
2Kristensen, Bødkergaard & Bibby
sample stimulus intensities and we will investigate the consequences to statistical inference, in
particular estimation and bias.
The need for flexible sampling leads to the concept of adaptive designs, which is by no means
a concept unique to cognitive experiments but to most types of clinical trials. That the design
is adaptive means that the design may change dynamically during the trial, usually based on
the observed design and possibly outcomes up to a certain time. For example, a biased coin
design (Efron, 1971) tries to make the groups in a randomised trial equally large by adapting the
randomisation probability depending on the current allocation of experimental units – thus using
the design but not the outcome at a given step. This is contrary to a classic, fixed design (also
see Dawid and Didelez, 2010) where the trial structure is determined before the trial begins – for
example by setting the probability to be allocated to either group to 1/2 and thereby running
the risk of very unbalanced groups when the number of randomised subjects is small. A classic
use of outcome adaptive designs occurs in group sequential designs (e.g. Jennison and Turnbull,
1999) where one will allow for a trial to stop at an interim stage if the obtained data show strong
evidence against the null hypothesis (stopping for efficacy) or if it seems likely that the trial will
have a inconclusive outcome (stopping for futility). It is generally accepted that designs using
the outcome for example in an interim analysis must account for this in the analysis phase, one
problem being the increased type I error rate due to multiple testing, another being the bias of the
estimates (confidence intervals must also be produced using special methods). For example, in a
drug trial with an interim analysis stopping for efficacy because the observed effect of the drug is
very large, the na¨ıve estimate of the drug effect will be biased upwards.
In psychometrics and psychophysics, the adaptation uses both the outcome and the design.
Simply put, the problem is that if the stimulus intensities are chosen too low or too high, the
observed accuracies will be almost all zeros or ones leading to poor estimates of the psychometric
function. Thus, we would like to sample where there is more information about the psychometric
function. As these accuracies are specific to a participant and there is no way to know the partic-
ipant’s level of accuracy before the experiment, the design needs to adapt based on the previous
intensities along with the previous performance. Various designs have been proposed as reviewed
for example in Treutwein (1995) and Leek (2001), the simplest of which adapt the stimulus intensity
at a given time from the previous intensity based on the performance a few trials back.
Estimates from adaptive designs will usually inherit the asymptotic properties of those from the
fixed sample design (e.g. Melfi and Page, 2000). However, this does not account for the behaviours
in small samples and while there is acknowledgement for the need to account for these in the
medical trial literature this does not seem common in psychometrics (cf. for instance section 3.4.2
and 5.4.2 of the textbook Kingdom and Prins (2016)). As also highlighted by Bretz et al. (2009)
testing and the control of type I error rates are much better understood than estimation in adaptive
designs.
The article proceeds as follows. We first give a brief overview of some standard methods
for adaptive stimulus allocation. We then introduce the basic setup for the paper introducing
the concepts of psychometric functions along with notation for designs and dependence schemes,
before investigating the likelihoods as the basis of inference and estimation. We perform these
investigations both in designs employing adaptive and non-adaptive allocation as well as under
within-subject independence and dependence through the inclusion of random effects. Finally, we
illustrate these points through a simulation study followed by a brief discussion.
1.1 Brief overview of adaptive methods
In the following we give a brief review of methods used in psychometrics and psychophysics for
constructing adaptive designs (see Treutwein, 1995; Leek, 2001, for a more complete and in-depth
treatment). As there is substantial overlap with the literature on dose finding designs, we include
a few references from the related literature and make some comparisons.
The simplest class of adaptive designs assigns the next stimulus intensity from the current
intensity based on the performance of the subject a few trials back. This includes the up-down
design (Dixon and Mood, 1948) in which the stimulus is increased from the current intensity if the
current response was incorrect and decreased if the response was correct. This procedure targets
the accuracy probability 1/2. If the purpose of the study is to estimate some other quantile of
the psychometric function, this may constitute a disadvantage. Other designs may be employed to
target different accuracies, e.g. the one-up-two-down design will decrease the intensity only after
two consecutive correct responses and target the probability 1/√2=0.71. A more general approach