In this paper we propose the use of statistical models of event history ana
lysis for investigating human sleep. These models provide appropriate tools
for statistical evaluation when sleep data are recorded continuously over
time or on a fine time grid, and are classified into sleep stages such as R
EM and nonREM as defined by Rechtschaffen and Kales (1968). In contrast to
conventional statistical procedures, event history analysis makes full use
of the information contained in sleep data, and can therefore provide new i
nsights into non-stationary properties of sleep. Probabilities of or intens
ities for transitions between sleep stages are the basic quantities for cha
racterising sleep processes. The statistical methods of event history analy
sis aim at modelling and estimating these intensities as functions of time,
taking into account individual sleep history and assessing the influence o
f factors of interest, such as hormonal secretion. In this study we suggest
the use of non-parametric approaches to reveal unknown functional forms of
transition intensities and to explore time-varying and non-stationary effe
cts. We then apply these techniques in a study of 30 healthy male volunteer
s to assess the mean population intensity and the effects of plasma cortiso
l concentration on the transition between selected sleep stages as well as
the influence of elapsed time in a current REM period on the intensity for
a transition to nonREM. The most interesting findings are that (a) the inte
nsity of the nonREM-to-REM transitions after sleep onset in young men shows
a periodicity which is similar to that of nonREM/REM cycles; (b) 30-45 min
after sleep onset, young men reveal a great propensity to pass from light
sleep (stages 1 or 2) into slow-wave sleep (SWS) (stages 3 or 4); (c) high
cortisol levels imposed additional impulses on the transition intensity of
(i) wake to sleep around 2 h after sleep onset, (ii) nonREM to REM around 6
h later, (iii) stage 1 or stage 2 sleep to SWS around 2, 4 and 6 h later a
nd (iv) SWS to stage 1 or stage 2 sleep about 2 h later. Moreover, high cor
tisol concentrations at the beginning of REM periods favoured the change to
nonREM sleep, whereas later their influence on a nonREM change became weak
and weaker. As sleep data are also available as event-oriented data in man
y studies in sleep research, event history analysis applied additionally to
conventional statistical procedures, such as regression analysis or analyi
s of variance, could help to acquire more information and knowledge about t
he mechanisms behind the sleep process.