While you’re waiting for Dan Koretz’ book on testing to arrive – I think eduwonkette and I should get some kind of consideration for shilling for this book so often here – here’s a brief skoolboy’s-eye view on testing. Actual psychometricians are welcome to correct what I have to say.
Tests are typically designed to compare the performance of students (whether as individuals, or as members of a group) either to an external standard for performance or to one another. Tests that compare students to an external standard are called criterion-referenced tests; those that compare students to one another are called norm-referenced tests. Even though criterion-referenced tests are intended to hold students’ performance up to an external standard, there is often a strong temptation to compare the performance of individual students and groups of students on such tests, as if they were norm-referenced.
A typical standardized test of academic performance will have a series of items to which students respond, generally either in a multiple-choice or constructed response format, which means that students are constructing a response to the item. There’s usually only one right answer to a multiple-choice item, whereas constructed-response items may be scored so that students get partial credit if they demonstrate partial mastery of the skill or competency that the item is intended to represent. For any test-taker, we can add up the number of right answers, plus the scores on the constructed-response items, to derive the student’s raw score on the test. A test with 45 multiple-choice items would have raw scores ranging from 0 to 45.
For individual test items, we can look at the proportion of test-takers who answered the item correctly, which is referred to as the item difficulty or p-value, which has nothing to do with the p-values used in tests of statistical significance, but rather the proportion (p) of examinees who got the item right. Some test items are more difficult than others, and hence items will have varying p-values.
Raw scores are rarely interpretable, in part because they are a function of the difficulty of the items. For this reason, they are typically transformed into scale scores, which are designed to generate a score that will mean the same thing from one version of a test to the next, or from one year to the next. The scale for scale scores is arbitrary; the SAT is reported on a scale ranging from 200 to 800, whereas the NAEP scale ranges from 0 to 500.
The process of transforming raw scores into scale scores is computationally intensive, generally using a technique known as Item Response Theory (IRT), which simultaneously estimates the difficulty of an item, how well the item discriminates between high and lower performers, and the performance of the examinee. An examinee who successfully answers highly difficult items that discriminate between high and low performers will be judged to have more ability, and hence a higher scale score, than an examinee who gets the difficult items wrong.
There’s no one right way to transform raw scores into scale scores, and it’s always a process of estimation, which is sometimes obscured by the fact that scores are reported as definite quantities. (A little skoolboy editorializing here…) The expansion of testing hastened by NCLB has placed a lot of pressure on states, and their testing contractors, to construct scale scores for a test that represent the same level of performance from one year to the next (a process known as test equating). Much of this is done under great time pressure, and shielded from public view. The process is complicated by the fact that states typically don’t want to release the actual test items they use, because then they can’t use them in subsequent assessments as anchor items that are common across different forms of a test, since students’ performance on such items could change due to practice. Some tests are vertically equated, which means that a given score on the fourth-grade version of a test represents the same level of performance as that same score on the fifth-grade version of the test. In a vertically-equated test, if the average scale score is the same for fourth-graders as it is for fifth-graders, we’d infer that the fifth-graders haven’t learned anything during fifth-grade.
Proficiency scores represent expert judgments about what level of scale score performance should describe a student as proficient or not proficient at the underlying skill or competency that the test is measuring. For example, NAEP defines three levels of proficiency for each subject at each of the grades tested (4th, 8th and 12th): basic, proficient, and advanced. Cut scores divide the scale scores into categories that represent these proficiency levels, with students classified as below basic, basic, proficient, or advanced. These proficiency scores do not distinguish variations in students’ performance within the category; one student could be really, really advanced and another just advanced, and whereas a scale score would record that difference, a proficiency score would simply classify both students as advanced. The fact that proficiency levels are determined by expert judgment, and not by the properties of the test itself, means that they are arbitrary; the level of performance designated as proficient on NAEP may not correspond to the level of performance designated as proficient on an NCLB-mandated state test. Many researchers (including Dan Koretz, eduwonkette, and me) are concerned that the focus on proficiency demanded by NCLB accountability policies has the unintended consequence of concentrating the attention of school leaders and practitioners on a narrow range of the test-score distribution, right around the cut score for the category of “proficient,” to the detriment of students who are either well below or well above that threshold. Such a focus is a political judgment, not a psychometric one, and there are arguments both for and against it.
I’ll update this as more knowledgeable readers weigh in. If experts in measurement were to judge proficiency thresholds for knowledge about testing, I’d probably be classified as basic; Dan Koretz is definitely advanced. For a lively and readable treatment of these kinds of issues, get his book!