Stop-and-go "phantom jams" are usually blamed on drivers overreacting to the car in front. Each small brake gets amplified as it passes back through a line of cars, a property called string instability. The argument for automated vehicles as wave dampers rests on that diagnosis, and so do the human-driver models used to train them. Yet the human models are fitted by regressing each follower's acceleration on its immediate leader's motion. That regression cannot distinguish three things: the follower reacting to the leader, both cars reacting to the same wave arriving from further ahead, and drivers who already prefer tight gaps sorting themselves into tight-gap positions. Network causal inference calls this the homophily–influence problem. Car-following research has not treated it as a confound.
We propose to estimate, for each latent driving regime, whether human following is string-unstable, after adjusting for these two confounds. We then pool the estimates across corridors with random-effects meta-analysis. The deliverable is a confounding-adjusted instability margin for each regime, together with a prediction interval for a corridor that has never been observed. That interval is the quantity a traffic-smoothing controller designer actually needs.
Core Claim
The claim is that the per-regime string-stability margins fitted from leader–follower pairs are biased by common exposure to downstream waves and by self-selection into gaps, and that the bias is large enough to flip some regimes between "stable" and "unstable." We further claim that the margins vary across corridors in a way that single-site calibration hides. For a car-following law $\dot v = f(s, \Delta v, v)$, where $s$ is the gap, $v$ is ego speed and $\Delta v$ is leader speed minus ego speed, the standard linear criterion says a homogeneous platoon is string-stable at an equilibrium when
$$\lambda = \tfrac{1}{2} f_v^2 - f_{\Delta v} f_v - f_s \ge 0,$$
where $f_s, f_{\Delta v}, f_v$ are the partial derivatives of the acceleration law. In words, sensitivity to the gap destabilizes the platoon, while sensitivity to relative speed and to one's own speed stabilizes it. Common exposure distorts the estimate of $f_{\Delta v}$, because leader and follower decelerate together whether or not the follower is responding to the leader. Sorting distorts $f_s$, because differences between drivers' preferred gaps masquerade as within-driver gap sensitivity. Both errors flow directly into $\lambda$.
What Exists and What Is Missing
The mechanism we build on is the regime-switching car-following model of Markov Regime-Switching Intelligent Driver Model for Interpretable Car-Following Behavior. Its model, FHMM-IDM, describes each follower with two hidden Markov chains, one for driver intent and one for the traffic scene, and ties each regime to its own Intelligent Driver Model parameters. On 100 HighD pairs it recovers five driving regimes and five scenarios, avoiding the "regression to the mean" of a single parameter set. That paper does not compute stability per regime. It fits a single site. It treats the immediate leader's motion as the cause of the follower's response. We keep its regime structure and change all three of those choices.
The analytic side of string stability is represented by Hybrid System Stability Analysis of Multilane Mixed-Autonomy Traffic. It proves stability through variance-based Lyapunov functions, under the assumption that humans follow one fixed Optimal Velocity Model. Our per-regime margins say whether that assumption is safe for each human mode. This motivates the work but does not supply any of its methods.
The causal design comes from network science. Social Networks and the Identification of Peer Effects and Influence Propagation in Social Networks: A Data Mining Perspective establish that naive correlation overstates contagion when ties are self-selected or exposure is shared. Formation of Social Ties Influences Food Choice: A Campus-Wide Longitudinal Study shows a practical fix: match people on their behavior before a tie forms and attribute post-formation divergence to the new partner. We borrow that incident-tie design directly.
On a highway, the natural "tie formation" is a lane change that hands a follower a new leader. Cut-ins are themselves a known failure point for smoothing controllers in the I-24 field test, Traffic Control via Connected and Automated Vehicles (CAVs): An Open-Road Field Experiment with 100 CAVs. The graph-neural approach of Modeling Interference for Treatment Effect Estimation in Network Dynamic Environment handles spillover on social graphs. Its ablation found that removing temporal history caused the largest error increase. That constrains our design: adjustment must condition on each follower's own trajectory history, not just on the current snapshot.
Random-effects pooling, which separates within-study noise $\sigma_i^2$ from between-study spread $\tau^2$ and weights each study by $1/(\sigma_i^2+\tau^2)$, is routine in clinical synthesis, for example Risk factors of epidural hematoma in patients undergoing spinal surgery: a meta-analysis of 29 cohort studies. To our knowledge it has never been applied to car-following parameters. Here it supplies the method, not the motivation.
Research Program
The work proceeds in three stages.
The first stage fits a regime-switching car-following model with exposure adjustment. We extend FHMM-IDM so that the follower's acceleration model includes the motion of the vehicles two and three ahead, and a lagged downstream speed field. HighD, NGSIM and I-24 track every vehicle, so these inputs are observable. With them included, the response to the immediate leader is identified net of the shared wave. This is the "common exposure" adjustment.
The second stage handles sorting. We assemble all lane-change events that give a follower a new leader and that leave at least ten seconds of follower history before the change. Followers are matched on their pre-event regime path and speed–gap trajectory, following the incident design of Gligoric et al. We then compare the response of followers who received an aggressive-regime leader with matched followers who received a calm one. Two quantities come out of this: the causal effect of leader regime on the follower's regime-transition probabilities, and a within-driver estimate of gap sensitivity $f_s$ that is free of between-driver sorting.
A placebo test runs alongside. Future leader accelerations that the follower could not yet see should carry no information once the adjustment set is included. If they do, confounding remains.
The third stage is synthesis across corridors. Each recording site counts as a "study": the HighD locations, US-101, I-80, and the I-24 trajectory days from From Sim to Real: A Pipeline for Training and Deploying Traffic Smoothing Cruise Controllers. For each site we compute $\lambda$ per regime at a common set of equilibrium speeds, with site-level standard errors from a block bootstrap over vehicles. We then pool across sites using restricted maximum likelihood estimates of $\tau^2$. We report the pooled mean, $I^2$, and, most importantly, the prediction interval for a new site. We use meta-regression on speed limit, truck share and country to explain the remaining heterogeneity. The number of sites is small, likely 10 to 20, so $\tau^2$ will be estimated imprecisely. We will report its profile-likelihood interval rather than hiding it.
Baselines
The naive reference is a single-regime IDM fitted per site on leader–follower pairs only, with fixed-effect pooling of its implied $\lambda$. This is how most simulators are calibrated today, so any proposed change must improve on it.
The strongest existing modeling competitor is FHMM-IDM itself, fitted per site exactly as published, with its per-regime $\lambda$ computed afterward. It will be fitted on identical data splits, with the same number of regimes and the same optimizer and iteration budget as our adjusted version. That way, any differences come from the adjustment rather than from capacity. No reported result was found on Lacuna for FHMM-IDM trajectory error, since the paper reports case studies only. We will therefore report its held-out negative log-likelihood ourselves rather than quote a figure.
For trajectory prediction as a sanity floor, we use the tuned constant-velocity Kalman filter of Inertial Single Vehicle Trajectory Prediction Baselines and Applications with the NGSIM Dataset. That paper reports a mean negative log-likelihood of 4.44 at a 5-second horizon on NGSIM, and 3.73 for its LSTM–Kalman hybrid. Our adjusted model must not be worse than 3.73 on the same NGSIM horizon. If it is, the adjustment is buying causal cleanliness at the cost of a model that no longer describes the data.
The decisive comparison is not predictive, however. On the bias of $\lambda$ in simulation and on interval coverage at held-out sites, the adjusted model beats FHMM-IDM if it reduces mean absolute bias of $\lambda$ by a margin larger than the bootstrap error. It must also bring the 90% prediction interval's coverage of a held-out site's directly measured amplification to within 10 points of nominal where FHMM-IDM with fixed-effect pooling is not.
Two ablations isolate the key ingredients: exposure adjustment without the cut-in design, and the cut-in design without exposure adjustment. A further ablation replaces random-effects pooling with fixed-effect pooling and leaves everything else unchanged.
Why not just…?
Why not just fit one hierarchical Bayesian model over all trajectories instead of a two-stage meta-analysis? This is a real alternative. A one-stage hierarchical model with site-level random effects is the principled version of what we do. We will run it as a baseline, sharing the adjusted likelihood. If its pooled $\lambda$ and prediction intervals agree with the two-stage result, the conclusion is that pooling method matters little, and the contribution rests on the confounding adjustment and the stability estimand. The two-stage route remains useful because corridors whose raw video cannot be released can still contribute per-site summaries.
Why not just use an existing multi-anticipation car-following model, which already looks several cars ahead? Including vehicles further ahead handles common exposure, and that is stage one of our design. It does nothing about sorting: drivers who choose tight gaps still contaminate $f_s$. The ablation that drops the cut-in design measures how much of the bias sorting accounts for. If that share is negligible, we will say so, and the method collapses to an anticipation model plus pooling.
Why not just measure wave amplification directly from trajectories, without any behavioral model? Direct upstream-to-downstream amplitude ratios are the ground truth we validate against, so we take them seriously. They mix driver instability with platoon composition, inflow at the boundaries, and lane changes. They also cannot be transferred to a corridor with a different vehicle mix or a different automated-vehicle share. A per-regime $\lambda$ can be transferred. The held-out-site experiment tests whether that extrapolation beats simply carrying over the measured amplification from the most similar corridor. If it does not, the modeling layer has not earned its keep.
Decisive Checks
The first check uses simulation with known truth. In Flow, Flow: A Modular Learning Framework for Mixed Autonomy Traffic, we simulate multi-lane traffic whose drivers switch among IDM regimes and anticipate two cars ahead. Each driver's lane choice is tied to their desired headway. Because true $\lambda$ is known in this setting, we can measure the bias of the naive model, of FHMM-IDM, and of the adjusted estimator under varying strengths of sorting and anticipation. The prediction is that FHMM-IDM's bias grows with both, and the adjusted estimator's bias does not.
The second check is leave-one-site-out on real data. We pool all sites except one and predict that site's per-regime $\lambda$ and its measured wave amplification. We then score coverage of the 90% prediction interval and the error in predicted amplification.
The third check is a quasi-experiment, used if trajectory access around the I-24 automated fleet allows it. When a controlled automated vehicle becomes a human driver's leader, the leader's speed profile is set by the controller rather than by any trait it shares with the follower. Responses estimated from these exogenous-leader events should match our observational estimates and diverge from the naive ones.
The proposal fails if the adjusted and naive $\lambda$ agree within error in both simulation and the quasi-experiment.
References
- Markov Regime-Switching Intelligent Driver Model for Interpretable Car-Following Behavior
- Hybrid System Stability Analysis of Multilane Mixed-Autonomy Traffic
- Social Networks and the Identification of Peer Effects
- Influence Propagation in Social Networks: A Data Mining Perspective
- Formation of Social Ties Influences Food Choice: A Campus-wide Longitudinal Study
- Traffic Control via Connected and Automated Vehicles (CAVs): An Open-Road Field Experiment with 100 CAVs
- Modeling Interference for Treatment Effect Estimation in Network Dynamic Environment
- Risk factors of epidural hematoma in patients undergoing spinal surgery: a meta-analysis of 29 cohort studies
- From Sim to Real: A Pipeline for Training and Deploying Traffic Smoothing Cruise Controllers
- Inertial Single Vehicle Trajectory Prediction Baselines and Applications with the NGSIM Dataset
- Flow: A Modular Learning Framework for Mixed Autonomy Traffic