Unobserved categories can be dropped from rating scales (or partial credit items) and the remaining category recounted during estimation. For intermediate categories only, recounting can be prevented and unobserved categories retained in the analysis. This is useful when the unobserved categories are important to the rating scale (or partial credit) logic or are usually observed, even though they happen to have been unused this time. Category transitions for which anchor Rasch-Andrich threshold-values (step calibrations) are supplied are always maintained wherever computationally possible, even when there are no observations of a category in the current data set.

 

Use "Rating Scale= .... Keep" when there may be intermediate categories in your rating scale (or partial credit) that aren't observed in this data set, i.e., incidental zeroes.

 

Use "Rating Scale= .... Ordinal" when your category numbering deliberately skips over intermediate categories, i.e., structural zeroes.

 

"Rating Scale= .... Ordinal" Eliminate unused categories and close up the observed categories.

 

"Rating Scale= .... Keep" Retain unused non-extreme categories in the ordinal categorization.

 

When "Rating Scale= .... Keep", missing categories are retained in the rating scale (or partial credit), so maintaining the raw score ordering. But missing categories require arbitrarily extreme Rasch-Andrich thresholds. If these threshold values are to be used for anchoring later runs, compare these thresholds with the thresholds obtained by an unanchored analysis of the new data. This will assist you in determining what adjustments need to be made to the original threshold-values in order to establish a set of anchor threshold-values that maintain the same rating scale (or partial credit) structure.

 

Example 1: Incidental unobserved categories. Keep the developmentally important rating scale (or partial credit) categories, observed or not. Your small Piaget scale goes from 1 to 6. But some levels may not have been observed in this data set.

 Models = ?,?,?, Piagetscale

 Rating Scale = Piagetscale, R6, Keep

 

Example 2: Structural unobserved categories. Responses have been coded as "10", "20", "30", "40", but they really mean 1,2,3,4  

 Models = ?, ?, ?, Tensscale

 Rating Scale = Tensscale, R40, Ordinal

 ; if "Rating Scale= .... Keep", then data are analyzed as though categories 11, 12, 13, 14, etc. could exist, which would distort the measures.

 ; for reporting purposes, multiply Facets reported raw scores by 10 to return to the original 10, 20, 30 categorization.

 

Example 3: Some unobserved categories are structural and some incidental. Rescore the data and use "Rating Scale= .... Keep". Possible categories are 2,4,6,8 but only 2,6,8 are observed this time.

 (a) Rescore 2,4,6,8 to 1,2,3,4  

 (b) Set "Rating Scale= .... Keep", so that the observed 1,3,4 and unobserved 2 are treated as 1,2,3,4

 (c) For reporting purposes, multiply the reported Facets scores by 2 using Excel or similar software.

 Models = ?,?,?,Evenscale

 Rating Scale = Evenscale, R8, Keep

 1 = original 2, , , 2

 2 = original 4, , , 4

 3 = original 6, , , 6

 4 = original 8, , , 8

 *

 

---

 

Incidental and Structural Zeroes: Extreme and Intermediate

 

For missing intermediate categories, there are two options.

 

If the categories are missing because they cannot be observed, then they are "structural zeroes". Specify "Rating Scale= .... Ordinal". This effectively recounts the observed categories starting from the bottom category, so that 1,3,5,7 becomes 1,2,3,4.

 

If they are missing because they just do not happen to have been observed this time, then they are "incidental or sampling zeros". Specify "Rating Scale= .... Keep". Then 1,3,5,7 is treated as 1,2,3,4,5,6,7.

 

Categories outside the observed range are always treated as structural zeroes.

 

When "Rating Scale= .... Keep", unobserved intermediate categories are imputed using a mathematical device noticed by Mark Wilson. This device can be extended to runs of unobserved categories.