Vous êtes sur la page 1sur 22

JOURNAL OF THE EXPERIMENTAL ANALYSIS OF BEHAVIOR 2017, 107, 321–342 NUMBER 3 (MAY)

SELECTION BY CONSEQUENCES, BEHAVIORAL EVOLUTION, AND THE PRICE


EQUATION
WILLIAM M. BAUM
UNIVERSITY OF CALIFORNIA, DAVIS AND UNIVERSITY OF NEW HAMPSHIRE

Price’s equation describes evolution across time in simple mathematical terms. Although it is not a the-
ory, but a derived identity, it is useful as an analytical tool. It affords lucid descriptions of genetic evolu-
tion, cultural evolution, and behavioral evolution (often called “selection by consequences”) at
different levels (e.g., individual vs. group) and at different time scales (local and extended). The impor-
tance of the Price equation for behavior analysis lies in its ability to precisely restate selection by conse-
quences, thereby restating, or even replacing, the law of effect. Beyond this, the equation may be useful
whenever one regards ontogenetic behavioral change as evolutionary change, because it describes evo-
lutionary change in abstract, general terms. As an analytical tool, the behavioral Price equation is an
excellent aid in understanding how behavior changes within organisms’ lifetimes. For example, it illu-
minates evolution of response rate, analyses of choice in concurrent schedules, negative contingencies,
and dilemmas of self-control.
Key words: Price Equation, behavioral evolution, induction, concurrent schedules, induced activity,
operant activity, reinforcement, negative contingencies, self-control

Often mathematics expresses clearly and This paper shows how Price’s equation may
concisely what many words cannot. Price’s be applied to behavioral evolution (often
equation is an example. George Price (1970, called “shaping by consequences”). The first
1972), using straightforward algebra, derived section gives an overview of Darwinian pro-
an equation that expresses evolution by natu- cesses to give a framework for thinking about
ral selection in precise mathematical terms. behavioral evolution. The second
Recognition of its significance has grown section presents three versions of the Price
steadily since it was published in 1970 (Frank, equation, applying to genetic evolution, cul-
1995; Panchanathan, 2011). Nowadays, it is tural evolution, and behavioral evolution. The
part of the standard toolkit for clarifying evo- third section presents examples that illustrate
lution at multiple levels, and discussions of the approach to behavioral evolution:
multilevel selection often rely on it response rates on variable-interval schedules,
(e.g., McElreath & Boyd, 2007). A large part choice in concurrent schedules, self-control,
of the Price equation’s power derives from its and negative contingencies.
generality; because it describes selection in the
most general terms possible, it applies to any Overview of Darwinian Processes
Darwinian process of evolution by selection
and even to inanimate entities (Price, 1972, Every Darwinian process includes three ele-
1995). Behavior analysis is a science in which ments: Variation, Recurrence, and Selection.
evolution by selection is central, because a Variation occurs within a pool or population.
basic premise of behavior analysis is that Examples are: (a) a gene pool, which is an
behavioral change depends on selection by abstraction consisting of all the genes in a pop-
consequences (Skinner, 1981). Thus, Price’s ulation regardless of their occurrence in par-
equation should be interesting and useful to ticular organisms; (b) a pool of genomes, in
behavior analysts. Indeed, it may be thought which each member of the population consti-
of as restating or even replacing the law of tutes a “vehicle,” carrying the genome around;
effect, including the “correlation-based” law of (c) a pool of cultural practices, in which the
effect (Baum, 1973, 2012). members of the cultural group also are like
vehicles, putting the practices into action; and
(d) the coat colors that vary among all the
Correspond with: William M. Baum, 611 Mason #504,
mice (the “vehicles”) that live in a particular
San Francisco, CA 94108, Email: wbaum@sbcglobal.net meadow. Although variation is commonly
doi: 10.1002/jeab.256 thought of as operating on punctate entities

© 2017 Society for the Experimental Analysis of Behavior


321
322 WILLIAM M. BAUM

like genes, nothing restricts variation to punc- term introduced by Segal (1972) to denote
tate entities, because the required variation the relation between a phylogenetically impor-
can be in any property of the vehicles of the tant event (PIE) like food and the (adjunctive)
population. For behavioral evolution, the behavior it induces (Baum, 2012). Although
“vehicles” are not organisms, but samples of “induction” has had other usages, Segal’s
behavior that vary, for example, in the rate of usage is particularly useful, because it applies
an operant response, as will become clear not only to adjunctive behavior, but also to
below. The Price equation (below) clarifies operant behavior, the main focus of behav-
this flexibility. ioral evolution. In Baum’s (2012) explication,
Recurrence means “occurring again” or “com- the contingency between a PIE like food and
ing up again.” In genetic evolution, recur- an operant activity like lever pressing enforces
rence results from reproduction, sexual or a correlation that causes the activity to be
asexual, that causes traits to recur from par- induced by the PIE it produces (e.g., Reid,
ents to offspring. Reproduction is only one 1958; “reinstatement”). This insight will be
mechanism of recurrence (Baum, 2001). useful when we come to examples of behav-
Recurrence is a more abstract and general ioral evolution below.
term for what is required in an evolutionary Selection depends on competition between
process—that types within a population recur different variants within the pool or popula-
from one time to the next. When we turn to tion. It is not the only force by which a popula-
cultural evolution, for example, the mechan- tion’s traits may change; drift due to small
isms of recurrence are imitation and rule fol- population size, for example, has little to do
lowing (i.e., instruction), or a combination of with competition (Sober, 1984). Evolutionary
the two. Such recurrence is sometimes change, however, depends on competition
referred to as “transmission.” Cultural prac- due to differences among variants in fitness or
tices recur from one time to another (often recurrence within a population.
within a generation) and from one person to Competition is the heart of selection. Since
another by one person modeling behavior for the pool or population typically is limited
another (e.g., a dance movement, a ritual, or a (e.g., by carrying capacity of the environment)
method of manufacture) or by one person tell- or only grows relatively slowly (e.g., a cultural
ing another, “Behave this way” (e.g., how to group), substitutable variants must compete
dress, how to approach the king, or how to do through time because increases in one forces
arithmetic), which constitutes rule giving that decreases in others. Their competition makes
induces rule following in the receiver (Baum, their fitness or recurrence relative rather than
2000). Cultural recurrence differs from absolute; the fitness of any variant depends on
genetic recurrence too because cultural recur- its competitors. Short-necked giraffes might
rence is not necessarily, nor even usually, from survive and reproduce well on their own, but
parent to offspring. One’s cultural “parents” if longer-necked giraffes recur more often by
go beyond one’s genetic parents. One may leaving more offspring, shorter-necked giraffes
acquire behavior from peers and from mem- disappear from the population. Different
bers of an older generation other than one’s alleles of a gene, different ways of worshiping
genetic parents, such as teachers, coaches, or making a canoe, and different activities that
employers, and ministers (Boyd & Richerson, produce different outcomes—pecking at one
1985). One may acquire behavior even from a key rather than another or spending time with
stranger viewed in an online video. family versus working—all compete because of
In behavioral evolution, recurrence means limitations on resources or time. If one
repetition of behavior from time to time. Evo- increases, its competitors must decrease. Fit-
lutionary change and the Price equation only ness is always relative.
require some kind of recurrence; the mechan- For evolutionary change, the variants that
ism of recurrence varies from one evolutionary may substitute for one another are the variants
process to another. For behavior, the tradi- that matter; other variants may be independ-
tional view ascribes recurrence to “strengthen- ent of them. Darker coloration may compete
ing” (Skinner, 1938; Thorndike, 1911/2012). with lighter coloration, but not necessarily be
An earlier paper suggested that the mechan- linked to larger body size. The practice of
ism of behavioral recurrence is induction, the making internal-combustion-driven cars may
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 323

compete with the practice of making electric- behavioral recurrence may be unknown, the
motor-driven cars, but neither may have any behavior selected by the contingency com-
bearing on methods of birth control. No mat- bined with induction may be considered fitter
ter what the mechanism, selection of one vari- than its competitors and to have more “des-
ant (trait, practice, or activity) over others cendants” at a later time.
occurs when competition results from differ-
ences in their fitness or differential recurrence Price’s Equation
through time. If larger body sizes recur more
often than smaller body sizes, larger body sizes Genetic Evolution
increase in the population at the expense of Price (1970) derived the equation by proving
smaller body sizes. If members of a cultural a theorem. He made three assumptions— vari-
group eat either with chopsticks or a fork, and ation, “transmission” (Price’s word for recur-
eating with a fork recurs (spreads) in the rence), and selection. First, variation: a
group more often than eating with chopsticks, population of N individuals at two time periods
eating with a fork increases in frequency at the or “generations,” Time 1 and Time 2. A trait or
expense of eating with chopsticks. If my driv- property, x, of these individuals varies among
ing on the highway to work recurs in my drives them, and the variable xi indicates the level of
to work more often than my driving on back x in individual i. At Time 1, the mean of
roads, driving on the highway increases at the x across the N individuals is denoted x. Second,
expense of driving on back roads. transmission: offspring resemble their parents,
Cultural practices constitute temporally which means that offspring of a parent with
extended patterns of behavior that accomplish trait-level xi also have trait-level xi. Third, selec-
outcomes important to the survival or cohe- tion: Each individual i may contribute surviv-
sion of the group that practices them; they are ing offspring to the population at Time 2. The
ways of getting vital “jobs” done (Guerin, fitness or relative recurrence of individual
1997). A cultural practice is an integral whole i depends on how many of its offspring survive
with parts that function together to bring to Time 2. An individual that contributes no
about their desirable result. For example, surviving offspring has a fitness of zero. Others
manufacturing a certain kind of ceramic pot have fitness proportional to the number of off-
requires finding and processing clay, working spring they contribute. The fitness of individ-
the clay, decorating the pot, and firing the ual i is symbolized by wi, which varies across
pot. The practice may be refined by substitut- individuals. Price’s great insight was to see that
ing a more effective variant for one of the the change in the mean of x in the population
parts—say, finding a new type of clay or a new from Time 1 to Time 2, symbolized as Δx,
mode of decorating the pot. The parts make a depends simply on whether individuals high in
whole that may change as the parts change. the trait x are also high in fitness w and indivi-
Similarly, an operant activity is an integral duals low in the trait x are also low in fitness
whole that produces results useful to the w—that is, whether xi and wi vary together—
organism (Baum, 2002). Indeed, cultural prac- their covariance.
tices are operant activities—they are distin- Figure 1 illustrates Price’s reasoning with a
guished only by their being a property of a simple example. The first row of the table
group. A skilled operant performance is shows a population of 10 individuals, with
refined the same way that a cultural practice is i ranging from 1 to 10. Suppose all these indi-
refined, usually by substituting a more efficient viduals possess either one of two variants of a
variant for one less efficient, as when an activ- certain gene—its alleles, A or B. Seven of the
ity like swinging a baseball bat becomes more individuals possess the A allele, and 3 possess
effective by the substitution of one muscular the B allele, as we see in the second row of the
motion for another. table. If an individual possesses the A allele,
The combination of contingency and induc- we set xi = 0, and if an individual possesses the
tion results in one operant activity having B allele, we set xi = 1, as shown in the third
higher recurrence (“fitness”) than the variants row of the table in Figure 1. The bottom row
that compete with it. We will take up examples gives various arbitrarily selected numbers of
later. Right now, we only need to recognize offspring contributed to the population at
that, even though the physiological basis of Time 2, wi, for the 10 individuals. For
324 WILLIAM M. BAUM

i= 1 2 3 4 5 6 7 8 9 10

Allele: A A A A A A A B B B

xi = 0 0 0 0 0 0 0 1 1 1

@T2,wi = 1 0 2 0 2 1 0 2 2 0

Fig. 1. A simple example illustrating Price’s reasoning.


In the top row of the table, i indexes individuals in a popu-
lation of 10 individuals. Each individual possesses either
the A allele or the B allele of a gene, as shown in the sec-
ond row of the table. If an individual possesses the A
allele, xi is set to zero, and if an individual possesses the B
allele, xi is set to 1.0, as shown in the third row of the
table. The 10 individuals vary in fitness, which is the num-
ber of offspring or copies of the allele they possess at Time
2, wi, shown in the bottom row of the table. The graph
below shows fitness wi plotted against xi (X symbols). The
squares represent the average fitnesses of the two alleles.
The slope of the line shows that covariance is greater than
zero and selection occurs.

example, Individual 1 contributed one off- Fig. 2. Two examples of covariance between fitness wi
spring, Individual 2 contributed no offspring, and xi. Top: an example with two alleles of a gene, as in
Individual 3 contributed two offspring, and so Figure 1, but with diploid organisms. Individuals homozy-
on. Since the number of offspring contributed gous in the A allele possess no copies of the B allele. Heter-
by the A-allele individuals total to 6, their ozygous individuals possess one copy, and individuals
homozygous in the B allele possess two copies. If fitness
mean fitness equals 6/7. Since the number of tends to increase with number of copies of B, the slope of
offspring contributed by the B-allele indivi- the line connecting the average fitnesses (squares) shows
duals total to 4, their mean fitness equals 4/3, that covariance is positive and selection occurs. Bottom: an
greater than for the A allele. In the graph example in which x is a continuous variable and xi may take
on any level within its range. The positive slope of the line
below the table, these fitnesses wi are plotted fitted to the fitnesses or recurrences wi at Time 2 shows that
against xi, and the line connects the two mean covariance is greater than zero and selection occurs.
fitnesses. The key point of the graph is that
the line has a positive slope; individuals’ fitness
wi covaries with individuals’ trait-level xi. look like that shown. The bottom graph shows
Figure 2 shows two additional examples of the situation in which x is a continuous varia-
covariance between fitness and a trait x. The ble like body size, height, or coat color. Then
top graph shows an extension of the reasoning xi can take on any number of levels, resulting
behind the graph in Figure 1 to diploid organ- in a scattering of points like that shown.
isms. The x-axis represents the number of This covariance is the heart of the Price
copies of the B allele that different genotypes equation:
possess: Individuals homozygous in A have no
copies of B; heterozygous individuals have one Δx = covðwi , xi Þ ð1Þ
copy of B; and individuals homozygous in B
have two copies of B. If the more copies of which states that the population change in
allele B an individual possesses the higher the x equals the covariance between w and
fitness, the covariance between wi and xi would x across individuals.
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 325

The full Price equation contains applies Equation 3 without any reference to
another term: group selection.
X Three points about Equation 3 should be
Δx = covðwi , xi Þ + wi Δxi =N ð2Þ apparent. First, the equation is true, because it
i is a derived identity (see Appendix A). It
describes or defines rather than predicts, but
where Δxi is the change in xi from Time 1 to it is a valuable analytical tool for thinking
Time 2 due to imperfect fidelity of recurrence; about selection, for example, group selection,
if recurrence were just the passing on of xi when a population is structured into compet-
from one time to the other, Δxi would be zero. ing groups. Its theoretical aspect lies in the
Since the second term on the right is a mean, assumptions made when applying the equa-
it is usually rewritten as an expected value tion. The basic assumptions appear to be
Ew(Δxi), leading to the simplest form of the minimal—variation in a population of indivi-
Price equation (see Price, 1970, 1972, and duals, recurrence, and covariance—but some
McElreath and Boyd, 2007, chapter 6, for may be hidden, as for example, the assump-
more detail): tion that the environment remains stable from
one time to the next. The concept of fitness is
Δx = covðwi , xi Þ + Ew ðΔxi Þ ð3Þ crucial to the equation. Fitness is relative and,
because it is relative, fitness is dimensionless
Equation 3 says that the population’s mean (Appendix A). As a result, the two sides of
change in x equals the dependence of fitness Equations 1, 2, and 3 have the same units. Sec-
on xi plus an expected value that gives the ond, it neatly includes all three ingredients of
mean change in x apart from the covariance. the Darwinian process: Variation appears in xi
The covariance represents selection; if relative as it varies from individual to individual or
fitness varies positively with xi, Δx is positive, vehicle to vehicle of the population; recur-
and the trait increases in the population. If rence appears in the term wi; and selection
the covariance is negative, the trait decreases, appears in the covariance between xi and wi—
and if the covariance is zero, the trait remains in differential recurrence across variation in xi.
stable. The second term on the right, which Third, Price’s equation is extremely general. It
expresses the tendency of x to change as it doesn’t specify the mechanism of recurrence,
recurs from Time 1 to Time 2, due to factors the nature of the individuals that vary in xi,
usually rare or negligible (e.g., mutation or whether xi is a discrete or continuous variable,
meiotic drive), is usually considered close to or the factors that favor higher xi over lower xi
zero for genetic evolution (El Mouden, André, (differential recurrence or fitness wi).
Morin, & Nettle, 2014; Henrich, 2004). Like the law of effect, like many scientific
Put in more general terms, Equation 3 says laws, the Price equation is a tautology. For
that, if recurrence is differential with respect example, Newton’s law F = ma is a tautology.
to various levels of a trait, and fidelity of recur- Rachlin (1971) argued that the choice match-
rence is high, then the trait recurs increasingly ing law might be seen as a tautology. Indeed,
(positive covariance) or recurs decreasingly natural selection may be viewed as a tautology
(negative covariance) in the population from (Sober, 1984). Just as an activity that increases
one time to the next. Suppose that the popula- when in contingency with a reinforcer is said
tion of mice that live in a certain meadow vary to be reinforced, and the reinforcer is known
in coat color and that darker coat colors are by its ability to increase activities when in con-
less visible to predators. This lower visibility tingency with them, so a variant in a popula-
means that coat color varies with fitness. If we tion that increases has higher fitness, and
think of x as darkness of coat color, the covari- fitness is known by its ability to increase var-
ance between darkness and fitness is positive, iants within a population. Tautology is not a
and Equation 3 says that mean coat color in defect. Tautological laws may be highly useful
that population grows darker from one gener- generalizations, conceptual frameworks, and
ation to the next. Since the mice in the explanatory frameworks (Sober, 1984). Among
meadow constitute a population not struc- other possibilities, a tautology may allow us to
tured into groups, however, this example grasp and apply selection by consequences.
326 WILLIAM M. BAUM

Group Selection for each group. It accounts for all change


Suppose that a group under consideration within the population at the level of the indivi-
is one group among many, because the popu- duals making up all the groups.
lation is structured into groups. A good exam- The two terms on right-hand side of Equa-
ple is a bee colony; many colonies might tion 4 represent two components of change in
inhabit a geographical area, and they compete x in the population: the first for selection
for resources and succeed or fail as wholes. between groups and the second for selection
Another example is a human social group, between members within groups, the expected
such as a baseball team, which competes with value term. For genetic evolution, we may
other teams and wins or loses as a whole. ignore the expected value of Δxig in Equation 4
Human groups compete both directly, as in as we did the expected value of Δxi in Equa-
warfare, and indirectly, as in exploiting limited tion 3, because the change would be due to
resources poorly or well. Since Equation 3 rare or negligible factors such as meiotic drive
does not depend on the sort of individuals, and mutation (El Mouden et al., 2014; Hen-
Equation 3 would also apply to selection rich, 2004). Equation 4 then simplifies to the
among bee colonies and other social groups sum of a covariance for between-group selec-
that function as integrated wholes. For the tion and expected value (or average) of covari-
sake of clarity, we index groups with the sub- ance for within-group selection:
script g, instead of i, and apply Equation 3 to     
get an expression for Δx, now the change in Δx = cov wg ,xg + Eg cov wig , xig ð5Þ
average trait x across all the groups in the pop-
ulation. The equation says that if the covari- An alternative form of the Price equation
ance between xg and fitness wg is positive, the that uses the algebraic relation between covari-
prevalence of x in groups with a higher mean ance and the regression coefficient β may fur-
of x increases, and such groups survive, grow ther clarify group selection (see Henrich,
in size, and multiply relative to groups with 2004, and McElreath & Boyd, 2007, chapter
lower mean x. For example, bee colonies that 6, for more detail):
better resist invasion by mites will tend to sur-         
vive and grow better than those with less resist- Δx = β wg , xg var xg + E β wig ,xig var xig
ance, and baseball teams that coordinate well ð6Þ
together will tend to win and gain resources
against teams that are less coordinated. Equation 6 makes clear that the depend-
In applying Equation 3 to group selection, ence (β) between w and x and the variance
however, the second term on the right, which in x must both be greater than zero for selec-
would be written Ew(Δxg), is no longer negligi- tion to occur. Each term on the right implies
ble, because the change in xg apart from a role for each of the three ingredients of
group selection includes changes in individual evolution: variation (var(x)), recurrence (w),
members due to selection at the level of indi- and selection (β).
vidual members. Recognizing that xg is a mean When applied to genetic evolution, Equa-
across members in the group that may change tion 6 clarifies why genes for altruistic and
from Time 1 to Time 2, we treat Δxg the same cooperative behavior are unlikely to be
way as we treated Δx (see Appendix A) and selected by genetic evolution on its own. Some
arrive at: traits may have both between-group and
       within-group components that act in
Δx = cov wg , xg + Eg cov wig , xig + Ew Δxig concert—both have positive β—such as body
ð4Þ size (helpful in intragroup and intergroup
combat) or tendency to stay close to the rest
where the subscript ig indicates member i in of the group (avoiding predators). Genes for
group g (see the derivation in Appendix A altruism or cooperation, however, though
and Henrich, 2004, and McElreath & Boyd, helpful for the group, reduce the fitness of
2007, chapter 6, for more detail). The second individuals that behave so—at least in the
term on the right of Equation 4 is really just short term (Baum, 2016). Thus, for altruism
the average across groups (Eg) of Equation 3 or cooperation, β is positive for between-group
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 327

selection, but negative for within-group selec- 2, denoted, as before, wi. As before, if covari-
tion. Since variance across groups var(xg) in ance between wi and xi is positive, the practice
particular is lowered by migration between spreads in the population. The second term
groups, the groups have to be practically com- on the right of Equation 3 might be consid-
pletely isolated genetically for group selection ered non-negligible if one wished to incorpo-
to overcome negative individual selection. For rate errors on the part of recipients, but
a bee colony, variance within the colony usually may be considered negligible because
var(xig) is small, because the workers are close transmission errors would likely be insignifi-
relatives, and no migration occurs between cant compared with selection (see El Mouden
colonies, so altruistic behavior dominates. For et al., 2014; supplemental Document S1; for a
human groups, however, negative β is likely to detailed derivation of Eq. 4 for culture and
be high for any individual incurring an imme- further discussion).
diate cost for helping or cooperating with
others, and within-group variance var(xig) is Cultural Group Selection
likely to be substantial. Even a small amount
of migration between groups will ensure that Equation 3 would be about the spread of a
between-group selection is smaller than the practice through a cultural group (e.g., spread
within-group selection, making Δx for that of planting hybrid corn among American
altruistic behavior negative. In contrast to farmers; Rogers, 1995). If we think of cultural
selection on genes, however, when we come to groups as parts of a larger population, the
cultural evolution and behavioral evolution, a groups may compete with one another for
means for altruistic behavior to be readily resources and members. Equations 4, 5 and
selected will become clear. Another way of 6 apply then, with the same changes in mean-
using the Price equation to describe individual ing of the variables. Equation 5 contains the
and group selection, known as “contextual same two components of intergroup selection
analysis,” is briefly explained in Appendix B. and within-group selection (El Mouden et al.,
2014; Henrich, 2004). If a cultural practice
benefits the group relative to other groups,
Cultural Evolution the more members practice it, or the greater
Equation 3 may be applied to cultural evolu- the recurrence (wg) in the group at Time
tion, too, with suitable changes in definition. 2, the more positive is the covariance in the
A cultural practice that might be considered first term in Equation 5. Whether the covari-
all-or-none (e.g., allowing marriage between ance in the second term is positive or negative
first cousins or not, eating pork or not, and depends on whether the practice is reinforced
primogeniture or not) would take the place of or is costly for the individual who
an allele that is present or not (as in Fig. 1). engages in it.
As with genetic evolution, however, quantita- To understand how an altruistic or coopera-
tive variation in a practice (e.g., a ritual per- tive practice can spread in a population, Equa-
formed daily or less often, for longer or tion 6 shows what happens if the within-group
shorter duration, with more or less of some covariance for members is negative. As with
ingredient, etc.) would do just as well for the genetic group selection, the variances var(xg)
variable x (as in Fig. 2, bottom). and var(xig) are crucial. In cultural group
Recurrence of a cultural practice from Time selection, biases in imitation and instruction
1 to Time 2 doesn’t go just from parent to off- (“transmission biases”) that promote conform-
spring, as in genetic evolution. Teachers, coa- ity lower var(xig) to near zero, even in the face
ches, ministers, and peers also both model of considerable migration between groups.
practices and instruct practices that have For example, if immigrants into a group tend
proven successful in gaining the proximate to adopt the practice that is most frequent or
tokens of ultimate fitness (reinforcers and above average in the group or tend to adopt
avoidance of punishers). Particularly if these the practice of prestigious or successful group
influential people are prestigious and display members, their conformity will keep var(xig)
the trappings of success (many reinforcers and low (Henrich, 2004). Low var(xig) allows the
few punishers), this modeling and instructing cost to the individual represented by the nega-
spur the recurrence of the practice at Time tive covariance (β) of the second term to be
328 WILLIAM M. BAUM

offset by the positive covariance (β) for the To think of behavioral evolution with the
group, which allows altruistic and cooperative Price equation, we require a population of vehi-
practices to spread in the population. The cles or individuals that vary in a trait x. If one
group in which members sacrifice for the were thinking in molecular terms, one could
group will gain more resources than other think of a population of responses that differed
groups, will increase in numbers at the in a trait like force, for example. Glenn, Ellis,
expense of other groups, will be more likely to and Greenspoon (1992) explained that an oper-
survive, and will be more likely to multiply ant like the lever press, as it actually occurs, con-
(Henrich, 2004; Richerson et al., 2016). stitutes a population of responses. One could
Equations 5 and 6 represent “multi-level” also think of a population of interresponse
selection, but just at two levels—groups and times that differed in duration.
individuals (McElreath & Boyd, 2007; chapter The law of effect. All of Equations 1 – 6 can
6). A practice may have many parts, and be applied to behavioral evolution. Depending
some of these parts may change without the on one’s view of Thorndike’s (1911/2012)
practice taking on a new identity. For exam- writings, Equation 1 either precisely restates or
ple, a manufacturing process or a ritual may replaces Thorndike’s original law of effect.
drop or add elements; the American Pledge Thorndike studied cats’ behavior in escaping
of Allegiance after many years changed to from an enclosure with a mechanism inside,
include the words “under God,” but it is still such as a loop of string or a lever, attached to
the Pledge of Allegiance, and process a catch on the door to the box. A bowl of food
improvement constantly upgrades manufac- usually sat outside. A trial began with place-
ture of automobiles, but the practice remains ment of the cat in the box, and when the cat’s
the manufacture of automobiles (Baum, actions caused the string to be pulled or the
2000, 2002). When practices change in their lever to be pushed, the door to the box
parts, one might consider a third level of opened, ending the trial. Thorndike noted
selection: within-practice selection. In Equa- that the context of confinement and food
tion 4, Δxig may be treated as change due to induced a variety of escape- and food-related
competition among alternative parts within activities. Instead of recording the activities,
the practice and expanded as before to result however, Thorndike measured the time until
in a three-term Price equation (McElreath & the cat escaped and considered its decrease
Boyd, 2007; chapter 6). with trials to indicate formation of an associa-
tion between the situation of confinement and
the action that resulted in escape. In doing so,
Behavioral Evolution he focused on successful actions and largely
Price (1995; written circa 1971) opens a ignored the other, unsuccessful actions that
general discussion of selection, “Selection has occurred before escape. Thorndike thought of
been studied mainly in genetics, but of course the law of effect as a law of association, like
there is much more to selection than just the laws of frequency and contiguity. It stated
genetical selection. In psychology, for exam- that satisfiers that followed a stimulus–
ple, trial-and-error learning is simply learning response sequence strengthened the associa-
by selection” (p. 389). If we substitute for tion between the stimulus and the response.
“trial-and-error learning” instead “shaping by He acknowledged that, when successful
consequences,” we see that Price was thinking actions occurred sooner, fewer alternative
of operant behavior. Thus far, we have seen actions occurred and wrote that successful
how the Price equation can clarify a variety of actions were “stamped in” while unsuccessful
selection processes. As far as I know, no one actions were “stamped out.” Thorndike
has applied the equation to behavioral evolu- included comments in his writings that indi-
tion. The entities that may vary would no cate he considered the change due to selec-
longer be individual organisms, because we tion (Donahoe, 1999), but he never explained
are dealing with the behavior of a single the analogy between behavioral evolution and
organism. An organism’s behavioral patterns genetic evolution at length.
vary, recur in time, and are selected when they Instead of focusing on an isolated response
recur differentially—i.e., when they vary in that gets more strongly connected to the stim-
fitness. ulus situation, Equation 1 assumes a
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 329

population of various responses or actions, in arising at Time 2. These new variants, which
keeping with evolutionary population thinking allow shaping and give rise to novelty, would
(Mayr, 1970). If we take Thorndike’s trials to be like mutations.
constitute individuals in Price’s reasoning Response rate. Elsewhere I have argued that
(Fig. 1) and take a block of trials as the popu- a multiscale molar view of behavior based on
lation at Time 1, then the different levels of xi time allocation among temporally extended
represent variation in responses or actions. activities is more powerful and more general
Some of these variants operate the mechanism than a molecular view (Baum, 2012, 2013;
in Thorndike’s box (e.g., clawing at the string) Baum & Rachlin, 1969). How can the multi-
and have higher “fitness” or recurrence wi scale view square with “selection by conse-
than others that fail to do so, because these quences” and Price’s equation?
successful actions correlate with escaping from The answer requires thinking in temporal
the box. For successful actions, let xi equal 1.0, terms. The varying entities or vehicles would be,
analogously to allele B. For unsuccessful for example, time samples within a larger time
actions, let xi equal zero, analogously to allele interval, and these time samples would consti-
A. From one block of trials to another (i.e., a tute the population—seconds within a minute,
population at Time 2), the successful actions minutes within an hour, hours within a day,
(e.g., “string-pulling”) increase in the popula- weeks within a year, and so on (Baum, 1973,
tion, and the unsuccessful actions decrease, 2012). The variable xi that differs across time
and covariance between xi and wi is positive samples may be occurrence or not of an
(i.e., Δx in Eq. 1 is positive). activity—if the time samples are short—or rate
Equation 1 similarly restates or replaces or time taken up by the activity—if the time
Skinner’s (1938, 1953) version of the law of samples are longer—or any other quantitative
effect, which Skinner (1981) explicitly com- aspect of an activity, such as force or specifics of
pared to genetic and cultural evolution. Like topography.
Thorndike, he focused only on successful The same considerations apply whether one
actions—say, those that succeeded in depress- adopts the multiscale view that behavior con-
ing a lever—which he defined as a class. In a sists of extended activities (Baum, 2013) or the
free-operant situation, the operant responses molecular view that behavior consists of dis-
that actually occur constitute a population of crete responses (Skinner, 1938, 1969). Not
responses, as explained by Glenn et al. (1992). only activities, but discrete responses also take
The population consists of responses both suc- up time; that is why response rate cannot
cessful and unsuccessful at pressing the lever, increase indefinitely, but always must reach an
as suggested by Catania’s (1973) phrase “func- upper limit (Baum, 1981, 2015; Baum & Davi-
tional operant,” which includes all responses son, 2014b; Herrnstein, 1970). Whether dis-
induced by the contingency with food. If we crete responses or extended activities, their
take Time 1 as a time interval of observation— total takes up the observation time interval
say, an hour-long session—then the various and thus limits the size of the population to
responses or actions may be equated with the observation time interval. The time sam-
levels of xi, and we can call any variants that ples constitute the varying individuals or vehi-
get the lever depressed “successful” with xi cles, and the observation time interval acts
equal to 1.0, whereas those that fail to depress analogously to carrying capacity to force com-
the lever we can call “unsuccessful,” with xi petition among substitutable variants (i.e., the
equal to zero, analogously to alleles B and A time samples). They compete because they
in Figure 1. If the mean fitness or recurrence take up time—whether discrete responses or
of the successful actions (w 1 ) is greater than different activities and parts within activities. If
the mean fitness or recurrence of the unsuc- one variant of x increases in frequency, others
cessful actions (w 0 ), comparing Time 1 to must decrease. Thus, whichever view one
another session at Time 2, then the covariance takes, the varying individuals or vehicles would
between xi and wi is positive and the operant be the same: smaller time samples within the
population evolved. Catania’s “functional larger observation interval. Figure 3 illustrates
operant” is better described by Equation 2, this approach graphically.
because the expected value term allows the Figure 3 applies the multiscale approach to
possibility of new variants—new levels of xi— hypothetical data. The top graph shows an
330 WILLIAM M. BAUM

Fig. 3. Hypothetical plot of an activity at two different times, T1 and T2. Top: Each horizontal bar represents an epi-
sode of the activity with pauses in between. Time moves from left to right. The activity occurs more frequently at Time
2 than at Time 1. Bottom: The two observation time intervals are divided into equal time samples (vertical lines), which
constitute the individuals in this example of behavioral evolution. The two circles on the left show time samples empty of
the activity. The two circles in the middle show time samples that were completely filled with the activity. The two circles
on the right show time samples in between the two extremes, partially taken up by the activity.

activity occurring at two different times, Time varying amounts of time, were broken into suc-
1 and Time 2. The horizontal axis represents cessive 2-min time samples, as in the lower
time going from left to right. Each horizontal panel of Figure 3, and the number of presses
bar represents an episode or bout of the activ- in each time sample counted. The 2-min sam-
ity. The activity is occurring at a higher rate at ples were sorted according to the number of
Time 2 than at Time 1; the bouts are longer lever presses in the sample, each sample repre-
and the pauses between bouts are shorter senting a certain press rate (presses per min).
(e.g., Shull, Grimes, & Bennett, 2004). The The solid line shows the frequency distribution
bottom graph in Figure 3 shows the two time of press rates. Most were in the range from
intervals divided into small, equal time sam- 3 to 12 per min, but occasionally rates above
ples. The time samples vary in the proportion 15 per min occurred. The modal response rate
of the sample taken up by the activity. The two was about 5 per min, but the mean (dotted
circled time samples on the left are empty, line) is higher, because the distribution is
containing none of the activity—the propor- skewed to the right.
tion of time taken is zero. The two circled time If the proportions of activity illustrated in
samples in the middle are completely filled by Figure 3 and shown in Figure 4 recur exactly
the activity—the proportion of time taken is from one time to another, mutation might be
1.0. The two circled time samples on the right irrelevant when considering response rate.
contain proportions in between the two Then Equation 1 would suffice. If new time-
extremes of 0.0 and 1.0. sample response rates occurred at Time
When the analysis illustrated in Figure 3 is 2, however, then the analogy to mutation
applied to actual data, the result is a frequency might apply, and one would describe the
distribution like that shown in Figure 4 change with Equation 3, to include the
(Baum, 2012). The data were gathered by expected value term expressing change in
John Van Syckel when he was a graduate stu- x (Δxi) due to distortion in recurrence. Equa-
dent at University of New Hampshire. Several tions 1 and 3 may be applied in at least
sessions of a rat’s stable performance on two ways.
variable-interval (VI) 30 s of food delivery, in Directional selection and stabilizing selec-
which lever presses produced foods after tion. If one changed the situation in Figure 4,
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 331

Fig. 4. Variation in response rate within a rat’s stable


performance on a VI 30 schedule. Stable sessions were
divided into 2-min intervals as in Figure 3 and sorted
according to number of presses in the interval. The jagged
line shows the frequency distribution of press rates. The
dotted vertical line indicates the mean.

say, by changing the VI schedule average dura-


tion, one might expect the whole distribution
to shift. The top panel of Figure 5 illustrates
such a shift, from Time 1 with a leaner sched-
ule to Time 2 with a richer schedule. The Fig. 5. Directional selection and stabilizing selection.
change in schedule causes increased covari- Top: directional selection results in a shift of the popula-
ance between higher rates and recurrence. tion’s frequency distribution of x from Time 1 to Time
2. Bottom: stabilizing selection keeps a population stable
Higher rates recur until the distribution shifts by selecting against deviations from the mean in both
to a new, higher mean at Time 2. A whole- directions. The black curve represents the population’s
distribution shift like this is termed “direc- frequency distribution of x. The vertical dotted line repre-
tional selection” (Mayr, 1970). sents the mean of x in the population. The gray curve
Equation 1 or 3 applied to directional selec- represents variation in fitness, decreasing in both direc-
tions away from the mean. The arrows indicate the direc-
tion means, for example, that if time samples tion of selection. The plus sign indicates positive
(the vehicles or individuals indexed by i) covariance between fitness and x. The minus sign indicates
including more of an activity in Time Interval negative covariance between fitness and x.
1 correspond to relatively more such time sam-
ples in Time Interval 2 (i.e., positive covari-
ance with wi in Eq. 1), then the activity (x) Selection does not cease, however, when a
increases from Time Interval 1 (Population 1) population stabilizes around a particular
to Time Interval 2 (Population 2)—Δx is posi- mean. Selection then acts to keep the popula-
tive. In the laboratory example, Figure 4, if the tion stable. The lower panel of Figure 5 illus-
rat’s rate of lever pressing (xi) varies, and the trates this second application of Equation 1,
covariance between xi and wi (recurrence) is known as “stabilizing selection” (Mayr, 1970).
positive, the mean rate across an hour (x) The black curve represents the frequency dis-
increases from one hour to the next (Fig. 5, tribution of x around the mean (dotted line).
top). If a person’s life changes by having chil- The gray curve shows how fitness varies with x.
dren, spending more time with family per day Fitness decreases with deviations from the
(xi) in one month may vary positively with rela- mean in both directions. The gray arrows indi-
tive recurrence per day (wi) in the next cate how selection drives the population back
month, and the monthly amount of time spent toward the mean, and the circled plus and
with family (x) increases from month to minus show the sign of the covariance
month (i.e., Δx is positive). between x and fitness above the mean
332 WILLIAM M. BAUM

(negative) and below the mean (positive). The top panel in Figure 6 shows the same
These positive and negative covariances stabi- data as in Figure 4 with the VI feedback func-
lize the population. tion added. The jagged solid line shows the
When a behavioral performance like that in frequency distribution of press rates. The filled
Figure 4 is stable, selection within the stable square indicates the average press rate. Each
performance acts as in the lower panel of press rate is associated with a certain food rate,
Figure 5 to keep the population of response calculated by adding up the number of food
rates stable. Selection must favor higher rate pellets delivered for the press rate and divid-
for rates below the modal rate and must favor ing by the total time at that press rate (num-
lower rate for rates exceeding the modal rate. ber of intervals times 2 min). The diamonds
Covariance between rate and fitness (recur- show these food rates plotted as a function of
rence) must be positive below the modal rate the press rates. The broken curve shows a
and negative above the modal rate, as in feedback function fitted to the diamonds:
Figure 5.
The multiscale view applies to behavior at 60
r= ð7Þ
any time scale, whether seconds, days, months, t + Ba
or years. The Price equation allows one to
think about selection acting at various time where r is food rate in pellets/minute (FPM),
scales. As with cultural group selection, Equa- t is the VI in minutes (equal to 0.52), B is the
tions 4 and 5 apply to operant activities at a press rate in presses/minute, and the parame-
level or time scale of selection that may ter a equals 0.95 (Baum, 1992). If we assume
include selection at a lower level or shorter that differential reinforcer rate (food pellet
time scale. The analog to group selection is rate here) selects among response rates (press
selection of whole activities extended in time, rates here), then the slope of the feedback
like playing tennis well or poorly and relating function around any particular press rate sug-
to one’s spouse well or poorly. The analog to gests the degree of selection in that region of
within-group selection is the selection of parts the curve. The slope is positive throughout,
referred to above, because every activity is but becomes progressively flatter as press rate
composed of parts that are themselves activ- increases. Selection for higher response rate is
ities on a smaller time scale (Baum, 2002, highest for low response rates and becomes
2012, 2013; Baum & Davison, 2004). Thus, the progressively less for higher and higher
second, expected value, term on the right in response rates. Intuitively, the curve must be
Equation 5 for behavioral evolution represents showing that covariance between response rate
selection among parts of an activity. A tennis and fitness is greater below the mean than
player’s serve may improve when the ball is above the mean, but how to make the connec-
thrown up in a new way, and that variant may tion to fitness?
be selected. Active listening may improve A feedback function by itself cannot predict
one’s relationship with one’s spouse and be stable performance, fitness, or change. It may
selected to replace a passive stance. be part of a theory, but something more is
Theory of response rate. Let us now see how needed, because the fitness of a response rate
the Price equation may mesh with some con- is not measured by the reinforcers it garners,
cepts about response rate—feedback functions as comparison of the frequency distribution in
(i.e., contingency) and induction (Baum, Figure 5 with the feedback function in
2012). Even after an organism has been Figure 6 shows. For VI performance, some fac-
exposed to a single VI schedule for enough tor must deter high response rates, and that
days to stabilize its performance—that is, factor must be either an effect of response rate
response rate varies only unsystematically from itself or an effect of reinforcers on the
day to day—we may assume that within the sta- response rate that wins them. In an earlier
ble performance, dynamics of selection still paper, I suggested that the limiting factor
occur. Otherwise, were the schedule to change, might be the growing cost of responding as
performance might not change. Looked at in response rate increases (Baum, 1981). In the
this light, the feedback function for a VI sched- present, another possibility presents itself: the
ule itself suggests dynamics within stable per- inducing effect of reinforcers on response rate
formance (Baum, 1973, 2012). (Baum, 2015).
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 333

where B(I) is induction, which might be


thought of as induced response rate, and r is
pellet rate. B(I) corresponds to Time 2 in
Price’s equation. The concave-downward
shape of the feedback function tells us that
covariance between relative recurrence (fit-
ness) and response rate is high for lower
response rates and decreases for higher
response rates. Selection for increased
response rate is high for lower response rates
and low for higher response rates.
The lower graph in Figure 6 shows the
results of applying Price’s equation to the per-
formance in the upper graph. The squares
show the feedback function (Eq. 7; the dashed
curve in the upper graph) transformed
according to Equation 8. This transformed
food rate, B(I), is shown by the squares. The
parameter c (equal to 92.4) and the exponent
s (equal to 3.56) were derived by requiring the
average induced response rate (filled circle)
to equal the average response rate in the
upper graph (filled square), on the grounds
that average response rate is stable. The large
exponent causes induction to be low for low
food rates but to increase rapidly as food rate
increases (Baum, 2015). Fitness (w) of each
Fig. 6. Selection dynamics within a rat’s stable perfor- press rate was calculated as the ratio of B(I) to
mance on a VI 30 schedule. Stable sessions were divided into press hratei and dividing by the mean; if the
ratio BBðI Þ were called v, then fitness was cal-
2-min intervals and sorted according to number of presses
in the interval. Top: the jagged line shows the frequency dis-
tribution of press rates (right vertical axis), the diamonds culated as ½v=v . The solid curve shows fitness
show corresponding food rates, the broken curve shows the
feedback function, and the filled square shows the mean
across press rates. It is less than 1.0 (dashed
across sessions. Bottom: squares show food pellets per min line) for low response rates, increases above
(FPM) transformed by a power function modeling induction 1.0 in the region from about 2.5 to 10 presses/
to presses per min (PPM). The filled circle shows the pre- min, and then drops below 1.0 again for rates
served mean performance (square in top graph; assuming above 10 presses/min. Selection favors
overall press rate is stable). The solid curve shows fitness
(right vertical axis) calculated by treating the induced press- response rates in the midrange and goes
ing as a second population. The broken line indicates fitness against both lower and higher response rates.
equal to 1.0—selection favors increasing rate only for the The analysis in Figure 6 suggests an explana-
rates with relative fitness greater than 1.0. tion of performance on VI schedules—why
response rate is never extremely high as in
To apply Price’s equation to selection within ratio schedules, in which a reinforcer is pro-
this stable VI performance, we need to make duced by a number of responses, and never
two assumptions. The first is that the variable extremely low. If we compare the fitness curve
xi in Equation 1 is press rate indexed by the 2- in Figure 6 with the fitness curve in the lower
min intervals (i). The second is that recur- panel of Figure 5, we see that, as in the illus-
rence of any press rate is directly related to its tration in Figure 5, covariance between
associated food rate. The relation between response rate and fitness is positive for
recurrence and food rate is induction (Baum, response rates below the mean and negative
2012, 2015; Baum & Davison, 2014b) and may for response rates above the mean. The
be given by a power function: dynamics are such as to select moderate rates
and select against both extremely high and
B ðI Þ = cr s ð8Þ extremely low rates, but the overall average
334 WILLIAM M. BAUM

rate remains the same from time to time. response rates don’t just compete with one
Thus, VI schedules will always maintain mod- another, but also with the other activities
erate response rates. induced.
In contrast, ratio schedules, for which the Choice in concurrent schedules. Equation 2
feedback function is a straight line, always may be applied to performance on concurrent
select higher response rate for all response VI schedules. Much research has focused on
rates (Baum, 1993, 2015). On ratio schedules, concurrent VI VI schedules in procedures that
response rate will always tend to be the maxi- entail maintaining a pair of schedules for
mum possible, which is correct to an approxi- many sessions, until performance appears to
mation (Baum, 1981; Fig. 10). Why rate is ever be “stable”—that is, varying from session to
less than the maximum—as with ratio session but with no apparent trend. In report-
“strain”—requires more assumptions ing such an experiment, the whole session is
(e.g., Baum, 1981, 2015). usually treated as a unit, and the measure of
This explanation of VI and variable ratio choice or preference between the two VI alter-
h i
(VR) performance may be compared to ones natives is computed as either B1B+1B2 or log BB12 ,
based on the molecular view that behavior
consists of discrete responses. These where B1 and B2 are time or behavior spent in
molecular-based explanations rely on differen- responding at the two VI schedules. For Equa-
tial reinforcement of interresponse times tion 3, such data analysis would indicate that
(IRTs; e.g., Morse, 1966). Since the probability the session constitutes the individual and that
of a reinforcer goes up with time in a VI the trait x is choice, measured one way or the
schedule, the longer between responses, the other. A block of earlier sessions would consti-
more likely the next response will produce a tute a population at Time 1, and a later block
reinforcer—that is, long IRTs are differentially of sessions would constitute Time 2. Selection
reinforced. Such selection would work to keep favoring some levels of x over others would
response rate from becoming extremely high, result in choice changing across sessions, until
but it fails to explain why VI response rate is choice becomes stable and stabilizing selection
not extremely low—low enough that every takes over (Fig. 5; see Baum, 2010, for a tuto-
response would produce a reinforcer. The loss rial in dynamics of choice). If analysis remains
in reinforcer rate at low response rates might at the level or time scale of whole sessions,
seem like the reason that response rate stays Equation 3 suffices, but if analysis proceeds
in the moderate range, but such an appeal to beyond whole sessions to consider smaller
molar relations illustrates the inadequacy of time scales within sessions, then Equation 5,
the molecular account. The present explana- the multiscale equation, applies.
tion based on selection of response rates Some experiments on choice have taken a
accounts for the moderate VI rates by showing multiscale approach (e.g., Aparicio & Baum,
selection against both extremely high and 2006, 2009; Baum & Davison, 2004, 2014a;
extremely low response rates. Davison & Baum, 2000). Taking log-ratio of
The treatment of response rate offered here behavior as x, these experiments suggest inter-
is at least incomplete, because it fails to incor- pretation in the light of Equation 5. Typically,
porate competition between the operant activ- each experimental session is composed of
ity (pressing) and other activities induced by seven components, each consisting of a differ-
the reinforcer (PIE; e.g., food). As a result, it ent pair of VI schedules, one for responding
fails to make the correct prediction of stable at the left key or lever, and one for responding
response rates across VI schedules. These sta- at the right key or lever. Blackouts separate
ble rates are often fitted with a hyperbola that the components, no signals indicate which
incorporates a form of competition by viewing pair of schedules is in effect, and each compo-
response rate as dependent on relative rein- nent lasts for a certain number of food deliv-
forcer rates (Baum & Davison, 2014b; Herrn- eries (e.g., 10 or 12).
stein, 1970). An earlier paper showed, For example, Davison and Baum exposed
however, that response rates across VI sche- pigeons to such conditions for 50 sessions per
dules are well described by competition condition (Baum & Davison, 2004; Davison &
between two power functions (Baum, 2015). A Baum, 2000). By the end of 15 sessions, per-
more correct model will recognize that the formance appeared stable at the time scale of
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 335

whole sessions, as indicated, for example, by


calculating log ratio of key pecks at the left
and right keys across whole sessions. The
remaining 35 sessions were taken as stable at
that time scale and pooled to measure perfor-
mance at smaller time scales. To do this, all
approximately 245 components from the
35 sessions could be taken to define popula-
tions of log behavior ratios within compo-
nents. If the data from six pigeons are
pooled, the population is about 1470.
Although performance appeared stable at the
time scale of whole sessions, the procedure
assured that evolution would occur within
components, because no signals indicated the
different food ratios and only the obtained
food deliveries themselves could guide
responding.
Analysis begins at the time scale of the com-
ponent, because choice, measured from the
beginning of a component up to the first food
delivery, started out near indifference and
changed from delivery to delivery as the com-
ponent continued. The top panel in Figure 7
shows a result from an experiment with
pigeons and these frequently changing com- Fig. 7. Behavioral evolution at different time scales
ponents (Baum & Davison, 2004). It shows the within pigeons’ key pecking on concurrent VI VI sche-
effect on interfood peck ratio (from the begin- dules. Sessions consisted of seven unsignaled components,
each with a different reinforcer ratio. Both graphs show
ning of the component to the first food, from evolution within components with continuing food deliv-
the first to the second food, etc.) of repeated eries from the same alternative. Top: evolution of inter-
food deliveries from the same alternative, food peck ratio at the just-productive key (P) to the not-
called “continuations.” Peck ratios were calcu- just-productive key (N). Each diamond shows the ratio of
all pecks at P to all pecks at N from the beginning of the
lated as pecks on the just-productive key component or the previous food delivery to the next food
(P) divided by pecks at the not-just-productive delivery. Bottom: evolution of peck ratio in the first
key (N) during the interfood interval. Choice changeover cycle within interfood intervals. The diamonds
increased up to about the sixth continuation and triangles represent the length of the first visit (#1),
and then leveled off at a preference close to which was nearly always to the just-productive key (P1),
and the length of the second visit (#2), which was to the
4:1. Although the graph only shows averages not-just-productive key (N2). The lengths in pecks/visit
and not distributions of individuals, the are indicated on the secondary (right) vertical axis. The
increase indicates directional selection of filled squares represent the ratio of the two visits (P1/N2).
choice at this time scale. To think of the Selection at the smaller time scale (bottom) acted in con-
cert with selection at the larger time scale (top).
increase in relation to Equation 5, we think of
xg as log peck ratio in one component, with
g indexing components. For six pigeons, 35 ses- 500 after six reinforcers). Each point in the
sions, and seven components, the population graph is the average of a population of log
size n starts at 1470 for the interval from com- peck ratios following a certain number of con-
ponent beginning to first food and necessarily tinuing food deliveries, and the sequential
decreases across interfood intervals, because interfood intervals correspond to Time 1, Time
sometimes alternative N produces food 2, Time 3, and so on. The first term on the
(a discontinuation). Though not shown, for right of Equation 5, cov(wg, xg), represents the
each sequential interfood interval, a frequency covariance between log peck ratios (xg) across
distribution of xg exists, with g ranging from components (the individuals, indexed by g)
1 to n (decreasing from 1470 to roughly 1200 with fitness (wg) for one interfood interval
after one reinforcer and ultimately to about (Time 1) to the next interfood interval
336 WILLIAM M. BAUM

(Time 2). As long as peck ratio increases from food was delivered, was longer than the second
food delivery to food delivery, this covariance visit in the component (zero on the x-axis),
is positive. either because of a carry-over from the previ-
These interfood peck ratios, however, tell ous component or simply because it was
only what happens at their time scale. No allo- induced by the light coming on at the begin-
cation of behavior is specified completely by ning of the component. The second visit
the ratio of summed behavior. At least two remained about the same throughout the com-
more parameters must be specified: the length ponent, at about 1-2 pecks. The brevity and
of visits to or bouts of the alternative activities. invariance of the second visits indicates that
All activities are episodic. Working may responding during the interfood interval took
occur for some hours each day, and time with on a fix-and-sample pattern (Aparicio & Baum,
family for a certain time too, varying from day 2006; Baum, 2002; Baum, Schwendiman, &
to day. Similarly, in an experiment with Bell, 1999). Since only the first visits increased,
pigeons’ choice, pecks accumulate in episodes we may conclude that only those visits evolved.
of pecking at the alternatives. Since interfood The line for the visit ratio confirms this reason-
log peck ratio was calculated by summing all ing by following the increase in first visits
pecks at each alternative, it says nothing about closely; the ratio increases from about 2.3 to
any regularity at the smaller time scale of about 6.0 (left vertical axis). The squares repre-
within-interfood-interval activities. The second sent the averages of x1g across components (for
term on the right in Equation 5 represents this i just equal to 1).
smaller time scale. A similar curve for log ratio of Visits 3 and
One way to think about log peck ratio at this 4 (the second changeover cycle) could be cal-
smaller time scale is to compare pairs of culated. It too increased, but not as much as
sequential visits. The first visit was nearly always the log ratio of Visits 1 and 2, because Visit
to the just-productive alternative (P) and the 3 was always shorter than Visit 1. In the light
second to the not-just-productive alternative of Equation 5, we may conclude that the
(N). We may call this pair of visits one change- covariance cov(wg, xg) at the time scale of
over cycle, and because the visits must strictly interfood intervals (group level) and the aver-
alternate between alternatives, we may call the age covariance Eg[cov(wig, xig)] at the time
changeover cycles the individuals within an scale of changeover cycles were both positive.
interfood interval, in analogy to the members Selection on changeover cycles (members)
within a group. If we have measured 10 sequen- acted in concert with selection on interfood
tial visits in each interfood interval, we have five log ratios (groups).
changeover cycles. For each changeover cycle, Given this clear qualitative result, we might
we may calculate the log ratio of pecks per visit be encouraged to calculate the actual fitnesses
in the first visit to pecks per visit in the second and covariances as choice changed. To do this,
visit. This log ratio would be xig in Equation 5, one may calculate from frequency distribu-
with g indexing the roughly 1200-500 interfood tions of the populations of choice measures or
intervals at each ordinal position in the compo- visits. Appendix C shows a worked example,
nent and i indexing the (five) changeover using the visit lengths (pecks/visit) of the first
cycles within the interfood interval. The analy- visit following food—the diamonds in the
sis is analogous to a population of 1200-500 lower graph of Figure 7. The negative curva-
groups with five members each. For illustra- ture seen there indicates that the changes in
tion, we focus on the log peck-per-visit ratio visit length were initially positive and
between the first visit and the second visit—the decreased across food deliveries. The changes
first changeover cycle. The bottom graph in in visit length (Δx, corresponding to Δx in
Figure 7 shows the first postfood visit (P1; bro- Eq. 1) and the covariances between visit
ken line and diamonds), the second visit (N2; length and fitness across food deliveries
broken line and triangles), and their ratio appear in Figure A2. The changes and covar-
(P1/N2; solid line and squares). Each diamond iances are equal, as they must be in accord
or triangle represents an average of visits. The with Equation 1.
first visit increased from about 5-6 pecks to The conclusion that selection acted in con-
about 10-11 pecks (vertical axis to the right). cert at the different time scales shown in
Even the first visit in a component, before any Figure 7 raises a question: In what situation
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 337

would selection at different time scales oppose self-control, a bad habit like smoking incorpo-
one another? rates both a short-term positive contingency
Self-control and impulsivity. Equation 6 for with small reinforcers (nicotine) and a long-
behavioral evolution illuminates problems in term negative contingency with large reinfor-
self-control, including altruism and coopera- cers, whereas the competing good habit incor-
tion. The conflict between impulsivity and self- porates both a short-term negative
control translates into a positive first term on contingency with small reinforcers and a long-
the right (self-control) and a negative second term positive contingency with large reinfor-
term (impulsivity). Avoiding a bad habit or cers. Drunkenness provides short-term enjoy-
cultivating a good habit (e.g., refusing a drink ment and long-term catastrophic losses,
or a piece of cake or visiting the dentist) on whereas sobriety entails short-term loss of
any specific occasion entails the cost (pun- enjoyment and long-term large reinforcers.
isher) of forgoing immediate enjoyment or of Which contingency wins determines the alco-
enduring immediate discomfort, and these are holic’s quality of life.
the parts of a whole extended pattern (xg; In negative punishment, an activity prevents
sobriety, dieting, or good health). The β of the a fitness-enhancing event (a good phylogeneti-
second term is negative, because the individ- cally important event). In a laboratory exam-
ual parts or occasions xig of activity xg incur ple, Sanabria, Sitomer, and Killeen (2006)
this immediate punishment. Thus, the exposed pigeons to a procedure in which
extended pattern (self-control, cooperation, pecking at a lit key was induced by following a
or altruism) has positive β, but the positive β brief lighting of the key with food (“autoshap-
must suffice to offset the negative β of the sec- ing”). The researchers arranged that pecking
ond term (for xig) if the good habit is to sur- the key would cancel the food delivery,
vive. This offsetting may occur if var(xig) is thereby negatively punishing key pecking. The
reduced to zero by following a rule that negative contingency between pecking and
enforces good behavior (refusing the drink or food resulted in large reductions in pecking
cake; doing the right thing) on particular across sessions, indicating that pecking was
occasions. Seen this way, altruism and cooper- selected against. The Price equation tells us
ation are examples of good habits that incur a that the negative punishment resulted in lower
cost in the short term (helping out) but are recurrence of key pecking in competition with
reinforced in the long term (Baum, 2016; other activities across sessions; the covariance
Rachlin, 1995, 2002). between rate of pecking and pecking recur-
Negative contingencies. In a negative con- rence was negative—the opposite of Figure 3.
tingency, one activity prevents a reinforcer or
punisher (i.e., a good or bad phylogenetically
important event) while all other activities allow Conclusion
the event to occur. In negative reinforcement,
an activity prevents injury or some other Although Price’s equation is not a theory to
fitness-reducing event, and that activity is be proven or disproven, it is a useful analytical
selected. For example, Herrnstein and Hine- tool and may be a useful basis for theory. It
line (1966) exposed rats to a procedure in can elucidate selection at different levels in
which pressing a lever canceled upcoming genetic evolution and at different time scales
electric shocks, thereby reducing the shock in behavioral evolution, including cultural evo-
rate to a lower level. Across sessions, lever lution and shaping of behavior in individual
pressing increased, as illustrated from T1 to organisms within a lifetime. The examples in
T2 in Figure 3. The Price equation tells us that the present paper aim to shed light on dynam-
recurrence of lever pressing increased across ics within concurrent VI VI performance and
sessions because the negative contingency allocation on single schedules. The analysis of
between lever pressing and electric shock concurrent performance shows evolution of
selected pressing in competition with other the fix-and-sample pattern that often emerges
activities. The covariance between rate of lever (Aparicio & Baum, 2006; Baum, 2002; Baum &
pressing and recurrence of lever pressing was Davison, 2004). The analysis of response rate
positive because of the selection imposed by suggests a path to a more complete under-
the negative reinforcement. In problems of standing of changing response rates than the
338 WILLIAM M. BAUM

older molecular view afforded. Other applica- Baum, W. M., & Rachlin, H. C. (1969). Choice as time
tions of Price’s equation may elucidate dynam- allocation. Journal of the Experimental Analysis of Behav-
ior, 12, 861-874.
ics of behavioral change in other contexts, Baum, W. M., Schwendiman, J. W., & Bell, K. E. (1999).
particularly if experiments are designed in Choice, contingency discrimination, and foraging the-
advance to allow the equation to be applied ory. Journal of the Experimental Analysis of Behavior, 71,
easily. 355-373.
Boyd, R., & Richerson, P. J. (1985). Culture and the Evolu-
tionary Process. Chicago: University of Chicago Press.
References Catania, A. C. (1973). The concept of the operant in the
analysis of behavior. Behaviorism, 1, 103-115.
Aparicio, C. F., & Baum, W. M. (2006). Fix and sample
Davison, M., & Baum, W. M. (2000). Choice in a variable
with rats in the dynamics of choice. Journal of the Exper-
environment: Every reinforcer counts. Journal of the
imental Analysis of Behavior, 86, 43-63.
Experimental Analysis of Behavior, 74, 1-24.
Aparicio, C. F., & Baum, W. M. (2009). Dynamics of
Donahoe, J. W. (1999). Edward L. Thorndike: The selec-
choice: Relative rate and amount affect local prefer-
tionist connectionist. Journal of the Experimental Analysis
ence at three different time scales. Journal of the Experi-
of Behavior, 72, 451-454.
mental Analysis of Behavior, 91, 293-317.
El Mouden, C., André, J. -B., Morin, O., & Nettle, D.
Baum, W. M. (1973). The correlation-based law of effect. (2014). Cultural transmission and the evolution of
Journal of the Experimental Analysis of Behavior, 20, human behaviour: a general approach based on the
137-153. Price equation. Journal of Evolutionary Biology, 27,
Baum, W. M. (1981). Optimization and the matching law 231-241.
as accounts of instrumental behavior. Journal of the Frank, S. A. (1995). George Price’s contributions to evolu-
Experimental Analysis of Behavior, 36, 387-403. tionary genetics. Journal of Theoretical Biology, 175,
Baum, W. M. (1992). In search of the feedback function 373-388.
for variable-interval schedules. Journal of the Experimen- Glenn, S. S., Ellis, J., & Greenspoon, J. (1992). On the rev-
tal Analysis of Behavior, 57, 365-375. olutionary nature of the operant as a unit of behav-
Baum, W. M. (1993). Performances on ratio and interval ioral selection. American Psychologist, 47, 1329-1336.
schedules of reinforcement: Data and theory. Journal Goodnight, C.J., Schwartz, J. M., & Stevens, L. (1992).
of the Experimental Analysis of Behavior, 59, 245-264. Contextual analysis of models of group selection, soft
Baum, W. M. (2000). Being concrete about culture and selection, hard selection, and the evolution of altru-
cultural evolution. In N. Thompson & F. Tonneau ism. The American Naturalist, 140, 743-761.
(Eds.), Perspectives in ethology (Vol. 13, pp. 181-212). Guerin, B. (1997). How things get done: Socially, non-
New York: Kluwer Academic/Plenum. socially; with words, without words. In L. J. Hayes &
Baum, W. M. (2001). Two stumbling blocks to a general P. M. Ghezzi (Eds.), Investigations in behavioral episte-
account of selection: Replication and information. mology (pp. 219-235). Reno, NV: Context Press.
Behavioral and Brain Sciences, 24, 528. Henrich, J. (2004). Cultural group selection, coevolution-
Baum, W. M. (2002). From molecular to molar: A para- ary processes and large-scale cooperation. Journal of
digm shift in behavior analysis. Journal of the Experimen- Economic Behavior and Organization, 53, 3-35.
tal Analysis of Behavior, 78, 95-116. Herrnstein, R. J. (1970). On the law of effect. Journal of the
Baum, W. M. (2010). Dynamics of choice: A tutorial. Jour- Experimental Analysis of Behavior, 13, 243-266.
nal of the Experimental Analysis of Behavior, 94, 161-174. Herrnstein, R. J., & Hineline, P. N. (1966). Negative rein-
forcement as shock-frequency reduction. Journal of the
Baum, W. M. (2012). Rethinking reinforcement: Alloca-
Experimental Analysis of Behavior, 9, 421-430.
tion, induction, and contingency. Journal of the Experi-
Mayr, E. (1970). Populations, species, and evolution. Cam-
mental Analysis of Behavior, 97, 101-124.
bridge, MA: Harvard University Press.
Baum, W. M. (2013). What counts as behavior: The molar
McElreath, R., & Boyd, R. (2007). Mathematical models of
multiscale view. The Behavior Analyst, 36, 283-293.
social evolution: A guide for the perplexed. Chicago: Uni-
Baum, W. M. (2015). The role of induction in operant versity of Chicago Press.
schedule performance. Behavioural Processes, 114, Morse, W. H. (1966). Intermittent reinforcement. In
26-33. W. K. Honig (Ed.), Operant behavior: Areas of research
Baum, W. M. (2016). Driven by consequences: The multi- and application (pp. 52-108). New York: Appleton-Cen-
scale molar view of choice. Managerial and Decision tury-Crofts.
Economics, 37, 239-248. Panchanathan, K. (2011). George Price, the Price equa-
Baum, W. M., & Davison, M. (2004). Choice in a variable tion, and cultural group selection. Evolution and
environment: Visit patterns in the dynamics of Human Behavior, 32, 368-371.
choice.. Journal of the Experimental Analysis of Behavior, Price, G. R. (1970). Selection and covariance. Nature, 227,
81, 85-127. 520-521.
Baum, W. M., & Davison, M. (2014a). Choice with fre- Price, G. R. (1972). Extension of covariance selection
quently-changing food rates and food ratios. Journal of mathematics. Annals of Human Genetics, 35, 485-490.
the Experimental Analysis of Behavior, 101, 246-274. Price, G. R. (1995). The nature of selection. Journal of The-
Baum, W. M., & Davison, M. (2014b). Background activ- oretical Biology, 175, 389-396.
ities, induction, and behavioral allocation in operant Rachlin, H. (1971). On the tautology of the matching law.
performance. Journal of the Experimental Analysis of Journal of the Experimental Analysis of Behavior, 15,
Behavior, 102, 213-230. 249-251.
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 339

Rachlin, H. (1995). Self-control: Beyond commitment. rate of variable-interval reinforcement. Journal of the
Behavioral and Brain Sciences, 18, 109-159. Experimental Analysis of Behavior, 81, 65-83.
Rachlin, H. (2002). Altruism and selfishness. Behavioral Skinner, B. F. (1938). Behavior of organisms. New York:
and Brain Sciences, 25, 239-296. Appleton-Century-Crofts.
Reid, R. L. (1958). The role of the reinforcer as a stimu- Skinner, B. F. (1953). Science and human behavior.
lus. British Journal of Psychology, 49, 202-209. New York: Macmillan.
Richerson, P. J., Baldini, R., Bell, A., Demps, K., Frost, K., Skinner, B. F. (1969). Operant behavior. In Contingencies of
Hillis, V., … Zefferman, M. R. (2016). Cultural group reinforcement: A theoretical analysis (pp. 105-132).
selection plays an essential role in explaining human New York: Appleton-Century-Crofts.
cooperation: A sketch of the evidence. Behavioral and Skinner, B. F. (1981). Selection by consequences. Science,
Brain Sciences, 39, e30. 213, 501-504.
Rogers, E. M. (1995). Diffusion of innovation, 4th ed. Sober, E. (1984). The nature of selection: Evolutionary theory
New York: Simon and Schuster. in philosophical focus. Cambridge, MA: MIT Press.
Sanabria, F., Sitomer, M. T., & Killeen, P. R. (2006). Nega- Thorndike, E. L. (1911). Animal intelligence. New York:
tive automaintenance omission training is effective. Macmillan. Facsimile reprint by Forgotten Books,
Journal of the Experimental Analysis of Behavior, 86, 1-10. 2012.
Segal, E. F. (1972). Induction and the provenance of oper-
ants. In R. M. Gilbert & J. R. Millenson (Eds.), Rein- Received: June 27, 2016
forcement: behavioral analyses (pp. 1-34). New York: Final Acceptance: April 14, 2017
Academic.
Shull, R. L., Grimes, J. A., & Bennett, J. A. (2004). Bouts of
responding: The relation between bout rate and the
340 WILLIAM M. BAUM
   
Appendix A Δx = cov wg , xg + Ew Δxg ðA3Þ
Derivation of the Basic Price Equation
where g indexes the groups in the population,
Assume a population or group of instead of the individual members.
N members at two time periods or The first term on the right in Equation A3,
“generations,” Time 1 and Time 2. A trait or the covariance, represents group selection
property of these members x varies among alone. We take the second term Ew(Δxg), the
them. The trait could be the presence (x = 1) mean change in a group apart from group
or absence (x = 0) of a particular allele (x = selection, as a mean of change in the individ-
0, 0.5, or 1.0 for diploidy), or a quantitative ual members of each group. For one group,
trait like body size or coloration. At Time 1, the we treat it the same way as we treated Δx. So,
P
mean of x across members is N1 Ni= 1 xi , denoted for one group
x: Each member i contributes yi surviving off-    
Δxg = cov wig , xig + Ew Δxig ðA4Þ
spring or copies to the population at Time 2—
i’s absolute recurrence. Since the sum of y is
not necessarily equal to N, because the popula- where i indexes members within the group.
tion might increase or decrease from Time We substitute Equation A4 into Equation A3
1 to Time 2, absolute recurrence is not a good and arrive at:
measure of fitness. In Figure 1, we could       
equate fitness with yi because the population Δx = cov wg , xg + Eg cov wig , xig + Ew Δxig
size (N = 10) didn’t change. Fitness or relative ðA5Þ
recurrence of a member i, wi, equals yi =y.
P
Since y equals N1 i yi , wi equals P Nyi
y
. The sum which includes the change in notation Eg to
i i indicate that the second term on the right is
of wi equals N, and the mean w equals 1.0. an expected value across groups. It is Equa-
That is, wi is a relative measure and dimen- tion 4 in the paper.
sionless. The mean of x at Time 2 x 0 equals:
X X
x0 = wi xi =N + wi Δxi =N ðA1Þ Appendix B
i i
An alternative approach to individual and
where the first term on the right is the portion of group selection, called “contextual analysis,”
the new x at Time 2 due to selection, and Δxi de-emphasizes covariance and instead parti-
is the change in xi for member i from Time tions variance into within-group and between-
1 to Time 2 due to imperfect fidelity of recur- group components, not entirely unlike Equa-
rence. Since each wNi is a weight (i.e., they sum tion 6 (Goodnight, Schwartz, & Stevens,
to 1.0), the second term on the right can be 1992). The “between-group” component, how-
rewritten as the expected value of Δxi, Ew(Δxi). ever, is conceived of as affecting variance
Using the definition of covariance among individual members as a result of their
P membership in the group. Contextual analysis
( i wi xi =N − w x), we may substitute for the
first term on the right covðwi , xi Þ + w x. Recal- would assume that the fitness of a bee colony
ling that w equals 1.0, and taking the differ- is just the sum of the fitnesses of the individual
ence Δx = x 0 − x leads to the simplest form of bees and would not assign a fitness to the col-
the Price equation (see Price, 1970, and ony in competition with other colonies. In
McElreath & Boyd, 2007, chapter 6, for more other words, contextual analysis treats groups
detail): as if they were collections of individuals and
makes no distinction between collections and
Δx = covðwi ,xi Þ + Ew ðΔxi Þ ðA2Þ groups as integral wholes. The key difference
between contextual analysis and Price’s
approach is that contextual analysis assigns all
selection, fitness, and change to individual
Derivation of the Group Price Equation
organisms’ phenotypes—that is, group change
We may first just represent group selection is just the sum of the changes in the members
in which the groups constitute individuals: as a result of selection acting on them. The
BEHAVIORAL EVOLUTION AND THE PRICE EQUATION 341

Fig. A1. Computation of fitness of first visits (pecks/visit) going from one interfood interval (Time 1) to the next
(Time 2) or first visit in a component (Time 1) to the first interfood interval (Time 2; top left graph). Diamonds show
the frequency distribution of pecks/visit for Time 1 (left vertical axis). Squares show the frequency distribution for Time
2. Each pair allows a calculation of fitness in accord with Price’s equation, from the beginning of the component to the
first food, from the first food to the second, and so on, up to the eighth food delivery. Triangles show mean fitnesses of
the different visit lengths (right vertical axis). The broken line in each graph is a regression line fitted to mean fitness
(triangles) as a function of pecks/visit. Its equation appears in the graph.

two approaches may be appropriate to differ- Price’s approach might make more sense
ent situations. Contextual analysis appears to when a population is structured into true
work well for collections of flour beetles and groups that are integral wholes. Since we are
of leafy plants (Goodnight et al., 1992), but here concerned with behavior, which is
342 WILLIAM M. BAUM

organized into activities that function as inte-


gral wholes, Price’s approach seems more
appropriate (Baum, 2002).

Appendix C
Figure A1 shows frequency distributions of
the first postfood visit (averages in Fig. 7,
lower panel). Each graph contains two fre-
quency distributions (Time 1 and Time 2): for
a number of same-alternative food deliveries
(“continuations”) and the next higher number
of food deliveries. The frequency distribution
for the smaller number of food deliveries is Fig. A2. Change in the mean length of the first visit
shown with diamonds, and the next one is following food (Δx) and covariance between visit length
shown with squares. For example, the upper and fitness plotted against the second ordinal food deliv-
ery of the graphs in Figure A1. The first point shows the
left graph shows the frequency distribution of change and covariance from no food to the first food
the first visit in the component (0 Food; dia- delivery, the second point shows the change and covari-
monds) and the frequency distribution of the ance from the first food to the second food delivery, and
first visit following the first food (1 Food; so on. Change in visit length (Δx representing Δx of Eq. 1;
squares), the second graph on the left shows square) and covariance (X symbol) are equal, in accord
with Price’s equation.
the frequency distribution of the first visit fol-
lowing first food (1 Food; diamonds) and the
first visit following the second food (2 Food;
squares), and so on. Only visits in the range length. The slopes roughly accord, however,
from 1 peck to 24 pecks were included; these with the trend one would expect from the neg-
constituted the vast majority. The fitness of a ative curvature of the visit lengths shown in
visit length going from one interfood interval the bottom graph of Figure 7.
to the next (from Time 1 to Time 2) equals Figure A2 shows the changes in mean visit
the ratio of the relative frequencies in the two length (Δx) as squares. They decrease as
frequency distributions (Appendix A). The fit- expected from Figure 7. The abscissae indicate
nesses of the different visit lengths are repre- the second of the pair of interfood intervals
sented on the right vertical axes and appear in paired in Figure A1. For example, the 1 on the
each graph as triangles. x-axis shows the mean change from 0-1 Food in
The broken line in each graph is a regres- the upper left graph of Figure A1, the 2 on the
sion line fitted to the fitnesses (triangles). Its x-axis shows the mean change from 1-2 Food in
equation is given in the graph. The line gives the second graph on the left, and so on. The X
only a rough indication of the relation symbols show the computed covariance
between fitness and visit length, because it between fitness and visit length for each pair of
takes no account of the differences in fre- interfood intervals (Time 1 and Time 2). As
quency, but only shows how average fitness of expected from Equation 1, the mean change
the different visit lengths varied with visit and the covariance for each pair are equal.

Vous aimerez peut-être aussi