Journal of Experimental Psychology

1962, Vol. 64, No. 5, 482-488 SERIAL POSITION EFFECT OF FREE RECALL
483
1
THE SERIAL POSITION EFFECT OF FREE RECALL RESULTS TABLE 1

BENNET B. MURDOCH, JR. The data were first analyzed to MEAN NUMBER OF WORDS RECALLED
determine if practice effects occurred
University nf Vermont over the four sessions. Analyses of Group Mean SD
variance showed that there was a 10-2 18 6.39 0.76
Recently Murdock (1960) has that the trends which did show up significant (P < .01) improvement 20-1 16 6.87 1.16
shown that in free recall RI, the total were not consistent enough to justify over the four sessions for Groups 10-2, 15-2 19 8.25 1.40
30-1 19 8.82 1.98
number of words recalled after one any clear-cut conclusions. However, 15-2, and 20-2; whereas the effect was 20-2 1.5 8.53 2.08
presentation, is a linear function of t, a rather definite picture of the serial significant at only the .05 level for 40-1 16 8.24 1.08
total presentation time. Nothing was position curve itself did emerge from Group 30-1 and was not significant
said about the serial position effect, the data. Therefore, the present (P > .05) for Groups 20-1 and 40-1.
though this is a well-known phe- article will be restricted to a quanti- However, the largest difference ob- cant differences were found between
nomenon of free recall (e.g., Deese & tive description and attempted ex- tained between the best and the worst Groups 10-2 and 20-1 (t = 1.39), be-
Kaufman, 1957). However, given planation of the serial position curve session for any one group was 1.13 tween Groups 15-2 and 30-1 (t = 1.00),
that there is a serial position effect, of free recall. words, and all other intersession dif- or Groups 20-2 and 40-1 (/ = 0.48).
the simple linear relationship between ferences were less than 1.0 words. The serial position curves are
RI and t is rather surprising. PROCEDURE Therefore, when this practice effect is shown in Fig. 1. Probability of recall
In the customary serial position divided into four sessions and any- is plotted as a function of serial
curve of free recall, probability of Six groups each had a different combina-
tion of list length and presentation rate. where from 10 to 40 serial positions position. For greater generality, we
recall is plotted as a function of serial These six combinations were 10-2, 20-1, 15-2, its effect on the serial position curves would also like to use the data from
position. This means, then, that the 30-1, 20-2, and 40-1; the first number indicates was negligible. studies by Murdock and Babick (1961)
area under the serial position curve is list length and the second number indicates Table 1 shows the means and SDs and Deese and Kaufman (1957). In
equal to RI, the number of words presentation time (in sec.) per item. Thus, of the number of words recalled per
10-2 means a list of 10 words presented at the Murdock-Babick study there were
recalled after one presentation. If R j a rate of 2 sec/item. Notice that the first list (Ri). Rich mean is based on 80 18 5s each tested on 80 different 25-1
is a linear function of t then it must two, middle two, and last two groups were lists per ,5 and from 15 to 19 & per lists. Iii the Deese-Kaufman study
follow that the area under the serial matched for t, total presentation time (20, group. As predicted, groups with the there were two groups of 16 .5s each;
position curve is also a linear function 30, and 40 sec., respectively). same total presentation time did not
For each group there were 80 different lists. one group was tested on 10 different
of t. However, it is not immediately differ significantly in mean number 10-1 lists and the other group was
The lists were constructed by randomly
apparent how the serial position curve selecting words from the (approximately) of words recalled. That is, no signifi- tested on 10 different 32-1 lists. The
varies with t in such a way as to 4000 most common English words (Thorn-
maintain this simple linear rela- dike-Lorge, 1944, G count of 20 and up), I.OOr
tionship. except that homonyms, contractions, and
The present experiment was de- archaic words were excluded.
signed as an attempt to determine Group testing was used. Lists were read
to 5s either at every beat (presentation rate
how the serial position curve varied of 1 sec/item) or at every other beat (pres-
with list length and presentation rate entation rate of 2 sec/item) of an Electric
while still maintaining this linear metronome set at a rate of 60 beats/min.
relationship. Unfortunately, at the After each list there was a recall period of
end of the experiment it was still not 1.5 min. The 5s wrote down as many words,
as they could remember in any order that
clear how this relationship came about they wished. Each recall period was ter-
or, for that matter, whether the rela- minated by a verbal "Ready" signal which
tionship was even linear after all. preceded the start of the next list by 5-10 sec.
The basic reason for this failure was All groups were given 20 lists per session and
four sessions; successive sessions were spaced
1
This study was supported by a research 2-7 days apart. Nothing was said about
grant, M-3330, from the National Institutes rehearsing while the lists were being presented.
of Health. The author would like to thank In all there were 103 5s, students of both
Ellen Lissner, Cynthia Marvin, and Frank sexes from the introductory psychology course
Warhurst for analyzing the serial position who were fulfilling a course requirement. SERIAL POSITION
data. Exact Ns by group are shown in Table 1.
FK;. 1. Serial position curves for the six groups.
482
484 BENNET B. MURDOCK, JR.
S E R I A L POSITION EFFECT OF FREE RECALL 485
serial position data were presented in The yo column gives the value of y
the original article as Fig. 1 (p. 182) lie very close to 7 words. Since the x
when x = 0. If y0 is subtracted from value is 7 words, the recency effect ex-
and we read the points from the two 1.00 this gives the probability that the t e n d s over the last eight serial positions.
curves as accurately as possible. word in the last serial position will be Another way of indicating the simi-
These three serial position curves are correctly recalled. The results for t h e larity among different lists is by the /;.
shown here as Fig. 2. six groups of the present experiment \verc c o l u m n of Table 2. In the Gompertz the
We have, then, nine different serial very similar to each other, and an constant h determines the rate of change.
position curves. In general, the analysis of variance of the n u m b e r cor- Since the values of h are all rather
rectly recalled showed that the groups similar this indicates that all curves have
curves seem to share certain general SERIAL POSITION
did not differ significantly (F = 1.61, a similar rate of change, and if they have
characteristics: a marked recency FIG. 2. Serial position curves for 10-1 df = 5/97, P > .05). The recall prob- a similar rate of change all curves should
effect, a flat middle section, and a and 32-1 lists (Deese & Kaufman, 1957) and abilities were rather high but they were level out at about: the same x value if the
primary effect which is more precipi- 25-1 lists (Murdock & Babick, 1961). not 1.00 (and had they been the Gom- numerical values of g do not differ too
tous though smaller in magnitude pertz would not be applicable); the greatly.
than the recency effect. The presence over the flat part on each serial position corresponding recall probabilities for t h e
The primacy effect appears to extend
of a flat middle section, or asymptote, curve. The constants g and h were ob- Murdock-Babick and Docse-Kaufman over t h e first three or four serial posi-
is clearest in the 40-1 list (Fig. 1), but tained by a least squares method de- data were clearly lower.
t i o n s . This can be seen in the serial
becomes less and less obvious as list scribed by Lewis (1960, pp. 82-88) using The inflection point occurs b e t w e e n position curves of Fig. 1 and 2, as all of
length decreases. Actually, in the two the last eight points of each serial posi- the second and third words from t h e end the curves seem to level out at about the
10-word lists the primacy and recency tion curve (except of the two 10-word of the list and appears to be essentially t h i r d or fourth serial position. The
lists where only the last four or five independent of list l e n g t h and presenta- primacy effect is so short-lived that the
curves may have intersected each tion rate. The evidence for t h i s con- curve is difficult to describe mathe-
other before an asymptote has been points could be used).
The evidence for this conclusion is clusion is given under the Xi c o l u m n of matically. Actually, it may well be
reached. Table 2, where .V; = — I n ( — In g)/ln // ( I n
shown in Table 2 under the r1 column. exponential. Scmilog plots of the first
More specifically, the recency effect In all cases the Gompertz equation is log base c). That is, ,v, is the inflection t h r e e or four points of the nine curves
can adequately be described by the Gom- accounted for more than. 95% of the point, that x value at which the decelera- (using 1.00 — w a s the asymptote for each
pertz double-exponential function. As variance, and the mean coefficient of tion changes from positive to negative. curve) gave reasonable approximations
given by Lewis (1960, p. 81) the equation determination (r2) was 97.79%. The Xi values range from 0.5 words to 2.1 to straight lines and the slopes were
is y = vgh*. Probability of nonrecall(y) Since in all nine cases g < l/e the words with a mean of 1.57 words. Since rather similar to one another. A group
was plotted as a function of list length recency effect is consistently an S shaped the last word in any list has an x value curve based on the mean (y~c) values of
minus serial position (x). Thus, the last curve. This characteristic can be seen of 0, a mean of 1.57 words places the i n - the individual curves was an excellent
word in a list would have an x value of 0, in the serial position curves of .Fig. 1 and flection point midway between the second
f i t ; the rate constant was 0.77 and the
the next to last word an x value of 1, etc. 2. Starting from the last serial position, and third words from the end of the l i s t . i n t e r c e p t was .27 (sec Murdock & Cook,
Both v and g were fractional and positive. each curve is initially positively de- Actually, both Dcese-Kaufman curves 1960). 1 lowever, the fact that this group
The asymptote v was determined from celerated and then soon becomes nega- appear to have inflection points nearer curve was based on only three points
the mean recall probabilities averaged tively decelerated. the end of the list than any of the other should make one hesitant about placing
curves. Otherwise, however, the inflec- too much confidence in it.
tion points cluster rather closely in t h e
TABLE 2 range of-1.5-2.1 words. Finally, the primacy and recency ef-
f e c t s arc spanned by a horizontal asymp-
VALUES FOR GOMPERTZ DOUBLE-EXPONENTIAL FUNCTION TO KIT -' The recency effect extends over t h e tote. The asymptote is considered to
SERIAL POSITION CURVES IN FREE RECALL last eight serial positions and appears to extend from Serial Position 5 up to the
be essentially independent of list l e n g t h last eight serial positions. That is, in a
Group V « yt h r' Xi and presentation rate. The evidence for 20-word list the asymptote would extend
this conclusion is given under the .v.or, from Serial Position 5 through Serial
10-2 .548 .100 .055 .574 97.3% 1.5 6.9 column, where x.tt is that value of x at
20-1 .852 .050 .043 .596 98.8% 2.1 7.9 Position 12, in a 30-word list from Serial
15-2 .622 .026 .016 .518 97.8% 2.0 6.5 which the curve is 95% down. That is, Position 5 through Serial Position 22, etc.
.10-1 .814 .032 .026 .546 99.3% 2.0 7.0 at this point forgetting is 95%, of t h e
20-2 .730 .048 .035 .552 95.9% 1.9 6.9 That the asymptote is essentially hori-
'asymptotic value. The 95% level serves
40-1 .88.5 .036 .032 .557 98.7% 2.0 7.1 zontal is suggested by the middle parts
25-l» .851 .134 .114 .634 98.3% 1.5 as a convenient criterion to m a r k t h e of the serial position curves of Fig. 1
8.1
10-lb .566 .206 .117 .431 98.3% 0.5 4.1 end of the recency effect. and 2.
32-1" .840 .270 .227 .644 95.7% 0.6 7.4 The mean of the x.ts c o l u m n is 6.88
words or, rounded off to the nearest whole A close examination of the serial
•From Murdock and Babick (1961). number, 7 words. Except for the Deese- position curves suggests that the
'From Deese and Kaufman (1957).
Kaufman 10-1 list all the values seem to Irend line may have a small positive
486 BENNET B. MURDOCK, JR. SERIAL POSITION EFFECT OF FREE RECALL 487

TABLE 3 horizontal, and the slight positive Fig. 1, p. 580). For another, it agrees particular the recency effect is more
slope to the curve is no greater than well with some unpublished curves culled pronounced than the primacy effect.
PREDICTED ^ND OBTAINED INCRKMKNTS from several experiments recently re- This experiment clearly shows that the
VOR ASYMPTOTE would be expected from the tail end
ported by Murdock (1960). Finally, the results of Bousfield et al. (1958) are not
of the recency effect. exact same trends are present in some due to t h e slower presentation rate
Group AX Diff.
Pred. Obt. memory-span data reported by Waugh per se.
\ DISCUSSION (I960, Fig. 3, p. 75). Why did Bousfield et al. (1958) find
20-1 8 .039 .100 .061
30-1 18 .022 -.004 -.026 We have presented data to show that However, the empirical curve of Fig. 3 primacy more pronounced than recency?
20-2 8 .019 .015 -.004 is not in agreement with results reported One possibility is their instructions.
40-1 28 .Oil the serial position curve of free recall is
.027 .038 by Bousfield, Whitmarsh, and Esterson Twice in their instructions they told 5s
25-1 13 .044 .036 -.008 characterized by a rather steep (possibly
32-1 20 i .046 .014 exponential) primacy effect, an S shaped (1958). These authors used S-, 10-, 20-, t h a t the words were to be recalled,
.032
\ recency effect, and a horizontal asymp- and 40-word lists all presented at a rate ". . . in the order in which they occur
tote extending between the primacy and of 2.5 sec/word, and consistently f o u n d in your memory." The stress on order
recency effect. An idealized curve for a the primacy effect more marked than the may have given 5s a set to recall the
slope rather t than a zero slope How.
recency effect. Both Bousfield et al.
24-word list is shown in Fig. 3. Its equa- words in the order presented, and Deese
ever this pos,togitive slope cou,d be due (1958, pp. 260-261) and Deese (1957, pp. (1957) has shown that instructions to 5s
t0
tion is
, ncfeVthat the recency effect is 581-582) suggest that the relatively slow are an i m p o r t a n t variable in determining
only 95 /0 do,_ QWn . ; e 5% of the efl[ect , 00 , 27 f presentation rate may have encouraged t h e shape of the curve. A second possi-
7s™*11* l.0 exert an effect on the - .772 (M2)-M(L~') rehearsal and thus led to the greater bility is the design used. Bousfield et a!.
(allegedly) horizontal asymptote. primacy effect. To investigate this (1958) used a counterbalanced design
Ihe proper . tegt of thig condusion> where L is list length and x is Serial possibility, we conducted an additional such that each 5 had only one list at each
then, « to deaetermme whether the ob _ Position 1, 2, 3, . . . L. The constants experiment with 35 5s using 10 20-2.5 length. Thus, in effect each list was
tamed mcrei^^ (if ,my) ig greater for the primacy effect were those of the lists. The 20-word length was selected (to 5) of u n k n o w n length, and this fact
group curve discussed above while the because the curves of Bousfield el: al. may have encouraged rehearsal in the
than the incr>.rement attributable to the constants for the asymptote and the
5% remaining from the recency effect _ (1958, Fig. 1, p. 258) seemed to show the order of presentation.
recency effect were the mean values of most pronounced primacy effect for t h i s In any event, under the conditions of
Ihe follov bwin aual is deals only the constants given in Table 2 for the six length list. As Bousfield et al. (1958) the present experiment there seems little
with lists of )f 2Q wordg or more . the 10 lists of the present experiment. apparently used a somewhat longer doubt that the serial position effect of
and 15 worc^j ,ists could not be uged The curve of Fig. 3 is an empirical recall period we used a 4-min. recall free recall is essentially as depicted in
because the ,iere were tOQ few pointg curve, not a rational curve. It is an period in this additional e x p e r i m e n t ; Fig. 3. Of course, as Deese (1957, p. 581)
For each of |f the gk ,igts the obtained attempt to describe the serial position otherwise the procedure was identical has noted, the serial position curve is
increment , wag found by fitti a effect of free recall quantitatively, not with that of the other experiments re- sensitive to the introduction of experi-
least squareVes regreasion Hne to the explain it. Not only does this empirical ported here. mental variables. However, it has been
curve represent the nine curves of Fig. 1 The results of the experiment are found that more items are recalled with
asymptote, ^ determining its siopei then and 2 quite well, but also it is consistent
nmltiplymg^.the glope fay ^ where Ax. shown in Fig. 4. As can be seen, in free recall than with ordered recall
with several other sets of data. For general the results are quite consistent (Dcesc, 1957; Waugh, 1961), so evidently
is the differ,l!rence between Serial Posi. one, it agrees with serial position curves with the empirical curve of Fig. 1, and in free recall is the preferred, perhaps even
tion 5 and tr!^he sev enth-from-last serial for 20-1 lists reported by Deese (1957, the more basic, method of recalling a list
position (A^ = 8 f o r the 20.word ligti of unrelated words.
-,1.00
Ax - 13 I01,or the 25.Word list, etc.). Finally, why does the serial position
1 he expectested ; ncrement was foimd by curve of free recall take the shape it
obtaining t\h& preclicted y value from g.80
.80 <
O
does? One possible explanation is in
•t-l-tf\ I ^/-MTu-^/ii. • r i Ul terms of short-term proactive and retro-
the ;z equation tor the two (£
values of active inhibition. That is, each word
(Serial Position 5 and
in a list is both preceded by anywhere
. -um-last serial position), Uot from 0 to (L — 1) other words and fol-
then subtra ^j For each Hst the
constants s^hown {n Tab)e 2 were uged lowed by anywhere from (L — 1) to 0
other words. Up to a point, the more
Ihe prediQiicted and obtained incre. preceding words the more short-term PI
ments are ghown in TaWe 3; the and the more succeeding words the more
difference tj between the predicted and 5 10 15 short-term RI. The PI and RI effects
obtained i:| increments was not sta. SERIAL POSITION
SERIAL POSITION
presumably summate to determine the
"pf.! 7Significant « = 0.72, d/ = 5). FIG. 3. Idealized serial position curve total inhibitory effects.
, t ie a!tasymptote does appear to be for 24-word list. .Fin. 4. Serial position curve for 20-2.5 lists. If this explanation is correct, recent
488 BENNET B. MURDOCK, JR. Journal uf Experimental Psycftotogv
1962, Vol. 64, No. S, 489-494

studies of the short-term retention of and a horizontal asymptote spanning the

individual items should provide an in- primacy and recency effect. Finally, it was THE SCALING OK SUBJECTIVE ROUGHNESS
dication of the course of PI and RI to be suggested that the shape of the curve may AND SMOOTHNESS '
expected. It has been found that, in well result from proactive and retroactive
inhibition effects occurring within the list
short-term memory, PI effects appear to itself.
S. S. STEVENS A N D J U D I T H R I C H H A R R I S
be greatest after about three prior words Harvard University
(Murdock, 1961). This agrees well with REFERENCES
the finding that the primacy effect levels BOUSFIELD, W. A., WHITMARSH, G. A., &
out after the first three or four serial ESTERSON, J. Serial position effects and
These experiments begun us an at- a logarithmic f u n c t i o n (Stevens & Gaianter,
positions. In short-term memory RI tempt to apply the method of magni- 1957).
the "Marbe effect" in the free recall of
effects appear to approach an asymptotic meaningful words. /. gen. Psychol., 19S8, tude estimation to a continuum for Since the available number of emery cloth
value greater than zero (Murdock, 1961; 59, 255-262. grits is limited, it is difficult to determine a
which the stimulus seemed to have no pure category scale, but when a sample of
Peterson & Peterson, 1959). This agrees DEESE, J. Serial organization in the recall of metric scale, only an ordinal scale of grits more uniformly spaced in subjective
well with the finding of a horizontal disconnected items. Psychol. Rep., 1957, 3,
grades of sandpaper, or emery cloth. roughness was used, t h e form of the category
asymptote in the serial position curve. 577-582.
Unexpected discoveries led on to more scale changed as predicted: it became much
Finally, an examination of the RI curve DEESE, J., & KAUFMAN, R. A. Serial effects loss curved than a logarithmic function when
of short-term memory even suggests an in recall of unorganized and sequentially engaging inquiries. At the outset,
plotted against, the ratio scale of subjective
S shaped curve (see proportion of correct organized verbal material. /. exp. Psychol., both a ratio scale and a category scale roughness. Ten Os judged grits 320, 120,
1957, 54, 180-187. of apparent tactual roughness were 80, 50, 40, 30, and 24 twice each on a seven-
recalls over different retention intervals,
LHWIS, D. Quantitative methods in psychology. determined with 12 grades (grits) of point scale. The average judgments were
Tabjes 1 and 3, Murdock, 1961, pp. 619- New York: McGraw-Hill, 1960.
620). This agrees well with the Gom- emery cloth. The relation between 1.17, 2.17, 3.04, 4.12, 5.17, 6.08, and 6.62.
MURDOCK, B. B., JK. The immediate reten- These values determine a line that is straighter
pertz recency effect suggested here. tion of unrelated words. /. exp. Psychol., the ratio scale and the category scale t h a n a logarithmic function—a line that is not
Thus, it would appear that all the main 1960, 60, 222-234. was typical of the relation found on far from the pure form of the category scale,
characteristics of the idealized serial posi- MURDOCK, B. B., JR. The retention of prothctic contimia. (For these first as evidenced by the tendency of Os to use each
tion curve shown in Fig. 3 are compatible individual items. /. exp. Psychol., 1961, 62, results, see Stevens, 1961a; for an category number approximately equally often
with the results obtained from the short- 618-625. earlier related study, see Duclelc cv (Stevens & Gaianter, 1957). It appears,
term retention of individual items, and MURDOCH B. B., J R . , & BAUICK, A. J. The Baker, 1956.) therefore, that: roughness behaves like a
these findings lend support to the idea effect of repetition on the retention of prothetic c o n t i n u u m , and that the pure
that the serial position curve of free individual words. Atner. J. Psychol., 1961, The next study was an exercise in which category scale is not a logarithmic function
recall is essentially a manifestation of 74, 596-601. two students, C. S. Harris and J. P. McMa- of t h e magnitude scale (Eisler, 1962).
short-term PI and RI effects. MURDOCK, B. B., JR., & COOK, C. D. On hon, asked 12 (>s to judge the smoothness of
fitting the exponential. Psychol. Rep., the stimuli instead of the roughness. The I'Yom these preliminary studies a
1960, 6, 63-69. ratio scale of smoothness approximated the surprising fact emerged: magnitude
SUMMARY PETERSON, L. R., & PETERSON, M. J. Short- estimations of roughness, and of
inverse, or reciprocal, of t h e ratio scale found
This experiment, was a study of the serial term retention of individual verbal items. for roughness, and the category scale of smoothness, turned out to give fairly
position effect of free recall. Curves were J. exp. Psyrhol., 1959, 58, 193-198. smoothness was the reverse, or the comple- straight lines when plotted in log-log
obtained for 10-2, 20-1, 15-2, 30-1, 20-2, and THORNDIKK, E. L., & LORGE, I. The teacher's ment, of the category scale of roughness.
•word book of 30,000 words. New York: These results resemble Torgerson's (1960) coordinates against the grit number of
40-1 lists, where the first, number indicates
list length and the second number indicates Teachers College, Columbia University, findings when he scaled both the apparent the emery cloths. Another instance,
presentation time per word. On the basis of 1944. lightness- and the apparent darkness of gray it seems, of the psychophysical power
papers.
the available evidence it was concluded that, WAUGH, N. C. Serial position and the law. (Grit number refers to the
under the conditions of the present experi- memory-span. Amer. J. Psychol., 1960, In terms of a linear scale of apparent
roughness, it turned out that the stimuli usednumber of openings per inch in the
ment, the serial position curve is characterized 73, 68-79.
by a steep, possible exponential, primacy were bunched rather t i g h t l y at the low screen employed to sift the abrasive
WAUGH, N. C. Kree versus serial recall.
effect extending over the first three or four ./. exp. Psychol., 1961, 62, 496-502. particles.) I f apparent roughness and
Nnooth) end of the c o n t i n u u m , so much so
words in the list, an S shaped recency effect that the two category scales were almost its reciprocal, apparent smoothness,
extending over the last eight words in the list, (Received October 6, 1961) logarithmic functions of the respective ratio are power functions of particle size, it
scales. Other studies of category scaling
suggest that the nearly logarithmic form becomes a challenging task to deter-
mine more accurately the exponents
found here is an accident of the stimulus spac-
ing, and that if an iterative procedure were involved. In the preliminary experi-
used to arrive at a "pure" category scale the ments, each involving 12 Os, the
curve for roughness would be less curved than
approximate exponents were — 1.5 for
'This research was supported by Grant roughness and +1.2 for smoothness
G-10716 from the National Science Founda- when measured against grit number.
tion (Report PAR-266). The next problem was to determine
489
r