89

An Individualistic Approach to Item Analysis

Georg Rasch University of Copenhagen

1. Introduction
Traditionally the properties of a psychological test are defined in terms of
variations within some specified population. In practice such populations
may be selected in various reasonable ways, and accordingly the properties
referred to—for example, the reliability coefficient —are not specific to the
test itself but may vary according to how the population is defined. Similarly,
the evaluation of a subject is usually linked up with a population by a stald-
ardization of some kind and is therefore not specific to the subject per se.
Our .Aim is to develop probabilistic models in the application of which the
population can be ignored. It was a discovery of some mathematical sig-
nificance that such models could be constructed, and it seemed remarkable
that data collected in routine psychological testing could be fairly well
represented by such models.
in a previous study (2) an attempt was made to build up a general frame-
work within which the models:of an ca..lier study ;3) ap - ear to be special
cases, and some properties of this general fr%rnework were recognized. Bat
only recently it become quite clear that this model is in fact the complete
answer to the requirement that stateincat.% -thout the parameters and ade-
quacv of a discrete probabilistic model be ohjecti"e in a seise to he fully
specified.
At present, at least, the theory leading to this result is rather invOlved,
and it is not going to be a main topic fcir this parer. However, it is intended
that the following discussion on one of the models in the earlier study (3),
the model for item analysis in case of only two possible answers, should
demonstrate the nature of the type of objectivity we are aiming at, thus
pointing to the more general problem to be treated elsewhere.

2. Data
The kind of situation to be considered is the following; in which a fairly
large number of subjects were given an intelligence test. two subtests,
N (completing, numerical sequences) and F (analyzing geometric figures),
are of particular interest. The time allowed for N was chosen so that very
few of the subjects could be expected to achieve an appreciably larger num-

Excemt from P SOCIAL- SCIENCE /7?“,)

 •• •

90 Georg Rasch
An individualistic Approach fo Item Analysis 91

ber of correct answers even with unlimited time. (This fact was ascertained
by independent experimental evidence.) Therefore items that were not on one repetition the results are usually somewhat improved. A large num-
answered were counted as incorrect, and one of the two responses, correct ber of repetitions have not been tried, mainly because the questions are
such that it seems almost certain that several of them will easily be recog-
(+) or incorrect (—), was recorded for each item. Test F was scored in a
similar way. nized after a few repetitions. Therefore the possibility of a direct approach
to the estimation of response probabilities seems remote.
To compensate, we have recourse to an assumption that may seem
rather bold, possibly even artificia namely that AN can be factored into a
3. Model subject parameter and an item parameter. This, combined with the two
The model to be suggested is based on three assumptions: other assumptions, produces a model that turns out to have rather remark-
able properties, some of which even lead to the possibility of examining
1. To each situation of a subject (v) having to solve a test item (1), there
how well the model represents the data (see Section 6).
corresponds a probability of a correct answer which we will write in the Provided the two kinds of parameters can be operationally defined, they
form also have a clear meaning, as seen by inserting 3.2 into 3.1:

(3.1) f+1v, = 1+ Api o. (3.3) P I+ v ' = + epet .

2. The situational parameter A,,, is the product pf two factors,
Thus, if the same person is given items with Ei approaching 0, then his
probability of giving a correct answer approaches 0 while his probability -
(32) = evei. of giving an incorrect answer tends toward unity. And that is true for every
person, provided the model holds. Similarly, when ei gets large, the proba-
pertaining to the subject, e i to the item.
bility of + tends toward 1 and the probability of — toward 0. Since with
3. All answers, given the parameters, are stochastically independent.
increasing Ei the items become easier, we may call Ei the degree of easiness
of item i.
Each of these assumptions calls for some comments.
On the other hand, giving the same item to persons with approaching 0,
I. For description of observations two apparently antagonistic types of
we get probabilities of correct answers tending toward 0, while if 6, increases
models are available, deterministic models (such as the law of gravitation)
indefinitely the probability tends toward 1. This holds for every item.
and stochastic models (such as Nlendel's laws of heredity). However, the
Thus we may colloquially call fp the ability of subject v with respect to the
choice of one type or the other does not imply that the phenomena ob-
kind of items in question.
served were causally determined or that they did occur by chance.
In the definition of and e i there is an inherent indeterminacy. In fact, if
Even if it were believed that certain phenomena could be "explained
„v = 1, . N and Ei, . k is a set of solutions to the equations
causally" (whatever such a phrase may mean), a stochastic model may be
preferable (as in thermodynamics).
Although adopting a probabilistic model for describing responses to an (3.4) epei = A.4
intelligence test, we have taken no sides in a possible argument about
responses being ultimately explainable in causal tterms. then, if G, e: is another set of solutions, the relation
2. In many psychophysical threshold experiments a subject is exposed to
the same stimulus a large number of times. Assuming that the repetitions do
(3.5) eLE;
not affect the judgments of the subject, this procedure gives the opportunity
of estimating each A.,, 1 separately and hence of studying directly how the
situational parameter varies with subject and with strength of stimulus. In must hold for any combination of v and i. Thus -
such a situation we may or may not observe the multiplicative rule laid
down in 3.2.
For the inteiligence tests we shall deal with, experience has shown that (3.6)
92 Georg fly Neh Approach 10 Item Analysis 93

must be a constant, say, and accordingly the general solution is

(3.11) {p{W, (i.)} = p{(+)}p{(i+)}
1 pf(t), (4) 1= p{(0}p{(0) etc.
3.7) x =
, € =- e„ a > 0 arbitrary.
a
If in the first of these equations we divide by p{ (4,)} and in the second by
The indeterminacy can be removed by the choice of one of the items, say p{ (?)), we get
i = 0, as the standard item having "a unit of easiness," in multiples of which
the degrees of easinesses of the other items are expressed.
(3.12)
p{(..), i)}
pf(0, (')} = p{(..),(
p{
By this choice, or an equivalent one, the whole set of et's and et's is p{(0}— =
fixed. In particular P {M}

(3.8) The ratio (3.12) of the two probabilities is the conditional probability of
a + answer to i, given a + answer toj. The notation is
that is, the parameter of a subject is a very simple function of his probability
of giving a correct answer to the standard item,
(3.13) P{ ; '+) (+)} Pf() (4.)}
P{( i+ )}
(3.9)
pf+lv,01
Apo=
1 — pf-F- I p,01
Thus the relations (3.12) can be written

being the "betting odds" on a correct answer. Now we may be able to find
(3.14) p{(01 (?;,)}=p{(.1) I (L)} =p{(4.)},
a person who has in fact his e =1. We may refer to him as a standard subject
(v = 0). And then the item parameter
that is, the probability of a plus answer to i is independent of whether the
(3.10) answer to j is + or —; it is just the probability of a plus answer to i.
Et = kGi
And of course the same holds for a minus answer to i. This is a specifica-
•‘
is the same simple function of the probability that the standard person gives tion of the statement that the answers to i and jare stochastically independent.
a correct answer to this item. Assumption 3 also requires that for each subject the answers to all
3. To some psychologists the assumption of stochastic independence at questions be stochastically independent. Technically this is expressed in
first sight appears to be rather startling, since it is well known that usually the equation
quite high correlation coefficients between responses to different items
are found. (3.15) p{(1),(+), (!0} =P{(+)} P{M}
Correlated items are, however, a consequehce of the assumption. With
moderate variation of e, say from 0.1 to 10, we will obtain quite high cor- and all its analogues. The content of this statement is that the probability of
relation coefficients. But, of course, if e is the same, or nearly the same, for a certain answer to an item or of a combination of answers to a set of items
all individuals, the correlations become zero. or nearly zero. Under this is unaffected by the answers given to the other items.
model the interitem correlations do not represent intrinsic properties of the
items, but are mainly determined by the variations in the person parameters.
Let p { ( 0} and j){ (! ) stand. for a person's probabilities of a positive
4. Comparison of two items
and negative response. respectively, to item i. Considering next his possible
respOnses to two items, i and j, they can also be allotted probabilities: As an introduction to the more general treatment of the model in Section 5
p{ ), )), and so on. Now our third assumption states, among other we will consider how two items can he compared.
things, that his responses to. i and j should be "stochastically independent." According to 3.11 and 3.3, the probability of correct answers to both
Technically this is expressed in the following relations: item i and item j is
 An Indiridualisiic Approach to lieu: Analysis 95
94 Georg Raseh

Accordingl:', by the relation

p{(1),(4) 161= ,P{G) I C}P{(4) le}
(4.1) e 2E, Ej Ei C
(1 + e e,) (1 + fej (4.10)
)
Ei+ Ej = n

for a subject with the parameter e. Similarly, the ratio (OE) is estimated independently of the subject parameters, the
distribution of which is therefore irrelevant in this connection. /
(4.2) P{(+),(') lel— (1
e ei Furthermore, we may get a check on the model by first stratifying the
6€.0' subjects according to any principle—educational level or socioeconomic
status or even , -andomly —and then applying 4.10 to each of the groups. For
(4.3) pf(!),(+) = the model to hold, the ratio Ede.; should be the same in all of the groups and
(1 +e€,)(1 +fei)' the variation of the estimates obtained should therefore concur with the
binomial distributions 4.9.
1
(4.4) p{ 1),(.i4
( 1e} = (1 + e€,) (1 + fei) • The appropriate test for this constancy has a remarkable property.
Denote the within-group c's and /1'S by cil , n9 , with g =1, . . h, and their
totals by c. and it.. Since the groups could be collected into one group of
With the notations
size n. to which 4.9 applies, we have
a, = 1f in case of answer + to item i
(4.5)
0 in case of answer — to item i (4.11) P{c. I n.} = )c.
• Ei Ej Ei Ej
and
(4.6) a. = ai ai, On the other hand, the joint probability of the numbers c,,, due to
their stochastic independence, is
the probabilities of a.= 0 and 2 are given by 4.1 and 4.4, while the proba- )n„ —c,
n ( E:
bility of a. = 1 is the sum of 4.2 and 4.3: (4.12) P • • ., Ch In" ..., 14} , g)
; .
(

Ei 7r Ej ei
9=1

pia.= 11 el = ( + e€,)(l
(€,
(4.7) + Ej) In consequence the conditional probability of c,, ch given the total c.,
+
obtained by dividing 4.11 into 4.12, becomes independent of E i and ej :

Now the conditional probability of ai = I provided a.= 1 is—analogous to

3.13—obtained by dividing 4.7 into 4.2. However, by that operation the
ft
=1
(4.13) ch I c.,n,, nn} =
common denominator and e in the numerators cancel and we are left with (C :)

p{a, = 1 I a.= 1,e} 1"--

It follows that as far as the items i and j are concerned, the testing of the
(4.8) ei model can be carried out in a way that is independent of all of the parameters.
In the formal derivation of the fundamental relation 4.8, subjects and
irrespective of the subject parameter e. items can of course be interchanged. Thus the comparison of two subjects
Considering, then, a number, n, of subjects, all of whom happened to have and v by ,:leans of a single item with parameter E leads to the conditional
a= 1, the probability that c of them have a i = 1 (and thus cif = 0) is given probability
by the binomial law:

c —c (4.14) p „= 1 = —
) G+ Sv
(4.9) Pic I n} = E, E; Ei Ej
96 Georg Rasch An Individualistic Approach to Item Analysis 97

where a,,, ar, and a. have a meaning similar to 4.5 and 4.6. Probability 4.14 We can obtain r = 1 in k different ways—
is independent of which item was used.
In principle, therefore, it should be possible to estimate the ratio in- al = 1, a2 = . = ak= 0,
dependently of the item parameters. In practice, however, this method
does not work, because the number of items—in contrast to the number of a1 = 0, a2 = 1, a3 =...=ak=0
subjects—usually is small. (5.6)

a, = a, =- . . . = 0, ak---- 1

5. Generalization to k items —with the probabilities

In generalizing the results of the preceding section we will first consider the
responses of an individual with parameter to k items. With the notation eE1 E2 fek
(5.7)
4.5 and the adaptation Y(e) Y(e) • 7(f)'

(5.1) a.= a l + . . . + ak the sum of which is the probability

of 4.5, we may condense 3.3 to “Ei . . .

(5.8) Pia.= I 61 =
Y(f)
(f€,)ai
(5.2) P{a, I — 1 + eei
We can obtain r= 2 in (!i) different ways, namely by taking any two of the
ai's to be 1, the rest of them being 0. The probabilities of these combinations
and the generalization of 4.1 through 4,4 to are
2 E1€2 2 E1E3 f2E2E3
5 f2Ek-lEk
(5.9)
Pia., • • aki el = Pia' I fl • • • Pfakl el Y(f ) ' Y(6) ' y(f) ' •' y(f) '
(5.3)
Ca. E ;:k and the sum of them is
k
fi (1 +fei ) (5.10) P{a. -= 2 f} —
C2 (f i eo + +ek-Jek)
i=i Y(f) •

In general a. = r can be obtained in CO different ways, by taking any r out.

recalling that a, is either zero or one. From this result we derive the proba- of the k ai's to be I, the rest of them being 0. The probabilities of these
bility that a, takes on a specified value r. If r = 0, every ai = 0, and thus combinations being

(5.4) P{a.= Ole} = y(e) ,

1
(5.11) er Et • • • Er fr Ei • • • Er- +1 e r Ek-r-• 1 • • • Ek
Y(f) y(e) 7(e)
where for short we write the probability of a. = r becomes

(5.5) U (i+cc)=Y(e) • (5.12) pla.= r I 0 =

y(f)
 An Individualistic Approach to Item Analysis 99
9S Georg Rasa,

where, for short,

ance with the distribution 5.3, which depends on e.
Also a random variable
it a,, the distribution of which 5.12 depends on and therefore it can be
used for the estimation. Rut what 5.17 tells is that the constellation of O's
(5.13) 'Yr = (El • • • Er) + • • • + (Ek - r + I • • •c nttcC 1's prOducing a., which also varies at random, has a distribution that
does not depend On From the fundamental principle it then follows that
In particular for r= k, 5.13 contains only one term,
once a, has been recorded, any extra information about which of the items
were answered correctly is, according to our model, useless- as a source of
(5.14) yk = E,E2 • • • E.
inference about 6 (but not for other purposes, as will presently be seen).
The capital discovery that such situations exist was made by R. A.
If in 5.12 we let r pass through the values 0,1, ..., k, all possibilities have
Fisher in 1922, and following his terminology we shall call a. a sufficient -
been exhausted and therefore the probabilities must add up to unity:
statistic or estanator for the parameter in question.
— —

In the present situation, however, tie sufficiency of a. needs a qualifica-

(5.15) ± pia.= r 161 = I. tion as being relative, since it rests upon tne condition that the c i's are known.
As long as such knowledge is not available, the sufficiency as such is not
Hence very helpful, but the important point of 5.17 then is that it depends solely
(5.16) y(e) = E 'Yr upon the E'S, not on e.
r From 5.17 we can therefore proceed, as we did from 4.8. to consider a
collection of subjects that all happen to have a. = r. Specifying by n„; the
that is, yr are the coefficients in the expansion of the product 5.5 in powers m of subject v and denoting by (n o); given v, the set of responses a -- co • • •, a rk,
of 6.* If the E ' S were known, the y r 's could be'computed and it would be that is,
e
possible from an observed a to estimate and td indicate the precision of
the estimate —for example, in terms of confidence intervals. Thus a. is what (5.18) (auj = (aV1 • • • 9 aVk ) 9

is called an estimator for How to compute an estimate from the estimator

is not our concern at present, but as an estimator a. has an important we can rewrite 5.17 in the form
property.
"ri Er.oz
On dividing 5.12 into 5.3 we obtain the conditional probability of the (5.19) pf(a„,)jap• =1.1 = v= I, n.
m's, given that their sum is r. Through this operation both the common 'Yr
denominator and the common power of cancel and we get
€ 71 ek'k The responses of the n persons being independent, their joint probability
(5.17) P{ai, ak I a.= r,e} = is obtai ,ed by multiplying the n probabilities of 5.19. Denoting for short the
Yr
whale set of n x k responses by ( (a,1 )) the double parentheses indicating
—

which is independent of 6, the parameter to beestimated. variation over both 7, and i— we get

(5.20)
el" Elf.k
In order to realize the significance of this result we can turn to an obvious Pi( (ari) ) I (a,,.= r)} _ '
'Y'rt
but fundamental principle of science, namely, that if we want to know some-
thing about a quantity—for example, a parameter of a model—then we have where
to observe something that depends on that quantity, something that changes
if the quantity varies materially. For the purpose of estimating the parameter (5.21) E ape.
of a person, the observations al; ak are available. On repetition of the
experiment they would, according to our theory. vary at random in concord-
Statement 5.20 i.nplies that, as a consequence of the model, we have to deal
with the total number of correct answers to each item or the n persons
in question..
*In algebra they are known as elementary symmetric functions of
100 Georg Rasch A., Individualistic Approach to lion Analysis 101

6. Separation of parameters sidered. However, this procedure is greatly simplified by the fact that all the
proaabilities to be added are equal, namely—according to 6.4—
Let us finally consider the responses of n individuals with the parameters
6„ to k items with the parameters El, ek . With the notation a„i
introduced in the last section, the model 5.2 now takes the form
n
p=.
(6.5)
y ( e.),(E,) )
(6.1) P{avile v,€,} = (evel)avi
I + ei '
Thus we have only to count the number of different ways in which it is
and on the assumption of stochastic independence of all of the responses algebraically possible to build up a zero/one matrix with the row totals of
avi, v= 1, . . . n, i= I, . k, the joint probability of the whole set ((a vi)) and the column totals of i= 1, . . k.
of them becomes Determining this number is a combinatorial problem that appears to be
rather difficult, but at present we -teed nothing more that a notation. For
n k
this number we write
P{((avi)) I (p) ,(€,)} = I l Pfarii eviei
y=i
n k (6.6) [ r ,, r,
fl (f,,E.)".4 ..., ski
[ l(si)J 1
I, = 1 1=1
(6.2)
Hn k
+ evo and then we have
.=, ,=,

In the numerator we notice that the parameter e„

occurs in k places, each
Ft CI,' TT EP
(6.7) /Ma,. = si) I (ep),(Ei)} = Pcs,r:) ) ] ( cv : ( El m
time raised to a power ao , which all together makes C„1,•, and that the
parameter Ei occurs in n places, each time raised to a power do , adding up
to a total power of a.i . If furthermore the denominator is denoted by This joint probability distribution of the row totals a„. and the column
totals a. ; contains just as many parameters observables, and the latter
n k
would therefore seem suitable for estimation purposes. How true this is
(6.3) (ev) , (c) = 1-1 fl ( I + evEi), becomes clear when we divide 6.7 into 6.4—or 6.5—to obtain the proba-
bility of the whole set of observations, on the condition that the totals of
we can simplify 6.2 to rows and columns are given. In fact, all parametric terms cancel, and we are
left with a conditional probability

=1
ti • II €4' i =1 (6:8) P{((avi)) I (a,.= ,(a.i = si)}
(6.4) pi( (api)) I (f.),(€i)} —
( e, ),(Ei)) •
[ (.s.:)
(r) 1
This formula is the generalization of 5.3 ton persons, but in consequence that is independent of all of the parameters.
of 6.4 we now have to derive the probability that . and (I A , . • a.k Therefore, once the totals have been recorded, any further statement as
take on two specified sets of values: r,, .,r„ and s 1 , .. regards which of the items were, answered correctly by which persons is,
In analogy to Section 5— in particular the logical chain of 5.11 through according to our model, tHeless as a source of information about the
5.13—we should find all possible ways of building up zero/one matrices parameters. (Another use that can be made of the a„ i 's will emerge at a
((ard) that have the same row totals = r„, v = 1, . and column totals later stage of our discussion.) Thus the row totals and the column totals are
i =1, . . k, state the probability of each realization, and add up all not only suitable for zstimating the parameters; they imply every possible
such probabilities to a total joint probability of the two sets of totals con- statement about the parameters that can be made on the basis of the ob-
102 Georg Rasch
An individualistic Approach to item Analysis- 103

serrations. ((a0 )). Accordingly we will, in continuation of the terminology With this- notation the surnr ,ation yields on analogy to 6.10:
introduced in Section 5, characterize the row totals ay. , v = 1, . . n and
k
the column totals a. , i = 1, . . k as a set of sufficient estimators for the Ytxd( (ev)) flefi
parameters f,, . . and e,, . Ek. i==-ii
As 6.7 contains both sets of parameters, a direct utilization of this
(6.13) pf(a.,= s,) I (ep) ,(Ei)1 =
v((ep ) ,(f , ))
formula would apparently lead to a simultaneous estimation of both sets.
However, in view of previous results (see the comments following 5.17) it and accordingly the e i 's might be estimated from the column totals provided
would seem appropriate to ask whether it is possible—also in this general e,'s were known.
case—to estimate the item parameters independently of the person param- Thus we might estimate the e's if the c's were known, and the E'S if the
eters and, if so, vice versa as well. were known! And both e:.timations would even be relatively sufficient.
In order to approach this problem we will first derive the distribution of In fact, on dividing 6.10 into 6.7 to obtain the conditional probability of a.;
the row totals, appearing as exponents of the irrespective of the values for given au., we get
of the column totals, by summing 6.7 over all possible combinations of
s,, . . sk . During this summation the denominator y((p), (e 1)) remains
constant, as do the terms e;v, v = 1, ... n in the numerator. Thus, on II €7.1
(6.14) (a i= si) I (ay. = rv) ,(e,) ,(€i)1 =[((s?) ]
y,r ,)( €,)
."{

introducing the notation

which is independent of- the parameters to be estimated. Similarly, the

(6.9) y.(rv )((€,)) = E
i
i
)
[ (
( si fl €. ekik division of 6.13 into 6.10 gives

we obtain
(6.15) pl(a„.=- ri,) I (a ,(e,) ,(ci)1 =[ (r'; ) ] y( Psi) (1

Y.0 -0((Ei)) • 11 cv
„=, which is independent of the e's. But, of course, as long as neither set of
(6.10) pf(a,, = r„) I (ev) ,( )1 —
y( ( e.) ,(0 ) parameters is known, these possihildies are of no avail.
It is one of the characteristic features of the model under consideration
that this vicious circle can be broken. the instrument being a reinterpretation
from which it is seen that the e,.'s might be estimated from the row totals of the formulas 6.14 and 6.15. In fact, as 6.14 depends on the E'S but not on
if the Ei'S —and therefore also the polynomials 6.9—were known. the i's, this formula gives the opportunity of estimating the E'S without
Similarly, we can sum 6.7 over all possible combinations of r„ . . r„, dealing with the 6's. Thus the objections to both 6.7 and 6.13 have been
keeping s 1 , sk fixed. Substituting, in 6.9, el, 6„ for E„ ek and in
eliminated. The unknown C's in these expressions have been replaced with
consequence interchanging the r's and the s's, we get observable quantities: the individual totals Similarly, in 6.15 the E'S of
f..7 and 6.10 have been replaced with the item totals a in consequence of

(6.11) y(R,>((ep)) = E
(r,,
Nrs which we can estimate the „;'s without knowing or simultaneously estimating
the E . S. Thus the estimation of the two sets of parameters can be separated
from each other.
!n this connection we can return to 6.8, noting that this formula is a
where, by the way, consequence of the model structure-3.3 and he stochastic independence—
irrespective of the values of the parameters of which the right-hand term is
independent. Therefore. if from a given matrix ((ti,, i )) we construct a quantity
(6.12) . that would be useful for disclosing a particular type of departure from the
r(sr,i'd -- - {((srvi )]. model, then its sampling distribution as conditioned by the marg,inalsa„, and
I04 Georg Rawl; /la Individualistic Approach to /tem AtattyAis 105

a.; will be independent of all of the parameters. Thus the testing of the model parameters. Similarly, any two stimuli can be compared independently of
can be separated from dealing with the parameters. all other parameters than those of the two stimuli, the parameters of all
We will not consider here the question of how to perform such testing in other stimuli as well as the parameters of the subjects having been replaced
practice and also that of turning the observed row and column totals into with observable numbers.
adequate estimates of the 's and the e's. It is suggested that comparisons carried out under such circumstances be
In Chapter 6 of the earlier study (3) these questions were dealt with by designated as "specifically objective." The same term would seem appropri-
simple methods which were taken to be acceptable approximations. In the ate for statements about the model structure that arc independent of all the
case of subtest N the observations passed the test for the model satis- parameters specified by the model, their unknown values being, in fact,
factorily, but the model failed completely in the case of subtest F (geometrical irrelevant for the structure of the model.
shapes). In the latter subtest the time allowance, for some technical reasons. Of course. specific objectivity is no guarantee against the "subjectivity"
had been cut below the optimal limit, but a reanalysis of the data (not of the statistician when he chooses his fiducial limits or when he judges
reported here) has shown that when allowance is made for the 'working which kind of deviations from the model he will look for. Neither does it
speed of each subject, the data fit the model just as well as for the numerical save him from the risk of being offered data marred by the subjective attitude
sequences. of the psychologist during his observations. Altogether, when introducing
. However, from a theoretical point of view the method used to test the the concept of specific objectivity, I am not entering upon a general philo-
model was unsatisfactory (see Rasch [3], Chapter 10, in particular pp. sophical debate on the meaning and the use of objectivity at large. At
181— 182). By now we are in the process of working out better methods, and present the term is strictiy limited to observational situations that can be
therefore for the time being we shall leave the dbcumentation of the applica- covered by the stimulus-subject-response scheme, to be described in terms
bility of the model with simply a reference to the earlier work. of a parametric model that specifies parameters for stimuli and for subjects.
What has been demogstrated in detail in the case of two response cate-
gories is thct the specified objectivity in all three directions can be attained
insofar as the type of model defined herein holds. Recently it has. been
7. Specific objectivity
shown that except for unimportant mathematical restrictions, the inverse
As regards the basic formulas 6.14 and 6.15, we have already noted that statement is also true: if only two responses are available, the obserVations
when they are applied to the total set of data they enable us to separate the mus, conform to the simpl:: model 3.3 if it is to be possible to maintain
estimation of one set of parameters from that of the other. However, formula specific objectivity in statements about subjects, stimuli, and model.
6.15 can also he applied to any subgroup of the total collection of subjects
having been exposed to the k stimuli. Thus the parameters of the subjects in
the subgroup can be evaluated without regard to the parameters of the other
subjects. 8. Fields of application
In particular, the parameters of any two subjects can be compared on The problems we have been dealing with in the present paper were formu-
their own virtues alone, quite irrespective of the group or population to lated within a narrow field, psychological test theory. However, with the
which for some reason they may be referred. Thus, as indicated in Section 1, discovery of specific objectivity we have arrived at concepts of suci -i gen-
the new approach, when applicable, does rule out populations from the erality that the original limitation is no longer justified. Extensions into
comparison of individuals. other fields of psychology, such as psychophysical threshold experinwrits
Similarly, formula 6.14 can be applied to any subset of the k stimuli, and and experiments on perception of values, offer themselves, but the stimulus-
accordingly their parameters can be evaluated without regard to the param- subjectresponse framework is by no means limited to psychology. Thus in
eters of the other stimuli. In particular, the parameters of any two stimuli a recent publication (1) .a Poisson model was employed in an investigation -
can be compared separately. of infant mortality in Denmark in the period 1931-60. In each year the
With these additional consequences, the principle of separability leads to number of infant deans (of all causes or of a particular cause) out of the
a singular objectivity in statements about both parameters and model number of boys and of girls, born in or out of wedlock, was recorded. In
structure. In fact, the comparison of any two subjects can be carried out in this case the years served as subjects, the combination of iex and legitimacy
such a way that no other parameters are involved than those of the two of the (..hildren as the stimulus, and the number of infant deaths out of the
subjects—neither the parameter of any other subject nor any of the stimulus number of children born as the respOnses.
 An Individualistic Approach to !tent Analysis 107
106 Georg Rawl,

easy at-1(7 the distributions showed an undesirable accumulation of many

From economics we can take household budgets as an example. The corect answers.
families serve as subjects, income and expenditures—classified into a few This happened in 1953, when only the barest scraps of the theory had
types—as stimuli, and the amount earned and the amount spent as the been developed, and yielding to considerable time pressure the test con-
responses. structor, consulting me on the statistical part of tie problem, severely cut
These examples may indicate that the framework covers a rather large down the time allowances so as to move the distributions to the middle of
field within the social sciences. Delineating the area in which the models the range: While succeeding in that, we spoiled the test, turning it into a
described here apply is a huge problem, the inquiry into which has barely mixture of a test for ability and a test for speed.
started. More recently, however, I have had the opportunity to reanalyze both
But already the two intelligence tests mentioned in Section 2 and dis- sets of data, grouping the subjects primarily according to their working
cussed at the end of Section 6 are instructive as regards the sort of diffi- speed, as given by the number of items done, and applying to each group the
culties we should be prepared to meet. For one of them. the numerical
technique of Chapter 6 of the earlier report (2). The result was startling:
sequences, the earlier analysis (3, Chapter 6) showed a perfectly satisfactory Within each speed group I found confirmation of the theory, and the relative
fit of the observations to the model —that is, in this case specific objectivity
difficulties of the items were independent of the working speed. Altogether,
can be obtained on the basis of the response pattern for each subject. For
with speed as ancillary information, specific objectivity can be attained in
the other test, the geometrical shapes, the analysis most unambiguously
regard to the properties which the tests really aimed at measuring.
showed that the separability did not hold. Inverting the final statement, we get the moral of this story: Observations
Neither did it hold for a different intelligence test, which was of the omni-
may easily be made in such a way that specific objectivity, otherwise
bus type, containing items that presumably called upon very different available, gets lost. For instance, this can easily happen when qualitative
intelligence functions. In this case, therefore, the data could not be expected observations with, say, five categories of responses for convenience are
to allow for a description comprising only one parameter for each subject. condensed into three categories. If the basic model holds for the five cate-
The items in the numerical sequences were quite uniform in that they re- gories, it is mathematically almost impossible for the three-category model
quired the testee to recognize a logical structure in a sequence of numbers. also to hold. Thus the grouping, tempting as it may be, will usually tend to
According to the analysis. the items were sufficiently uniform—although of slur the specific objectivity.
very different levels of difficulty —to allow for a description of the data by In concluding, therefore, I must point out that the problem of the relation
one parameter for each subject as well as for each item. The items of the of data to models is not only one of trying to fit data to an adequately chosen
geometric figure test were constructed just as uniformly as the numerical model from our inventory to see whetncr it works; it is also holt , to make
sequences, and therefore it was somewhat of a surprise that the results observations in such a way that specific objectivity obtains.
turned out quite adversely.
To this material I could add observations on two other tests, constructed
with equal care. One was a translation of the idea of Raven's matrix test References
into letter combinations, at the same time substituting the multiple choice i. MATTHIESSEN, P.C. Infant Mortality in Denmark, 1931 —60. Copenhagen: Statistical
by a construction, on the part of the testee, of the answer. For this test the Department, 1965.
results were just as good as for the numerical sequences. The other test 2. RAscii. G. "On General Laws and the Meaning of Measurement Psychology," in
consisted of a set of verbal analogies where the number of answers offered Proceedings if the Fourth 1lc rke/ey Symposium on Alathematical Statistics and Pntha-
bili'y \lc'. IV, 321-34, Berkeley: Univ. of California Press, 1961.
was practically infinite, with the effect that the multiple choice was in fact
3. Prob.-hilistic Models for Some Intelligence and Attainment Tests. Copenhagen:
eliminated. Here the results of the testing were just as disappointing as for Danish Institute for Educational Research, 1960.
the figure test.
This contrast, however, led to the solution of the mystei -y. The difference
between the two pairs of tests was due not to construction principles but to
the administration of the tests. For all fou: tests the adequate time allowance
was determined by means of special experiments. On applying them to
random samples of 200 subjects, it turned out that the number of correct
answers formed a convenient distribution for the letter matrix test and for
the numerical sequences, but verbal analogies and the figure test were too