Development and evaluation of methods to examine sources of heterogeneity in DIF and response shift when analyzing patient-reported outcome data
Subjective concepts such as fatigue, anxiety, or quality of life are increasingly being used in biomedical research. These unobservable subjective concepts are measured by means of questionnaires, usually self-reported by patients. However, the analysis of these data can be challenging. Indeed, the direct comparison of questionnaire scores between different groups of individuals or over time assumes that all individuals interpret the questions in the same way and that this interpretation does not change over time. Yet, this assumption may be unmet. For example, some individuals may interpret the measured concept (and the associated items) differently because of their cultural, environmental, and personal characteristics, but also because of the experiences they went through.
This phenomenon is known as differential item functioning (DIF). In addition, following a major health event, some individuals may also change their interpretation of the items composing the questionnaire (response shift, RS). Both DIF and RS can threaten the accurate interpretation of patient-reported outcome measures. In addition, they are also phenomena that deserve to be studied. This thesis aims to propose and evaluate, by simulations, methodological developments to study the sources of heterogeneity of these two phenomena. This manuscript is composed of three main works: (i) the development of a method aiming at individualizing RS detection, (ii) the development of a method for detecting DIF in the presence of two binary covariates, and (iii) the study of the psychometric properties of the post-traumatic growth inventory (a phenomenon potentially related to RS)
Application of Rasch Measurement Theory to clinical research: Practical and theoretical aspects of a metrological approach for the evaluation of clinical trial endpoints
Metrology benefited to physical measures, on which most clinical trial endpoints are based on. As Patient-Reported Outcomes (PROs) measures are more and more used in clinical trials, it is urgent that they reach the same level of credibility and interpretability than traditional trial endpoints. The objective of this work was thus to explore how metrology can benefit to the evaluation of treatments as per the PRO endpoints from clinical trials. The Rasch model has previously shown numerous advantages in fulfilling metrological requirements. A first step of this work was to conceptualize the trial in the case of a PRO endpoint through a scientific model, based on metrological vocabulary: the clinical trial as a measurement system. From this, we list and categorize measurement uncertainty sources in clinical trials, in order to improve its expression with the trial PRO results. Finally, we present the results from a simulation study which suggest that calibration of PRO measures does not negatively impact the results from the trial. Proposals to improve measurement of patient experience in clinical trials are then made based on this work.
Analysis of the place of phase IV studies for drug risk assessment in vulnerable populations
During clinical development, the risk to the patient is monitored. However, when a health product is placed on the market, knowledge of its benefit-risk balance is only fragmentary. The post-marketing surveillance system, which makes it possible to update this balance, is based essentially on spontaneous reporting of adverse effects. What is the need for phase IV studies, in addition to clinical trial data and pharmacovigilance evaluation? The objective of our research is to study the complementary approach of phase IV studies and their interest in relation to phase III studies and pharmacovigilance data, by focusing on an example of a population that is little studied in clinical trials, the elderly, and old drugs, benzodiazepines and an effect occurring during long-term use, drug dependence. Our work shows that post-marketing studies are essential to know the risk and dependence profile of a drug, which evolves throughout its "life". Predicting possible situations of dependence, detour, etc. requires a good evaluation of the benefit-risk balance not only for medical use in the MA, but also in all off-label use situations, including non-medical use. These data are essential to understand the obstacles to the application of recommendations or regulations on proper use.
Discover older theses
Applicability of randomised trials results in general medicine
Many clinical practice guidelines are based on randomised controlled trials performed in secondary or tertiary care settings and general practitioners questioned their relevance for primary care patients.
The first objective was to compare the intervention effect estimates between randomised controlled trials performed in primary care and randomised controlled trials performed in secondary or tertiary care settings by using a meta-epidemiological approach. There was no difference in intervention effect estimates between the two types of randomised trials with a ratio of odds ratio of 0.98 (95% confidence interval 0.88 to 1.08). Nevertheless, the main medical fields encountered in this study were not fully representative of the medical conditions encountered in primary care with many studies on psychiatry or addictology (38.2%) or pneumology (13.2%) and very few in endocrinology or cardiovascular diseases.
The second objective was to compare the characteristics of type 2 diabetes patients of general practices to those included in the randomised trials on which clinical practice guidelines are based. Primary care patients differed from patients included in randomised trials in many important aspects. They were older (mean±standard deviation 68.8 ±1.1 years vs 59.9 years [standardised difference 0.8]), had higher BMI (31.5 (6.93) kg/m2 vs 28.2kg/m2 [standardised difference 0.48]). They also had more hypoglycemic (80.7% vs 45.4% [standardised difference 0.89] than randomised trials patients, but less cardiovascular history (myocardial infarction: 7.6% vs 23.1% [standardised difference -1.14]).
Our results show, that there is a need for study including a wider range of patients to ensure a better generalisability of the results and improve the use of study findings to daily practice.
Methodological and statistical aspects of within-person randomized trials evaluating a topical treatment in dermatology
To assess a topical treatment in dermatology, we may use a within-person (also called split-body) design. Randomization units are no longer patients but rather lesions or body sites. Experimental and control treatments are then applied simultaneously on the different lesions. This study design reduces the inter-observation variability, and therefore the number of patients to be included in the trial. However, this study design presents methodological constraints, notably a risk of intergroup contamination, named the “carry across effect”. The objective of this PhD thesis was to study the methodological and statistical aspects of within-person randomized clinical trials. The first part of this work presents a synthesis of the methodological aspects to be considered when using this design. This work was conducted by a group associating dermatologists and biostatisticians. In the second part, we performed a methodological review of within-person randomized trials published between 2017 and 2021. The methodological aspects identified in the first work were assessed and discussed in this review. The third work focused on the statistical analysis. Using a Monte-Carlo simulation study, we compared the statistical properties of several analysis methods (mixed models and generalized estimating equations). Finally, as an illustration, the last part presents a within-person randomized trial protocol for evaluating topical sirolimus in superficial lymphatic malformations. The within-person design has multiple methodological and statistical specificities that must be taken into account when planning and analyzing the study. Dermatologists and methodologists must also consider patient acceptability.
Characterization of social cognition abilities in behavioral addictions through the prism of gaming and gambling disorders
This thesis focused on social cognition (SC) in the two behavioral addictions (BA) included in international classifications: gaming (GmD) and gambling (GbD) disorders. Two systematic reviews exposed the scarcity of studies linking SC and BA. Nevertheless, studies included in the reviews suggested the presence of an alteration of some of the SC’s components. Additionally, patient-reported outcomes confirmed the presence of interpersonal difficulties. These elements demonstrated the necessity to explore the SC abilities of patients with Gmd or GbD, in order to improve knowledge on addictive processes and propose alternative treatments focused on these difficulties. This thesis presented three studies on gamblers or gamers, whether or not they have an addiction. The first study showed specific allocation of attention toward social information in poker players compared to controls. The second study demonstrated a link between difficulties in identifying facial emotion and GbD and specificities in social metacognition in GmD. Finally, preliminary results of a study regarding GbD patients at the beginning of their treatment showed the importance of taking into account patient-reported outcomes in SC. Those results were discussed in light of clinical et scientific aspects, and put in perspective with future possible research.
Causal inference from observational data: development and applications for critical care
The increasing amount of observational data, especially in critical care, leads to the consideration of causal inference statistical methods. This thesis presents two works that highlight the current challenges of these statistical methods used on observational health data. The first analysis evaluates the impact of barbiturates in a population of trauma brain injured patients prospectively included in the open critical care cohort AtlanRéa. The evaluation of the impact of barbiturates was possible by respecting the assumptions of causal inference and by using a method based on propensity scores: inverse probability weighting.
Beyond the results of this analysis, which showed an increase in mortality in the group treated with barbiturates, we were faced with the problem of the violation of the positivity assumption. We then compared different statistical methods of causal inference in a context of violation of the positivity assumption, which can be associated with an extrapolation issue. The methods predicting the occurrence of the outcome are the most robust in these situations. In this context of accumulation of health data, a perspective of optimization of the use of statistical methods in the framework of causal inference will reside in the use of machine learning algorithms to avoid the problems of model specification.
Multilevel joint modelling of target lesions dynamics and survival : application to the prediction of the response to immunotherapy in bladder cancer
Treatment evaluation in oncology relies on time-to-death and longitudinal measurements of the Sum of Longest Diameters (SLD) of target lesions. Both processes and their association can be analyzed together using a nonlinear joint model. However, using a composite marker such as SLD neglects the heterogeneity in lesion dynamics, which might be exacerbated under immunotherapy.
The main objective of this PhD was to develop multilevel nonlinear joint models of tumor dynamics and their impact on survival, to better characterize all the source of variability in the response to treatment. We relied on data from a phase 2 (IMvigor210) and a phase 3 (IMvigor211) clinical trials of 300 and 900 advanced or metastatic Urothelial Carcinoma (UC) patients, treated with atezolizumab immune checkpoint inhibitor. In a first nonlinear joint model, we showed the impact of tumor location on its dynamics and association with survival. In particular, the liver lesions dynamics was strongly associated with the risk of death as compared to other location. Then, we showed the ability of HMC Bayesian algorithm implemented in Stan software to provide unbiased and precise estimation of the parameters of a nonlinear joint model of SLD and survival, with reasonable sensitivity to prior information. Finally, we developed a Bayesian hierarchical joint model of individual lesions and survival. An additional level of random effect was integrated, specific to the lesion, to quantify the inter-lesions variability under immunotherapy. Using individual dynamic prediction approaches, we showed the benefit of the individual lesions follow-up to identify most at risk patient as compared to SLD follow-up.
This work paves the way for a better understanding of the inter and intra-patient variability in response to new immunotherapy treatments.
Management of complex patients in Pediatric Odontology: Benefit / Risk of practices and improvement pathways, focus on EMONO
The widespread use of MEOPA in dental practices is relatively recent. A review of practices and the legal framework related to the use of nitrous oxide in pediatric odontology seems so particularly relevant. In addition, the emergence of derived uses, outside the context of care or during, leads to increased vigilance towards it. It therefore seems appropriate to identify the effects felt and sought by children during inhalation of MEOPA in the context of dental care. During this thesis, we recalled the context of French use of nitrous oxide and then analyzed it against the benefit / risk of the various other sedations used in pediatric odontology. We then developed a presentation of the different works carried out, their materials and methods and their results. The whole was then discussed and future prospects considered