6

are exceptions to this pattern of usage which suggest that the ground squirrel call system does not denote predator type. Squirrels being closely pursued by carnivores sometimes produce whistles, and chatters are given to distant hawks (Owings and Virginia, 1978; Leger et al., 1980; Owings and Leger, 1980). Similar patterns of call usage have been described in Belding's ground squirrels (Robinson, 1980, 1981; Sherman, 1985) and marmots (Blumstein, 1995a,b; Blumstein and Arnold, 1995; Blumstein, in press). The data on alarm call production in sciurids are consistent with the idea that these signals do not describe predator classes directly, but rather encode differences in response urgency perceived by the caller. This will vary quite reliably with predator type, as fast-moving hawks usually afford less time for escape than the relatively slow approach of a carnivore. Ground squirrels and marmots thus have call systems that are exquisitely well-matched to the hunting tactics of their two principal classes of predator; the signals are designed to allow receivers to judge the time available for fleeing to a burrow refuge, which is arguably information of greater functional importance than a taxonomic description of the approaching threat.

Signals with the property of functional reference must also meet a perception criterion. They should be sufficient, in the absence of the eliciting stimulus and of other normally available cues, to permit receivers to select appropriate responses. This property has been termed 'context independence' (Evans et al., 1993a; Macedonia and Evans, 1993). The most common technique for assessing perception of putative referential signals is playback experiments, in which recorded sounds are presented to conspecific receivers (e.g., Seyfarth et al., 1980b; Macedonia, 1990; Evans et al., 1993a). This approach, by design, strips away the contextual cues that might normally be provided by the non-vocal behaviour of the sender. Appropriate responses reveal that such information is not essential, but they do not demonstrate that it is unimportant (see below).

Theoretical papers discussing the issue of external reference in animal signals have traditionally employed a simple dichotomous classification scheme in which signals are considered to be either affective or referential (Marler, 1977, 1978, 1984; Gouzoules et al., 1985). More recent work has modelled the properties of animal signals as points falling along a continuum, with signals that principally reflect the motivational state of the sender, such as the distress calls of precocial birds (Collias and Joos, 1953; Abraham, 1974) and the cries of human infants (Lester, 1985), at one end, and affect-free referential signals, such as machine-generated speech, at the other. Categorizing a signal as functionally referential is equivalent to postulating a threshold value on such an underlying continuum, and then demonstrating empirically that the properties of the signal are such that it is exceeded. There is a degree of unavoidable arbitrariness inherent in partitioning any sort of variation in this fashion. It is also true that whether or not a signal meets the criteria described above will be dependent not only upon the characteristics of the system being studied, but also upon extraneous factors such as the number of animals available for study (and hence the level of statistical power), and the sensitivity of the response assays employed. The advantage of this approach is that it allows us to address issues, such as the development and phylogenetic distribution of referential signals, that would otherwise be intractable.

V.

A STRATEGY FOR EXPERIMENTAL ANALYSIS

It is important to note that no single observation or experimental test is sufficient to identify the property of functional reference. Instead, we require a program of research that involves studies of both production and perception (Fig. 1). It is consequently not straightforward to obtain compelling evidence for referential signalling and this, combined with the very small number of systems that have so far been studied, suggests that we may have underestimated the incidence of this phenomenon.