Skipping few questions

Thanks IanStewart for this detailed analysis. I think by reading your guidance I should be able to better evaluate myself and choose whatever can give me a higher score.

I do not recommend the strategy of skipping RC.

1. Every RC passage will likely have both easy and difficult questions.

If you get the easy questions wrong (by skipping), you will likely be penalized heavily even if you subsequently answer many, likely easier, questions correctly afterwards.

2. Getting a series of questions wrong will be heavily penalized regardless of difficulty level.

3. If you skip an "easier" RC passage and get the answers wrong, you will be heavily penalized.

Bottomline:

You will likely have much to lose if you adopt a blanket strategy of skipping RCs. The test is now shorter and there is less room for error.

I advise instead to adopt smart guessing (e.g., eliminating some answer choices), if at all you want to "skip".

I never advise anyone to skip an entire set of questions in a row on an adaptive test.

Just to clear up two common misconceptions:

- getting a series of questions wrong is not heavily penalized regardless of difficulty level. A series of wrong answers is normally heavily penalized only because that series will very often include wrong answers to easy questions. But the difficulty level is critically important. If every question in a series of wrong answers is extremely hard, a series of wrong answers actually won't hurt a test taker much at all. When the algorithm is working out your score, it takes no notice of whether your wrong answers occurred in a row or if they were spread out.

The reason a string of wrong answers in a row is damaging most of the time (but not always) is because the difficulty level of each question often starts to plummet with each wrong answer. So the later questions in that string are often easy, and wrong answers to easy questions hurt a lot. But it's specifically in RC where that principle is irrelevant, since RC questions are not adapting individually based on your answers. If you get your first three RC questions wrong, it doesn't become any more likely that the fourth RC question will be easy.

- the test is shorter now, but only because the number of diagnostic questions has been reduced. Test takers have the same room for error as before.

I intentionally left out some obvious stuff, such as:

- Wrong answers to a series of easy questions will be more heavily penalized than a series of not-easy questions.

- Wrong answers to "difficult" questions will attract less penalty simply because the testtaker is being compared to the performance of other testtakers. Since a majority of testtakers will get difficult questions wrong, the penalty will be less severe.

- The test is shorter because the number of experimental questions has been reduced.

We do not have direct knowledge of how the algorithm works, but some of us have PhD-level training in research and research methodology. Much of what people quote is based on intuition of how things should work, rather than how things actually work. Research methodology is one such area.

I would urge extreme caution in advising testtakers to intuitively skip some RCs in hopes of maximizing on other sub-sections. RCs and their questions come in groups of questions. The difficulty levels of these questions will be unknown quantities. We cannot advise beyond stating that, in general terms, some RC passages will have overall higher difficulty levels compared to others. But the relative mix of difficulty levels are unknown parameters.

In addition, the scoring algorithm may account for the fact that RC questions come in bunches as opposed to coming in singles. The algorithm may have built-in compensation mechanisms. We do not know.

Let us reserve quick and intuitive judgements. The scoring algorithm is quite sophisticated. It relies on potentially hundreds of thousands of data points spread over years and decades. The parameters are internal knowns and external unknowns. They can be slowly tweaked in externally unknown ways and still retain overall reliability.

My own observation is that the test demonstrates some unacceptable levels of reliability, eg, large swings in scores over short periods of time for some special groups of testtakers. But this topic is separate from the subject above.

I recommend that testtakers not ruin themselves by attempting to outsmart the algorithm. Instead, use safe and tried normal testtaking strategies. This does not include, in my view, randomly skipping entire bunches of RC questions.

Posted from my mobile device

Hello,
I am struggling to complete verbal on time. On an average I get short of time to solve 3-4 questions and thus no matter I need to skip these 3-4 questions so that I dont leave any question unmarked. No doubt I am preparing myself for this situation not to happen regularly(any suggestion please to improve? ) but say, I have 8 minutes with 7 questions remaining ( 1 RC and other SC/CR left). What should I do now?

This isn't the case. The precise mathematical model underlying the GMAT algorithm is well-known; it's a 3-parameter logistic IRT model, and you can read about it in many academic journal articles (and it's not based on thousands of data points, it's based on probability theory). I've coded algorithms identical to the GMAT algorithm, so any judgments I make about it are neither quick nor intuitive. If I'm ever speculating on this forum about the algorithm, I'm generally careful to point that out, but most of what I write about it is based on the math the algorithm uses, and is not speculation.

I do agree with you that guessing at an entire RC passage is going to be a bad idea for most people. But it will be a good idea for some test takers. I think the only test takers who should consider it are people who:

- normally don't score in the above-average range in Verbal (so normally score below V30 or so)
- absolutely know they will need to guess at 4 or more questions somewhere in Verbal to finish on time
- absolutely know that they spend appreciably more time on average on RC questions than on SC and CR questions

If someone spends, say, 3 minutes on average per RC question (including time to read the passage), but only 1.5 minutes per SC and CR question, that person can either guess at one RC passage (let's assume four questions) and answer everything else on the test, or can answer every RC question and guess at seven CR/SC questions. The opportunity to answer three additional questions is almost always going to be worth the risk you take by guessing at random (potentially easy) RC questions.

But I'd also advise anyone considering doing this, or considering any new pacing strategy, to do empirical tests. You can test a new strategy out using the official practice tests, to see what effect that strategy has on your score. If the strategy leads to a lower score than before, abandon it, and if it leads to consistently higher scores, keep using it. Optimal strategy depends on your personal skills, so there is no one strategy that is advisable for everyone.

Modern probability theory, such as Bayesian Probability Theory, estimates likelihoods of future events based on a selection of past events. Inferences continue to updated as new likelihoods are calculated.

The more popular statistical estimation, used in significance testing, is not based on construction of likelihood estimates. It is based on the concept of comparing an outcome event to some arbitrary null event and then testing significance levels based on arbitrary cut-offs. Both methods have their shortcomings, and this more popular method faces much criticism. The Bayesian thinking is an improvement on significance testing, but it too faces criticism, a primnary one being that it relies on a past updated estimation that in turn is based on another previous estimation.

You cannot construct probability distributions without data.

A question such as the following is a badly formed question:

"I run out of time at the end. I will get one or more RCs, some SC and CR. Should I just skip all the RC and answer the SC and CR?"

Given the probabililistically adaptive nature of the test, this question is poorly posed and has no good answers.

What a testtaker will see as their next question and its associated difficulty level is directly correlated to what they did in the past, namely, how they answered questions in the past. Not all SC/CR/RC will be the same (different difficulty levels), etc. There is no good answer such as "just skip the RC and try your best to answer the SC and CR. This is as bad an answer as the original question.

Now, if we are considering absolute worst-case scenarios where we know that the test will heavily penalize testtaker who does not answer all the questions. Yes, in this case, adopt "strategic guessing" or completely "skipping" (i.e., simply randomly marking answers because there is no real way to pass over a question without choosing an option) can be employed. But this is an obvious point that maximizes only one outcome: How should I finish the test without being penalized for not finishing?

Due to the probablistic nature of the adaptive test, I continue to be reluctant to provide generic advice on "skipping" questions. Those who deal with probabilities will know that one should reserve making concrete statements in the face of uncertain events.

I would like to see actual test evidence regarding the outcomes of events related to "skipping" as well as data on alternative events that would have transpired when such "skipping" techniques were not employed.

There are 3-4 questions per RC passage, it takes about 2-3 minutes to read the passage in depth, and 1-1.5min per each question, so skipping one RC question wouldn't save you 4-5 minutes. I recommend that you make educated guesses rather than skipping questions at random since it would be near impossible for you to determine the difficulty level of the question until you have actually read it and tried it. Getting an easy question incorrect will hurt your score far more than getting a difficult one incorrect. Also, be sure to select an answer and submit when skipping. Leaving a question completely blank will hurt your score more than getting it incorrect.