The role of sampling in clinical trial design

Abstract

A treatment's recovery rate depends upon the percentage of clients who received the treatment and recovered. This rate is not logically interpretable as the personal probability of recovery of any individual client assigned to this treatment unless the rate is 0% or 100%. So clinical trials need to be designed to help us learn how to distinguish before treatment the sorts of clients who recover in response to each available form of treatment from those who do not. This requires our developing sufficiently comprehensive sampling of clients and client covariates as part of the design of clinical trials, which would be more likely and efficiently achieved were there centralized programmatic planning and coordination of the development of these aspects of clinical trial design.

RÉSUMÉ

Le taux de guérison d'un traitement dépend du pourcentage de clients qui reçoivent le traitement et qui guérissent. Ce taux n'est pas logiquement interprétable comme la probabilité de guérison d'un client individuel assigné à ce traitement tant que ce taux n'est pas 0% ou 100%. C'est pourquoi les essais cliniques doivent être conçus pour nous aider à distinguer avant traitement le type de clients qui guérissent en réponse à chaque type disponible de traitement de ceux qui ne guérissent pas. Cela demande de développer de manière suffisamment exhaustive l’échantillonnage des clients et des variables covariées comme partie du plan de recherche, ce qui pourrait être plus probablement et efficacement réalisé lorsque sont centralisées la planification programmatique et la coordination du développement de ces aspects de la conception de l'essai clinique.

Abstract

A taxa de recuperação de tratamento depende da percentagem de clientes que receberam o tratamento e se recuperaram. Esta taxa não é logicamente interpretável como a probabilidade de recuperação pessoal de qualquer cliente individual atribuído a este tratamento a menos que a taxa é de 0% ou 100%. Assim, os ensaios clínicos devem ser concebidos para nos ajudar a aprender a distinguir, antes do tratamento os tipos de clientes que se recuperam em resposta a cada forma de tratamento disponíveis e os que não fazem. Isso requer o nosso desenvolvimento suficientemente abrangente de amostragem de clientes e covariancia do cliente como parte da concepção dos ensaios clínicos, o que seria mais provável e eficientemente alcançado foram o planeamento programático centralizado e coordenação do desenvolvimento destes aspectos da concepção do ensaio clínico.

Abstract

Il tasso di guarigione a un trattamento dipende dalla percentuale dei pazienti che hanno svolto la terapia e sono guariti. Questo tasso non è interpretabile logicamente come la probabilità personale di guarigione di ogni paziente assegnato al trattamento, tranne che questo tasso sia 0% o 100%. Quindi i trial clinici devono essere strutturati per aiutarci ad imparare come distinguere prima del trattamento i tipi di pazienti che guariscono in relazione ad ogni forma di terapia disponibile da quelli che non migliorano. Questo richiede lo sviluppo di un campionamento sufficientemente globale dei pazienti e delle loro covariate come parte del disegno di ricerca dei trial clinici, che sarebbe più probabilmente ed efficientemente raggiunto se ci fosse una programmazione centralizzata e un coordinamento dello sviluppo di questi aspetti nel disegno di ricerca del trial clinico

Abstract

Keywords:

• aptitude-treatment interaction research
• outcome research
• statistical methodology
• client sampling
• client-covariate sampling

Notes

1. The essential clinical purpose in doing two-phase (input-outcome) RCTs is to resolve the issue of to what treatment it is best to assign each client. Most RCTs have so far been two-phase rather than multi-phase (input-process-outcome) studies. To resolve the equally important issue of how the therapist ought best to manage the next phase of treatment in the light of how the client has so far responded is the essential clinical purpose in doing multi-phase input-process-outcome studies, which cannot be done justice to here (see, e.g., Davison, 2000; Krause & Lutz, 2009b; Lambert et al., 2002; Lutz et al., 2009; Ruberg et al., 2010).

2. There is the logical possibility of there being persons who would recover if left untreated but not if given treatment j, and so an alternative kind of U, but this may not so obviously be an empirical possibility. For those who prefer the latter and so at least some plausible example, the following hypothetical situation (and also Barlow,2010) may be useful. Perhaps for some persons after a severe psychological trauma, enough time spent in non-stressful circumstances, and without any felt prospect of being subsequently subjected to stressful circumstances, there occurs an adequate diminution of post-traumatic stress disorder (whether this is by gradual forgetting, consistent extinction trials, repression, or however). Such a person would be a U. However, if for such persons the prompted remembering of the circumstances of the trauma (or of events in her or his life preceding it that might have contributed to making these circumstances so traumatic) interferes with this natural course of diminution, then this stressful intervention treatment (a treatment j) may prevent the prior sort of recovery but may not itself produce recovery. Such a person would be a U who when given treatment j would fail to recover because the treatment interfered with the processes of natural recovery and either provided no effective substitute for them or was itself impeded by them. If this treatment did provide such an effective substitute for and also was not impeded by such natural processes of recovery, then the person would be a T _j &U.

3. Whereas knowing these percentages and their underlying frequencies is crucial for public health purposes. Public health policy concerns the investment of public resources so as to minimize the social cost of ill health, which necessarily requires consideration of how many clients of what sorts have what maladies, how these numbers are trending over time and place, what treatments (including simply not intervening) are cost-effective enough for what sorts of clients with which of these maladies, and for what maladies and sorts of clients are new treatments most sorely needed to control the social cost of illness. If (a) enough persons do or are likely to suffer from a particular malady and their doing so has a social cost exceeded by few enough competing social problems in the light of society's present tolerance for social problems and resources for dealing with them, (b) a large enough percentage of these persons recover from this malady after receiving a particular treatment j who currently would not otherwise (T _j &~Us, T _j &~T _k s, etc.), and (c) the net marginal impact on total social costs of providing treatment j is favorable or at least socially tolerable, then investing public health resources in treatment j is socially worthwhile. Population surveys of the prevalence, incidence, outcome, and social cost of each malady when left untreated and also clinical outcome surveys of ordinary practice (rather than merely RCT estimated) cost-effectiveness of alternative available treatments for each malady are necessary for rationally developing public health policy, because only in the light of such information can clinical trials of new treatments for particular maladies be rationally prioritized for public health resources’ investment. Thus, considerable ambiguity as to the expectable cost-effectiveness for individual clients is (and simply has to be) tolerable for public health purposes but properly is not tolerable for clinical purposes. With client and client covariate sampling taken seriously the traditional RCT would be a proper tool for informing both public health and clinical policy

4. This ought not to be presumed to be simply a correlational matter, with each full-range covariate correlating with one or more outcome variables. It is also possible that configurations of several covariates’ levels are associated with particular levels (or configurations of levels) of the outcome variable(s) even though none of the individual predictor variables correlates at all with the outcome variable(s). In other words, the covariates may be quite irregularly associated with or interactive in their effects on outcomes (see Krause,2010; Krause & Howard, 2002) and so are a matter of non-linear association rather than linear correlation (see Lutz et al., 2005; Ruberg et al., 2010; Strobl et al., 2009; West & Thoemmes,2010).

5. Some comment is deserved on the issue of counterfactual or potential responses (see, e.g., Flanders, 2006; VanderWeele & Hernán, 2006; West & Thoemmes,2010), in that each client receives one particular treatment and not another, and so there is no way to tell empirically whether any given client would have recovered if that client had received the other treatment. This logical obstacle is routinely finessed statistically by making the assumption of sufficient relevant similarity (uniformity, exchangeability) between the groups of clients who receive different treatments (see, e.g., Dawid, 2000), which assumption holds at the mathematical limit (i.e., as a “statistically expected value”) on the basis of randomly assigning to its comparison groups whatever clients are available for inclusion in an RCT but is unjustified for any single RCT (Krause & Howard, 2003; West & Thoemmes,2010, p. 23). So, instead of focusing on individual clients, the logic of the RCT has focused on groups of clients, on each group's mean outcome. However, in order logically to bring the benefits of this maneuver down to the individual client level, it is necessary to consider distinct types of clients so that each client of each particular type responds to any given treatment in the same way that every other client of that type does, e.g., either all those of a particular type recover or none does. Covariates are needed to predictively distinguish the various such outcome-homogeneous types of clients if clinical decision-making is to be thoroughly benefitted from such uniformities of response by client type. Such perfectly outcome-homogeneous types of clients are thus the logical stand-in for individual clients and would allow us to make unambiguous inferences about individual clients from clinical trial data, which does not require potential response theory to see but is seen by those who are working at such theory (see West & Thoemmes,2010).

6. This is further complicated by the Us included in these comparison groups, because the ~T _j &Us and ~T _k &Us inflate these treatments’ apparent effectiveness (insofar as these treatments do not prevent these Us’ recovery). Furthermore, the unreliability of random assignment for equating the proportions of these types of clients in the comparison groups of an individual RCT (as pointed out above) makes even the difference in percentages of recovered clients between its included treatment groups also unreliable. Because natural-course control groups are at present impossible or unethical to achieve and other sorts of control groups cannot reasonably be assumed to produce results comparable to natural course (Krause & Lutz, 2009a), the detection and study of natural-course recoverers in the population of persons with a given malady is essential for the development of covariates predictive of untreated recovery and so for the pre-treatment recognition of Us. Thus, a natural history adjunct to treatment comparison experiments is essential for treatment effectiveness research for both clinical and public health purposes if we are to learn when, if at all, natural-course recovery occurs in each malady, what is its nature, and how it is impacted by various treatments. The taking, studying, and collating of meticulous comparable clinical histories is certainly a proper part of such research.