Bayesian regression techniques for determining true measurements

Abstract

Techniques for estimating a true measurement from a Bayesian regression on an observed measurement of received responses.

Claims

1. A non-transitory computer-readable medium having instructions stored thereon that, responsive to execution by an electronic system, cause said electronic system to perform operations to estimate difficulty or ability, said operations comprising: receiving a correct or incorrect response for each item and each test-taker and a total score for each test-taker on a multi-item test taken by a plurality of test-takers, wherein a correct response is identified by the numeral 1 and an incorrect response by the numeral 0; estimating a probability of a correct response to each item by each test-taker from a Bayesian regression on the response of the test-taker to the item taking the form of a weighted average of the item response of the test-taker and an average response to the item over the plurality of test-takers, the respective weights being the square of a correlation between the item response and the total score of the test-takers over the plurality of test-takers and one minus the squared correlation; determining a logit of the Bayesian-estimated probability of the correct response for each item and each test-taker; estimating a difficulty of each item as minus an average of the logits over the plurality of test-takers; determining an average of the item difficulties over the plurality of items; estimating an ability of each test-taker as a sum of the average item difficulty and an average of the logits over the plurality of items; and iteratively performing the actions of estimating the probability of the correct response, determining the logit, estimating the item difficulty, determining the average item difficulty, and estimating the ability, wherein the most recently estimated ability replaces the total test score on each iteration, wherein the actions of estimating are iteratively performed until a change in each test-taker's ability estimate is less than a predetermined amount, wherein a reduced test length is required to achieve a given non-zero internal validity in comparison to estimating an ability of each test-taker by equally weighting all item responses.

2. The non-transitory computer-readable medium according to claim 1, further comprising: estimating a new difficulty for each item using a Rasch model; and substituting an average of the new Rasch difficulties for the average item difficulty to obtain a new ability estimate for each test-taker.

3. A non-transitory computer-readable medium having instructions stored thereon that, responsive to execution by an electronic system, cause said electronic system to perform operations comprising: receiving a correct or incorrect response for each item and each test-taker and a total score for each test-taker on a multi-item test taken by a plurality of test-takers, wherein a correct response is identified by the numeral 1 and an incorrect response by the numeral 0; estimating a probability of a correct response to each item by each test-taker from a Bayesian regression on the response of the test-taker to the item taking the form of a weighted average of the item response of the test-taker and an average response to the item over the plurality of test-takers, the respective weights being the square of a correlation between the item response and the total score of the test-taker over the plurality of test-takers and one minus the squared correlation; determining a logit of the Bayesian-estimated probability of the correct response for each item and each test-taker; estimating a difficulty of each item as minus an average of the logits over the plurality of test-takers; determining an average of the item difficulties over the plurality of items; estimating an ability of each test-taker as a sum of the average item difficulty and an average of the logits over the plurality of items; and iteratively performing the actions of estimating the probability of the correct response, determining the logit, estimating the item difficulty, determining the average item difficulty, and estimating the ability, wherein the most recently estimated ability replaces the total test score on each iteration, wherein said estimating said ability of each test-taker has greater internal validity for a given test length than estimating an ability of each test-taker by equally weighting all item responses.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) Embodiments of the present technology are illustrated by way of example and not by way of limitation, in the figures of the accompanying drawings and in which like reference numerals refer to similar elements and in which:

(2) FIG. 1 illustrates the regression of the sample proportion P on the population proportion .

(3) FIG. 2 illustrates the regression of the population proportion on the sample proportion P.

(4) FIG. 3 illustrates Bayesian (-on-P) as a percentage of frequentist error margins in a two-option case.

(5) FIG. 4 illustrates Bayesian (-on-P) error margins in the two-option case.

(6) FIG. 5 shows a flow diagram of a -on-P point and interval estimation.

(7) FIG. 6 shows credibility-interval coverage of 1.96 standard errors in most applications of -on-P estimation.

(8) FIG. 7 shows a comparison of -on-P with other estimates in a particular example involving three options.

(9) FIG. 8 shows a comparison of critical with Bayesian and conventional margins of error.

(10) FIG. 9 shows a comparison of samples sizes required to produce commonly used error margins.

(11) FIG. 10A and 10B show a flow diagram of delta single-parameter logistic modeling methods, in accordance with a number of embodiments of the present technology.

(12) FIG. 11 shows a table comparing item difficulties and item discrimination data for various modeling techniques.

(13) FIG. 12 shows a graph comparing Rasch and delta single-parameter logistic model validities.

DETAILED DESCRIPTION OF THE INVENTION

(14) Reference will now be made in detail to the embodiments of the present technology, examples of which are illustrated in the accompanying drawings. While the present technology will be described in conjunction with these embodiments, it will be understood that they are not intended to limit the invention to these embodiments. On the contrary, the invention is intended to cover alternatives, modifications and equivalents, which may be included within the scope of the invention as defined by the appended claims. Furthermore, in the following detailed description of the present technology, numerous specific details are set forth in order to provide a thorough understanding of the present technology. However, it is understood that the present technology may be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present technology.

(15) In this application, the use of the disjunctive is intended to include the conjunctive. The use of definite or indefinite articles is not intended to indicate cardinality. In particular, a reference to the object or a object is intended to denote also one of a possible plurality of such objects.

(16) Embodiments of the present technology are directed toward simple regression of a true measurement on an observed measurement. The direction of this regression (true on observed) corresponds to Bayesian analysis, whereas the opposite direction corresponds to classical (or standard) estimation in statistics. In one embodiment, the true measurement may be a population proportion, and the observed measurement may be a corresponding sample proportion. In another embodiment, the true measurement may be the probability of a correct response to an item by a particular person, and the observed measurement may be the actual response by that person to the item. The simple regression involves a correlation between the variables on the two sides of the equation.

(17) In population proportion-on-sample proportion regression embodiments of the present technology, an estimator of population proportions is developed that is even more efficient than the Fienherg-Holland estimator, particularly in a small sample (n<500), and demonstrates that all shrinkage coefficients are estimators of squared correlation coefficients in population-on-sample regression. The margins of error for the population proportions, provided they follow a Dirichlet distribution, are shown to be 1.96 times their standard deviations in virtually all realistic applications. These techniques relate to the technical field of poll-taking, particularly utilizing statistics which encompass the estimation of population proportions from the regression of population on sample proportions. Using an unbiased estimator of the square of the correlation between the population and sample proportions in a Bayesian context produces not only point estimates of the population proportions but also credibility intervals that are narrower than corresponding conventional confidence intervals.

(18) Simple regression analysis in statistics is a procedure for estimating the linear relationship between a dependent variable and an independent variable in a given population. The relationship for standardized variables is expressed as an equation for a straight line in which the coefficient of the independent, or regressed-on, variable in the equation is determined from a sample. While the dependent variable may vary, the regressed-on variable is fixed. This variable is the sample statistic, mean or proportion, in population-on-sample regression. This is why population-on-sample regression corresponds to Bayesian estimation. The opposite is true of sample-on-population regression, which corresponds to frequentist estimation in which the population parameter, mean or proportion, is fixed.

(19) A point-estimation advantage of the population-on-sample regression procedure is that it generally avoids the problem of relative-frequency estimates equal to zero or one by reasonably adjusting them away from these extreme values. This inward adjustment is the result of regression toward the mean, which for relative frequencies is greater than zero and less than one.

(20) Population-on-sample regression shares the efficiency advantage of Bayesian over traditional estimation, which shows itself in a reduction of least-squares risk functions with a corresponding shortening of confidence intervals. In mental test theory, for example, standard errors of estimate, used to determine confidence intervals for true scores, are shorter than standard errors of measurement, used to determine confidence intervals for observed scores. The shrinkage coefficient in this case is the squared true-observed-score correlation, and the standard error of estimate is equal to the standard error of measurement time this correlation.

(21) The method comprising embodiments of this invention provides for both point and interval estimation for arriving at usable results of a poll. The approach to each follows both frequentist and Bayesian tracks within a regression framework. Development of either the frequentist or the Bayesian point estimator makes no distributional assumption about the population proportions. Distributional assumptions come into play only in the treatment of interval estimation. Normally a regression approach will require three or more observations to accommodate the need to estimate the slope and the intercept of the regression line from data. However, because the mean requires no estimation in the case of single-sample proportions, being the reciprocal of the number of options or categories, the Bayesian as well as the frequentist point estimator developed herein can apply to data involving only two (binomial or multinomial) observations.

(22) The focus here physically has been on a single sample obtained from a single population. Conceptually, the sample may be one of many that the population can produce or the population may be one of many that can produce the sample. The first possibility underlies the frequentist approach and the second the Bayesian approach to statistical inference. The method constituting embodiments of this invention has adopted the Bayesian approach showing that it can lead through regression to considerably more efficient estimation of population proportions than the frequentist approach, especially for samples no larger than 500.

(23) Although reference is made explicitly to proportions, the method applies equally to other forms of expressing such results of a poll, for example, percentage, fractions, and decimal fractions, with appropriate adjustments known by those skilled in the art.

(24) The following steps develop the regression point and interval estimators in the frequentist (P-on-) case (Step 1) and in the Bayesian (-on-P) case (remaining steps).

(25) The Regression of P on

(26) The frequentist approach to point estimation via regression corresponds to the traditional estimation procedure in which, for a sample of size n, nP (an integer) has a binomial or a multinomial distribution with E.sub.t(P.sub.kt)=.sub.k for each option k of a total of K options, t indexing the sample. The regression expressing this or the Bayesian approach involves, for a single sample t, the mean P and .sub. over options

(27) $\left({}_{}=\stackrel{\_}{P}=\left(1/K\right)\underset{k=1}{\overset{K}{.Math.}}{P}_{\mathrm{kt}}\right),$
the standard deviations S.sub.P and .sub. over options

(28) $\left(S_p=\sqrt{\left(1/K\right)\underset{k=1}{\overset{K}{.Math.}}{\left(P_{\mathrm{kt}}-{}_p\right)}^2}\right),$
and the correlation coefficient .sub.P over options. To assure that E.sub.t(P.sub.kt)=.sub.k, the slope coefficient in the regression

(29) $\begin{array}{cc}{P}_{\mathrm{kt}}=\left(\frac{S_p}{{}_{}}\right){}_p{}_k-\left(\frac{S_p}{{}_{}}\right){}_p{}_{}+\stackrel{\_}{P}+{}_{\mathrm{kt}}& \left(3\right)\end{array}$
must be equal to one, the population and sample means .sub.90 and P being equal to 1/K, so that P.sub.kt=.sub.k+.sub.kt, where .sub.kt denotes sampling error. Since E.sub.t(.sub.kt)=0, E.sub.t(P.sub.kt)=.sub.k, as in the traditional binomial or multinomial estimation procedure. The regression implication of E.sub.t(P.sub.kt)=.sub.k then is that the correlation .sub.P between .sub.k and P.sub.kt must be equal to the ratio of their standard deviations, .sub. and .sub.P:

(30) $\begin{array}{cc}{}_p=\frac{{}_{}}{S_p}& \left(4\right)\end{array}$

(31) This result resembles a basic result of classical mental test theory (Gulliksen, 1950) in which represents a true and P an observed score. The next section will use this result to obtain an estimator of .sub.P.sup.2.

(32) Step TwoAn Estimator of .sub.P.sup.2

(33) In this as in the previous section, for each option k, .sub.k as the regressed-on variable is assumed to be fixed while P.sub.kt can vary over samples (t=1, 2, . . . ). Ordinarily the exact value of .sub.P.sup.2 is unknown. Because P.sub.kt is a proportion, however, .sub..sup.2 is expressed in terms of .sub.P.sup.2, the expected value of S.sub.P.sup.2, for substitution into Equation (4) to yield an estimator of .sub.P.sup.2 in which .sub.P.sup.2 is replaced by S.sub.P.sup.2:

(34) $\begin{array}{cc}{S}_P^2=\frac{1}{K}\underset{k=1}{\overset{K}{.Math.}}{\left(P_{\mathrm{kt}}-{}_p\right)}^2& \left(5\right)\end{array}$
where .sub.P is the mean of the K values of P.sub.ktin sample t. Without further assumptions or conditions, the following derivation leads to the sample estimator of .sub.P.sup.2 in Equation 15.

(35) In Equation 5, .sub.P, equal to 1/K, is the population as well as the sample mean proportion so that S.sub.P.sup.2 with K rather than (K1) in the denominator, is an unbiased estimator of .sub.P.sup.2:

(36) $\begin{array}{cc}{}_P^2=E_t\left\{\left(\frac{1}{K}\right)\underset{k=1}{\overset{K}{.Math.}}{\left(P_{\mathrm{kt}}-{}_p\right)}^2\right\}& \left(6\right)\end{array}$
If .sub.k=.sub.+.sub.k, where

(37) $\left(1/K\right)\underset{k=1}{\overset{K}{.Math.}}{}_k=0,\mathrm{then}{}_{}^2=\left(1/K\right)\underset{k=1}{\overset{K}{.Math.}}{}_K^2.$
As noted in the preceding section, P.sub.kt=.sub.k+.sub.kt. The expected values of .sub.kt and .sub.k.sub.kt (equal to .sub.k times the expected value of .sub.kt) are equal to zero. Substitution first of .sub.k+.sub.kt for P.sub.kt, and then of .sub.+.sub.k in Equation 6 thus, with .sub.=.sub.P, leads to

(38) 0$\begin{array}{cc}{}_P^2={}_{}^2+\left(\frac{1}{K}\right)\underset{k=1}{\overset{K}{.Math.}}{}_k^2& \left(7\right)\end{array}$
where, for each option k, .sub.k.sup.2 is the sampling variance
.sub.k.sup.2=(1/n).sub.k(1.sub.k) (8)

(39) Substitution of .sub.+.sub.k for .sub.k in the computation of the mean of Equation 8 over the K values of .sub.k.sup.2 produces

(40) $\begin{array}{cc}\frac{1}{K}\underset{k=1}{\overset{K}{.Math.}}{}_k^2=\frac{{}_k\left(1-{}_k\right)}{n}-\frac{{}_{}^2}{n}\mathrm{since}\underset{k=1}{\overset{K}{.Math.}}{}_{}{}_k={}_{}\underset{k=1}{\overset{K}{.Math.}}{}_k=0\mathrm{and}\underset{k=1}{\overset{K}{.Math.}}{}_k^2=K{}_k^2,\mathrm{Equation}7\mathrm{thus}\mathrm{becomes}& \left(9\right)\\ {}_P^2={}_{}^2+\frac{{}_{}\left(1-{}_{}\right)}{n}-\frac{{}_{}^2}{n}\mathrm{or},\mathrm{with}\frac{1}{K}\mathrm{for}{}_{},& \left(10\right)\\ {}_P^2={}_{}^2+\frac{K-1}{{\mathrm{nK}}^2}-\frac{{}_{}^2}{n}& \left(11\right)\end{array}$
Solution of Equation 11 for .sub..sup.2 finally yields the expression of .sub..sup.2 in terms of .sub.P.sup.2:

(41) $\begin{array}{cc}{}_{}^2=\left(\frac{n}{n-1}\right){}_P^2-\left(\frac{K-1}{n-1}\right){\left(\frac{1}{K}\right)}^2& \left(12\right)\end{array}$
The formula for .sub.P.sup.2 is thus

(42) $\begin{array}{cc}{}_P^2=\frac{\left(\frac{n}{n-1}\right){}_p^2-\left(\frac{K-1}{n-1}\right){\left(\frac{1}{K}\right)}^2}{S_p^2}& \left(13\right)\end{array}$
so that the estimator of .sub.P.sup.2 is

(43) $\begin{array}{cc}{\hat{}}_P^2=\frac{\left(\frac{n}{n-1}\right)S_P^2-\left(\frac{K-1}{n-1}\right){\left(\frac{1}{K}\right)}^2}{S_P^2}& \left(14\right)\end{array}$
or, since

(44) $S_P^2=\left(1/K\right)\underset{k=1}{\overset{K}{.Math.}}{P}_k^2-{\left(1/K\right)}^2,$

(45) $\begin{array}{cc}{\hat{}}_P^2=1-\frac{K\left(1-\underset{k=1}{\overset{K}{.Math.}}{P}_{\mathrm{kt}}^2\right)}{\left(n-1\right)\left(K\underset{k=1}{\overset{K}{.Math.}}{P}_{\mathrm{kt}}^2-1\right)}& \left(15\right)\end{array}$
where

(46) $\underset{k=1}{\overset{K}{.Math.}}{P}_{\mathrm{kt}}^2$
is the sum of squares of the K proportions computed from the sample t of size n. Equation 14 shows that {circumflex over ()}.sub.P.sup.2 is not only an increasing function of S.sub.P.sup.2, K, and n but also an unbiased estimator of .sub.P.sup.2, the S.sub.P.sup.2 in the denominator being fixed.

(47) Step ThreePoint Estimation: Estimation for P

(48) The regressions underlying the developments in this and the preceding section are opposite in direction. Both involve P and . The development in the preceding section considered as fixed and P as variable. In the development here, however, the reverse is true: P is fixed, and is variable. In this development, the fixedP.sub.k sample (k=1, 2, . . . , K) comes from a single population, which is one of any number of possible populations, with their correspondingly different .sub.k values. Although the expected value of variable P.sub.k is .sub.k, the expected value of a variable .sub.k is not P.sub.k, but a value {circumflex over ()}.sub.k somewhere between P.sub.kand 1/K. Whereas the direction of regression assumed in the proceeding section worked for the development of a formula for {circumflex over ()}.sub.P.sup.2, the regression direction taken here is particularly appropriate for the estimation of an unknown .sub.k, assumed variable, from a known P.sub.k, assumed fixed for each option k:

(49) $\begin{array}{cc}{\hat{}}_k={}_P\left(\frac{{}_{}}{S_P}\right)P_k-{}_P\left(\frac{{}_P}{S_P}\right)\stackrel{\_}{P}+{}_{}& \left(16\right)\end{array}$
where {circumflex over ()}.sub.k is the regression estimate of .sub.k, .sub.P is the correlation between P.sub.k and .sub.k for the population sampled from, and .sub.P and .sub. are the means and S.sub.P and .sub. are the standard deviations over options of P.sub.k and .sub.k, respectively. Since .sub.P.sup.2=.sub..sup.2/S.sub.P.sup.2, this equation simplifies to
{circumflex over ()}.sub.k=.sub.P.sup.2P.sub.k.sub.P.sup.2P+.sub.(17)
where .sub. is equal to P so that, with both .sub. and P denoted by ,
{circumflex over ()}.sub.k=.sub.P.sup.2P.sub.k+(1.sub.P.sup.2)(18)
or, since =1/K,

(50) $\begin{array}{cc}{\hat{}}_k={}_P^2{P}_k+\frac{1-{}_P^2}{K}& \left(19\right)\end{array}$
Estimation of the population proportion .sub.k corresponding to the observed proportion P.sub.k thus requires knowledge only of .sub.P.sup.2. If .sub.P.sup.2=1, {circumflex over ()}.sub.k=P.sub.k; if .sub.P.sup.2=0, {circumflex over ()}.sub.k=1/K. Generally, in practice, {circumflex over ()}.sub.k will be somewhere between P.sub.k and 1/K.

(51) Since P.sub.k is assumed fixed, substitution of {circumflex over ()}.sub.P.sup.2 for .sub.P.sup.2 in Equation 19 yields an estimate of .sub.k that is not subject to sampling variation.

(52) FIG. 1 illustrates the regression of P on and FIG. 2 the regression of on P. In both figures, the population proportions are fictitious since their actual values are unknown. Knowledge of these values is unnecessary because the only requirements for estimation are the sample proportions in FIG. 1 and the regression line in FIG. 2. The vertical lines define 95% confidence intervals. Based on n=100, the value of {circumflex over ()}.sub.P.sup.2 is 0.88, the slope of the regression line in FIG. 2. In addition to their different slopes, the two regression lines notably have different intercepts: 0 in FIGS. 1 and 0.03 in FIG. 2. Except when {circumflex over ()}.sub.P.sup.2=1, -on-P regression produces population-proportion estimates that are greater than zero and less than one.

(53) Two examples provide data to illustrate the -on-P regression procedure. The first, cited by Tull and Hawkins (1993, pp. 745-746) in the spirit of R. A. Fisher's classic tea-testing illustration of the Student t-test, was a Carnation taste test comparing Coffee-mate to real cream. Of 285 participants who claimed to be able to distinguish between two cups of coffee presented them, one containing Coffee-mate and the other containing cream, 153 were correct and 132 were incorrect, the corresponding proportions being 0.54 and 0.46. With {circumflex over ()}.sub.P.sup.2=0.42, the Bayesian-estimated proportions were 0.52 and 0.48, respectively. These proportions more strongly than their frequentist counterparts support the conclusion that people could not tell the difference between Coffee-mate and real cream.

(54) In the second example, a large school district tested three different textbooks for first-year high school algebra. The first was the book used for the past several years; the second and third were new books containing questions taken from recent versions of a statewide mathematics examination. The question format differed in these books, being open-ended in the second book and multiple-choice (as in the statewide examination) in the third book. Two hundred students in different classes used each book. Of the 450 students who passed the statewide examination, 130 had used the first book, 158 the second book, and 162 the third book. The corresponding proportions were 0.29, 0.35, and 0.36, respectively. Substituting these proportions, which sum to 1.00, in Equation 13, together with K=3 and n=450 yields 0.48 for the value of {circumflex over ()}.sub.P.sup.2, and using this value for {circumflex over ()}.sub.P.sup.2 in Equation 19 yields {circumflex over ()}.sub.k values of 0.31, 0.34 and 0.35 for the three books, respectively. The first and third values notably differ (by 0.02 each) from their uncorrected counterparts while the second, being closer to the mean of 0.33, shows a difference of only 0.01, to two decimal places. If the books had been equally effective, the expected proportions within the passing group would equal to 0.33 for students using all three books.

(55) Step 4AInterval Estimation: Two-option Case

(56) Reported survey results often include half the size of the 95% confidence interval as the so called margin of error. For K=2, the procedures developed in this embodiment of the invention involve intervals different from the conventional ones. The confidence or credibility intervals appropriate for the K=2 procedures developed here are functions of the standard error of measurement, applicable to the egression of P on , or the standard error of estimate, applicable to the regression of on P (Kelley, 1923, 1927). Both standard errors involve the assumption of homoscedasticity: Values of the standard errors of measurement are equal for all values of .sub.k, and values of the standard error of estimate are equal for all values of P.sub.k (k=1, 2, 3 . . . , K). In the case of proportions, as opposed to means of multi-valued variables, this assumption makes sense only when K=2.

(57) Though of less practical value, the standard error of measurement, .sub.P=S.sub.P{square root over (1.sub.P.sup.2)}, produces confidence intervals directly comparable to the conventional ones. Estimates of .sub.P are obtainable by using {circumflex over ()}.sub.P.sup.2 for .sub.P.sup.2 in the formula for .sub.P.

(58) The Carnation data provide an example. The two observed proportions, 0.54 and 046, were inaccurate by an amount equal to 0.06. As the conventional 95% margin of error, this value (0.06) is 1.96 times , {square root over (0.5(10.5)/285)}. Use of the standard error of measurement would produce a 95% confidence interval of the same size, to two decimal places. Substituting the value of 0.04 for S.sub.P and the {circumflex over ()}.sub.P.sup.2 value of 0.42 for .sub.P.sup.2 yields .sub.P=0.03, or (multiplying 0.03 by 1.96) a 95% error margin of 0.06.

(59) Confidence intervals determined from the standard error of measurement are directly comparable to conventionally determined confidence intervals because both are based on the assumption of a fixed and a variable P. The standard error of estimate, applicable in the regression case of a fixed P and a variable , has the same formula as the standard error of measurement with the exception that .sub. replaces S.sub.P:

(60) 0${}_{\left(\hat{}-\right)}={}_{}\sqrt{1-{}_P^2}.$
Since

(61) ${}_P^2=\frac{{}_{}^2}{S_P^2},$
the standard error of estimate will, except when .sub.P.sup.2=1, be smaller than the standard error of measurement by a factor of .sub.P. Conceptually, the standard error of estimate should be smaller than the standard error or measurement because the difference (P) contains a varying component representing bias that is absent in the difference ({circumflex over ()}-). Credibility intervals for a variable it will therefore generally be smaller than corresponding confidence intervals for a variable P. When K=2, the estimate of .sub.{circumflex over ()} corresponding to .sub.p is equal to the estimate of .sub.p multiplied by {circumflex over ()}.sub.P.

(62) In the Carnation example, with {circumflex over ()}.sub.P.sup.2=0.42, the standard error of estimate is {square root over (0.42)} times 0.03 (the standard error of measurement), or 0.019, so that the 95% margin of error (1.96 times 0.019) is 0.04. This (rounded from 0.037) is considerably smaller than the conventional error margin of 0.06. The sample of 285 would, in fact, have to be 417 larger (a total of 702 respondents) to achieve the same 0.04 margin of error conventionally. Since both the confidence and the credibility intervals overlap the chance proportion of 0.50, the data do not support the claim that tasters can tell cream from Coffee-mate.

(63) For the credibility interval to be comparable to the conventional 0.06 confidence interval, it must also contain 95% of the area under its frequency curve. The next section will investigate the extent to which this is the case not only here but also more generally.

(64) QuestionDo 1.96 Standard Errors Constitute 95% Credibility Intervals in -on-P Estimation?

(65) The answer, generally, is yes, as this section demonstrates.

(66) Corresponding to the assumption of a binomial distribution for sample proportions is the assumption of a beta distribution for population proportions. One distribution is the conjugate of the other. Dirichlet distributions are correspondingly conjugates of multinomial distributions. Such assumptions of conjugate distributions are common in Bayesian analysis (e.g., Good, 1965, Chapter 3). Under the beta-distribution assumption, not only does the 0.04 credibility interval of the Carnation example contain 95% of the area under its frequency curve but also, as FIG. 6 shows, 1.96.sub.{circumflex over ()} credibility intervals ranging from 0.01 to 0.1 of point estimates ({circumflex over ()}) between 0.05 and 0.50 will generally contain 95% of possible values. (No subscript for is necessary here because a beta distribution involves only two proportions, and 1). The coverage shown in FIG. 6 is based on calculations, not Monte Carlo sampling. FIG. 6 shows confidence-interval proportions for values of {circumflex over ()} not only between 0.05 and 0.50 but also, though indirectly, between 0.50 and 0.95. Because beta distributions having mean values between 0.50 and 0.95 are mirror images of beta distributions having mean values between 0.05 and 0.50, the credibility-interval proportion of {circumflex over ()} is equal to the credibility-interval proportion for 1{circumflex over ()}, provided that both distributions have equal standard deviations (.sub.{circumflex over ()}).

(67) FIG. 6 shows credibility-interval proportions as a function of beta-distribution means and standard deviations because these are the parameters involved in the determination of credibility intervals. Beta distributions, however, are functions of two parameters, a and b, related to beta-distribution means ({circumflex over ()}) and standard deviations (.sub.{circumflex over ()}), as follows:
a=({circumflex over ()})(a+b) (20)
and
b=(1{circumflex over ()})(a+b) (21)
where

(68) $\begin{array}{cc}a+b=\frac{\hat{}\left(1-\hat{}\right)}{{}_{\hat{}-}^2}-1& \left(22\right)\end{array}$
For the Carnation data, a+b=0.52(1-0.52)(0.019).sup.2 1, or 690, so that a=0.52(690), or 359, and b=(1-0.52)690, or 331, and for these values of a and b the interval between 0.52-1.96(0.019) and 0.52+1.96(0.019) contains 95% of the area under the beta-distribution frequency curve. FIG. 6 shows this result in the row for 0.50 (close to 1-0.52) and the column for 0.020 (close to 0.019).

(69) The two standard errors, the standard error of measurement and the standard error of estimate, differ not only in the lengths of the confidence or credibility intervals that they produce but also in one other important respect: While the standard error of measurement, computed from S.sub.P and {circumflex over ()}.sub.P.sup.2 (itself a function of S.sub.P), is subject to sampling variation due to the possible variation of P for each option over samples, the standard error of estimate does not change because under the -on-P assumption governing its use each P remains constant while only can change for each option. In both these respects, the standard error of estimate is superior to the standard error of measurement for use in the determination of credibility or confidence intervals for population proportions.

(70) The results shown in FIG. 6 are applicable to Dirichlet as well beta distributions because a beta distribution describes each Dirichlet proportion if all the other proportions are aggregated as its complement.

(71) Step 4BInterval Estimation: Case of Two or More Options

(72) The standard error of estimate provides the basis for a common credibility interval, applicable particularly for proportions when K=2. The assumption of a Dirichlet prior distribution for the population proportions makes possible the determination of a different interval for each proportion when K3. (When K=2, the two intervals are the same.) If designates the total of the parameters of a Dirichlet distribution, then the posterior variance of .sub.k for option k is

(73) $\begin{array}{cc}\mathrm{Var}\left({}_k.Math.P_k\right)=\frac{{\hat{}}_k\left(1-{\hat{}}_k\right)}{+n+1}& \left(23\right)\end{array}$
where P.sub.k is the observed proportion for option k. Use of this equation, with {circumflex over ()}.sub.k from Equation 19 estimating the expected value of .sub.k, requires knowledge of .

(74) Since E(.sub.k|P.sub.k)=({circumflex over ()}{circumflex over ()}.sub.k+nP.sub.k)/(+n) for a Dirichlet distribution, the Direchlet shrinkage coefficient corresponding to {circumflex over ()}.sub.P.sup.2 here is n/(+n), and so for the -on-P procedure

(75) $\begin{array}{cc}{}_{.Math.P}=\frac{n\left(1-{\hat{}}_P^2\right)}{{\hat{}}_P^2}& \left(24\right)\end{array}$
or, from Equation 15,

(76) $\begin{array}{cc}{}_{.Math.P}=\frac{K\left(1-\underset{k=1}{\overset{K}{.Math.}}{P}_k^2\right)}{K\underset{k=1}{\overset{K}{.Math.}}{P}_k^2-1-\frac{K-1}{n}}& \left(25\right)\end{array}$
According to Fienberg and Holland (1973), the minimax value of is {square root over (n)} and the maximum-likelihood estimator of is

(77) $\begin{array}{cc}{}_{.Math.P}=\frac{K\left(1-\underset{k=1}{\overset{K}{.Math.}}{P}_k^2\right)}{K\underset{k=1}{\overset{K}{.Math.}}{P}_k^2-1}& \left(26\right)\end{array}$

(78) Using the values of computed from Equations 25 and 26 as well as {circumflex over ()} in Equation 23 with data from the three-textbook example presented earlier produced the 95% credibility or confidence intervals in FIG. 7, which includes corresponding traditional, frequentist results.

(79) FIG. 7 shows separate error margins for the three different textbook groups, as well as their different population-proportion estimates under the four different procedures, corresponding in the case of the non-frequentist procedures to their different values of : 21, 232, and 487, respectively, for minimax, Fienberg and Holland, and -on-P. The -on-P procedure produced the narrowest margins of error. This result is not surprising since, as Equations 25 and 26 make clear, the .sub.|P for the -on-P procedure is greater than the .sub.|P for the maximum-likelihood procedure, the difference diminishing as n gets large. According to Equation 24, the for the minimax procedure ({square root over (n)}) will also be smaller than the .sub.|P for the -on-P procedure unless {circumflex over ()}.sub.P.sup.2/(1{circumflex over ()}.sub.P.sup.2){square root over (n)}, which is not the case here. Since the standard error of the three non-frequentist procedures ranged between 0.01 and 0.02, the credibility intervals, as FIG. 6 shows, all have approximately 95% coverage.

(80) Depending on which of the four estimation procedures they use, investigators looking at the study's result might come to entirely different conclusions. Since 0.33 bordered or lay outside the confidence or credibility interval of the older textbook in all but the -on-P procedure, investigators using the three other procedures might conclude that the older textbook was less effective than the two new ones. An investigator using the -on-P procedure, however, would not reach that conclusion. All three credibility intervals produced by that procedure, despite being generally narrower than the other by 0.02, contain the chance proportion 0.33. The conclusion following from this result is that further study is necessary before selecting a textbook for general use.

(81) Comparison of Bayesian and Frequentist Procedures in the Two-option Case

(82) FIGS. 3 and 4 provide a broader view of the frequentist and Bayesian (-on-P) regression procedures. Limited to K=2, these figures show, for a range of sample sizes and P values, Bayesian margins of error as percents of frequentist margins of error (FIG. 3) and actual Bayesian margins of error (FIG. 4). The three high curves in each figure represent high {circumflex over ()}.sub.P.sup.2 values, approaching one for n=500, while the bottom curve (for P=0.54) represents a comparably low {circumflex over ()}.sub.P.sup.2 value. When {circumflex over ()}.sub.P.sup.2 is high, Bayesian and frequentist margins of error are very nearly equal, as are corresponding point estimates; when {circumflex over ()}.sub.P.sup.2 is low, Bayesian margins of error are low and point estimates are close to the mean, relative to their frequentist counterparts. For P values that are very close to the mean, {circumflex over ()}.sub.P.sup.2 can be so low that Bayesian point estimates are for all practical purposes equal to the mean, with margins of error effectively equal to zero. This is the case for P=0.52 when n500.

(83) Is -on-P an Empirical or a Purely Bayesian Procedure?

(84) In regressing the observed proportion P toward the mean, 1/K, the squared correlation {circumflex over ()}.sub.P.sup.2 resembles the shrinkage coefficient w in Fienberg and Holland (1973) or 1B B in Efron and Morris (1973) and Morris (1983). Because the development using w or I - B involves empirical Bayesian estimation, this resemblance suggests that -on-P regression may also be empirical Bayesian. This is not the case, however.

(85) The -on-P procedure is a regression, not an empirical Bayesian, procedure. The difference is important. While estimates in both the -on-P and the Fienberg and Holland (1973) procedures are expected values of given P, both and P may vary in empirical Bayesian estimation while only may vary in estimation by -on-P regression. If P as well as were to vary in -on-P regression, then the credibility intervals computed from the standard error of estimate would be too small, as the coverage proportions in FIG. 6 would be too large. The more apt comparison is with pure Bayesian estimation because in both this and -on-P regression P is fixed while only it may vary. Because the regressed-on variables are fixed in regression estimation, the coverage proportions in Table 1 and their corresponding credibility intervals are accurate.

(86) Shrinkage Coefficients as Slope Coefficients in Regression

(87) Shrinkage coefficients may be interpretable as slope coefficients in regression. Stigler (1990) made this observation in relation to the work originated by James and Stein (1961), involving means. The squared correlation .sub.P.sup.2 is in fact the single-sample binary (0-1) data counterpart to the multi-sample shrinkage coefficient 1B cited by Morris (1983) under the empirical Bayes assumption of normal distributions for both sample and population means. If for m samples, with m>2, .sub.{circumflex over (X)}.sup.2 and .sub..sup.2 are the respective variances of these distributions, then according to Morris, 1B=.sub..sup.2/(.sub..sup.2+.sub.{circumflex over (X)}.sup.2), which is the square of the correlation between and {circumflex over (X)}. In view of Equations 23 and 24, the shrinkage coefficient w in the Fienberg and Holland (1973) procedure is a consistent estimator of the squared correlation between and P, the shrinkage coefficient developed here ({circumflex over ()}.sub.P.sup.2) being an unbiased estimator of it.

(88) In the James-Stein case, .sub.{circumflex over (X)}.sup.2=1, so that the risk function of the traditional estimator {circumflex over (X)} is equal to one for each value of . If has a normal distribution, then the posterior variance of given {circumflex over (X)} is equal to 1B and, as Efron and Morris (1973) observed, the risk function of the estimator (1B){circumflex over (X)}, which is the posterior mean of given {circumflex over (X)} assuming to have a mean of zero, is smaller than the risk function of {circumflex over (X)} by an amount equal to B. The estimator (1B){circumflex over (X)} is the James-Stein estimator if B is replaced by

(89) $\left(m-2\right)\underset{i=1}{\overset{m}{.Math.}}{\stackrel{\_}{x}}_i^2,$
whose expected value is equal to B because, with .sub.{circumflex over (X)}.sup.2=1 and the variance of X,

(90) $\underset{i=1}{\overset{m}{.Math.}}{\stackrel{\_}{x}}_i^2$
has a X.sup.2.sub.m distribution with negative first moment equal to 1/(m2). The James-Stein shrinkage coefficient

(91) $1-\left(m-2\right)\underset{i=1}{\overset{m}{.Math.}}{\stackrel{\_}{x}}_i^2$
is thus interpretable as an unbiased estimator of the square of the correlation between and {circumflex over (X)}, or as an unbiased estimator of the slope coefficient in the regression of on X, the intercept being equal to zero.

(92) Since {circumflex over ()}.sub.P.sup.2 is also an unbiased estimator, it corresponds in single-sample proportions estimation to the James-Stein shrinkage coefficient in multi-sample means estimation.

(93) Bayesian Versus Conventional Margins of Error in the Estimation of Population Proportions

(94) The use of Bayesian (-on-P) estimation of a population proportion requires an amended definition of margin of error. In the conventional or classical estimation, the margin of error depends only on sample size. In Bayesian estimation, the margin of error varies not only with sample size but also with the estimated population proportion obtained from the observed sample proportion. The margin of error that is meaningful in Bayesian estimation is the difference between 0.50 and the estimated population proportion. This margin of error is called the critical margin of error. When the estimated population proportion is 0.53, for example, the critical margin of error is 0.03 (0.53-0.50). If 0.53 is the Bayesian estimate of the population proportion in a sample large enough to produce a margin of error equal to 0.03 for a Bayesian-estimated population proportion of 0.53, then the conclusion from this result is that the population proportion is equal to 0.53 plus or minus 0.03 or, in other words, that the population proportion is marginally larger than 0.50. If for the same sample size the Bayesian population proportion estimate is larger than 0.53 (or smaller than 0.47), then that estimate plus or minus its error margin will exclude 0.50, as illustrated by the accompanying graph of a sample of size 300 (FIG. 8).

(95) For conventional and Bayesian estimation of population proportions, FIG. 9 shows a table that compares samples sizes required to produce commonly used error margins. In the case of Bayesian estimation, the error margins are critical margins of error. For every error margin, Bayesian estimation requires a smaller sample than conventional estimation.

(96) In other embodiments of the regression of a true measurement on an observed measurement to be described below, the true measurement may be the probability of a correct response to an item by a particular person and the observed measurement may be an actual response by that person to that item. The probability of a correct response regressed on an observed measurement may be implemented in a delta single-parameter logistic modeling technique. Delta single-parameter logistic modeling includes receiving a response for each item and each test taker, and a total score for each test-taker. The correlation between the item response and the total test score or other ability measure over the plurality of test-takers is determined. A Bayesian-estimated probability of a correct response for each test-taker to each item is determined. The logit of this probability estimate is determined. The difficulty of an item is estimated as a function of the average logit over the number of test-takers. A test-taker's ability is likewise estimated as a function of the average logit and the average difficulty over the number of test items.

(97) Referring now to FIGS. 10A and 10B, a delta single-parameter logistic modeling method, in accordance with a number of embodiments of the present technology, is shown. It is appreciated that method may be implemented in hardware, software, firmware and/or a combination thereof. The method may begin with receiving correct or incorrect item responses (x) for each test taker and each test item, and a total test score (X) for each test-taker, at 110. In one implementation, the test may be a multi-item test taken by a plurality of test takers. In one implementation, the value of the response (x) may be 1 if correct and 0 if incorrect.

(98) At 115, a correlation () between the item response (x) and the total test score (X) over the plurality of test takers is determined. The correlation () incorporates item discrimination, which is the extent to which an item measures what it is supposed to measure (what the test as a whole measures). At 120, an initial Bayesian estimate of the probability (P) of a correct response for each test-taker to each item is determined as the weighted average of x and the proportion of test-takers who answer the item correctly, wherein the weights are .sup.2 for x and one minus .sup.2 for the proportion of test-takers who answer the item correctly.

(99) At 125, a logit () of the probability (P) is determined, wherein the logit of a probability is the natural logarithm of the ratio of the probability to it complement (i.e., one minus the probability). At 130, a difficulty of an item is estimated as minus the average logit () over the number of test-takers. At 135, a given test-taker's ability () is estimated as the sum of the average difficulty (b) and the average logit () over the number of test items. At 155, the difficulty of the item and the test-taker's ability are output. In one implementation, the results may be output by storing on a computing device readable medium (e.g., computer memory), displaying on a monitor (e.g., computer screen), and/or the like.

(100) At 140, the processes of 115-135 may optionally be iteratively repeated, replacing the correlation between item response and total test score with the correlation between item response (x) and test-taker ability (), one or more times or until a change in the given test-taker's ability () is less than a predetermined amount. In one implementation, the processes of 115-135 may be iteratively repeated until no estimate of given test-taker's ability () differs from its estimate on the preceding cycle by more than 0.004. In such an implementation, a given test-taker's ability () estimates are used until they settle into stable values, when the iterative process ends.

(101) At 145, new difficulty (b) estimates for use in item banks and test-form equating may optionally be determined using a simple Rasch model, At 150, the process is to retain these Rasch b estimates while adding their average to each estimate obtained in processes 115-140 and subtracting from it the average b estimate obtained in these processes (115-140).

(102) The processes of 115-135 are referred to herein as the delta single-parameter logistic model. The processes of 115-140 implement the delta single-parameter logistic model providing an improved test-taker's ability (). The processes of 115-150, which combine the delta single-parameter logistic model for the test taker's ability () and the Rasch model for the item difficulty (b), is referred to herein as the hybrid model.

(103) Embodiments of the present technology will be further elucidated in the following description, which will begin with data adjustment, moving from the test level to the item level via the point-biserial correlation r.sub.xX between item response (x) and test score (X). In classical test theory, this correlation, with the difference between the mean test score of examinees who get the item right and the mean test score of examinees who get the item wrong in the numerator (McNemar, 1962, p. 192), not only measures but in fact captures the precise meaning of item discrimination. The section following the next will describe parameter estimation in the fit of the Rasch model to item-response data adjusted by r.sub.xX. (Throughout this description, lower-case x will refer to an item and upper-case X to a test.)

(104) Logits: Observed Scores vs. True-score Estimates

(105) The logit of a proportion or probability P is the natural logarithm of the ratio of P (1P). If P.sub.iq is the probability of a correct response to item i by examinee q, then for a two-parameter logistic model the logit of P.sub.iq is equal to a.sub.i(.sub.qb.sub.i). The next section will need an empirical counterpart of P.sub.iq that has a computable logit. In the Gulliksen version of classical test theory, the observed response x.sub.iq, equal to zero or one, qualifies as such a counterpart, but x.sub.iq has no computable logit. The Kelley version of classical test theory suggests a solution to this problem: a weighted average of an observed score and the observed score mean for a single item,
p.sub.iq=r.sub.xX.sup.2x.sub.iq+(1r.sub.xX.sup.2)x.sub.i (27)
where x.sub.i is the item difficulty, equal to the mean of x.sub.iq over q, and r.sub.xX, as noted earlier, is a measure of item discrimination. Unless r.sub.xX is equal to one, which is not possible for a point-biserial correlation, or x.sub.i is equal to zero or one, P.sub.iq will always have a computable logit.

(106) The use of p.sub.iq rather than x.sub.iq as the empirical counterpart of P.sub.iq makes rational as well as mathematical sense, especially for multiple-choice items. On a zero-to-one scale, the value of one is arguably too large a measure for a correct item response in this case. Not only may the knowledge or skill measured by an item be at least partially irrelevant but also a correct response does not necessarily reflect that knowledge or skill. Guessing may play a role. The effect of guessing, however, also depends on the item's discrimination. An examinee who gets the item right is more likely to have a high test score and thus be more generally knowledgeable on the subject tested if the item is highly discriminating than otherwise. This difference is properly reflected in the value of p.sub.iq. If x.sub.iq=1 for each of two items that vary in discrimination, the value of p.sub.iq will be closer to one for the more discriminating than for the less discriminating item. A like argument applies in the case of incorrect answers to the two items. A correct answer has a greater positive effect and an incorrect answer a greater negative effect on p.sub.iq if made to a more discriminating than to a less discriminating item in contrast to the uniformly equal positive and negative effects on x.sub.iq. For a multiple-choice item and arguably for any item scored on a zero-to-one scale, p.sub.iq will generally be a more precise reflection of relevant item knowledge than x.sub.iq.

(107) Fitting the Model to Data

(108) Transformation of the single-parameter logistic model of Equation (1) to a logit form permits the accommodation of item discrimination and difficulty in model fit. With the addition of the indices i and q, Equation(1) has the logit form

(109) 0$\begin{array}{cc}\ln \left(\frac{P\left({}_q;b_i\right)}{1-P\left({}_q;b_i\right)}\right)={}_q-b_i& \left(28\right)\end{array}$
the left side corresponding empirically to

(110) $\ln \left(\frac{p_{\mathrm{iq}}}{\left[1-p_{\mathrm{iq}}\right]}\right),$
the logit of p.sub.iq. If

(111) ${}_{\mathrm{iq}}=\left(\frac{p_{\mathrm{iq}}}{\left[1-p_{\mathrm{iq}}\right]}\right),$
then empirically the logit form of the single-parameter logistic model is a regression equation like X=T+E of classical test theory, .sub.iq denoting error:
.sub.iq=(.sub.qb.sub.i)+.sub.iq (29)
so that, as ordinary least-squares estimates assuming the average to be equal to zero,
{circumflex over (b)}.sub.i=.sub.i (30)
and
{circumflex over ()}.sub.q=.sub.q+b(31)
.sub.i being the mean of .sub.iq over q, .sub.q the mean of .sub.iq over i, and b the mean item difficulty.
All computation of item statistics in this and subsequent sections are over q.

(112) Since {circumflex over (b)}.sub.i is a mean and .sub.q a sum of means, estimates of their sampling variances are, respectively,

(113) $\begin{array}{cc}{\hat{}}_{b_i}^2=\frac{\underset{q=1}{\overset{N}{.Math.}}{\left({}_{\mathrm{iq}}-{\stackrel{\_}{}}_i\right)}^2}{N\left(N-1\right)}\mathrm{and}& \left(32\right)\\ {\widehat{\hat{}}}_{{}_q}^2=\frac{\underset{i=1}{\overset{N}{.Math.}}{\left({}_{\mathrm{iq}}-{\stackrel{\_}{}}_q\right)}^2}{n\left(n-1\right)}+\frac{\underset{i=1}{\overset{n}{.Math.}}{\hat{}}_{b_i}^2}{n^2}& \left(33\right)\end{array}$

(114) Despite their apparent equivalence, differing only by a scale conversion, .sub.q and .sub.qb.sub.i cannot both be true scores, like T. Just as a true score in classical test theory must be both test- and examinee-dependent, so a true score for an item in modern test theory must be both item- and examinee-dependent. This is the case for .sub.qb.sub.i, but not for .sub.q, which is only examinee-dependent, though both .sub.q and .sub.qb.sub.i have equal standard deviations and correlations with other variables.

(115) For a single sample, in fact, the correspondence between T and .sub.qb.sub.i is striking if, following common practice, the mean of is assumed to be zero. in this case, just as the mean of T (T), being equal to the proportion of examinees (X) who get the item right, is the measure of item difficulty in classical test theory, so the mean of .sub.qb.sub.i over examinees is equal to minus the measure of item difficulty (b.sub.i) in modern test theory. While X decreases, minus the mean of .sub.qb.sub.i over examinees increases with increasing item difficulty.

(116) The accommodation of varying item discrimination by using the true-response estimates p.sub.iq to estimate b.sub.i and .sub.q occurs without the use of the parameter a.sub.i. The resulting single-parameter logistic model retains the parameter-separation but not the unweighted-scores property of the Rasch model, a price paid for by the accommodation of varying item discrimination. As weighted scores, the true-response estimates p.sub.iq reflect varying item discrimination in contrast to the observed responses x.sub.iq, which do not. The combination of a weighted average like the Kelley estimation model of classical test theory with the Rasch model of modern test theory extends the usefulness of the single-parameter logistic model empirically to tests consisting of items that vary substantially in discrimination.

(117) The Parameter a and Item-Test Correlation

(118) In classical test theory, as noted earlier, the correlation between the response x (0 or 1) to an item and total test score X provides a measure of the discrimination of the item. According to Lord and Novick (1968, p. 378, Equation 16.10.7), in the case of the two-parameter normal-ogive model with assumed to have a standard normal distribution (.sub.=0 and .sub.=1), the slope parameter a has the following relationship to a correlation similar to r.sub.xX: a=r.sub.x/{square root over (1r.sub.x.sup.2)}, where r.sub.x is a biserial correlation between and x equal to either zero or one depending on the value of a latent item-knowledge variable (y) having a standard normal distribution (.sub.y=0 and .sub.y=1). The numerator in the formula for this correlation, like the numerator in the corresponding formula for r.sub.xX, involves the difference between the mean of examinees who get the item right and the mean of examinees who get the item wrong (McNemar, 1962, p. 189). Since r.sub.x/{square root over (1r.sub.x.sup.2)} is a direct function of r.sub.x, a in the two-parameter normal-ogive model is a measure of item discrimination in the tradition of classical test theory.

(119) In this tradition, Birnbaum (1968, pp. 402-403) in fact justified the need to measure discrimination in item response models by citing standard deviations of actual item-test correlations (r.sub.xX) that were too large to occur by chance. This illustration is of interest here for two reasons: Birnbaum recognized (a) the near-equivalence of X and and (b) the role of dichotomous item response measures (x) in the measurement of item discrimination (. . . item-test biserials approximate item-ability biserials . . . ). The remainder of this section examines the relationship between item discrimination and the parameter a in Birnbaum's two-parameter logistic model.

(120) Like the normal-ogive model, the two-parameter logistic model must make two assumptions in order to have a.sub.i as a slope parameter of (.sub.qb.sub.i). A straightforward way to expose the two assumptions in this case is to change the scale of by multiplying (.sub.qb.sub.i)+.sub.iq through by value of a.sub.i (different from a regression coefficient) that will make the standard deviation of the product of a.sub.i and .sub.iq equal to one (S.sub.a.sub.i.sub..sub.iq=1) while assuming, as is traditionally the case, that S.sub.=1. Just as X=T+E in classical test theory implies that S.sub.T=r.sub.XTS.sub.X, so in modern test theory .sub.iq=(.sub.qb.sub.i)+.sub.iq implies that S.sub.=r.sub.S.sub.. Since S.sub.a.sub.i.sub..sub.iq=1 and S.sub.=r.sub.S.sub.=1,

(121) $\begin{array}{cc}\begin{array}{c}{a}_i{\mathrm{.Math.}}_{\mathrm{iq}}=\frac{{\mathrm{.Math.}}_{\mathrm{iq}}}{S_{{\mathrm{.Math.}}_{\mathrm{iq}}}}\\ =\left(\frac{1}{S_{}\sqrt{1-r_{}^2}}\right){\mathrm{.Math.}}_{\mathrm{iq}}\\ =\left(\frac{r_B{S}_{}}{S_{}\sqrt{1-r_{}^2}}\right){\mathrm{.Math.}}_{\mathrm{iq}}\\ =\left(\frac{r_B}{\sqrt{1-r_{}^2}}\right){\mathrm{.Math.}}_{\mathrm{iq}}\end{array}& \left(34\right)\end{array}$
which shows that the value of a.sub.i corresponds to the value presented by Lord and Novick (1968, p. 378) and by Lord (1980, pp. 31-32) in the normal-ogive case: r.sub./{square root over (1r.sub..sup.2)}, like x being a two-valued variable measuring the response to an item. (In fact, r.sub.=r.sub.x.) In Equation (12), the second value (.sub.iq/S.sub..sub.iq) is the standardized value of .sub.iq, whose standard deviation is equal to one. The two assumptions are S.sub.a.sub.i.sub..sub.iq=1 and S.sub.=1.

(122) Iteration from r.sub.xX to r.sub.x

(123) The near-equivalence of X and cited by Birnbaum (1968) suggests the use of iteration to estimate .sub.q and b.sub.i by way of .sub.iq. The first step in the iteration is to use r.sub.xX to estimate .sub.iq, as described earlier. Subsequent steps use r.sub.x, the values being obtained from the preceding step. Iteration continues till the estimates of both and b stabilize. Since r.sub.x=r.sub.x(b), the result of this procedure corresponds precisely at the item level to Kelley's at the test level, (b) and T both being true scores, as indicated earlier. Since the parameter a is a function of r.sub.xor r.sub. (the two being equal), the use of r.sub. to measure item discrimination is tantamount to the use of a to do so in the two-parameter normal-ogive and logistic models. Test analysts who wish to follow the Verhelst and Glas (1995) approach to the accommodation of varying item discrimination in a single-parameter logistic model can also use r.sub. in a=r.sub.xX/{square root over (1r.sub.xX.sup.2)} or r.sub.x in a=r.sub.x/{square root over (1r.sub.x.sup.2)} to help determine the constant value of a in their model. The second of these is the equation used in the following numerical example.

(124) Numerical Example

(125) This section describes the use of simulated data to compare the single-parameter model in its traditional Rasch form with the form described here, as well as a number of variations of each. The data consisted of individual item responses 0's and 1's) on 10-, and 30-item tests, each administered to 1,000 examinees. The values were randomly selected from a standard normal distribution. The b values for the 10-item test were 1.5, 0.75, 0, 0.75, and 1.5, each repeated once. These ten values were duplicated in the 20-item test and triplicated in the 30-item test. Items created to have each b value were also created to have either of two values of r.sub.x. One is the maximum possible value for its difficulty, point-biserial correlations having maximum values less than one, and the other is 0.144, chosen to make the mean r.sub.x for a test equal to 0.400. Table 1 shows these r.sub.x values in the third-to-last row.

(126) Data creation. Using the values and item specifications just described, the following regression equation yielded the probability of a correct response, P.sub.iq, for each item and each examinee: P.sub.iq={square root over (P.sub.i(1=P.sub.i))}r.sub.x.sub.q+P.sub.i, where P.sub.i is the mean over examinees of 1/(1+e.sup.(.sup.q.sup.b.sup.i.sup.)). This equation for P.sub.iq accommodates both item difficulty (P.sub.i) and item discrimination (r.sub.x) in the determination of the probability of a correct response. Comparison of P.sub.iq with R.sub.iq, a random number uniformly distributed between 0 and 1, resulted in the simulated response: 1 if R.sub.iqP.sub.iq, 0 otherwise. This procedure determined the response of each examinee to each item on each of the three tests.

(127) Comparison of models. To facilitate comparison with other models, the model described here will be called the delta model because it involves an increment to the single-parameter logistic model to account for item discrimination. In addition to the Rasch model, these other models are the Rasch K model, a single-parameter model described by Verhelst and Glas (1995) that has a constant (unfitted) discrimination parameter (a) that may differ in value from item to item, the biserial delta model that uses biserial instead of point-biserial correlations for r.sub.x in the delta model, and a hybrid model involving Rasch model estimates of b and delta model estimates of . The focus of the comparisons will be on the correlations of b and estimates with their true values.

(128) Estimation procedures. The estimation procedures differed for the Rasch and the delta models. Estimation for the Rasch and Rasch K models used the maximum-likelihood procedure described by Wright and Panchapakesan (1969), with Newton-Raphson iteration. Involving least squares, estimation for the delta and biserial delta models used Equations (8) and (9) together with the iteration procedure described in the preceding section. Iteration continued till the difference between successive estimates was equal to zero, to two decimal places. The constant discrimination parameter a used in the Rasch K model for each item was equal to r.sub.x/{square root over (1r.sub.x.sup.2)}, the values being the ones estimated separately for each test with the delta model. Both the Rasch and the delta estimation procedures involved joint (b and ) estimation without any distributional assumptions.

(129) Estimation for the hybrid model used the Rasch procedure for b and, with the b values fixed at their Rasch estimates, the delta procedure for .

(130) Results. The five models lined up differently for the estimation of b and the estimation of . One tied at the top for b but not for , one tied at the top for but not for b, and one tied at the top for both. Two did not fare well for either b or . The b comparisons involved only the items in the 10-item test, the items in the other tests being specification replicates of these items. The comparisons involved all three tests.

(131) Referring now to the table in FIG. 11, the b results are shown. The top part of the table displays the difficulty and the bottom part the discrimination data for the ten items. The fourth row, labeled True in the left column, contains the values of b and the third-from-bottom row contains the values of r.sub.x used to create the item responses. The remaining rows contain data obtained by parameter estimation. The numerically labeled columns in the top part of the table display the b estimates ({circumflex over (b)}) for the ten items obtained from the use of each of the five models, as well as the mean of these estimates and the correlation, r.sub.{circumflex over (b)}b, between them and their true values. The r.sub.{circumflex over (b)}b correlations are substantially lower for the biserial delta, Rasch K, and delta models than for the Rasch and hybrid models, being close to one for the latter two. The mean b estimates are also substantially farther from the true mean for the delta, biserial delta, and Rasch K models than for the Rasch and hybrid models, whose common mean (0.01) is very nearly equal to the true mean (0.00). The row labeled contains the a estimates for the ten items, along with their mean. The bottom row contains values of the t-Fit statistic described by Wright and Masters (1982, Ch. 5) and Masters (1988). These statistics are interpretable more or less like t statistics, highly negative values expected for items of high discrimination and highly positive values expected for items of low discrimination. The data bore out these expectations, as comparison of the t-Fit values with the a estimates in the second-to-last row and the r.sub.x values in the third-to-last row shows.

(132) The line-up of the Rasch and delta models was just the opposite for estimation, the hybrid model holding the same position in both comparisons. Measuring internal validity, the correlations of the estimates with their true values, which increased with test size, were substantially higher in all three tests for the delta and hybrid models (0.807, 0.895, and 0.916) than for the biserial delta model (0.790, 0.860, and 0.895) and the Rasch (0.740, 0.848, and 0.883) and Rasch K (0.679, 0.790, and 0.859) models. Referring now to FIG. 12, results for the delta and hybrid models (top curve) and the Rasch model (bottom curve) are shown. Of particular interest, seen clearly in this figure, is that the delta and hybrid correlation for the 20-item test was higher than the Rasch correlation for the 30-item test.

(133) Discussion. Although the focus here is on the comparison between the Rasch model and the delta model, consideration of the internal-validity standings of the other models is also informative. In yielding unfitted though reasonably differing slopes for the item response curves, the Rasch K model yielded b and estimates that had lower internal validities than those of the Rasch model. The biserial delta model sharpened the r.sub.x differences among the items but, in so doing, distorted both the b and the estimates with a consequent degradation in the internal validities of both. In any case, use of either of these models in practice would be expected to produce no better internal validities for either b or than their Rasch and delta counterparts.

(134) Because a trade-off exists between test length and item discrimination in their effects on the internal validity of a test (the higher the average item discrimination the lower the test length needed to achieve a specific internal validity), use of a model that weights items directly according to their discrimination measures should produce estimates having higher internal validities than use of a model that does not, all else equal. Estimates of should, accordingly, have a higher internal validity when produced by the delta model than when produced by the Rasch model from the same test data. That difference is precisely what the results here show for each of the three tests. The difference is in fact so great that, as noted earlier, use of the delta model on the 20-item test produced estimates that had a higher internal validity than use of the Rasch model on the 30-item test.

(135) The superiority of the delta over the Rasch model applies only to the estimation of . In the estimation of b, the Rasch model is superior to the delta model with respect to both the internal validities and the means of the estimates. This difference is especially important for test developers who maintain item banks and use anchor items in equating tests.

(136) When one model is better for and the other for b estimation and when the estimation of both are important, as they both are, which model is a test developer or analyst to use? That question motivated the addition of the hybrid model to the study. The parameter separation that exists in both the Rasch and the delta model, suggested that the hybrid model might share the advantages of both. The results confirmed this expectation. Use of the hybrid model in each test resulted in estimates that had the internal validity of the delta model and b estimates that had the internal validity, as well as the mean, of the Rasch model. In the line-up of all five models, the hybrid model was clearly the best in both and b estimation.

(137) To study the effect of differences in item discrimination on parameter estimation, the data were created in an attempt to maximize these differences while maintaining a reasonable mean value. The t-Fit statistics, shown in the bottom row of the table in FIG. 11, indicate that this attempt was successful. Although the dispersion of item discrimination measures in actual data is not likely to be so large as in the data studied here, some dispersion will always occur in practice. The differential effects found in this simulation study may correspondingly be smaller in the real world, but they will nevertheless exist there. While retaining the Rasch model's advantages in the estimation of item difficulties, use of the hybrid model will make the existence of real-world differences in item discrimination less of a challenge than it would be with use of the Rasch model.

(138) Test developers or analysts reading the description in this specification may wonder why they should fit a single-parameter model to accommodate varying item discrimination when the option exists simply to use the two-parameter logistic or normal-ogive model for the same purpose. One important eason may be this: Single-parameter models do not permit the crossing of item response curves that two-parameter models do. Another reason, at least as important: Like the delta model, two-parameter models may distort the distribution of b estimates in the process of increasing the precision of estimation. The first is a theoretical, the second a practical problem. The hybrid model avoids both.

(139) Embodiments of the present technology can be used in various applications of single-parameter logistic models, which yield the probability of an outcome as a function of the ability of the performer and the difficulty of the job whenever the purpose of these applications is to obtain estimates of performer ability and job difficulty. Use of embodiments of the present technology will result in improved estimation of performer ability and, combined with the Rasch model, improved estimation of job difficulty, as well. Examples of other applications include baseball, involving a player's ability to hit and the difficulty of hitting a pitcher, and spelling bees, involving the spelling ability of a contestant and the difficulty of a word to be spelled. It is, however, appreciated that embodiments of the present technology are not limited to these examples.

(140) The foregoing descriptions of specific embodiments of the present technology have been presented for purposes of illustration and description. They are not intended to be exhaustive or to limit the invention to the precise forms disclosed, and obviously many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the present technology and its practical application, to thereby enable others skilled in the art to best utilize the present technology and various embodiments with various modifications as are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the claims appended hereto and their equivalents.