Perfect score + GPA applicants are actually pretty rare even at Harvard

Card references SAT scores out of 2400 several times in the lawsuit. For example, he mentions that “Harvard’s class of 2019 had mean and median SAT scores of 2241 and 2270.” I’d expect that Card has access to all the SAT scores for the class of 2019 and is computing some basic stats on them such as mean, median, and number scoring 2400; rather than Harvard making a special point of highlighting perfect scoring applicants.

A quote form the Harvard Statement of Material Facts is below. It states that 2+ academic rating includes some applicants with “perfect” grades and testing. And it implies being one of the rare few that get a 1 academic rating (fewer than 0.5% of applicants in lawsuit sample) involves criteria beyond just having perfect stats.

"An applicant receiving a “2+” academic rating is typically an applicant with perfect, or near-perfect, grades and testing, but no evidence of substantial scholarship or academic creativity.

In many circumstances, an applicant receiving a “1” academic rating has submitted academic work of some kind that is reviewed by a faculty member. "

I agree that the higher admit rate primarily relates to scores being correlated with other parts of the application rather than score itself. However, I doubt that 2400 scores typically have “much higher average intelligence” than the numerous other Harvard applicants who score a few points shy of 2400. Instead I’d expect the bulk of the perfect scores are persons who took the test several times, until they managed a 800 on each section at least once, which superscored to 2400 Superscoring would explain why the number of 2400’s in Harvard’s applicant pool is so large in relation to the total number scoring 2400 as listed by CollegeBoard.

The Arcidiacono regression coefficients suggest the primary driver in the academic component of admission is academic rating, rather than test scores. After controlling for academic rating, the additional contribution of AI was quite small. The small component that does remain may largely relate to +/- academic ratings distinction, which were not controlled for.

Of course academic stats are an important component of academic rating. The percentages in different academic deciles who received a 2 (or better) academic rating is below, as listed by Arcidiacono. There is clearly a correlation, but it’s also clearly not a mechanical function of an AI threshold.

Top Decile – 98%
2nd Decile – 94%
3rd Decile – 84%
4th Decile – 70%
5th Decile – 51%
6th Decile – 26%
7th Decile – 8%
8th Decile – 1%

Academic rating explained 9% of variance in admissions decisions (Card model), and academic index explained ~half of variance in academic rating. Harvard OIR’s model was able to explain a larger portion of variance in academic rating through grades and scores than Arcidiacono was able to in his full model that including several functions of academic index, gender, race, concentration, hook status and many other factors. Harvard’s OIR model’s regression separated grades and scores, which allows a greater relative weight onconverted GPA (or whatever GPA is listed in file) and lesser weight on scores; while Arcidiacono’s model forced the AI score heavy weighting (2/3 of AI is scores), which may suggest Harvard is internally places a greater emphasis on grades.

@Data10 -

I don’t think we disagree about the role of perfect scores in Academic 1s. I thought I was pretty clear that perfect scores were neither necessary nor sufficient to garner an Academic 1, but that (I speculated) they may be a component. As your quote shows, “in many circumstances” Academic 1s will have had faculty review, implying clearly some will not have. Just a speculation on my part, and it’s a small speculation at that. On thinking some more, perhaps those 361 Harvard perfect SAT scores do in fact represent superscores, implying that Harvard attracted somewhat fewer single sitting perfect 1600s than the 361 out of 550ish possible would imply.

About “much higher average intelligence” in the perfect scorer group, well, we can quibble, but surely you recognize that a small portion of the population is “beyond” the test. That’s why for most demographics, you get the fat tail at the right limit and truncation of the curve. As I said, there are kids in the 1600 group who would have scored 1800+ if the test were made harder. No one in the 1550 group can test up there on that harder test, by definition. All we are trying to do is estimate average intelligence from scores, and when the group you are estimating is at the right limit, surely you can see that the “true” distance between the 1550 group and the 1600 (and beyond) group is quite a lot larger than 50 points.

Harvard seems to reflect this understanding, as that mere 50 point difference apparently about doubles the unhooked white and Asian group admit rate to around 30% based on 10 years of data. Add the condition of perfect GPA, and I bet the admit rates are about 40-50%, almost making Harvard - at least to my mind - a “match” school for true 1600/4.0 candidates. Pretty powerful (if I am correct) for otherwise unhooked kids. I cited my data sources in my first post.

Regression coefficients and model specifications are fun, but I think the real juice in these documents is the raw data to the extent they are provided or can be inferred.

The SAT is composed of quick and simple multiple choice questions, emphasizing basic 9th/10th grade knowledge, rather than complex thought. You need to be consistent without careless error and fast to get a perfect score ( without superscore). Some persons who are “beyond” the test may get a perfect score, and others will not get a perfect score. In the most recent SAT, two careless errors in the math section would drop your score down to 1550. It doesn’t take many careless errors to get a 1550. It’s not a good intelligence test and certainly not a good test of complex thought/logic. The SAT questions probably do not resemble the questions on any Harvard exam, given in any class.

You are comparing a 50 point difference in total score to the change in admit rate for a 50 point difference per section. These are not equivalent. Note that my earlier post mentioned Harvard’s class of 2019 had a median SAT score of 2270, so 50 points lower per section means scores below the median scores of the entering class. Grades and course rigor are correlated with test scores, which influences admissions. However, there are also various other contributing factors besides academics. If I guess that 2400 is similar to academic decile ~10 and 2250 decile to ~7, then the lawsuit lists the following percentage earning 2+'s in various sections is below. Those are some substantial differences, particularly for what others say about the candidate in the counselor and teacher LORs, and alumni interview. Even if Harvard was test blind and did not consider scores in the admission process, large difference in admit rates between 2250 and 2400 applicants would be expected.

Typical AI decile for 2400 – 98% Academic, 42% Counselor, 46% LORs, 64% Interview, 36% EC

Typical AI decile for 2250 – 71% Academic, 29% Counselor, 31% LORs, 44% Interview, 28% EC

@Data10

Actually I am not. At least I don’t think so. See the graph on page 7 (8 in the document as filed) here: http://samv91khoyt2i553a2t1s05i-wpengine.netdna-ssl.com/wp-content/uploads/2018/06/Doc-421-145-Admissions-Part-II-Report.pdf

If you look at the “75” bar, that is equivalent to 50 points less per section than perfect. Just eyeballing it looks like whites are admitted at ~11% rate, and Asians at ~6%, for a guesstimated blended rate around 9%.

Now look at the “80” bar, which is perfect per section, and you observe white admit rates ~33%, and Asian ~26%, for a blended rate north of 30%.

I do keep mentioning the 1600 scale though, and of course more properly these are 2400 scale data. So, tell me if you think I am misreading the graph, but it looks like 2250 = 9% blended white and Asian admit rate and 2400 = 30+% blended rate, more than triple.

At the “77.5” bar (approximate), which would be 2325 on the old scale and 1550 on the new, the white admit rate is ~16% and the Asian ~10%, guesstimated at 15% blended. That’s half the rate of the perfect scorers for those groups. And it’s just a 50 point difference. Of course, this differential is localized and not applicable to any other part of the curve.

I admit it is all just eyeballed. The Card report is interesting, but he sort of lost me with the model’s incorporation of the Personal Rating (which is clearly manipulated directly - Harvard was really pretty ham-handed here, and let’s not talk about the Overall Rating… looks like someone in the admissions office learned how to use Goalseek…).

I do recommend some reading in the intelligence measurement field if you are interested. PM if you want some recommendations.

I agree. I apologize for the error. This is the type of careless error I was referencing, with the SAT level math (and graph interpretation) questions.

Nevertheless, my point remains about a variety of other sections being correlated with higher scores, such that a large difference in admit rate between score ranges is expected, even if scores were not considered in admissions decisons. My earlier post mentioned that LORs and interview seemed particularly correlated with high scores, with relatively large difference in percentage receiving high 2+ ratings in the different academic deciles (which has 2/3 weighting for scores). A similar pattern occurs for smaller differences in academic deciles corresponding to the discussed score difference. For example, 46% of top academic decile receive a 2+ LOR, 41% of 2nd, decile, 36% of 3rd decile. .

Getting 2+ in these LOR and interview sections, with the larger correlation, seems particularly important in explaining admission decisions in Card’s model. As summarized below, if academic rating is not considered in Card’s model, it can still explain the majority of variance in admission decisions. However, if LOR and interview are not considered, then the model does not even explain 1/3 of variance in admission decisions.

Card’s Full Model: – Explains 64% of variance in decisions
Card’s’ Model, without Academic Rating: – Can explain 53% of variance in decisions
Card’s Model, without LORs and Interview: – Can explain 32% of variance in decisions

In the Harvard OIR graph, there is also is a similar rate of decrease in admit rate for other score ranges, rather than something really unique about getting a 2400 vs getting a near 2400. For example, if scores decrease from 775x3 = 2325 to 750x3 = 2250, then it appears that admit rate drops by almost half in the graph. If score decreases from 2250 to 725x3 = 2175, again admit rate appears to drop by almost half. There is a notable drop in admit rate for each score range, yet the Arcidiacono regression coefficients for AI remain low, again suggesting that among the ~42% of applicants with similar 2 academic ratings (ignoring +/-), the driving force in admissions decisions is primarily the other sections of the application that are correlated with varying score, rather than the score itself.

@Data10

Sure, of course, and I couldn’t agree with you more. Higher scores are going to be positively correlated with all the other things we care about, such as LoR, Interview, EC, etc. That’s because higher scores are positively correlated with intelligence, which itself is correlated with all those other aspects of a person’s application. It’s a tough pill to swallow, because there is overwhelming evidence that intelligence - or at least the time path of the growth of reasoning ability - is basically fixed very early in life, is largely under genetic control, and appears impossible to alter in any durable way. Nevertheless, that’s the way it is.

I agree with your observations about the SAT not being a very good individual measure of intelligence - in fact, no group administered task is going to be very good I’m afraid as regards the individual. Yet, in aggregate, we observe largely the same mean differences by sex and race, as well as the same variances by group, in SAT scores as with more specialized IQ scores. In fact, we could do something very simple, like recording the speed with which a subject can hit a button in response to a green light (a reaction time task); how quickly a subject can apprehend a word fleetingly flashed on a screen (e.g., the minimum time in milliseconds a subject can still “see” the word); and how long a string of numbers a subject can recall and repeat backwards correctly (reverse digit test); and we will see again those same mean differences and variances, for the most part. This all strongly suggests - really since Spearman and the early work on the g factor 100 years ago now - that all these tasks are measuring the same underlying process within the brain.

So, that the SAT is not particularly complex is true, but largely immaterial. Smarter people in general will do better, although of course there is always going to be individual variance as a function of measurement error. Speed, especially on the math sections, is not really an issue for the top of the distribution. I’ve worked a little with profoundly gifted kids, for instance, and some can do things like SATM in less than half, sometimes even a quarter, of the allotted time, well before 9th grade. The problem is to get them to check their work - conscientiousness is not particularly correlated with intelligence and I could tell you some very funny stories about that! (There are also reliable differences - mostly by sex - in certain conscientiousness measures, but that’s a whole different issue…)

MODERATOR’S NOTE: OK, please move on. This is turning into a debate about methodology, and CC is not a debate site, per the Terms of Service.

^ Fair enough, how about the question that perhaps some of the data-driven posters can look into: are admit rates for unhooked 1600/4.0 applicants to Harvard in excess of 40%? It appears to me that they must be, or at least were over the past decade or so, but I’d love input on this.

Let’s keep it simple, and forget about all the other holistic stuff and get to the simple correlation between perfect score + GPA and admit decision.

Maybe. Maybe not. At the end of the day, who cares? It is what it is. This is simple a party game along the lines of What Are My Chances?

To paraphrase Bette Davis in What Ever Happened to Baby Jane?: But you can’t, Blanche. You can’t forget all about the other holistic stuff.

Well, Harvard certainly appears to care. After all, it went to great lengths with the Obama DOJ and in the courts to try to shield all these data.

Yes, because Harvard is the only college in the US that is not fully transparent with its admissions process. 8-|

Unless anyone here is a past or present unhooked 1600/4.0 applicant (which I doubt), I’m missing why John/Jane Q. Public cares. But whatever, I guess.

Looking at the scattergrams in Naviance for our school, I see 30% rate of acceptance to Harvard for kids with perfect SAT (4 out of 12) for the past few years. However, I happen to know the guy who got in last year, and he’s an ISEF and Regeneron winner, and the year before this one of two acceptances was an athletic recruit and another was I’m not sure who exactly but likely also a nationally-ranked academic superstar (there were a few). Granted this is a small sample, but based on this, I don’t think my kid (1600/4.0, state-level academic achievements) should even bother applying.

“are admit rates for unhooked 1600/4.0 applicants to Harvard in excess of 40%?”

@SatchelSF My guess is that it is below 40%, from my observation of kids who applied to Harvard, from CC’s result pages and from some back of envelope calculations.
BTW, for class 2019/application year 2014, the ACT was probably taken in 2013 in which there were 1162 ACT 36C. So, roughly 55% of perfect ACT/SAT scorers seem to put an application in for Harvard. In 2018, number of ACT 36C has gone up to 2760 (talking about score inflation and growing ACT popularity). I cannot find the new SAT 1600 numbers this year, but my guess is around 650. Assuming most kids do not take another format (let alone try a perfect score) if they score perfect one already, you have about 3400 kids a year with a SAT/ACT perfect score. That will bring your “unhooked 1600/4.0 applicants” close to 2500. And if 700 admission offers go to the unhooked and more than a third of them 250 are 1600/36/4.0 scorers, you would be looking at more like a 10% acceptance rate, still far better than the 2% for the average unhooked.

At the risk of stating the obvious, that data pool is too small from which to extrapolate, and while a specific HS may be extremely diligent about updating Naviance, this is not always the case.

My guess is that 30% acceptance rate is closer to the actual number, but guessing is all any of us can do.

@jzducol -

Very helpful, but I am a little troubled with your 10% estimate, unless admissions has very recently gotten much tougher (and/or grade and score inflation has increased even more than it already has). Which could be! We do have the data from the source I linked ostensibly showing that for classes 2007-2016, perfect SAT scorers (2400 scale) had a blended 30% admit rate for unhooked whites and Asians, and that is before applying the restriction of 4.0 GPA (of course, correlation is likely to be fairly high). That’s from where I inferred a potential 40% figure. (I think you are assuming perfect correlation between 4.0 and 1600 SAT/36C ACT in your 2500 applicant estimate?) I do not know if those data in the doc I linked include ACT perfect scorers as well, but I doubt it as there are significantly more ACT combinations that get you to 36C while SAT 1600 (2400) has only 1, and 36C seems to map to a range of 1570-1600 currently.

We do have the admit rate by academic decile information in the current documents (6 years of data), and there would appear to be higher than 10% admit rates for unhooked applicants in the highest decile, of whom there are approximately 2300 per year: ~15% for whites and ~13% for Asians. (Tables 5.1R and 5.2R in the Arcidiacono rebuttal report.) This is a mechanical index derived primarily from scores (2/3) and GPA (1/3) - I think - but in any event I would have to imagine that the entire top decile is not populated by 1600/4.0 candidates! Perhaps there is a higher admit rate for 1600/4.0 versus 36C/4.0, but in any event it would look like the 15% and 13% figures would have to be floors. And again, we have that older data 2007-2016 showing 30%+ just for perfect scorers (with no GPA restriction, and with a question as to whether it includes ACT 36C).

Anyway, don’t you just wish the raw data would somehow get out so we could crowdsource this!

Naturally. My point was rather that most admitted kids with perfect scores may also be hooked in some way.

I haven’t seen Harvard publish this, besides the lawsuit numbers. However, some other selective college do publish more detail, which are in the ballpark of the Harvard lawsuit numbers. Perfect SAT + 4.0 is probably not that different from perfect score stats, with the vast majority of perfect SAT scoring applicants having high GPAs. For example, in 2014 (near time of referenced Harvard class), Brown had a 23% admit for applicants with a perfect 36 score on the ACT, as listed at https://web.archive.org/web/20141002221234/https://www.brown.edu/admission/undergraduate/about/admission-facts . Stanford mentions a 31% acceptance rate for applicants who scored a perfect 2400 in ~2008-12 (see https://alumni.stanford.edu/get/page/magazine/article/?article_id=66225) . The admit rate is lower today, and 1600s are far less rare than 2400s, so I’d expect both numbers would be notably lower today. Parchment (includes self reported) showed a similar rate of admission among perfect SAT + perfect GPA applicants for most selective, holistic admission schools. Even Caltech, which has a reputation for being less holistic, rejected the majority of perfect SAT+GPA applicants.

As emphasized in my earlier post, I doubt that the near perfect score is the driving force behind the increased admission rate, so while it is an interesting statistic; I don’t think it is especially meaningful to applicants at such colleges. That is, it is more important to be an applicant who has really incredible LORs, ECs, interview, and presents the full package that it is to be an applicant who scores a 1600 instead of 1550. The 4.0 + 1600 score applicants who do not present this are likely to be rejected at a wide variety of selective, holistic colleges; rather than just random decisions similar to the overall admit rate for that score. For example, a while back there was a thread about a well verified 2400+4.0 applicant who was rejected by the full HYPSMC group. Another poster claimed to be rejected by all Ivies with perfect stats, although as I recall he later admitted that he was accepted to “only Cornell”. MIT explains this more eloquently than I can on their website:

I always wonder if there is any difference between an SAT1600 vs ACT36 in the eyes of AOs. Do they know 1600 is harder to get than a 36? If most kids submit one or the other then the number of perfect scores (including both SAT/ACT) has tripled in the last ten years. If the percentage of them applying to Harvard stays the same then perfect score acceptance rate could be a third of what it was ten years ago. And the problem is that there is no sign this increasing trend of perfect scorers is going to plateau anytime soon.

AO’s can see ACT subscores, so they know whether an applicant’s 36 composite is the more rare 36 on all 4 ACT sections, or the more common mix of 36s on some ACT sections and 34-35 on others. However, I think the more important question is would they care about the difference between a perfect subscore 36 vs non-perfect subscore 36 vs 1600 SAT? I’d expect that this level of score difference has very little impact on admissions decisions, yet there may be more significant differences in admission rates between these groups due to correlations, particularly correlations with who chooses to take and submit SAT vs ACT.

You would think so, but it’s repeated endlessly on CC Harvard could fill their class with perfect stats.