Misfit Statistics for Rating Scale Categories

Analysis of ordinal observations has seldom included investigation into whether each invited category is performing as intended. Even in Rasch analysis such investigation has not been routine, partly because diagnostically useful fit statistics were not available. Now, analyses based on recent versions of BIGSTEPS indicate that two useful sets of statistics have been formulated.

Average Measure Difference

Implicit in the use of ordinal observations is the specification that the higher the category number, the more of the latent variable is evidenced. Thus, on average, "better" performers should produce higher ratings than "worse" performers. Also "easier" items should manifest higher ratings than "harder" items.

The observation Xni=x is modelled as governed by the difference between person n's ability Bn and item i's difficulty Di. The effect off the measure difference Bn-Di is observed as Xni=x. Examining the entire data set, the average measure difference (AMD) modelled to produce an observation x is

sum [over (Xni=x)] (Bn - Di) / sum(1) for i=1,L n=1,N

The AMD for each category can be computed. Since "more" of the rating scale is modelled to reflect "more" of the underlying variable, the AMDs are expected to increase up the rating scale. This pattern can be seen in Example 1, a well-behaved rating scale. As the categories ascend from 0 to 4, the AMDs increase from -2.34 to 2.21 logits.

Category           Average             Average
Number             Measure             Measure
                Difference (AMD)    Difference (AMD)
----------------------------------------------------
0                -2.34                -2.34
1                -1.56                -1.56
2                 0.12                 1.57
3                 1.57                 0.12
4                 2.21                 2.21
----------------------------------------------------
               Example 1:             Example 2:
              Well-behaved            Problematic
              rating scale            rating scale

Example 2, a problematic scale, shows a different pattern. The AMDs are ascending for the most part, except AMD for category 3 which, at 0.12, is less than that for category 2 at 1.57. This suggests that category 2 is not "less" than category 3 in practice, despite the scale designer's intention. A common cause of this is the use of the central option of a five category Likert scale to signify "No Opinion" or "Don't Know". "Don't know" either investigates behavior on a dimension different from the other categories or enables the respondent to escape from answering the question.

Here are some remedies to disordered AMDs, such as Example 2:

1) Some or all of the observations in categories 2 or 3 can be treated as missing. Indeed, if category 2 is off-dimension or used idiosyncratically, then it is not measuring the desired dimension and all observations in category 2 could be treated as missing.

2) Closer examination of the definitions of categories 2 and 3 may indicate that reversing their order would maintain the ordinally ascending meaning of the scale. Simply recode all 2's as 3's and all 3's as 2's.

3) Third, the difference between a "2" and a "3" may not be clear to respondents, e.g., the difference between "often" and "nearly always". Then categories "2" and "3" can be joined into one category, numbered 2, so that category 4 now becomes 3.

When combining or deleting categories, aim toward equalizing the category frequencies as much as possible, so that each category contributes about equally to the measurement process.

Observed / Expected Mean-Square Fit Ratios

AMDs can be correctly ordered, but the categories themselves still be used haphazardly. The modelled raw-score variance of an observation, Xni, on a rating scale is

Vni = sum from k=0 to m (k-Eni)^2 Pnik

The observed squared residual of Xni is

(Xni - Eni)^2

Summing these variances across the data and partitioning by rating scale category, the variance explained by ratings in category x is modelled to be

Mx = sum over all Xni of (x-Eni)^2 Pnix

The observed residual sum of squares due to ratings of Xni=x is

Ox = Sum over (Xni=x) (x-Eni)^2

When the data fit the model, the modelled variance approximates the residual sum of squares. Differences are diagnostic of misfit.

The INFIT statistic, Vx, summarizes their agreement for category x:

Vx = Ox/Mx

This fit ratio has a mean-square form with expectation 1.0, and range 0 to infinity. Values greater than 1.0 indicate improbable category use. Values less than 1.0 indicate overly predictable category use.

The squared standardized residual for an observation of Xni=x is

Znix^2 = (x-Eni)^2/Vni

Summing these terms across the data and partitioning by rating scale category, the contribution of category x is modelled to be

M'x = Sum (Znix^2 Pnix)

The observed sum of squared standardized residuals for observations of Xni=x is

O'x = Sum for (Xni=x) Znix^2

Again, when the data fit the model, the observed sum approximates the modelled sum.

The OUTFIT mean-square for observations in category x is the ratio of observed to expected sum-of-squared standardized residuals, Ux:

Ux = o'x/M'x

This fit ratio is also a mean-square with expectation 1.0 and range 0 to infinity. Values greater than 1.0 indicate improbable category use. Values less than 1.0 indicate overly predictable category use.

The Table shows the results for the familiar "Liking for Science" data set. AMDs exhibit the desired monotonically ascending pattern. The OUTFIT mean-squares, however, show some unwanted behavior. The bottom category with mean-square 1.20 is used approximately as modelled. The central category with mean-square .69 is overly predictable. This suggests that some children responding to this survey avoided making other than obvious choices. One child responded in the central category to every item. The top category, with mean-square 1.47, manifests improbable observations. A few children liked activities that they were expected to dislike. These activities included "watching rats" and "finding old bottles". From the perspective of measuring "Liking for Science", these idiosyncratic ratings are off-dimension and so perturb the measuring system. Measurement would be improved by recoding these inconsistent ratings as missing.

The INFIT mean-squares, which are more sensitive to idiosyncratic usage of adjacent categories, are within their typical range.

Category      Count   AMD    INFIT Mean-square   OUTFIT
0 "dislike"    378    -.87      1.09              1.02
1 "neutral"    620     .13       .86               .69
2 "like"       852    2.21      1.00              1.47

Misfit Statistics for Rating Scale Categories. Linacre JM. … Rasch Measurement Transactions, 1995, 9:3 p.450

Rasch Books and Publications
Invariant Measurement: Using Rasch Models in the Social, Behavioral, and Health Sciences, 2nd Edn. George Engelhard, Jr. & Jue Wang	Applying the Rasch Model (Winsteps, Facets) 4th Ed., Bond, Yan, Heene	Advances in Rasch Analyses in the Human Sciences (Winsteps, Facets) 1st Ed., Boone, Staver	Advances in Applications of Rasch Measurement in Science Education, X. Liu & W. J. Boone	Rasch Analysis in the Human Sciences (Winsteps) Boone, Staver, Yale
Introduction to Many-Facet Rasch Measurement (Facets), Thomas Eckes	Statistical Analyses for Language Testers (Facets), Rita Green	Invariant Measurement with Raters and Rating Scales: Rasch Models for Rater-Mediated Assessments (Facets), George Engelhard, Jr. & Stefanie Wind	Aplicação do Modelo de Rasch (Português), de Bond, Trevor G., Fox, Christine M	Appliquer le modèle de Rasch: Défis et pistes de solution (Winsteps) E. Dionne, S. Béland
Exploring Rating Scale Functioning for Survey Research (R, Facets), Stefanie Wind	Rasch Measurement: Applications, Khine	Winsteps Tutorials - free Facets Tutorials - free	Many-Facet Rasch Measurement (Facets) - free, J.M. Linacre	Fairness, Justice and Language Assessment (Winsteps, Facets), McNamara, Knoch, Fan
Other Rasch-Related Resources: Rasch Measurement YouTube Channel
Rasch Measurement Transactions & Rasch Measurement research papers - free	An Introduction to the Rasch Model with Examples in R (eRm, etc.), Debelak, Strobl, Zeigenfuse	Rasch Measurement Theory Analysis in R, Wind, Hua	Applying the Rasch Model in Social Sciences Using R, Lamprianou	El modelo métrico de Rasch: Fundamentación, implementación e interpretación de la medida en ciencias sociales (Spanish Edition), Manuel González-Montesinos M.
Rasch Models: Foundations, Recent Developments, and Applications, Fischer & Molenaar	Probabilistic Models for Some Intelligence and Attainment Tests, Georg Rasch	Rasch Models for Measurement, David Andrich	Constructing Measures, Mark Wilson	Best Test Design - free, Wright & Stone Rating Scale Analysis - free, Wright & Masters
Virtual Standard Setting: Setting Cut Scores, Charalambos Kollias	Diseño de Mejores Pruebas - free, Spanish Best Test Design	A Course in Rasch Measurement Theory, Andrich, Marais	Rasch Models in Health, Christensen, Kreiner, Mesba	Multivariate and Mixture Distribution Rasch Models, von Davier, Carstensen

Forum

Rasch Measurement Forum to discuss any Rasch-related topic

Go to Top of Page
Go to index of all Rasch Measurement Transactions
AERA members: Join the Rasch Measurement SIG and receive the printed version of RMT
Some back issues of RMT are available as bound volumes
Subscribe to Journal of Applied Measurement

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.

The URL of this page is www.rasch.org/rmt/rmt93j.htm

Website: www.rasch.org/rmt/contents.htm