reza wrote:Hi Ian,
As I mentioned in the post above, in my website, I tried to simplify CAT which means it is not going to be 100% accurate but I believe that is better than knowing nothing and acting uniformly as if it is a paper-and-pencil test. However, the statement of "not being accurate at all" is not accurate either!
As I repeated a lot in this very thread, I NEVER said the last questions are not important at all and you should leave them out. But I would never sacrifice my first 10-20 questions for those last ones. Why?
What you said about the string of wrong answers is correct in an exam with unlimited number of questions which the termination point comes when the system is sure about your ability. But in the real test, you have a limited number of questions to answer. So if you start off with wrong answers, let's say due to nervousness, you may never have the chance to prove your ability only because the number of questions are limited and the section will end long before you can come back.
The opposite is true. Imagine, you get lucky and for some reasons you know the answers of the first 40 questions, how many question should you do wrong for the system to know that you were not as smart as it estimated initially? Another 20? Since there are about 50 questions in each section, the system will not have the opportunity to find out your real ability. I do not DENY that it will punish you for the wrong answers as you can see in the graph in my website, you can lower your score with bad performance. But the punishment does not justify why someone should blow the first questions to get the last ones right!
Finally, I NEVER said that the experimental questions will ALWAYS come at the end. but it is more likely it will not be among the first 10 questions. why? a simple logic!
the system does not have a predefined number of experimental questions to give. You may get 5 or 15 or none. it depends on your performance. If you are too volatile, you might get fewer since the system needs to ask more of real questions to estimate your ability. In fact, it has no value to give an experimental question to someone that even the system does not know how good he is! Right?
Now, considering that there is very limited number of questions to ask, and you have no idea how good the examinee is, would you give all the experimental questions early in the exam? what if you need to ask more of the real stuff due to patchy performance? the system would not give away the "precious" limited questions unless it is sure it can afford it.
Well, we agree on a few points. CATs are surely different from pen-and-paper tests, and a test-taker should understand going in that the test will continue to get harder until questions are too difficult to answer, no matter what your ability, unless you are a top-level performer. That is, almost everyone will struggle to answer many questions on the test, and this is not a sign that you're doing badly. The goal of the test is to discover the difficulty level at which you answer 60% of questions correctly, so beyond a certain point in the test, it is likely you will answer roughly 40% incorrectly, unless you are at one extreme on the ability scale.
It's also true that the GMAT is not an 'ideal' CAT; ideal CATs are not of fixed length, and terminate when the algorithm's confidence level in your ability exceeds a predetermined value. However, the GMAT has been designed to be sufficiently long that this confidence level will in almost all cases be acceptable, and questions with unsuitable statistical properties (poor discrimination, for example, or extreme difficulty values, whether too high or too low) will be discarded from the question pool to ensure this. If the test wasn't certain to attain this confidence ('information') level, the test would have no validity. The fact that the GMAT is of predetermined length also does mean that earlier questions are
slightly more important than later ones; in particular, it means that it barely matters if you need to guess at one or maybe two questions at the end. A long string of wrong answers is especially damaging, however, regardless of whether the test is of predetermined length; this is a consequence of the IRT principles behind the algorithm.
There is published information about how experimental questions function on the exam and you can read about this in GMAC research reports. The number is predefined, and some of the questions will be inserted early in the test. Experimental questions are divided into pools of between 7 and 14 questions each. Each test-taker, before they begin a test, is assigned one pool, and these questions are randomly inserted in the test. Because the test is designed to achieve a certain minimum information level, the experimental questions, logically, can be inserted early in the test, long before the algorithm has determined the test-taker's score, because there will be a sufficient number of appropriate real questions later in the test to determine that score.