe. A second claim is that the constraint should be tuned, not imposed: how strongly to enforce it, along with class weighting and decision thresholds, should be chosen by expected cost. Then the structure can relax when data contradict it, as in the Costco-style high-wage evidence the file mentions. The swarm optimizer is the search engine here, not the contribution. We say this plainly because a reviewer would rightly doubt that particle swarm search beats random search on a few dozen knobs.

What Prior Work Does and Does Not Do

Three predictive papers on the IBM HR dataset (1,470 records, about 16% leavers) set the accuracy landscape, and they disagree with each other. Guerranti and Dimitri find plain logistic regression best, at 85.01% AUC-ROC. Wenhui Li reports a tabular Transformer at 0.7452 AUC-ROC and 0.3849 AUCPR, with XGBoost far lower. Mohiuddin et al. report 89.12% accuracy for XGBoost with SMOTE oversampling and then build salary and overtime what-if advice on top. All three treat "JobSatisfaction" or "OverTime" as flat columns, and all three optimize symmetric metrics. Their large spread on one dataset suggests that preprocessing and tuning choices, not architecture, drive the numbers. That motivates searching over those choices explicitly under a decision objective. Teng et al. show that turnover spreads through peer and manager departures (91% of exits follow prior ones). This constrains our semi-synthetic simulator: it must include contagion, or it will be too easy.

The mediation literature supplies the structure we encode. Wong et al. give the demands-to-burnout-to-turnover path, including a work-overload-to-burnout coefficient of 0.590, and the null pay moderation. Jiahui Li et al. show a chain in which a stressor reaches satisfaction only through burnout. Lee et al. find that salary helps retention but cannot offset a toxic communication climate. These studies are cross-sectional and stop at intention scales, which is why we use them to calibrate a simulator rather than as ground truth about real exits.

The decision layer reuses Olaya et al.. That work ranks customers for treatment by expected causal profit rather than by predicted risk, and shows this beats effect-only ranking on marketing data. We carry the idea to workforce interventions with heterogeneous costs and a choice among actions: a raise, workload relief, or nothing. The PSO literature (Ji et al., Wang et al.) supplies the search method. Those papers compare PSO only against untuned models or genetic algorithms, never against random search at equal budget, so we treat its advantage as unproven.

Research Program

We first build the semi-synthetic workforce. Employees get demand variables (overtime, travel, commute), a latent burnout score driven by demands, a salary, and a satisfaction score. Quit hazards come from a structural model whose demand-to-burnout slope is set from Wong et al.. Peer-departure contagion follows Teng et al.. Marginal feature distributions are matched to IBM HR. Observed features include only noisy proxies for burnout, mimicking IBM's single-item work-life-balance and environment-satisfaction scores. We generate a family of worlds indexed by one parameter $\gamma$, the strength with which pay buffers the burnout path. At $\gamma = 0$ the world matches the nurse finding; at large $\gamma$ it matches the Costco-style claim that wages retain everyone. Because interventions are simulated, every policy's true cost is computable.

Second, we build the structured model as a gradient-boosted classifier with XGBoost interaction constraints. Salary may interact with non-burnout features but not with the burnout-proxy group, and a strength knob mixes the constrained and unconstrained models. A two-stage variant first predicts the burnout proxy from demands, then predicts attrition from that proxy plus salary.

Third, we select all settings by cost. The search covers tree depth, learning rate, class weight, oversampling ratio, constraint strength, and the per-action treatment thresholds. Each candidate is scored on held-out folds by

$$ J = \frac{1}{n}\sum_{i=1}^{n}\Big[c_{a_i} + R_i\,\Pr(\text{leave}_i \mid a_i)\Big], $$

where $a_i$ is the action the policy assigns employee $i$, $c_{a_i}$ is that action's cost, $R_i$ is the replacement cost, and the quit probability is the true simulated one. Replacement costs are set at $4,000, following [Guerranti and Dimitri](/work/a-comparison-of-machine-learning-approaches-for-predicting), for general staff, and at $88,000 for a nursing-like scenario, following Zhu et al.. PSO runs this search with a fixed evaluation budget.

Fourth, we check on IBM HR whether the data themselves show the pattern. Specifically, we ask whether the salary–attrition association weakens among employees with high overtime and low work-life-balance scores. We report the IBM results with the caveat the file stresses: that dataset is synthetic.

Baselines

The do-nothing reference is the cost of intervening on no one. Next to it we report a classifier that predicts "stay" for everyone, which already scores about 84% accuracy and shows why accuracy is the wrong yardstick here.

The strongest predictive baseline is logistic regression as run by Guerranti and Dimitri, which reports 85.01% AUC-ROC and 87.96% test accuracy on IBM HR. It is the strongest because it tops the reported numbers and is the model an HR analyst would trust. We also run XGBoost with SMOTE in the style of Mohiuddin et al., which reports 89.12% accuracy, since that is the setup behind current what-if advice. All predictive baselines get the same folds and the same tuning budget, measured in model fits, under the same optimizer. Each is then paired with the best cost-optimal threshold policy, so they are not handicapped on the decision step.

The strongest decision baseline is the expected-causal-profit ranker of Olaya et al., built on the unconstrained model. We found no reported result on Lacuna for expected-causal-profit ranking on workforce attrition data, so we run it ourselves.

Three ablations isolate the ingredients: the structural constraint with AUC-based tuning, cost-based tuning without the constraint, and PSO replaced by random search at equal budget. Beating the field means the structured, cost-tuned policy achieves lower true cost $J$ than every baseline at $\gamma = 0$, with confidence intervals across simulator seeds that exclude zero difference. At matched AUC, its gap should come from fewer raises going to high-burnout employees.

Why not just…?

Why not just tune for AUC and then pick the cost-optimal threshold? This is the most serious alternative, and in worlds where pay does buffer burnout it may be just as good. A threshold changes who gets an intervention but not which intervention the model thinks will work. That second question is where a flat model goes wrong in a $\gamma = 0$ world. The comparison between the AUC-tuned-plus-threshold baseline and the full method, run across the $\gamma$ sweep, settles it. If the two are equal at $\gamma = 0$, the structural claim fails.

Why not just use random search or Bayesian optimization instead of particle swarms? For a search space of roughly ten knobs, random search may well match PSO. The cited PSO studies (Ji et al., Wang et al.) never test against it. We therefore make no claim for PSO and report the equal-budget comparison as an ablation. If random search ties, the recommendation is simply to use random search.

Why not just feed overtime and demands straight into the classifier and skip the burnout structure? A flexible enough model could in principle learn the salary-by-burnout interaction unaided. With 1,470 records and 16% leavers, though, it may not. The flat XGBoost result in Wenhui Li (0.2875 AUCPR) suggests these models are data-starved. We test this by varying simulated sample size: the constraint should matter at small samples and wash out at large ones. If it never matters, the structural idea is unnecessary.

Proposed Experiments

The decisive experiment sweeps $\gamma$ from zero to strong pay buffering at IBM-sized and ten-times-larger samples, with both replacement-cost regimes. We report true cost $J$, regret against the oracle policy, AUCPR, and the share of the raise budget spent on employees in the top burnout quartile. The prediction that would falsify the proposal is equal cost between the structured and unconstrained policies at $\gamma = 0$, or a large loss for the structured policy at high $\gamma$ despite tuning of the constraint strength.

A second check asks whether cost-based tuning picks different models than AUC-based tuning. This shows whether the disagreements among Guerranti and Dimitri, Wenhui Li, and Mohiuddin et al. matter for decisions at all. The third check is the IBM HR stratified analysis of the salary effect by burnout proxy, which we present as suggestive only.