normalizing hsk scores

[kudra]

There are lots of threads here asking about how to convert a %-right score into a normalized score (as reported by HSK). The normalized scores are the scores with known cutoffs for the various grades. Most threads conclude that it is a black box inaccesssible to us mere mortals.

Here is what I propose -- surely lots of people take timed practice tests for which they know their raw scores, then shortly after take the real hsk, and find out their normalized score. So we just need to tablulate all that, and draw some graphs – linking the practice test raw scores to the actual normalized scores reported by the HSK. We could even do some statistics and curve fitting if we wanted, but the tables and graphs would probably give people a good idea how their raw scores on practice test will translate into HSK normalized scores. I think this should be fairly stable since I think the scores are related to a standard sample group, not each years percentiles. As mentioned here, "These scores show the relative position of the candidate in a standard HSK sample group."

For example I know what % I got right in the practice test on each section. And I know the normalized scores as reported by HSK. (As a bonus for modeling I know approximately how many I filled in at random on the actual test because I ran out of time) If I had enough of this kind of data for takers of the HSK Elem/Inter I could figure out the approximate cuttoffs.

So if you want to contribute, cut out the form below. Keep track of how you do on practice tests. When you find out your HSK normalized score fill in the form with your numbers (Use a new form for each practice test even though the same HSK). pm me the results. I will tabulate them, draw some graphs, and tables, probably an excel spreadsheet. I will strip off personal identifying info, i.e. I will not match your data I report with your member name. This way you can see what you can expect, and see how many more right you need on practice tests to get to your goals.

Sample Form (Copy this, and replace with your numbers -- these are number right. Add any execptional deviations from real test conditions as I did in the Note section.)

----------------

Sample Test -- #1 from An Intensive Course of HSK (Elementary and Intermediate)

Practice Test

Listening

1(1-15) -- 8

2(16-35) -- 10

3(36-50) -- 6

Sentence Structure

1(51-60) -- 6

2(61-80) -- 15

Reading

1(81-100) -- 9

2(101-130) -- 12

Cloze

1(131-154) -- 15

2(155-170) – 0

Notes on practice test: did not time, i.e. used as much time as wanted, but no help.

Actual

Random fill at end, and left blank

Listening – 5

Sentence Structure – 7

Reading – 15

Cloze (ABCD) -- 6

Close character fill in left blank -- 15

Actual reported score

Listening – 48

Sentence Structure – 47

Reading – 45

Cloze -- 32

Total -- 173

Elem/C

[uchihak]

Hi Kudra,

Thanks for making this post.

I've always wondered too about how to estimate your final HSK score. It will be nice if the mock tests can include a conversion table from the raw scores like the mock SATs.

[yonglin]

Those of you who are just getting your scores in now, might also like to have a read of this thread and help out in compiling data about HSK results to make it easier to correlate test scores to real scores.

imron, may I kindly ask what the purpose of collecting such data is?

BLCU has publihed endless numbers of hsk-related statistical tables, and I believe that everyone who ever aspired to take the HSK would have picked up at least one preparatory book which lists all grade boundaries.

For instance, I used the table in 速成强化教程 An Intensive Course of HSK (Elementary and Intermediate level). In this book, grade boundaries for all sections are stated in %. Sure these will tend to vary a little bit between different sessions, but the difference should be fairly small...?! I did one practise test in the book before I took the HSK (the exercises were so boring!!), and I got exactly the same scores on the real test as the practise test.

Did I miss something out?

[imron]

The purpose is to help people better correlate practice test scores to real test scores so that when they take a practice test they have a better idea of how that will map into a real HSK score. Often the score someone gets on a practice test is quite different from what they get on the real test.

[kudra]

yonglin -- i have that book. i thought the tables told how to go from reported score to percentile rank in the population of test takers. i dont know how to go from % correct answers, ie raw score to a 'reported score'. what table did you use? thanks(tapping on palm)

[yonglin]

i have that book. i thought the tables told how to go from reported score to percentile rank in the population of test takers. i dont know how to go from % correct answers, ie raw score to a 'reported score'. what table did you use? thanks(tapping on palm)

Hmm... when I wrote this post, I was under the impression that the scores in the upper table on p.4 corresponded to the % of questions you got right on each section (i.e., that they just made the questions in this way). However, it must be the case that they normalize these scores afterwards, since I personally got 63 (odd number) on my tingli for instance. My bet is that there is not much difference between the raw (% of questions right) and the normalized (percentile of the test-taking population) around level 5. Had I not cheated while I did the practise tests, my percentage scores would (roughly) have corresponded to the percentile rankings in that table (heheh~~~ quite complicated that sounded). For instance, say I got 31/50 questions right on my tingli (=62%), then I looked at the table and thought "Aye! That's a 5!" since the boundary for a 5 is 56-64, not paying attention to that these numbers actually represented the percentile rankings.

My bet is that they try making the test so that the percentile scores correspond to the percentage of questions you get right, but that some smaller adjustments are made once the scores are in to make the normal distribution more sound.

For instance, this would agree with the scores posted above

Listening

1(1-15) -- 8

2(16-35) -- 10

3(36-50) -- 6

24/50 = 48% (actual 48 )

Sentence Structure

1(51-60) -- 6

2(61-80) -- 15

21/30 = 70% (actual 47)

Reading

1(81-100) -- 9

2(101-130) -- 12

21/50 = 42% (acutal 45)

Cloze

1(131-154) -- 15

2(155-170) – 0

15/40 = 38% (actual 32)

The only score which is significantly off here is the yufa section. But I think this has more to do with the practise questions in this book being unreasonably easy (I remember that I actually got an 8 on this practise test, whilst my reported score was a 7).

Personally, I think that this person realistically should have expected a 3 (no offence), since 38% on the zonghe is dangerously close to the 底线 at 36 (i.e., getting one more question wrong would make it impossible to get more than a 3). As we all know: once one score drops below the band for a 3, you cannot get more than a 3 on the whole.

So my conclusion is that you can use the table on p.4 of this book to represent the % scores (at least if you're in the bottom-middle end of the distribution), although care should be taken with the yufa section in particular, which is likely to overestimate your abilities.

The reason that I'm still not posting my practise scores is because I never took a test seriously (without rewinding the tape), and I didn't save the notebook in which I wrote my answers for the practise tests (too much books to carry around my China travels already).

[Sarevok]

I kept records of quite a few of my practice tests scores, should I use the best ones for this? Also my case is quite different, because I did much better on the actual test in June than on any of my practice tests (阅读96, yay ), which is something that never happened to me before (I took the 初中等 exam 3 times in total, but for my first try I never got the results... not that it matters that much, because I probably couldn't get even 3级 at that time), the %right from the practice tests were roughly corresponding to the actual normalized scores before then...

But still, I was not good enough, as I got a total score of 336, while you need 337 for 8级 Actually there were 4 points missing, not just this one... 3 more in 听力 and one more in 综合 and I could have done it. I'll probably go for the 高等 next time, as it should be better to struggle for the lowest grade there than for the highest in 初中等, even though the test is much more difficult overall...

[kudra]

I kept records of quite a few of my practice tests scores, should I use the best one for this?

I would use the last one, or possibly all of them, but not the best one alone.

As for the raw % corresponding to the % of the population -- it never occurred to me to think of this. They don't mention this anywhere that I can see. I might come out in the analysis, which would suggest that they designed the questions thus. Still, it looks like we have only anecdotal evidence, n=2.

As for expecting a 3, I was quite happy with it. I had signed up for the beginning test, but was informed 2 weeks before that it would not be offered, so I opted to take the Elem/Inter test. I had not been reviewing writing characters by hand, and thought I'd concentrate on the other aspects with 2 weeks left instead of trying to cram writing characters. (no offense taken)

[Pravit]

Aren't they changing the test format next year entirely anyway?

If it helps, although I don't have the exact figures, the % of questions I got right on practice tests was almost the same as my actual score. For example, I was getting around 80 - 85% questions right in listening comprehension on practice tests, and I got a score of 86 for listening on the actual test. The only exception to this rule might be the fill-in-the-blank section, where I got an actual score that was a few(say around 5-10) points higher than my percentage on practice tests. I used the Intensive Course of HSK book to prepare. Due to my poor performance on practice fill-in-the-blank, I was expecting to get only a 7, but I managed to pull an 8