Reviewed International Conferences

[INVITED]

Estimation and validation in population modeling: comparison of simulations and model-based approaches for predictions.

Prague M.

Published in :

Summer Sim, Montreal, Canada. (July 24th-27th 2016).

Abstract

...

Bibtex

@conference{prague2016Summersim, title={Estimation and validation in population modeling: comparison of simulations and model-based approaches for predictions.}, author={Prague, M.}, year={2016}, booktitle={Summer Sim, Montreal, Canada.} }

[INVITED]

Inverse-probability-weighted semi-parametric estimation of treatment effect in cluster randomized trials with missing data

Prague, M.

Published in :

Society for clinical trials, Montreal, Canada. (May 15th-18th 2016).

Abstract

...

Bibtex

@conference{prague2016SCT, title={Inverse-probability-weighted semi-parametric estimation of treatment effect in cluster randomized trials with missing data }, author={Prague, M.}, year={2016}, booktitle={Society for clinical trials, Montreal, Canada.} }

Estimating the Marginal Effect of Interventions to Reduce Spread of Communicable Diseases: What can be gained from Contact Network Information?.

M. Prague, P. Staples, JP. Onnela, E. Tchetgen Tchetgen and V. De Gruttola.

Published in :

ENAR, Austin, USA. (March 6th-9th 2016).

Abstract

We develop methods to leverage information on contact networks to improve efficiency and validity of analysis of data from cluster-randomized trials (CRTs) designed to investigate the impact of prevention interventions. These methods make use of flexible approaches for network generation that are based on degree-corrected stochastic block models and that are potentially applicable for a wide range of transmissible diseases. A semi-parametric doubly robust estimator is useful for estimating the marginal effect of the intervention while adjusting for imbalance in covariates and missing information. Simulations of CRTs show that adjusting for network features increases efficiency, in some cases, by more than 30% and leads to considerable increase of the statistical power while retaining coverage near the nominal value of 95%. We identify the major network features driving these improvements, such as the number of first- and second-degree contacts, the number of closed sub-communities, and the shortest path to an infected individual. Because network information is difficult to collect and networks are likely to be modeled with considerable uncertainty, we perform a sensitivity analysis evaluating the degree to which it is profitable to use partial or approximate information. Adjustments for covariates measured with error based on regression calibration are investigated.

Bibtex

@conference{prague2016ENAR, title={Estimating the Marginal Effect of Interventions to Reduce Spread of Communicable Diseases: What can be gained from Contact Network Information?}, author={Prague, M. and Staples, P. and Onnela, JP. and Tchetgen Tchetgen, E. and De Gruttola, V.}, year={2016}, booktitle={ENAR, Austin, USA.} }

Accounting for Informative Missingness, Interaction and Interference in Cluster Randomized Trials.

M. Prague, R. Wang, E. Tchetgen Tchetgen, A. Stephens and V. De Gruttola

Published in :

Society for clinical trials, Washington DC, USA. (March 17th-20th 2015).

Abstract

We investigate the use of semiparametric methods for estimation of intervention effects on clustered outcomes motivated by cluster-randomized trials (CRTs) of effectiveness of HIV prevention interventions. We review Inverse probability weighted (IPW) and Augmented Generalized Estimating Equation (A-GEE) methods for dealing with informative missingness and consider data that are missing at random (MAR), interactions between treatment and other baseline covariates, and complex correlation structures. As neither method alone corrects for imbalance in covariates, we propose an augmented IPW (AIPW-GEE) that weights by the inverse of the probability of observation and allows different regression models for explanatory covariates in each intervention group. We demonstrate reduction of bias and improvement in efficiency by simulation studies and by investigation of large sample properties. We also assess the impact of interference covariates; interference arises when the observation and outcome of an individual may depend on covariates of other individuals. Through investigation of the choice of the correlation structure in presence of interference, we conclude that biases arise unless a working independence structure is used for the AIPW-GEE, regardless to the true correlation structure. An R package (augmentedIPW) implementing this method both for continuous and binary outcome will be described. The method is illustrated using data from CRT of an HIV and sexually transmitted infection risk reduction-intervention in South Africa. Using AIPW-GEE instead of standard GEE strengthens the magnitude and the significance of the intervention effect for some outcomes such as the percentage of protected sexual intercourses during the last three months. Modification is not unidirectional, for some other outcomes, effect was reduced. In summary, we show that adjusting for the presence of data that are MAR, interactions between treatment and baseline covariates, and interference can improve inference.

Bibtex

@conference{prague2015SCT, title={Accounting for Informative Missingness, Interaction and Interference in Cluster Randomized Trials.}, author={Prague, M. and Wang, R. and Tchetgen Tchetgen, E. and Stephens, A. and De Gruttola, V.}, year={2015}, booktitle={Society for clinical trials, Washington DC, USA.} }

[INVITED]

Comparison of GEE-based methods in cluster-randomized trial with missing data when outcome depend on other patients covariates.

Prague M., R. Wang, E. Tchetgen Tchetgen and V. DeGruttola

Published in :

Joint Statistical Meeting, Boston, USA (August 2nd-7th 2014)

Abstract

We investigate the use of semiparametric regression for estimation of intervention effects on clustered outcomes motivated by cluster-randomized trials (CRTs) of effectiveness of HIV prevention. Interest lies in inference for the marginal parameter describing these effects in presence of data that are subject to being missing at random (MAR). We focus on complex correlation structure designs with patients nested in sub-clusters, and with outcomes and missingness patterns that may depend on other subjects’ covariates. Through simulation studies, we investigate the impact of the choice of the correlation structure in unweighted and inverse probability-weighted (IPW) GEE and demonstrate biases that arise unless a working independence structure is used for the GEE. To make use of knowledge of the correlations, we propose the use of an augmented term that depends on other patients’ covariates whose selection may be guided by the correlation structure; this term is intended to improve the efficiency of the estimation of the treatment effect.

Bibtex

@conference{prague2014JSMcluster, title={Comparison of GEE-based methods in cluster-randomized trial with missing data when outcome depend on other patients covariates.}, author={Prague, M. and Wang, R. and Tchetgen Tchetgen, E. and DeGruttola, V.}, year={2014}, booktitle={Joint statistical meeting, Boston, USA}}

From descriptive to mechanistic models to study causal effects : application to the e ect of HAART on CD4 count.

Prague M., D. Commenges, J.M. Gran, O. Aalen and R. Thiébaut

Published in :

Joint Statistical Meeting, Boston, USA (August 2nd-7th 2014)

Abstract

The problem of assessing the effect of a treatment on a marker in observational studies raises the difficulty that attribution of the treatment may depend on the observed marker values. This problem has been treated using marginal structural models relying on the counterfactual formalism. Another approach to causality is based on dynamical models where causal influence has been formalized in the framework of stochastic processes with linear increments models. This can be further extended in continuous time with mechanistic models, particularly based on ordinary differential equations with random effects, and allows incorporating biological knowledge. We show that a continuum can be established between descriptive and mechanistic modeling. Although inference in mechanistic models is challenging and requests specific methods, these models can yield more powerful and reliable results with qualitative interpretation such as mediation. The different approaches are illustrated by estimating the effect of highly active antiretroviral treatment (HAART) on CD4 count in a simulation study, then, in an observational study of HIV infected subjects (ANRS CO3 Aquitaine Cohort).

Bibtex

@conference{prague2014JSMmeca, title={From descriptive to mechanistic models to study causal effects : application to the effect of HAART on CD4 count.}, author={Prague, M. and Commenges, D. and Gran, J.M. and Aalen, O. and Thiébaut, R.}, year={2014}, booktitle={Joint Statistical Meeting, Boston, USA}}

Using mechanistic models to analyze the effect of interleukins 7 treatment in HIV infected patients

A.Jarne, R. Thiébaut, M. Prague and D.Commenges

Published in :

International Biometric Society, Florence, Italy. (July 6th-11th 2014)

Abstract

...

Bibtex

@conference{jarbe2014IBS, title={Using mechanistic models to analyze the effect of interleukins 7 treatment in HIV infected patients}, author={Jarne, A and Prague, M. and Commenges, D. and Thiébaut, R.}, year={2014}, booktitle={International Biometric Society, Florence, Italy}}

Mechanistic versus marginal structural models for estimating the effect of HAART on CD4 counts.

D. Commenges, Prague M. and R. Thiébaut

Published in :

Medical Research Council Conference on Biostatistics, Cambridge, UK (April 24th-26th 2014)

Abstract

...

Bibtex

@conference{commenges2014MRC, title={Mechanistic versus marginal structural models for estimating the effect of HAART on CD4 counts}, author={Prague, M. and Commenges, D. and Thiébaut, R.}, year={2014}, booktitle={Medical Research Council Conference on Biostatistic, Cambridge, UK}}

From in vivo to in vitro quantifification of antiretroviral drugs effects based on dynamical models of HIV.

Prague M., D. Commenges and R. Thiébaut

Published in :

HIV Dynamics and evolution, Utrecht, Netherlands (May 8th-11th 2013).

Abstract

Population dynamics of HIV and CD4+ T cells can be modeled with Ordinary Differential Equations (ODE). We aim at quantifying thein vivo effect of combinations of antiretroviral drugs treatment (cARTs) by a function of the effects of the antiretroviral drugs (ARVs) in the combination. To estimate the ARVs effects we must have a large dataset and it is desirable to add external information to ensure identifiability. An adequate modeling of in vitro assays yields such information.
Recent single-round infectivity assays allowed quantifying the dose-response curves in vitro (Shen et al., Nat. Med., 2008): the instantaneous inhibitory potential (IIP) has been established as a measure of ARVs activity. The IIP of cARTs can be viewed as a function of ARV’s IIP based on known interactions (Jilek et al., Nat. Med., 2012). Bliss independence is a convenient assumption to build dynamical models. Random effects account for inter-individual variability of IIPs that may result from host and virus genomics (Sampah et al., PNAS, 2011). Finally, more flexibility is provided by estimating an in vitro to in vivo conversion factor. We used a Bayesian techniques for estimating the ARVs effects: cARTs effects follow by computation.
This approach is applied to a dataset of 350 patients from four clinical trials (ALBI, PUZZLE, PREDIZISTA, ZEPHIR). As a start, only the eight main ARVs in these trials are considered (AZT, 3TC, D4T, DDI, RTV, LPV, APV and DRV). First analysis show that we may rank cARTs in agreement with previously published studies results. We propose an extension of our method to the analysis of the effect of cARTs in HIV-1 infected patients followed in the Aquitaine cohort. Our approach opens the perspective of individualizing treatment, a step toward “personalized medicine”.

Bibtex

@conference{prague2013HIVDyn, title={From in vivo to in vitro quantifification of antiretroviral drugs effects based on dynamical models of HIV.}, author={Prague, M. and Commenges, D. and Thiébaut, R.}, year={2013}, booktitle={HIV Dynamics and evolution, Utrecht, Netherlands} }

Toward information synthesis with mechanistic models of HIV dynamics.

Prague M., Commenges D. and Thiébaut, R.

Published in :

33st Annual conference of the International society for Clinical Biostatistics, Bergen, Norway (August 21st-25th 2012)
Statistics in Health French research group days, Rennes, France. (Sept. 2012)

Abstract

Parameters in mechanistic models based on ODE (Ordinary Differential Equations) have an intrinsic meaning. Thus, HIV modelling should lead to similar estimated values among clinical trials for some parameters such as the virus proliferation rate even if patients' histories and treatments differ. In the perspective of optimizing treatment, we aim to build a model which forecasts the patient treatment response in several studies. To validate it, we will present, in a Bayesian framework, a methodology for combined estimation of parameters over several clinical trials.
We use the "Activated T cell model" with random effects on parameters. A pharmacodynamic function links the treatment dose to the effect of several antiretroviral drugs. To account for non-identifiability, a Bayesian approach allows introducing prior information using data from the literature. In view of the numerical complexity, we use a Maximum a Posteriori (MAP) estimator instead of classical MCMC. The EMRODE algorithm (Estimation in Models with Random effects based on Ordinary Differential Equations) allows computing the MAP. We analyse sequentially the different studies by taking as prior the updated posterior of previous analyses.
We applied the methodology on two clinical trials (Albi ANRS 070: n=150 untreated patients starting dual nucleosides therapy and Puzzle ANRS 104: n=40 heavily pre-treated patients starting salvage therapy). Initial separate analyses show good prediction abilities of the model and fair agreement between parameters posterior distributions among studies. Combined analyses improve the fits and the predictions.

Bibtex

@conference{prague2012ISCB, title={Toward information synthesis with mechanistic models of HIV dynamics.}, author={Prague, M. and Commenges, D. and Thiébaut, R.}, year={2012}, booktitle={33st Annual conference of the International society for Clinical Biostatistics, Bergen, Norway}}

Bayesian MAP Estimation in Models with Random effects based on Ordinary Differential Equations applied to Treatment Monitoring in HIV

Prague M. and Commenges D.

Published in :

Eurandom Workshop on Parameter Estimation for Dynamical Systems (PEDS II), Eindhoven, Neetherland (June 4th-5th 2012)

Abstract

Studies of dynamical models based on Ordinary Differential Equations (ODE) considerably enhanced the knowledge about the interaction between HIV and the immune system. Parameters in these models bring valuable information about cells birth and death rates in the population. Moreover, random effects to take into account inter-individual variability open the perspective of treatment individualization based on ODE equilibrium properties.
We use the “Activated T cell model” with random effects on several parameters. A pharmacodynamic function links the treatment dose to the effect of several antiretroviral drugs. We adopt a Bayesian approach because of problems in practical identifiability in such a model. Priors are elicitated so as to cover previously published evaluations. The numerical complexity leads us to choose a Maximum a Posteriori (MAP) estimation method rather than a full Bayesian estimation by MCMC. We will describe the NIMROD algorithm (Normal approximation Inference in Models with Random effects based on Ordinary Differential equations) based on a Newton-like algorithm proposed by Guedj et al. . Model prediction abilities will be illustrated on two clinical trials.
ODE equilibrium can be characterized by the basic reproductive number R0. When R0 ≤ 1, infection is controlled. To optimize a treatment, no cost function is necessary; we only have to control the probability that R0 is below one. A Bayesian algorithm such as Metropolis-Hastings allows us to calculate the a posteriori distribution of R0 for a specific observed patient. We will present an adaptive drug dose tuning algorithm, which converges toward the critical dose characterized by R0 = 1 when information about the patient increases.

Bibtex

@conference{prague2012PEDSII,
 title={Bayesian MAP Estimation in Models with Random effects based on Ordinary Differential Equations applied to Treatment Monitoring in HIV.}, author={Prague, M. and Commenges, D. }, year={2012}, booktitle={Eurandom Workshop on Parameter Estimation for Dynamical Systems (PEDS II), Eindhoven, Neetherland}}

Treatment monitoring of HIV infected patients : optimal drug dose control.

Prague M., Commenges D., Drylewicz J. and Thiébaut R.

Published in :

3rd Conference of the International Biometric Society Channel Network, Bordeaux, France (April 11th-13th 2011).

Abstract

Highly active anti-retroviral treatments lead to an undetectable HIV viral load in most patients, the virus however cannot be eradicated. Thus, the main issue is to reduce side effects of treatment that patients have to take for a lifetime. We tackle the problem of monitoring the treatment dose. We aim at finding the minimum dose that guarantees an undetectable viral load, we call it the critical dose.
We propose an adaptive treatment strategy. At different time points, biological measurements give additional information about the patient’s immune system (CD4 cell count) and the viral replication (viral load). This information is used to readjust the dose. Subsequently, the general idea is to observe how the patient will react to this readjusted dose and then tune it again in an iterative manner.
Our approach takes advantage of mechanistic models based on ordinary differential equations. We use the fact that such dynamical systems have a unique non-trivial equilibrium with a non-zero viral load: it is characterised by a reproductive number greater than one. Thus, if the reproductive number is below one infection would tend to extinction. Unlike usual control processes, our method does not depend on a cost function weighing side effects and antiviral efficiency.
We apply Bayesian theory to estimate the posterior distribution of the patient’s biological parameters according to the collected data. Priors are chosen as posteriors arising from the analysis of previous clinical trials data. We dynamically update the parameters with a Metropolis- Hastings algorithm. This allows us to produce realisations from the posterior distribution of the reproductive number. We apply a Robbins-Monro algorithm to tune the dose so as to have a high probability that this reproductive number is below one.
We prove a theoretical consistency of the strategy. Whatever the given initial drug dose, at every time point, there is a unique optimal dose for the patient to be prescribed. Our Metropolis- Hastings-Robbins-Monro algorithm converges to this optimal dose. Then, as further information is available due to data collection, the updated optimal dose approaches by above the critical dose. Therefore, we never under-estimate the drug dose and we avoid potential harm for the patient. We also performed a simulation study. For all simulated patients, the final tuned dose was close to the critical dose. The method seems to work not only asymptotically but also for a restricted number of observations.

Bibtex

@conference{prague2011ChannelNetwork, title={Treatment monitoring of HIV infected patients : optimal drug dose control.}, author={Prague, M. and Commenges, D. and Drylewicz, J. and Thiébaut, R.}, year={2011}, booktitle={3rd Conference of the International Biometric Society Channel Network, Bordeaux, France} }