Joint models in both directions and at scale: jmBIG and JMbdirect for dynamic prediction

Atanu Bhattacharjee · Royal Statistical Society

Joint models in both directions and at scale: jmBIG and JMbdirect for dynamic prediction

Longitudinal biomarkers and time-to-event outcomes that act on each other, fitted on cohorts of any size

Atanu Bhattacharjee

Senior Lecturer and Academic Statistician, School of Medicine, University of Dundee

Royal Statistical Society

Time-to-event data: the building blocks

Event times are right-censored; survival is described by the survival curve or the hazard, estimated by Kaplan–Meier and related to covariates by Cox regression.

time since study entry end of study ● event○ censored

Censoring

Follow-up may end before the event; we observe the earlier time and which one it was. Censoring is assumed non-informative.

Survival and hazard

Share still event-free, and the instantaneous event rate among those at risk.

Kaplan–Meier

Product of conditional survival over event times; nⱼ at risk, dⱼ events at tⱼ.

Cox proportional hazards

Covariates act multiplicatively on an unspecified baseline hazard; β by partial likelihood, hazard ratio exp(β).

Kaplan EL, Meier P. JASA 1958; 53: 457–481. Cox DR. J R Stat Soc B 1972; 34: 187–220.

Longitudinal data: mixed models and dropout

Repeated measurements are correlated within a person, irregularly timed and measured with error; the linear mixed model separates the population trajectory from each person's deviation.

time population average individual trajectories

Yᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ + εᵢ(t),  bᵢ ~ N(0, D),  εᵢ(t) ~ N(0, σ²)

Fixed and random effects

β: population trajectory. bᵢ: person-specific intercept and slope; D induces the within-person correlation. σ²: measurement error.

Missing at random

Dropout depends only on observed values: likelihood inference from the mixed model stays valid.

Missing not at random

Dropout depends on the unobserved trajectory, typically because the event itself ends follow-up. The mixed model alone is biased; selection, pattern-mixture or shared-parameter (joint) models are needed.

Laird NM, Ware JH. Biometrics 1982; 38: 963–974. Rubin DB. Biometrika 1976; 63: 581–592. Diggle P, Kenward MG. Appl Stat 1994; 43: 49–93.

Time-varying covariates in the Cox model, and their limits

The extended Cox model uses the covariate value at each event time. That is valid for external covariates; a biomarker measured on the patient is internal, error-prone and observed only at visits.

What the data look like for one patient time since baseline biomarker event no further measurements true trajectory mᵢ(t) observed values Yᵢ(t) last value carried forward

Value known at every event time

In practice the last observation is carried forward between visits.

No measurement error

Plugging in noisy values attenuates the association towards zero (Prentice 1982).

An external covariate

An internal marker exists only while the patient is event-free, so its path carries information about the event.

hᵢ(t) = h₀(t) exp{ γᵀwᵢ + α yᵢ(t) },  yᵢ(t): observed covariate value at t

Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data, 2nd ed. Wiley 2002. Prentice RL. Biometrika 1982; 69: 331–342.

Four ways to use repeated measurements

ApproachHow the marker entersMeasurement errorInformative dropout
Time-dependent CoxLast observed value carried forwardIgnoredIgnored
Two-stageFitted trajectory from a mixed model, plugged into a Cox modelPartly removedIgnored
LandmarkingValue at landmark time t, Cox model fitted on those still at riskIgnoredConditioned on survival to t
Joint modelLongitudinal and survival submodels share random effects bᵢModelledModelled

Joint modelling fits the marker and the event in one likelihood, so the trajectory, its uncertainty and the dropout it causes are estimated together.

Informative dropout: measurements stop because of the event or the patient's state, so the missing values are related to the outcome.

Why joint modelling matters

Fitting the marker and the event together, in one model, answers what each separate analysis gets wrong.

Less bias in the marker–event association

The true trajectory, not its noisy measurements, enters the event model.

Correct marker trajectories

Dropout caused by the event is modelled, so the trajectory is not distorted by who remains.

Efficiency

Both outcomes inform the shared parameters, so treatment effects on each are estimated more precisely.

Individual prediction

A patient's own measurements update their own risk as follow-up continues.

Hickey GL et al. BMC Med Res Methodol 2016; 16: 117. Alsefri M et al. BMC Med Res Methodol 2020; 20: 94. Wang JL, Zhong Q. Annual Review of Statistics and Its Application 2025; 12.

How joint models began

In the 1990s, AIDS and cancer studies needed to link repeatedly measured markers, such as CD4 counts, to survival. A few papers built the idea of fitting the two together.

19941996199720002004 DeGruttola & TuCD4 progressionand survival Faucett & ThomasBayesian joint model,Gibbs sampling Wulfsohn & Tsiatisjoint likelihood,EM algorithm Henderson, Diggle& Dobson: latentprocesses, drop-out Tsiatis & Davidianoverview of theframework

DeGruttola V, Tu XM. Biometrics 1994; 50: 1003–1014. Faucett CL, Thomas DC. Statistics in Medicine 1996; 15: 1663–1685. Wulfsohn MS, Tsiatis AA. Biometrics 1997; 53: 330–339. Henderson R, Diggle P, Dobson A. Biostatistics 2000; 1: 465–480. Tsiatis AA, Davidian M. Statistica Sinica 2004; 14: 809–834.

The joint model: two models that share the patient

One model follows how each patient's marker changes over time. A second model describes the risk of the event. Both share the patient's own deviation from the average, so a patient whose marker runs high also carries the matching risk.

Marker model average path over time + this patient's own deviation + measurement noise Event model baseline risk over time × effect of baseline factors × effect of the true marker now the patient's own random effects

Yᵢ(t) = mᵢ(t) + εᵢ(t),  mᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ

hᵢ(t) = h₀(t) exp{ Wᵢᵀγ + α mᵢ(t) }

Yᵢ(t) observed marker, mᵢ(t) true marker, bᵢ random effects, εᵢ(t) measurement error, hᵢ(t) event hazard, α association: log hazard ratio per unit of the true marker.

What feature of the marker drives risk?

The link between marker and event is a clinical question: is it the level today, how fast it is changing, how much has built up, or how far the patient sits from the average?

Current value

Risk depends on the true level of the marker now.

α mᵢ(t)

Current slope

Risk also depends on how fast the marker is changing.

α₁ mᵢ(t) + α₂ mᵢ′(t)

Cumulative effect

Risk depends on accumulated exposure, the area under the trajectory.

α ∫₀ᵗ mᵢ(s) ds

Patient deviation

Risk depends on the patient's own departure from the average level and slope.

αᵀbᵢ

Each term enters the event model as hᵢ(t) = h₀(t) exp{ Wᵢᵀγ + term }, a straight-line effect on the log-hazard scale.

Fitting a joint model

The random effects are never observed, so the likelihood averages over every value they could take. Two families of methods do this.

Maximum likelihood (EM)

Treat the random effects as missing data: fill them in from their expected values given the data, update the parameters, repeat. The integral is approximated by Gauss–Hermite quadrature or Laplace approximation.

Wulfsohn & Tsiatis 1997; Rizopoulos 2012.

Bayesian (MCMC)

Place priors on all parameters and draw random effects and parameters together by Markov chain Monte Carlo; uncertainty in predictions comes directly from the draws.

Faucett & Thomas 1996; Ibrahim, Chen & Sinha 2004.

L(θ) = Πᵢ ∫ p(Yᵢ | bᵢ; θ) p(Tᵢ, δᵢ | bᵢ; θ) p(bᵢ; θ) dbᵢ

θ: all parameters; the marker and event contributions are independent given the patient's random effects bᵢ.

Extensions and software

After 2000 the single-marker, single-event model grew towards the data clinicians hold, and software made it routine.

Standard joint model one marker, one event Several markers correlated random effects across markers Brown, Ibrahim & DeGruttola 2005; Chi & Ibrahim 2006 Competing risks cause-specific hazards for several event types Elashoff, Li & Li 2007 Latent classes subgroups with their own trajectory and risk Proust-Lima et al. 2014 Software JM 2010likelihood, R JMbayes 2016Bayesian, predictions joineRML 2018several markers, MCEM FastJM 2022linear-time, biobank jmBIG 2024partitioned big data JMbdirect 2026feedback, flexible links

Brown ER et al. Biometrics 2005; 61: 64–73. Chi YY, Ibrahim JG. Biometrics 2006; 62: 432–445. Elashoff RM et al. Stat Med 2007; 26: 2813–2835. Proust-Lima C et al. Stat Methods Med Res 2014; 23: 74–90. Rizopoulos D. J Stat Softw 2010; 35(9): 1–33; J Stat Softw 2016; 72(7): 1–46. Hickey GL et al. BMC Med Res Methodol 2018; 18: 50. Li S et al. Comput Math Methods Med 2022; 1362913.

Three open issues in joint modelling

Three assumptions of the standard joint model limit its use in routinely collected multimorbidity data. My two pieces of work address them.

1. Scale

Fitting is slow; joint models have mostly been used in small to moderate studies, while health records hold hundreds of thousands of patients.

2. One direction

The marker drives the risk of the event, but events never change the marker that follows.

3. A straight-line link

Each unit of the marker adds the same amount of log-risk, so thresholds and plateaus are missed.

jmBIG (2024) → issue 1

Bayesian joint models fitted on patient subsets and combined.

JMbdirect (2026) → issues 2 and 3

Feedback from events to the marker, and flexible spline or neural links between marker and risk.

Li S et al. Comput Math Methods Med 2022; Bhattacharjee A et al. BMC Med Res Methodol 2024; 24: 172; Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230.

At scale: joint models on routinely collected data

Health-record data bring volume, heterogeneity and computational cost. jmBIG splits the patients into subsets, keeps each patient's longitudinal records with their survival record, fits the joint model to every subset and combines the estimates.

Full cohort N patients, all visits and events Subset 1 Subset 2 ⋯ Subset S joint model → θ̂₁ joint model → θ̂₂ ⋯ joint model → θ̂ₛ Combined estimates θ̄ Dynamic prediction

θ̄ = (1/S) Σₛ θ̂ₛ,   SE(θ̄) = (1/S) Σₛ SE(θ̂ₛ)

Bayesian routes: posterior means and standard errors averaged across the S subsets. Likelihood route: coefficients and Hessian updated from all subsets together.

Volume, heterogeneity and computational cost of health-record data, and the subset estimation: Bhattacharjee A, Rajbongshi BK, Vishwakarma GK. jmBIG. BMC Med Res Methodol 2024; 24: 172. Efficient implementation has lagged behind, limiting joint models to small and moderate studies: Li S et al. Comput Math Methods Med 2022; 1362913.

How the data are linked, split and merged

A joint model reads two files: many rows per patient for the marker, one row per patient for the event, linked by the patient identifier. jmBIG splits by patient, so each subset keeps every row of its patients, and merges the subset estimates.

Longitudinal file idtimemarker 1t₁₁y₁₁ 1t₁₂y₁₂ 1t₁₃y₁₃ 2t₂₁y₂₁ 2t₂₂y₂₂ ⋮⋮⋮ Survival file idtimeevent 1T₁δ₁ 2T₂δ₂ Patient i all marker rows + (Tᵢ, δᵢ) linked by id Subset 1: patientswith all their rows Subset 2… Subset S… θ̂₁ θ̂₂ θ̂ₛ Merged estimates θ̄

θ̄ = (1/S) Σₛ θ̂ₛ

Bayesian routes: posterior means and standard errors averaged across subsets. Likelihood route: coefficients and Hessian updated from all subsets. A patient's rows are never split across subsets.

Bhattacharjee A, Rajbongshi BK, Vishwakarma GK. BMC Med Res Methodol 2024; 24: 172.

One million patients in under an hour

Standard workstation (32 GB RAM) Trusted research environment, health records 01020304050 01020304050 minutes minutes 10100,0001,000,000 10100,0001,000,000 25.6513.0411.215.03 45.2329.2420.5417.49

● Bayesian MCMC (JMbayes2 engine)

■ Semiparametric EM (FastJM engine)

▲ Likelihood EM (joineRML engine)

◆ Bayesian HMC (Stan engine)

Processing time in minutes by number of patients (log scale); values at 1,000,000 patients labelled. A cohort of 1,000 patients with repeated visits and time to death took 2.3, 1.2, 1.2 and 2.72 minutes respectively. Source: Bhattacharjee, Rajbongshi and Vishwakarma, BMC Med Res Methodol 2024;24:172, Table 1.

Issue 2 in practice: conditions that drive each other

A feedback loop between conditions Type 2 diabetes HbA1c trajectory Cardiovascular disease Chronic kidney disease (eGFR) raises risk worsens control raises risk predicts CV death

Scale

Routinely collected health records hold repeated biomarker measurements on hundreds of thousands to millions of patients.

Direction

A cardiovascular event can permanently shift the glycaemic trajectory; a first cardiovascular hospitalisation accelerates eGFR decline.

Prediction

Risk for each patient has to update as new biomarker values and new events arrive.

From one direction to both directions

Standard joint model Biomarker trajectory Event hazard α The event never changes the marker path Bidirectional joint model (JMbdirect) Biomarker trajectory Non-terminal event e.g. cardiovascular event Terminal event e.g. death κ₀, κ₁ f₁ f₂ Events feed back into the marker path

Two changes to the standard model: the link between marker and risk is allowed to bend, and events that happen along the way change the marker that follows.

Feedback: each event moves the marker's level and slope

Some events change the marker that follows: a cardiovascular event can shift HbA1c for good, and cancer recurrence makes CA-125 jump.

time t marker mᵢ(t) Tᵢⱼ κⱼ₀ level shift slope + κⱼ₁ path without the event non-terminal event j

Level shift and slope change

An immediate jump once the event has happened, and faster or slower progression afterwards.

Clinical scale

CA-125 after ovarian cancer recurrence; eGFR decline after a first cardiovascular hospitalisation.

What the data must show

At least one measurement after the event for the jump, two for the slope; about five per event type across the cohort.

mᵢ(t) = Xᵢ(t)ᵀβ + Zᵢ(t)ᵀbᵢ + Σj ∈ Jᵢ(t) { κⱼ₀ I(t ≥ Tᵢⱼ) + κⱼ₁ (t − Tᵢⱼ)₊ }

Try it: an event that changes the marker

Move the sliders. A non-terminal event at time T shifts the marker's level and slope; the risk of the terminal event follows the new path.

Dashed: path and risk if the event had no effect on the marker. Solid: with feedback.

Illustration of the JMbdirect feedback term

Risk does not always rise in a straight line

The standard model assumes each unit of the marker adds the same amount of log-risk. Real risk often has a threshold or a plateau. Replacing the straight line with a smooth curve lets the data show the shape.

marker value log-risk threshold straight line flexible curve

Smooth spline curve

Built from many small pieces joined smoothly; a penalty stops it from wiggling more than the data support.

Neural additive curve

A sum of S-shaped steps that can capture sharp thresholds, kept in check by a penalty.

hᵢₖ(t) = h₀ₖ(t) exp{ Wᵢᵀγₖ + fₖ(mᵢ(t)) }

fₖ(m) = Σₗ αₗ Bₗ(m)   or   Σq σ(aq m + bq)

What the simulations show

Allowing feedback and a flexible link gave the sharpest and best-calibrated predictions, ahead of the established joint models; the gain grows with sample size and with the strength of feedback.

Do predicted risks match what happens? predicted risk observed risk flexible link + feedback:on the ideal line linear, no feedback:risks pulled to the middle dashed: perfect agreement

Sharper ranking of patients

Highest discrimination of the six models compared.

Ahead of established joint models

two-stagejoineRMLJMbayes2JMbdirect

Gains grow

More patients and stronger feedback widen the advantage; a straight-line model does not catch up with more data.

Simulation with competing risks and informative dropout; full figures in the backup slides. Source: Bhattacharjee A. BMC Med Res Methodol 2026; 26: 230.

From one disease to a network of morbidities

In a patient with several conditions, each risk depends on several markers, and an event in one condition can change the markers of another.

Shared risk

The risk of each condition depends on the current state of every marker, not its own alone.

Spillover

An event in one condition shifts the marker of another: renal decline after stroke, systemic markers after cancer recurrence.

Prediction

Each patient's risk for each condition, carrying the effect of every earlier event.

hᵢₘ(t) = h₀ₘ(t) exp{ Wᵢᵀγₘ + fₘ( mᵢ₁(t), …, mᵢM(t) ) }

m*ᵢₖ(t) = mᵢₖ(t) + κⱼₖ₀ I(t ≥ Tᵢⱼ) + κⱼₖ₁ (t − Tᵢⱼ)₊ ,   k ≠ j

How the two frameworks fit together

Biomarker and event histories JMbdirect feedback κ and flexible links fₖ, subject-level fitting jmBIG pathway partitioned, parallel estimation for large cohorts Individual dynamic predictions πᵢₖ(u | t)

Why separate estimation

Feedback terms and flexible links need each patient's whole trajectory, which cannot be split across data shards; they use their own routines: splines in the M-step, neural-component optimisation and feedback-augmented trajectory updates.

One workflow

The same model specification routes to the partitioned pathway when the cohort is large, with predictions returned in the same form.

Dynamic prediction updates with every visit

At each clinic visit, the patient's own measurements so far update their chance of staying event-free over the coming years.

follow-up time marker event-free probability new visit t u before the new visit after the new visit

Landmark t, horizon u

The prediction is made at time t, using everything recorded so far, for the window from t to u.

Every measurement counts

The whole marker history to t enters through the patient's random effects, with its uncertainty.

Updates at each visit

A new value changes the random effects and so the whole predicted curve.

πᵢ(u | t) = Pr(Tᵢ > u | Tᵢ > t, Dᵢ(t)) = ∫ Pr(Tᵢ > u | Tᵢ > t, bᵢ) p(bᵢ | Dᵢ(t)) dbᵢ,   Dᵢ(t): data up to t

Try it: a prediction that updates at every visit

Add the patient's visits one at a time. Each new value updates the patient's own trajectory and the predicted chance of staying event-free over the next five years.

Grey dashed: population average. Blue: this patient's fitted trajectory. Orange: event-free probability from the latest visit; earlier predictions fade.

Illustration: random-intercept-and-slope model, shrinkage update

Why joint models for dynamic prediction

A dynamic prediction needs the patient's whole marker history, corrected for noise, projected forward over the prediction window. Only the joint model provides all of these in one model.

What the prediction needsBaseline CoxTime-varying CoxLandmarkingJoint model
Updates as new measurements arrive✗✓✓✓
Uses the whole marker history✗✗✗✓
Corrects for measurement error✗✗✗✓
Valid survival probability for an internal marker✓✗✓✓
One coherent model across all prediction times✓✓✗✓
Also predicts future marker values✗✗✗✓

Internal marker: measured on the patient and ending at the event. Landmarking: a separate Cox model at each prediction time, using the last value.

Kalbfleisch JD, Prentice RL. 2nd ed. Wiley 2002. van Houwelingen HC. Scand J Stat 2007; 34: 70–85. Rizopoulos D. Biometrics 2011; 67: 819–829. Rizopoulos D et al. Dynamic predictions with time-dependent covariates using joint modeling and landmarking. arXiv:1306.6479. Putter H, van Houwelingen HC. Landmarking 2.0. arXiv:2103.16252.

Dimitris Rizopoulos: joint models into practice

Software anyone can use, predictions that update per patient, ways to judge and combine them, and their use to schedule clinical decisions.

2010 JMjoint models in R,maximum likelihood 2011 Dynamic predictionspatient-specific survivalupdated with each value 2012 The textbookJoint Models for Longitudinaland Time-to-Event Data 2014 Bayesian model averagingweights that change bypatient and by time 2016 JMbayesBayesian fitting, predictiondiscrimination, calibration 2019 Personalised schedulesbiopsy timing in prostate canceractive surveillance (PRIAS) Now JMbayes2several mixed-type markers,competing risks, multi-state

Rizopoulos D. J Stat Softw 2010; 35(9). Biometrics 2011; 67: 819–829. CRC Press 2012. Rizopoulos D et al. JASA 2014; 109: 1385–1397. Rizopoulos D. J Stat Softw 2016; 72(7). Tomer A et al. Biometrics 2019; 75: 153–162.

Landmarking or joint modelling?

Landmarking

Landmark: a fixed time t at which prediction is made.

Select patients still event-free at t, take their marker value at t, fit a Cox model for the window from t to the horizon u; repeat for each landmark.

Simple and robust to how the marker evolves; uses only the value at t, without measurement error or the shape of the history.

Joint model

One model for the whole marker path and the event; prediction at any t comes from the patient's posterior random effects.

Uses every measurement and its uncertainty, and gives predictions that are coherent across landmarks.

Depends on correct specification of the trajectory and of the link between marker and hazard.

Landmarking is the simple, robust benchmark; the joint model uses the full history and its uncertainty. Reporting both shows how much the marker history adds.

Checking dynamic predictions

A prediction is useful when it ranks patients correctly and when the predicted risks match what happens. Each is checked within the window from the landmark t to the horizon u.

Discrimination

Time-dependent AUC: the chance that a patient with the event in the window was given a higher risk than one without it.

Pr(π̂ᵢ < π̂ⱼ | Tᵢ ∈ (t, u], Tⱼ > u)

Prediction error

Brier score: mean squared gap between predicted event-free probability and observed status.

E[(I(T > u) − π̂(u|t))²]

Calibration

Predicted against observed risk: slope 1 and intercept 0 mean the predictions are right on average.

slope βcal, intercept αcal

Rizopoulos D. Biometrics 2011; 67: 819–829. Steyerberg EW et al. Epidemiology 2010; 21: 128–138.

Where joint models are heading

Many markers, many diseases

Multimorbidity as a network: each hazard driven by several trajectories, each event reshaping several markers.

Feedback by design

Events that change the biomarker path written into the model, rather than absorbed into random-effect variance.

Health-record scale

Partitioned and parallel estimation, so cohorts of a million patients fit in under an hour.

Flexible links

Splines and neural additive terms that reveal thresholds and plateaus while keeping a likelihood and uncertainty bands.

Deep-learning survival

Dynamic-DeepHit and recurrent-network models scale to very large cohorts; joint models keep explicit, interpretable structure and pointwise uncertainty.

Prediction in the clinic

Risk that updates at every visit, built into dashboards and health-record systems for patient-level decisions.

Take-home messages

A repeatedly measured biomarker should not go straight into a Cox model: measurement error, gaps between visits and dropout bias the result.

Joint models fit the marker and the event together through the patient's own random effects, and give predictions that update at every visit.

Joint models now scale to health-record cohorts, allow events to change the marker, and let risk follow non-linear shapes.

References (1)

Cox DR. Regression models and life-tables. J R Stat Soc B 1972; 34: 187–220.

Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. JASA 1958; 53: 457–481.

Laird NM, Ware JH. Random-effects models for longitudinal data. Biometrics 1982; 38: 963–974.

Rubin DB. Inference and missing data. Biometrika 1976; 63: 581–592.

Diggle P, Kenward MG. Informative drop-out in longitudinal data analysis. Appl Stat 1994; 43: 49–93.

Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data, 2nd ed. Wiley; 2002.

Prentice RL. Covariate measurement errors and parameter estimation in a failure time regression model. Biometrika 1982; 69: 331–342.

DeGruttola V, Tu XM. Modelling progression of CD4-lymphocyte count and its relationship to survival time. Biometrics 1994; 50: 1003–1014.

Faucett CL, Thomas DC. Simultaneously modelling censored survival data and repeatedly measured covariates: a Gibbs sampling approach. Stat Med 1996; 15: 1663–1685.

Wulfsohn MS, Tsiatis AA. A joint model for survival and longitudinal data measured with error. Biometrics 1997; 53: 330–339.

Henderson R, Diggle P, Dobson A. Joint modelling of longitudinal measurements and event time data. Biostatistics 2000; 1: 465–480.

Tsiatis AA, Davidian M. Joint modeling of longitudinal and time-to-event data: an overview. Stat Sin 2004; 14: 809–834.

Ibrahim JG, Chen MH, Sinha D. Bayesian methods for joint modeling of longitudinal and survival data. Stat Sin 2004; 14: 863–883.

Brown ER, Ibrahim JG, DeGruttola V. A flexible B-spline model for multiple longitudinal biomarkers and survival. Biometrics 2005; 61: 64–73.

References (2)

Chi YY, Ibrahim JG. Joint models for multivariate longitudinal and multivariate survival data. Biometrics 2006; 62: 432–445.

Elashoff RM, Li G, Li N. An approach to joint analysis of longitudinal measurements and competing risks failure time data. Stat Med 2007; 26: 2813–2835.

van Houwelingen HC. Dynamic prediction by landmarking in event history analysis. Scand J Stat 2007; 34: 70–85.

Rizopoulos D. JM: an R package for the joint modelling of longitudinal and time-to-event data. J Stat Softw 2010; 35(9): 1–33.

Rizopoulos D. Dynamic predictions and prospective accuracy in joint models. Biometrics 2011; 67: 819–829.

Rizopoulos D. Joint Models for Longitudinal and Time-to-Event Data. CRC Press; 2012.

Proust-Lima C et al. Joint latent class models for longitudinal and time-to-event data: a review. Stat Methods Med Res 2014; 23: 74–90.

Hickey GL et al. Joint modelling of time-to-event and multivariate longitudinal outcomes. BMC Med Res Methodol 2016; 16: 117.

Hickey GL et al. joineRML: a joint model and software package. BMC Med Res Methodol 2018; 18: 50.

Papageorgiou G et al. An overview of joint modeling of time-to-event and longitudinal outcomes. Annu Rev Stat Appl 2019; 6: 223–240.

Lee C, Yoon J, van der Schaar M. Dynamic-DeepHit. IEEE Trans Biomed Eng 2020; 67: 122–133.

Li S et al. Efficient algorithms for a semiparametric joint model: massive biobank data. Comput Math Methods Med 2022; 1362913.

Bhattacharjee A, Rajbongshi BK, Vishwakarma GK. jmBIG. BMC Med Res Methodol 2024; 24: 172.

Bhattacharjee A. JMbdirect: joint modeling with semi-parametric link functions for bidirectional feedback. BMC Med Res Methodol 2026; 26: 230.

Rizopoulos D, Hatfield LA, Carlin BP, Takkenberg JJM. Combining dynamic predictions using Bayesian model averaging. JASA 2014; 109: 1385–1397.

Rizopoulos D. The R package JMbayes. J Stat Softw 2016; 72(7): 1–46.

Tomer A et al. Personalized schedules for surveillance of low-risk prostate cancer patients. Biometrics 2019; 75: 153–162.

Backup: prediction accuracy of the six models

AUC at 5 years (higher is better) Integrated Brier score, 0–5 years (lower is better) Linear, no feedback Linear, feedback Spline, no feedback Spline, feedback Neural, no feedback Neural, feedback 0.700.750.800.85 0.130.140.150.160.17 0.720.750.820.850.810.84 0.1680.1560.1410.1340.1430.136

AUC at 5 years: probability that a patient with the event by year 5 receives a higher predicted risk than one without it.

Integrated Brier score: mean squared difference between predicted survival and observed status, averaged over 0–5 years.

Coverage of 95% intervals: 0.88 for linear without feedback, 0.95 for spline with feedback.

n = 500, competing events, piecewise baseline hazard; means over 1,000 replicates. Blue: no feedback term; orange: feedback term included. Source: Bhattacharjee, BMC Med Res Methodol 2026;26:230, Table 6.

Backup: calibration and comparison in figures

Calibration slope βcal (ideal 1) and intercept αcal AUC at 5 years on identical data 0.700.800.901.00 0.600.700.800.90 Linear, no feedbackLinear, feedbackSpline, no feedbackSpline, feedbackNeural, no feedbackNeural, feedback Two-stagejoineRMLJMbayes2JMbdirect 0.740.820.930.990.920.98 α 0.18α 0.12α 0.06α 0.02α 0.07α 0.03 0.670.710.730.85 Comparators: linear link, no feedback; JMbdirect: spline link with feedback

n = 500, competing risks, non-linear true association, moderate feedback, landmark t = 0, horizon u = 5; means over 1,000 replicates with the same train and test splits. Source: Bhattacharjee, BMC Med Res Methodol 2026;26:230, Tables 5 and 9.

1 / 36