arXiv subjects

Daniel Commenges

Publications and source records attributed to Daniel Commenges.

At least 19 recordsLinked to original sources

arXiv2020

Robust and Efficient Optimization Using a Marquardt-Levenberg Algorithm with R Package marqLevAlg

Implementations in R of classical general-purpose algorithms for local optimization generally have two major limitations which cause difficulties in applications to complex problems: too loose convergence criteria and too long calculation time. By relying on a Marquardt-Levenberg algorithm (MLA), a Newton-like method particularly robust for solving local optimization problems, we provide with marqLevAlg package an efficient and general-purpose local optimizer which (i) prevents convergence to saddle points by using a stringent convergence criterion based on the relative distance to minimum/maximum in addition to the stability of the parameters and of the objective function; and (ii) reduces the computation time in complex settings by allowing parallel calculations at each iteration. We demonstrate through a variety of cases from the literature that our implementation reliably and consistently reaches the optimum (even when other optimizers fail), and also largely reduces computational time in complex settings through the example of maximum likelihood estimation of different sophisticated statistical models.

stat.CO↗

arXiv2019

Effects of interventions and optimal strategies in the stochastic system approach to causality

We consider the problem of defining the effect of an intervention on a time-varying risk factor or treatment for a disease or a physiological marker; we develop here the latter case. So, the system considered is $(Y,A,C)$, where $Y=(Y_t)$, is the marker process of interest, $A=A_t$ the treatment. A realistic case is that the treatment can be changed only at discrete times. In an observational study the treatment attribution law is unknown; however, the physical law can be estimated without knowing the treatment attribution law, provided a well-specified model is available. An intervention is specified by the treatment attribution law, which is thus known. Simple interventions will simply randomize the attribution of the treatment; interventions that take into account the past history will be called "strategies". The effect of interventions can be defined by a risk function $R^{\intr}=\Ee_{\intr}[L(\bar Y_{t_J}, \bar A_{t_{J}},C)]$, where $L(\bar Y_{t_J}, \bar A_{t_{J}},C)$ is a loss function, and contrasts between risk functions for different strategies can be formed. Once we can compute effects for any strategy, we can search for optimal or sub-optimal strategies; in particular we can find optimal parametric strategies. We present several ways for designing strategies. As an illustration, we consider the choice of a strategy for containing the HIV load below a certain level while limiting the treatment burden. A simulation study demonstrates the possibility of finding optimal parametric strategies.

stat.ME↗

arXiv2019

Causality without potential outcomes and the dynamic approach

Several approaches to causal inference from observational studies have been proposed. Since the proposal of Rubin (1974) many works have developed a counterfactual approach to causality, statistically formalized by potential outcomes. Pearl (2000) has put forward a theory of structural causal models which gives an important role to graphical models and do not necessarily use potential outcomes. On the other hand, several authors have developed a dynamical approach in line with Granger (1969). We analyze prospective and retrospective causal questions and their different modalities. Following Dawid (2000) we develop criticisms about the potential outcome approach and we show that causal effects can be estimated without potential outcomes: in particular direct computation of the marginal effect can be done by a change of probability measure. Finally, we highlight the need to adopt a dynamic approach to causality through two examples, "truncation by death" and the "obesity paradox".

stat.ME↗

arXiv2018

Dynamic Modelling of Multivariate Dimensions and Their Temporal Relationships using Latent Processes: Application to Alzheimer's Disease

Alzheimer's disease gradually affects several components including the cerebral dimension with brain atrophies, the cognitive dimension with a decline in various functions and the functional dimension with impairment in the daily living activities. Understanding how such dimensions interconnect is crucial for AD research. However it requires to simultaneously capture the dynamic and multidimensional aspects, and to explore temporal relationships between dimensions. We propose an original dynamic structural model that accounts for all these features. The model defines dimensions as latent processes and combines a multivariate linear mixed model and a system of difference equations to model trajectories and temporal relationships between latent processes in finely discrete time. Dimensions are simultaneously related to their observed (possibly multivariate) markers through nonlinear equations of observation. Parameters are estimated in the maximum likelihood framework enjoying a closed form for the likelihood. We demonstrate in a simulation study that this dynamic model in discrete time benefits the same causal interpretation of temporal relationships as models defined in continuous time as long as the discretization step remains small. The model is then applied to the data of the Alzheimer's Disease Neuroimaging Initiative. Three longitudinal dimensions (cerebral anatomy, cognitive ability and functional autonomy) measured by 6 markers are analyzed and their temporal structure is contrasted between different clinical stages of Alzheimer's disease. Keywords: causality, difference equations, latent process, longitudinal data, mixed models, multivariate data.

stat.AP↗

arXiv2016

Dealing with death when studying disease or physiological marker: the stochastic system approach to causality

The stochastic system approach to causality is applied to situations where the risk of death is not negligible. This approach grounds causality on physical laws, distinguishes system and observation and represents the system by multivariate stochastic processes. The particular role of death is highlighted, and it is shown that local influences must be defined on the random horizon of time of death. We particularly study the problem of estimating the effect of a factor $V$ on a process of interest $Y$, taking death into account. We unify the cases where $Y$ is a counting process (describing an event) and the case where $Y$ is quantitative; we examine the case of observations in continuous and discrete time and we give a typology of cases where the mechanism leading to incomplete data can be ignored. Finally, we give an example of a situation where we are interested in estimating the effect of a factor (blood pressure) on cognitive ability in elderly.

stat.ME↗

arXiv2016

Modeling CD4+ T cells dynamics in HIV-infected patients receiving repeated cycles of exogenous Interleukin 7

Combination Antiretroviral Therapy (cART) succeeds to control viral replication in most HIV infected patients. This is normally followed by a reconstitution of the CD4$^+$ T cells pool; however, this does not happen for a substantial proportion of patients. For these patients, an immunotherapy based on injections of Interleukin 7 (IL-7) has been recently proposed as a co-adjutant treatment in the hope of obtaining long-term reconstitution of the T cells pool. Several questions arise as to the long-term efficiency of this treatment and the best protocol to apply. We develop a model based on a system of ordinary differential equations and a statistical model of variability and measurement. We can estimate key parameters of this model using the data from INSPIRE, INSPIRE 2 $\&$ INSPIRE 3 trials. In all three studies, cycles of three injections have been administered; in the last two studies, for the first time, repeated cycles of exogenous IL-7 have been administered. Our aim was to estimate the possible different effects of successive injections in a cycle, to estimate the effect of repeated cycles and to assess different protocols. The use of dynamical models together with our complex statistical approach allow us to analyze major biological questions. We found a strong effect of IL-7 injections on the proliferation rate; however, the effect of the third injection of the cycle appears to be much weaker than the first ones. Also, despite a slightly weaker effect of repeated cycles with respect to the initial one, our simulations show the ability of this treatment of maintaining adequate CD4$^+$ T cells count for years. We were also able to compare different protocols, showing that cycles of two injections should be sufficient in most cases. %Finally, we also explore the possibility of adaptive protocols.

stat.AP↗

arXiv2015

Information Theory and Statistics: an overview

We give an overview of the role of information theory in statistics, and particularly in biostatistics. We recall the basic quantities in information theory; entropy, cross-entropy, conditional entropy, mutual information and Kullback-Leibler risk. Then we examine the role of information theory in estimation theory, where the log-klikelihood can be identified as being an estimator of a cross-entropy. Then the basic quantities are extended to estimators, leading to criteria for estimator selection, such as Akaike criterion and its extensions. Finally we investigate the use of these concepts in Bayesian theory; the cross-entropy of the predictive distribution can be used for model selection; a cross-validation estimator of this cross-entropy is found to be equivalent to the pseudo-Bayes factor.

math.ST↗

arXiv2014

Inference with penalized likelihood

This work studies the statistical properties of the maximum penalized likelihood approach in a semi-parametric framework. We recall the penalized likelihood approach for estimating a function and review some asymptotic results. We investigate the properties of two estimators of the variance of maximum penalized likelihood estimators: sandwich estimator and a Bayesian estimator. The coverage rates of confidence intervals based on these estimators are studied through a simulation study of survival data. In a first simulation the coverage rates for the survival function and the hazard function are evaluated. In a second simulation data are generated from a proportional hazard model with covariates. The estimators of the variances of the regression coefficients are studied. As for the survival and hazard functions, both sandwich and Bayesian estimators exhibit relatively good properties, but the Bayesian estimator seems to be more accurate. As for the regression coefficients, we focussed on the Bayesian estimator and found that it yielded good coverage rates.

math.ST↗

arXiv2012

A universal approximate cross-validation criterion and its asymptotic distribution

A general framework is that the estimators of a distribution are obtained by minimizing a function (the estimating function) and they are assessed through another function (the assessment function). The estimating and assessment functions generally estimate risks. A classical case is that both functions estimate an information risk (specifically cross entropy); in that case Akaike information criterion (AIC) is relevant. In more general cases, the assessment risk can be estimated by leave-one-out crossvalidation. Since leave-one-out crossvalidation is computationally very demanding, an approximation formula can be very useful. A universal approximate crossvalidation criterion (UACV) for the leave-one-out crossvalidation is given. This criterion can be adapted to different types of estimators, including penalized likelihood and maximum a posteriori estimators, and of assessment risk functions, including information risk functions and continuous rank probability score (CRPS). This formula reduces to Takeuchi information criterion (TIC) when cross entropy is the risk for both estimation and assessment. The asymptotic distribution of UACV and of a difference of UACV is given. UACV can be used for comparing estimators of the distributions of ordered categorical data derived from threshold models and models based on continuous approximations. A simulation study and an analysis of real psychometric data are presented.

math.ST↗

arXiv2012

The stochastic system approach to causality with a view toward lifecourse epidemiology

The approach of causality based on physical laws and systems is revisited. The issue of "levels", the relevance to epidemiology and the definition of effects are particularly developed. Moreover it is argued that this approach that we call the stochastic system approach is particularly well fitted to study lifecourse epidemiology. A hierarchy of factors is described that could be modeled using a suitable multivariate stochastic process. To illustrate this approach, a conceptual model for coronary heart disease mixing continuous and discrete state-space processes is proposed.

math.ST↗

arXiv2010

Extending The Range of Application of Permutation Tests: the Expected Permutation p-value Approach

The limitation of permutation tests is that they assume exchangeability. It is shown that in generalized linear models one can construct permutation tests from score statistics in particular cases. When under the null hypothesis the observations are not exchangeable, a representation in terms of Cox-Snell residuals allows to develop an approach based on an expected permutation p-value (Eppv); this is applied to the logistic regression model. A small simulation study and an illustration with real data are given.

stat.ME↗

arXiv2009

A general definition of influence between stochastic processes

We extend the study of weak local conditional independence (WCLI) based on a measurability condition made by Commenges and Gégout-Petit (2009) to a larger class of processes that we call D'. We also give a definition related to the same concept based on certain likelihood processes, using the Girsanov theorem. Under certain conditions, the two definitions coincide on D'. These results may be used in causal models in that we define what may be the largest class of processes in which influences of one component of a stochastic process on another can be described without ambiguity. From WCLI we can contruct a concept of strong local conditional independence (SCLI). When WCLI does not hold, there is a direct influence while when SCLI does not hold there is direct or indirect influence. We investigate whether WCLI and SCLI can be defined via conventional independence conditions and find that this is the case for the latter but not for the former. Finally we recall that causal interpretation does not follow from mere mathematical definitions, but requires working with a good system and with the true probability.

math.ST↗

arXiv2008

Statistical models, likelihood, penalized likelihood and hierarchical likelihood

We give an overview of statistical models and likelihood, together with two of its variants: penalized and hierarchical likelihood. The Kullback-Leibler divergence is referred to repeatedly, for defining the misspecification risk of a model, for grounding the likelihood and the likelihood crossvalidation which can be used for choosing weights in penalized likelihood. Families of penalized likelihood and sieves estimators are shown to be equivalent. The similarity of these likelihood with a posteriori distributions in a Bayesian approach is considered.

math.ST↗

arXiv2008

Likelihood for generally coarsened observations from multi-state or counting process models

We consider first the mixed discrete-continuous scheme of observation in multistate models; this is a classical pattern in epidemiology because very often clinical status is assessed at discrete visit times while times of death or other events are observed exactly. A heuristic likelihood can be written for such models, at least for Markov models; however, a formal proof is not easy and has not been given yet. We present a general class of possibly non-Markov multistate models which can be represented naturally as multivariate counting processes. We give a rigorous derivation of the likelihood based on applying Jacod's formula for the full likelihood and taking conditional expectation for the observed likelihood. A local description of the likelihood allows us to extend the result to a more general coarsening observation scheme proposed by Commenges & Gégout-Petit. The approach is illustrated by considering models for dementia, institutionalization and death.

math.ST↗

arXiv2007

A general dynamical statistical model with possible causal interpretation

We develop a general dynamical model as a framework for possible causal interpretation. We first state a criterion of local independence in terms of measurability of processes involved in the Doob-Meyer decomposition of stochastic processes, as in Aalen (1987); then we define direct and indirect influence. We propose a definition of causal influence using the concepts of ``physical system''. This framework makes it possible to link descriptive and explicative statistical models, and encompasses quantitative processes and events. One of the features of this paper is the clear distinction between the model for the system and the model for the observation. We give a dynamical representation of a conventional joint model for HIV load and CD4 counts. We show its inadequacy to capture causal influences while on the contrary known mechanisms of HIV infection can be expressed directly through a system of differential equations.

math.ST↗

arXiv2007

Bivariate linear mixed models using SAS proc MIXED

Bivariate linear mixed models are useful when analyzing longitudinal data of two associated markers. In this paper, we present a bivariate linear mixed model including random effects or first-order auto-regressive process and independent measurement error for both markers. Codes and tricks to fit these models using SAS Proc MIXED are provided. Limitations of this program are discussed and an example in the field of HIV infection is shown. Despite some limitations, SAS Proc MIXED is a useful tool that may be easily extendable to multivariate response in longitudinal studies.

stat.AP↗

arXiv2007

A Latent Process Model for Dementia and Psychometric Tests

We jointly model longitudinal values of a psychometric test and diagnosis of dementia. The model is based on a continuous-time latent process representing cognitive ability. The link between the latent process and the observations is modeled in two phases. Intermediate variables are noisy observations of the latent process; scores of the psychometric test and diagnosis of dementia are obtained by categorizing these intermediate variables. We propose maximum likelihood inference for this model and we propose algorithms for performing this task. We estimated the parameters of such a model using the data of the five-year follow-up of the PAQUID study. In particularThis analysis yielded interesting results about the effect of educational level on both latent cognitive ability and specific performance in the mini mental test examination. The predictive ability of the model is illustrated by predicting diagnosis of dementia at the eight-year follow-up of the PAQUID study bsed on the information of the first five years.

math.ST↗

arXiv2006

A Newton-Like Algorithm for Likelihood Maximization: The Robust-Variance Scoring Algorithm

This article studies a Newton-like method already used by several authors but which has not been thouroughly studied yet. We call it the robust-variance scoring (RVS) algorithm because the main version of the algorithm that we consider replaces minus the Hessian of the loglikelihood used in the Newton-Raphson algorithm by a matrix $G$ which is an estimate of the variance of the score under the true probability, which uses only the individual scores. Thus an iteration of this algorithm requires much less computations than an iteration of the Newton-Raphson algorithm. Moreover this estimate of the variance of the score estimates the information matrix at maximum. We have also studied a convergence criterion which has the nice interpretation of estimating the ratio of the approximation error over the statistical error; thus it can be used for stopping the iterative process whatever the problem. A simulation study confirms that the RVS algorithm is faster than the Marquardt algorithm (a robust version of the Newton-Raphson algorithm); this happens because the number of iterations needed by the RVS algorithm is barely larger than that needed by the Marquardt algorithm while the computation time for each iteration is much shorter. Also the coverage rates using the matrix $G$ are satisfactory.

math.ST↗