The Efficacy of Warm's Weighted Mean Likelihood Estimate (WLE) Correction to Maximum Likelihood Estimate (MLE) Bias

Thomas Warm (1989) reports that "Lord (1983) found that maximum likelihood estimates of θ [person ability] are biased outward" and then he restates Lord's expression for the size of this bias:

where, for dichotomous Rasch items,
J = Σ P_θi (1-P_θi ) (1-2P_θi )
I = Σ P_θi (1-P_θi )
summed for all items, i = 1,L in the test, where P_θi is the Rasch-model probability of success of ability θ on item i.

The corrected estimate, θ_WLE = θ_MLE + (J / ( 2 * I² ) ), which is almost always closer to the item mean than θ_MLE.

How effective is this bias correction? Warm uses a Monte Carlo study to demonstrate its effectiveness, but an exact algebraic investigation can be conducted.

I posited a test of 25 items, with its item difficulties uniformly spaced 0.2 logits apart. Figure 1 shows the locations (x-axis) of the items on the 25-item test. The item difficulties are centered on 0 logits.

Applying the MLE method of Wright & Douglas (1996) for estimating θ from known item difficulties, a Rasch ability estimate, M(s) is obtained for each possible raw score, s=0-25, on the test of 25 items. Since the estimates corresponding to s=0 and s=25 are infinite, they are substituted by estimates corresponding to s=0.3 and s=24.7 score-points. The MLE ability estimates are shown in Figure 1.

Figure 1. MLE and WLE for 25 dichotomous items.

Warm's bias correction is applied to each MLE estimate, M(s), to produce a Warm's Mean Likelihood Estimation (WLE) value, W(s). See Figure 1. WLE estimates are more central than the MLE estimates, except for estimates corresponding to scores of 0.3 and 24.7, where the MLE estimates are used unchanged.

Under Rasch model conditions, each raw score, s, on a given set of items, corresponds to one estimated ability θ(s), but each true (generating) ability corresponds to all possible raw scores. For 25 items, there are 2^25 = 33,554,432 possible different response strings. According to the Rasch model, each of these response strings has a finite probability of being observed for each generating ability.

Probability of response string n for ability θ = Pnθ = Π exp( (xni (θ - di) ) / (1 + exp(θ - di) )

for i = 1 to 25, where xni is the scored 0,1 response to item i in response string n, and di is the difficulty of item i.

Response string n has a raw score of s = Σ xni for i = 1 to 25. Score s has an MLE estimate of Mn = M(s) and a WLE estimate of Wn = W(s).

The expected values of the estimates corresponding to each generating value can now be computed:

Expectation (MLE(θ)) = Σ Pnθ Mn for n = 1 to 2^25

Expectation (WLE(θ)) = Σ Pnθ Wn for n = 1 to 2^25

These values are plotted in Figure 1 for θ in the range -6 logits to +6 logits. The WLE ogive corresponds to the identity line with the generating values for most of its range. The MLE ogive is slightly less central (as predicted by Fred Lord). We can see that the WLE bias correction is effective over the entire range of MLE estimates for non-extreme scores (-4 to +4 logits). The biggest bias correction is 0.23 logits at a generating value of 3.6 logits, as shown in Figure 2. This is less than half the size of the standard error of each estimate which is close to 0.5 logits for most of the range. We can also see that, for "true" generating abilities within 2 logits of the center of the test, the MLE bias is less than 0.1 logits, and so negligible for practical purposes.

Figure 2. Detail of Figure 1 showing MLE bias.

Similar investigations for tests of length 2 to 24 items demonstrated that the WLE bias correction is effective for tests of 7 dichotomous items or more.

Polytomous Items

We can apply the same logic to Rasch-model polytomous items.

Bias (θ_MLE) = - J / ( 2 * I² )

J = = Σ Σ (P'_θikP"_θik)/P_θik = Σ( (Σk³P_θik ) - 3(Σk²P_θik )(ΣkP_θik ) + 2(ΣkP_θik )³ )

I = Σ( (Σk²P_θik ) - (ΣkP_θik )² )

where P_θik is the Rasch-model probability of a person of ability θ being observed in category k of item i, and the summations are over i=1, L and the polytomous categories k=0,m.

The corrected estimate, θ_WLE = θ_MLE + (J / ( 2 * I² ) ), which is almost always closer to the item mean than θ_MLE.

The results of this investigation are shown in Figure 3 with items 0.1 logits apart, and thresholds 1 logit apart. Using the estimation procedure in Linacre (1998), the results are similar to the findings for dichotomous items in Figure 1.

Warm's bias correction is seen to be efficacious for the correction of MLE bias across the useful measurement range of the items, but MLE bias is also seen to be inconsequential for most practical purposes.

Figure 3. MLE and WLE for 12, 4-category, items.

John M. Linacre

Linacre J.M. (1998) Estimating Rasch measures with known polytomous (or rating scale) item difficulties: Anchored Maximum Likelihood Estimation (AMLE), RMT 12:2 p. 638

Lord, F. M. (1983). Unbiased estimators of ability parameters, of their variance, and of their parallel-forms reliability. Psychometrika, 48, 2, 233-245.

Warm T.A. (1989) Weighted Likelihood Estimation of Ability in Item Response Theory. Psychometrika, 54, 427-450.

Wright B.D., Douglas G.A. (1996) Estimating measures with known dichotomous item difficulties. Rasch Measurement Transactions, 10:2, 499

Linacre J.M. (2009) The Efficacy of Warm's Weighted Mean Likelihood Estimate (WLE) Correction to Maximum Likelihood Estimate (MLE) Bias, Rasch Measurement Transactions, 2009, 23:1, 1188-9

Rasch Publications
Rasch Measurement Transactions (free, online)	Rasch Measurement research papers (free, online)	Probabilistic Models for Some Intelligence and Attainment Tests, Georg Rasch	Applying the Rasch Model 3rd. Ed., Bond & Fox	Best Test Design, Wright & Stone
Rating Scale Analysis, Wright & Masters	Introduction to Rasch Measurement, E. Smith & R. Smith	Introduction to Many-Facet Rasch Measurement, Thomas Eckes	Invariant Measurement: Using Rasch Models in the Social, Behavioral, and Health Sciences, George Engelhard, Jr.	Statistical Analyses for Language Testers, Rita Green
Rasch Models: Foundations, Recent Developments, and Applications, Fischer & Molenaar	Journal of Applied Measurement	Rasch models for measurement, David Andrich	Constructing Measures, Mark Wilson	Rasch Analysis in the Human Sciences, Boone, Stave, Yale
in Spanish:	Análisis de Rasch para todos, Agustín Tristán	Mediciones, Posicionamientos y Diagnósticos Competitivos, Juan Ramón Oreja Rodríguez

Go to Institute for Objective Measurement Home Page. The Rasch Measurement SIG (AERA) thanks the Institute for Objective Measurement for inviting the publication of Rasch Measurement Transactions on the Institute's website, www.rasch.org.

Coming Rasch-related Events
June 23 - July 21, 2023, Fri.-Fri.	On-line workshop: Practical Rasch Measurement - Further Topics (E. Smith, Winsteps), www.statistics.com
Aug. 11 - Sept. 8, 2023, Fri.-Fri.	On-line workshop: Many-Facet Rasch Measurement (E. Smith, Facets), www.statistics.com