The 50 million dollar lie
garyrubinstein.teachforus.org
garyrubinstein.teachforus.org
But the first chart in the blog post, with the big mass of blue on it, is surely not the strong evidence of weak correlation, that the author is making it out to be?
There could still be a strong correlation in that data, even if it looks like a blob of blue - because there are so many data points on that chart, that we can no longer tell where the density of points is. Its possible that there is a dense line of points in there that still gives a strong correlation, that we just can't see.
A heatmap with some level of binning, or a hex bin chart, or 2d histogram, would surely be more appropriate in these circumstances? Something like this: http://matplotlib.org/examples/pylab_examples/hexbin_demo.ht...
Maybe along with some direct math analysis of the correlation. Am I missing something?
You seem to be confusing correlation and significance. Having lots of data points makes the correlation more significant, it does not make the correlation stronger.
No, that's not correct. You can't conclude anything from looking at a big blob because you don't know the density of points at different places in the blob. This is the point the guy you replied to was making.
As an extreme example, imagine a billion data points that fit perfectly on a straight line. Then superimpose a million data points randomly on top of it. What does it look like? A big blob. But almost every point is highly correlated with that straight line.
I agree that a scatter plot is not the best for showing data with that many data points, but frankly it's kind of irrelevant. The point the author was making was just that it isn't that hard to cook up some data that is not highly correlated, but will be if you bin and average it.
See the discussion here: http://news.ycombinator.com/item?id=4027337
I stand by the claim made in my blog post. Don't use scatterplots, use a density plot instead.
Incidentally, according to a comment the author made, the correlation is actually 0.3. That's far better than his graph suggests.
I might add that discussions like these, Illustrate a broader problem with the hacker news community. Rather than discussing the facts of the story, many of the posts are instead about the less relevant detail about how the author chose to present them, despite the fact the the authors point is still effectively made (and if the 0.3 were included there wouldn't have been any doubt)
BTW: tmoertel did some analysis on the data: http://news.ycombinator.com/item?id=5125851 and the correlation appears to be around 0.2, which is pretty weak.
It is. If you look at the density in the upper right of the cloud, it is clearly clustered around the x=y line. The correlation may not be very strong, but without extra information about how strong of an effect would constitute clear evidence in favor of including the value add metric, we can't say anything more about it.
> Do actual phenomena often yield "deceptive" blobby scatterplots with dense lines hidden inside them?
Absolutely. You're looking at a misapplication of a plotting technique - a scatterplot with substantial overplotting. The author may be convinced that there's no correlation there, but by presenting this plot they provide no evidence in support of their point.
Um, that's not much to go on. Can you give some examples?
But the relation, whatever it is describing, is pretty weak since the ellipse isn't very irregular (it's only "slightly" irregular). http://www.emeraldinsight.com/content_images/fig/08702801090... Think Kepler's laws.
Edit: The first post in the series shows a scatter of the direct (non-delta) scores, and it looks like San Fransisco in the morning: http://garyrubinstein.teachforus.org/2012/02/26/analyzing-re...
Not very clear what you are try to say:
(1) What "standard deviation"? Perhaps you mean the standard error of the correlation coefficient is large relative to the its estimated value. Given the large number of data points it is likely to be quite small. Another warning flag is the failure to report either the correlation coefficient or its standard error.
(2) What does "ellipse isn't irregular" mean? Given all the overlapping data points the shape of the plot is entirely driven outliers. To my eye this looks what you would get from plotting a bivariate normal distribution.
(3) Kepler? The linked graph? How is this helpful in understanding what's going on?
The first graph has a straightforward interpretation: you are making two imperfect measurements of each teacher's performance at two different times. There is noise in both measurements and teacher performance may have actually changed between measurements.
Typo: should have been "standard distribution"... Less colloquially: "a two-variable normal distribution."
"What does "ellipse isn't irregular" mean?"
Eccentricity. Sorry `fer the double negative.
"Kepler? The linked graph? How is this helpful in understanding what's going on?"
If you take the scattered data and draw an ellipse containing a fixed probability (presumably represented with some confidence, by the data shown), the ellipse would be more eccentric for the same enclosed probability, the higher the correlation. It's a single statistic and is analogous to Kepler's 2nd law ("a line joining a planet and the Sun sweeps out equal areas during equal intervals of time."), except with probability rather than area... (To contain the same probability, you have to adjust the ellipse's size, while holding its enclosed probability constant, and accounting for the new eccentricity, which isn't under our control. If the area is fixed and the distance changes, solve for speed... I think visually. Alternatively, think V=IR.; if V is constant and R is known, solve for I.). The link was just illustrative of any such scatter plot. It was the best I could find, but I agree, it could have been better.
I agree that the article is low on quantitative statistics though.
http://en.wikipedia.org/wiki/Multivariate_normal_distributio...
Interesting thing is the (positive) means and variance seem similar which is consistent with underlying random variable being the same.
Everyone wants to use statistics as a tool to bolster their point of view but don't want to bother to actually learn enough statistics to deal with the situation they are trying to analyze.
I will make a strong claim: If doctor effectiveness were this stable, and it were the only data I had available, I would choose a doctor on this metric, ceteris paribus.
I found a scatterplot by the author that uses smaller marks so it looks a lot less blob-ish and more heat-map-ish:
http://garyrubinstein.teachforus.org/2012/02/26/analyzing-re...
I'm trying to track down the original data to verify this myself, but a couple of quick points:
There definitely seems to be good evidence of a relationship. It might be "weak" but that's not the same thing as "ineffective". Considering the number of data points, it makes it unlikely that the trend occurred by chance. This plot also makes it clear why the averaging made it look so pretty: the average in just about any percentile group in 08-09 matched 09-10 regardless of how finely it's divided (to a point).
It's certainly the case that there's a lot of "noise". The question is how to interpret it. One might conclude that there is a lot of "measurement noise" -- that teacher evaluations are inherently messy and rather uninformative except as a larger trend. There seems to be reasonable support for this! Students aren't the same, even from year to year. Of course it could "signal noise" - maybe teachers themselves change from year to year. Perhaps one year you're more motivated, the next your not, or vice versa.
To understand the implications of this noise, we also need to ask what we're using it for. For instance, if this were a plot of the correlation between money donated to Watsi and quality-of-life, you'd probably think you were doing pretty well! You might point out that a particular quality-of-life metric is flawed or noisy, or that the quality-of-life depends on lots of things besides health, but likely you'd be happy that you're clearly causing some benefit.
So when we ask whether the data suggest the viability of value-added measurements in aggregate, to me that seems to be a yes. But if we ask whether the data suggest using these measurements to make a decision on a per-teacher basis, that depends on the context of the alternatives! If this is the only thing we have to go by, perhaps it's not so bad.
I would love to see how stable the other parts of the teacher evaluation fit in. If you combine multiple noisy signals, you often get a much better picture of what's going on! This accounts for only 1/3 of the teacher's evaluation. I would like to think that if you appeared not to be a value-added teacher, but your principals and coworkers spoke highly of you, that you'd still get that raise.
Also, I'm looking forward to multiple years of this data - if your score accounts for 35 percent of the variance from one year to the next, multiple years might look a lot better. Perhaps in deciding teacher tenure, scores over 5-10 years could be really useful.
http://www.chrisstucchio.com/blog/2012/dont_use_scatterplots...
He is certainly aware of my post - he even commented on it and I offered him some constructive suggestions (such as adding alpha if he wants to use scatterplots).
I'm beginning to think that maybe his goal is to make VAM look bad, rather than to actually clarify and explain the data.
(I also left a comment on his blog expressing my concerns with his plot. Strangely, I seem to be hellbanned on his blog - the post is invisible when I view the post in an incognito window.)
I grabbed the data set [1], loaded it into R, and tried to recreate the first graph in the original post. Observations:
1. The value-added scores are quantized so, as some people have suggested, it's hard to see the true density relationships in the original plot. There's a lot of overlapping data points.
2. I re-plotted the data using jittering to more accurately show the density distribution. I also added a GAM smoother to show the trend. Result: https://docs.google.com/file/d/0B2Tqdso8AsFNS3lWQTdPUGpHZlk/...
3. I also plotted the change in each teacher's value-added score from one year to the next. The resulting histogram: https://docs.google.com/file/d/0B2Tqdso8AsFNeUNiM3hHUGwzdlk/...
4. If we model the relationship as next_year_value_added = this_year_value_added + random_error, here's the fitted model:
Call:
lm(formula = va_0910 ~ va_0809, data = ds_cmp)
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.009157 0.002315 3.956 7.67e-05 ***
va_0809 0.820085 0.017646 46.475 < 2e-16 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Residual standard error: 0.2153 on 8691 degrees of freedom
Multiple R-squared: 0.1991, Adjusted R-squared: 0.199
F-statistic: 2160 on 1 and 8691 DF, p-value: < 2.2e-16
[1] http://www.ny1.com/content/top_stories/156599/now-available-...
You use subject/grade/teacher level, but you get a different answer if you use aggregated teacher level.
(I used unweighted average for each teacher/school/year.)
In his original scatter plot, because of the big ball in the middle it's easy to handwave and say, "look a random blob" -- but even a slightly longer look seems to indicate there is a positive correlation in the data. His averaging that he does at the end, IMO, makes it clear.
But anyway, each one is free to reach his own conclussions.
In other words: without the averaging, you are seeing the full implications of the data: slight positive correlation, but a lot of scatter--i.e., not much meat there. With the averaging, you are seeing only the positive correlation, with all other information filtered out--i.e., you are seeing only part of the picture, and it's the part that, surprise, surprise, makes the person who ordered the data collected look good.
Sure, you could add a volatility measure as well, but the trend is clear -- you'd rather have your child by a teacher with a high teacher score than a low one. Or are you saying that over 12 years of schooling you don't think you'd have any preference having your child taught by teachers with a score in the lowest 5%ile versus a teacher with a predicted score in the top 5%ile? Of course there's some volatility, but over time, it's seems pretty obvious to me.
Nobody is arguing that a small correlation doesn't exist. They are arguing that a small correlation isn't very meaningful.
Or are you saying that over 12 years of schooling you don't think you'd have any preference having your child taught by teachers with a score in the lowest 5%ile versus a teacher with a predicted score in the top 5%ile?
If you base those percentiles on a metric that has a very low correlation with performance, then there won't be much of a difference in outcome.
Or are you saying that you want to pick who to fire and who to promote based on a metric that we objectively know has only a very weak correlation with reality?
What alternatives do we have? Do you know of a more accurate metric?
The status quo in teaching is to evaluate teachers based on qualitative observations from management, same as it is in most other industries. As the same blogger points out in another post http://garyrubinstein.teachforus.org/2013/01/13/50-million-3... "Conspicuously missing from the various weighting schemes they compare is one with 100% classroom observations. As this is what many districts currently do and since this report is supposed to guide those who are designing new systems, wouldn’t it be scientifically necessary to include the existing system as the ‘control’ group? As implementing a change is a costly and difficult process, shouldn’t we know what we could expect to gain over the already existing system?"
The status quo also has the advantage of not being ridiculously expensive and complicated. Ask Bill Gates to release the data if you're curious about its performance, but given what I've seen, I'd be shocked if it was terribly different than these complicated new metrics.
So replace a metric with 0.1 correlation with one having 0.0 correlation?
If you are advocating that we should use the school principal's opinion rather than VAM, why do you believe opinion is superior? Do you have evidence that principal's opinion has a higher correlation with student outcomes than VAM?
Two problems with this:
1. The other metric almost certainly doesn't have an 0.0 correlation, but we don't know that for sure since the data wasn't released for whatever reason.
2. The metric with 0.1 correlation (or whatever the number is)... keep in mind the context. What is the correlation with? Test scores, something that can be and is often rigged, and is at best tangentially related to the ultimate goal which is inherently qualitative in nature. I think the comparison to programming is instructive. If you had 1/3 of your salary determined by the LOC you wrote or the number of bugs you closed, you would game the system and maximize your salary, even if you weren't providing value by doing that and even if those naive metrics did very weakly correlate with performance in some big study prior to implementing the monetary incentive. What would be gained in that scenario?
If you are advocating that we should use the school principal's opinion rather than VAM, why do you believe opinion is superior? Do you have evidence that principal's opinion has a higher correlation with student outcomes than VAM?
One of the points made in the blog I linked to is that it's unfortunate that they didn't release that data. And as I said above, it's not straightforward to assume that these weak correlations are actually meaningful in practice.
The author is arguing that because the measurement is noisy, we should ignore it. This is silly - it just means a single class-year's point estimate is noisy, and multiple class-years must be combined to form an accurate estimate.
If you had 1/3 of your salary determined by the LOC you wrote or the number of bugs you closed, you would game the system and maximize your salary...
Indeed - if 1/3 of my salary was determined by the number of kids in my class who can read better than when they are predicted to read at this age, I'd definitely try to make sure that reading skills improved.
Conversely, if a stupid metric such as student/principal opinion were used, I'd focus on jokes and friendliness over education.
(Well, actually I didn't back when I taught. But that sure didn't help my student evaluations...)
Not quite..
Indeed - if 1/3 of my salary was determined by the number of kids in my class who can read better than when they are predicted to read at this age, I'd definitely try to make sure that reading skills improved.
Keep in mind that "score on a reading test" and "reading skills" are not the same things. For instance, one would be significantly improved by spending valuable class time teaching students tricks about how to succeed on the standardized reading test. It's similar to the incentives you'd get by determining salary by LOC - even if LOC correlates with performance, basing pay on it encourages all kinds of nonsense that doesn't actually benefit anyone.
So in that context, we have a metric that very loosely correlates with this clearly flawed marker of success. And we're supposed to spend ungodly sums of money implementing this strategy? Come on.
Conversely, if a stupid metric such as student/principal opinion were used, I'd focus on jokes and friendliness over education.
Well, they do propose to use student and principal evaluations..
True. Score on a reading test might be lower than actual reading skills due to a lack of test taking skills.
Then again, I can't think of any measurable quantity more highly correlated with reading skills than scoring high on a reading test. Do you have evidence that "principal's opinion of teacher" is better?
...we have a metric that very loosely correlates...
You are confused. GR shows that year on year correlations are weak (0.3), not that correlation between reading test and reading skills are weak.
You are correct that a certain nonzero amount of teaching test skills will be effective at improving scores. I see no reason to believe this amount will be large - do you?
Again, standardized tests are not the only gameable metric. Principal/student opinion is as well. Why do you believe they are less gameable and more accurate than directly measuring student ability?
It's like.. how much money and effort do you think is worthwhile to spend on something with such uncertain prospects? There are tons of great grant proposals to funding agencies which are rejected. Maybe we would be better off funding some of those. Opportunity costs exist, and it seems like undue sums of money are being wasted on this stuff.
It's not that conspicuously absent as they don't use 100% weighting for anything. That said, they do attempt to maximize ability to predict state test scores, and in that optimization classroom observation plays the smallest role of the three metrics (2-9%).
Given they appear to have to done the analysis across all weightings, I seriously doubt they'd hide this data.
I do find it odd that teachers would rather be judged based on one or two people observing their classroom and ignoring actual student output, rather than raw numbers.
As a developer I'd much rather be evaluated on some metric I could optimize for (I think some metric measuring feature value/bugs/fix rate/etc...), rather than just my manager watching me code/debug a couple of times per month.
As a developer I'd much rather be evaluated on some metric I could optimize for (I think some metric measuring feature value/bugs/fix rate/etc...), rather than just my manager watching me code/debug a couple of times per month.
Okay, I really don't understand this mindset. You'd really prefer it is, say, 1/3 of your salary was determined by the number of lines of code you wrote? Or the number of bugs you closed? Or some equally inane metric that only loosely correlates with your actual performance and can be easily gamed?
What am I missing?
I was a quant trader for a while. A considerable chunk of my salary was determined by profit, hardly unreasonable.
Measuring developers in other areas is difficult because of heterogeneous goals - last year I built a search engine, this year I'm statistically tracking user behavior. Hard to compare one to the other.
Education does not suffer this problem - last year a teacher taught 30 kids to read. This year she taught 28 kids to read. The goal is always maximizing the fraction of kids who can read.
If it's only loosely correlated, measuring value-added then punishing / rewarding teachers based it won't be very accurate. It's like lines of code - programmer talent is probably weakly correlated with lots of lines of code, but you don't want to reward or punish them based on LoC because it will lead to pathological behavior.
You should look at other things to alter (which are less easy to game once you start trying to control them), like class size, course materials, assessment style, how the teacher actually teaches, etc.
Giving employee evaluations that are 5% performance and 95% random chance is a recipe for an organization that no capable intelligent person would consider for a career.
Not necessarily. The very weak correlation means that the difference in this score only accounts for a very small amount of the difference in outcome; and even that claim assumes that the observed correlation is evidence of a causal link, which is not necessarily the case. You can't usefully conclude anything from this data except that more investigation is needed.
And on a separate note - there's a lot of anti-Gates sentiment flying around lately. Whether his methods are the best or not, he's using his own money. I find it really difficult to give any criticism directed at Bill Gates' foundation any credence. And that not knowing much at all about what he does.
Maybe people would like Bill more if he did a Larry Ellison and bought a yacht instead.
This looks to me like someone trying to make the report look good so that Bill Gates will get positive press, think this was worthwhile, and fund them again.
Also, that's probably why the headline calls it a "lie". This is something of a pattern.
The blog author saying there is no clear trend in his random blob of points is misleading. The trend is there, it is just difficult to tell the density of points in his scatter plot because of the marker size and plot size.
Except for that in this case the noise is more important than the trend. Think about it, if you're firing or sanctioning perhaps 30%+ of teachers each year for no reason, then only complete morons would go into teaching.
It's the same as airport security, where a .1% false positive rate is unacceptable, whereas a 75% false negative rate is just fine.
So I'm thinking they're not too worried about this.
This is comparing apples and oranges.
I know that hackers, in particular, have real problems with "Argument from Authority", but stats is one place where it's really, really easy to go wrong, so knowing how much formal training someone has can be an important indicator.
Anyway, here's the data for anyone interested: http://www.ny1.com/content/top_stories/156599/now-available-...
As for my formal stats training: None beyond High School. That said, I've worked with and interviewed many with far more training. One thing, I learned is that most people don't have great intuition for statistical problems and it's not terribly highly correlated with years of schooling. I've met PhD's in economics and statistics who have made basic conceptual errors as well as those with less training who were more reliable.
So while training may be correlated with accuracy, you should still demand a well-reasoned argument and think critically about it.
If you have no such evidence then the onus will be on you to make your argument more complete, more coherent, and more comprehensive. If you have evidence (note: evidence, not proof) then you can be a little less rigorous in what you say, and rely on people giving you the benefit of the doubt while they work through the argument.
What I see a lot of is long, apparently good arguments, that then turn out not to be as complete or coherent. they sometimes just don't hang together.
Significant amounts of formal study in a subject is evidence that someone might just have a better understanding. After spending a lot of time on the internet I'm tired of having to wade through every single argument in detail looking for all the possible chinks.
Maybe that's just impossible. Maybe every person has to redo every analysis for every argument. Seems like a complete waste of almost everyone's time. What about "Don't Repeat Yourself" or "Don't re-invent the wheel." I guess we are doomed to reinvent the wheel in every single discussion.
I would argue that the whole point of HN is to think through arguments in other domains and build intuition by reasoning through problems and arguments. Otherwise, what's the point? No one is going to arrive at this thread and scan the top-rated comments for the solution to his school district's problems.
In cases where actual decisions are being made where the analyses are much more thorough and fully validating much more expensive, other techniques are available. First, one generally builds an awareness of the strengths of each team member which suggests where errors may be more likely. Additionally, one can check a random set of the most likely problem areas. Perhaps, most importantly, while everyone won't re-do every analysis, it's highly unlikely that an any important analysis will only be done once. So one can expect that the high-level results are generally consistent.
I have not tried to look at the numbers. Here is my opinion.
My experience of school was that a small minority of teachers are truly excellent, and a small minority were horrible. The truly excellent ones are not distinguished so much by what happened in their class as by what happened in the following classes, and what classes they left people excited about taking. The terrible ones, by contrast, showed up as poor performance all around.
The described data set has a clear correlation. Whether or not the correlation is meaningful in individual cases is much less clear. My initial approach would be to try to fit a hierarchical Bayesian model to a data set, then use the resulting model to come back with predictions about individual teachers. Teachers whom, after several measurements, are overwhelmingly identified as terrible should be removed. If you find a population of teachers who show up as superstars, they should be subjected to further study to see if we can predict the quality of incoming teachers, and to see if we can learn lessons from them that improve other teachers.
However this is a well-studied problem. I'm sure someone has tried something like this. I am sure that there are a lot of vested interests. I have not attempted to evaluate work in this area, and I have no opinion on how good it is.
That said, my training -- read Friedman's intro text in high school (just freetime reading on my own -- so not formal per se). Two quarters of stats as an undergrad. Two quarters in grad school. All of which at least 15 years ago -- so mostly forgotten anyways. :-)
http://ed2worlds.blogspot.com/2013/01/gates-foundation-waste...
Key quote:
"...it has been known for more than 60 years now that correlating averages of groups grossly overstates the strength of the relationship between two variables"
Apparently, this was also used in the past to show high correlation between African-Americans and illiteracy.
The second issue is that the author does not offer a better solution to the problem. Some information is almost always better than no informantion in decision making. Statistics in human-based samples are always tricky, and its tough to get strong correlations easily. As an aside, this is part of why clinical trials are so tricky and expensive.
(I am not a stats professional, but am a biomedical research scientist with some training in human-population based stats).
There seem to be a lot of comments about the charts on this page. Arguably basic statistics should be a part of every education - is that happening in the USA?
http://garyrubinstein.teachforus.org/2012/03/06/analyzing-re...
Some qualitative comments (from the UK)
An experienced and resourceful teacher may be allocated more 'problem' students than a relatively new teacher as a matter of policy. Teaching students at the extremes of the ability range is harder than the median as the reasons for (say) severe under achievement can be very diverse (learning difficulties, home situation, undiagnosed medical issues - diabetes in my recent experience with one student - and so on) and each student may need a different approach.
I have taught in a few settings over the years. The best teaching happens in my experience in settings where people work together. Performance related pay based on dodgy statistics in an atmosphere of public ridicule does not seem the best way to encourage sharing of practice and teamwork!
"Good luck with that" as I believe the late Mr Jobbs used to say to competitors.
That is exactly what makes this study interesting. All the students were assigned randomly! Much like in a randomized medical study this allows the establishment of a causal relationship.
Two thoughts...
1) Ethics: You put some children with dyslexia or with challenging behaviour with newly trained teachers with little experience? Did their parents have any say in this? Was there an ethics process? Was there support available for newly trained teachers meeting major needs in their randomly allocated classes? Amazing.
2) Correlation does not imply a causal relationship. You need a theory. Theory in this area is not positivist, it is fuzzier.
As I said before, 'good luck with that'.
2) Randomized experiments can establish causation if done properly. This is how medical treatments are evaluated. The theory is simple: teachers that were more effective in the past will be more effective in the future and using objective metrics (tests, classroom observation and surveys) we can measure teacher effectiveness. This certainly positivist but more important likely repeatable.
You seem like you are interested in this but I get the sense you already have strong views. Maybe you should put those aside for a moment and take a closer look at this new and interesting study you might learn something:
I will read the reference, however, to those of us in Europe, please remember that this methodology resembles a group of earnest steam punks analysing flies in amber with brass and ebony microscopes on an ice floe that is breaking up... I'm suggesting that we suspect positivist descriptions do not capture the important values in education. 90% of what children learn is learned outside school I suggest.
I am prepared to be astounded with contrary evidence.
Also, see this blog for some more discussion along the same lines: http://ed2worlds.blogspot.com/2013/01/gates-foundation-waste...
Seems pretty clear that Gates' study presents the data in a very misleading manner to make a very weak correlation look like a very strong correlation.
Graphs are not statistics. Why doesn't he give us a correlation coefficient on his big blue blob graph?
http://www.metproject.org/downloads/MET_Validating_Using_Ran...
Take a look at the FULL report. All the statistics behind the graph you criticize are in Table 1 on page 10.
The next 40 pages seem to be a pretty competent description of both the strengths and weaknesses of their study.
Found it following the link: http://garyrubinstein.teachforus.org/2012/02/26/analyzing-re...
In teacher evaluation you also have critically important contextual factors. What course is being taught? Are the students motivated? What's the classroom like? A lot of ELL kids in the class; does the teacher speak their language?
Many people see some amount of value in these imperfect metrics because they a) seek to measure the critical output (programmers are paid to make code and teachers are paid to make kids learn), and b) simple numeric metrics aren't biased by human fallibility. Popularity amongst peers won't have any impact on LOC or student test scores.
The author feels they shouldn't be used as 33% of the overall evaluation, which is the suggestion made by Gates.
In terms the specific points the author made about Gates' evidence in favor of their suggestion of 33%, I think he's basically right. I haven't read the original study and so I'm assuming for sake of discussion that he's conveying it correctly.
Over saturated scatterplots are not very useful, and in the hands of someone with no training or knowledge of basic statistics they may be dangerous, but they're not wrong. A trained consumer of research, seeing a plot like the author's first, will only be able to conclude that the correlation is not perfect. That rules out only a subset, and a subset that's fairly uncommon outside of the physical sciences, of the possible relationships between two variables. Not very useful, but not misleading unless there's an assumption that the number of data points is relatively small (which may be the case in some research contexts).
Averaging points and then taking the fake aggregated points and plotting those is wrong. It's using a tool to show variability in the wrong way. Averaging hides the variability that is supposed to be shown in a scatterplot. Why not average down to 2 points and draw the line then? Because that's wrong.
I'm thankful that one or two other people here understand, but a bit dismayed at the kneejerk groupthink.
And I guess I'll throw in my two cents into the inevitable teachers-vs-data debate. I believe teachers may be too quick to toss aside performance analysis because they are so close to the personal stories and circumstances...no line chart can seem to account for the student who was improving greatly but then missed an insurmountable number of classtime due to problems at home.
Most humans are unwilling/unable to see their lives and work measured as averages and trend lines and deterministic outcomes, and I think teachers are acutely of this belief due to the nature of their work. And unfortunately, this blinds them to the possibility that while there is no cookie-cutter solution that always works, there are surely some measures/guidelines that often work, or at least mitigate even the most extreme circumstances. As it is, there's not much incentive or opportunity to see "the bigger picture" outside of what happens in one's own classroom year after year.
That said, this is not even close to being mainly the fault of the teachers. Among the other problems is that administrators are just as prone to shortsightedness, and "the numbers" can be wielded by a petty/mediocre principal/district officer to apply pressure to a teacher. It's no wonder that teachers are quick to distrust analyses, given the possibility that they're abused for political means.
So, essentially, it's a giant crapfest with no easy solutions. From the perspective of Gates' efforts, we can only hope that he pours in so much money and advocacy that there's eventually a shift towards best practices...though it's safe to say there will still be unintended consequences that aren't all favorable.
The more we rely on objective numbers, the less a good teacher has to fear from a bad principal. The principal has no ability to make the teacher's students do worse than their statistical predictor, after all.
On the other hand, many other popular metrics (e.g. principal's opinion/coworker's opinion) can easily be tweaked by a principal who wants to put pressure on a teacher.
Grouping data points does mask volatility. This doesn't make the effect any less real on the societal level. If we seek to improve student performance, skewing the pool of teachers towards the better ones will do that.
Yet there is also an enormous amount of volatility which is also real. Ideally we would just create better predictors. Constraining ourself for the moment to this set, since they have high error rates on the individual level, it's also reasonable to assert that in relying on this metric, we will make unfair decisions for many teachers.
The right solution depends on how much you value optimizing student outcomes vs. optimizing fairness as well as second order effects such as driving away teachers. My bias would be to use these scores for economic incentives to attempt to ensure that retention for the best teachers is higher than for the worst. Without additional work, termination likely isn't warranted based on this data.
To those who argue that pay differentials require better evidence, I would suggest comparing the validity of performance evaluations across any other job. It's imperfect but better for society than doing nothing.
For long-term fairness, it's important that there isn't a single value-added model. Competitive pressure can do wonders to improve model quality and it's also fairer since those who don't do well on one model can move to a different rubric.
X|X
---
X|X
Where the axes are scores above and below 600 in math on the first and the second time taking the test. There are some individuals who do better or worse - but just because we can draw a detail-obfuscating graph of data doesn't mean there is no detail in the data, or no correlation, or that SAT scores won't help us predict future SAT scores. It just means we can draw a detail-obfuscating graph, and that there's not a perfect correlation.
One point the author does have is that these scores are not perfectly predictive, so bad performance one year shouldn't mean that the teacher gets i.e. fired. OK. He seems to be protesting using information to draw any conclusions at all. Perhaps it would be more effective to show what fallacious or irrational conclusions are being falsely drawn from this data and used to fire teachers, and then object to those specific instances of irrational actions. This value-added measure has not been shown to be intrinsically unreliable or wrongheaded but maybe some applications of it are.
1) How is a "teacher" defined? Just(hash firstname lastname)? If so...
2) When a teacher teaches multiple subjects per year, which one is chosen? Or are they both represented, so that "teacher" actually means (hash firstname lastname subject)?
https://github.com/tmoertel/nyva-cursory
Anyway, here are the variables I used to identify comparable observations:
## these variables identify comparable teacher observations
id_vars <- c("subject", "grade", "dbn",
"teacher_name_first_1", "teacher_name_last_1")
I was pretty curious how the author specifically munged his data, since then we could put to rest the speculation about the degree of correlation.
There's such a high correlation between an individual student's score from one year to the next, and the number of data points is so small, that trying to tweeze out the teacher's contribution in a statistically meaningful way is fairly hopeless.
What I'd like to see are correlates with a teacher's own scores on the tests of the material they're supposed to be teaching (oh you think they all can get perfect scores easily? Hah!) and whether teachers banned from giving homework do any worse, and whether dumping the enforced curriculum in favor of letting students study at their own pace makes any difference. Given how little what the teachers actually do seems to make, it would be logical to at least dump the things which make school dull and unpleasant.
The thing is, it's easy to find out who the best teachers are. Simply ask the students, parents, and staff. They all know who the good ones and the bad ones are - and that can't be gamed.
Are you claiming that there's something special about teaching that makes teacher performance much more difficult, or even impossible, to quantify? If so, what is the point of teaching if one cannot measure results?
Your solution to "ask the students, parents, and staff" is precisely a measurement method (albeit a more qualitative than quantitative one), and moreover, one that can be gamed easily by anyone who's socially shrewd. Qualitative measures like that is exactly why we have so many crappy politicians.
The same company decided to rate programmers based on the bug list. A big chart was posted on the wall that showed the daily bug count. It wasn't a week before huge fights erupted over the bug count - over what was and was not a bug. The programmers quickly gamed it. They'd hide bugs, they'd refuse to fix one bug if that fix would produce some other minor bugs (e.g. if the bug was "feature X missing", then feature X is added, but had a couple issues with it, then X would not get fixed). They'd even add "bugs" blamed on other programmers and then "fix" them to get the credit.
Management gave up on that after two weeks and pulled the banner down.
The fact that one company used a couple of bad objective metrics doesn't mean all objective metrics are bad. They are used with a great deal of success in many fields. Sales people are paid on commission, traders are paid proportionally to (risk adjusted) profits, etc. All it means is that if you set the wrong goals for your company, you'll probably succeed at the wrong thing.
So tell us - is "maximize the % of students capable of multiplying/dividing 4 digit numbers" the wrong goal? If so, what is the right goal, and why can't it be measured?
You use the time that you used to use for teaching reading and devote it to teaching methods for multiplying and dividing numbers up to 4 digits in a standardized test setting.
You might say this isn't gaming the test, it's just a stupid way for schools to optimise their results for that metric. But that metric is mechanical; there's no measurement for deep understanding of the principles, just success or failure for applying mechanical rules by rote.
Most US standardised test settings use machine marked multiple choices. I'm told that there are lessons for 'bubbling in' - lesson on how to fill in the multiple choice answer bubbles to ensure fewer mis-marked answers.
Clever - you caught me. I got lazy and didn't feel like typing up all the goals of a 3'rd grade education system into an HN comment, preferring only to provide a simple example and hoping that a reasonable reader would extrapolate.
Mea culpa - I'll stop assuming reasonable readers in the future.
Or are you actually suggesting that the school system might forget to include reading when defining their goals?
Can all the goals of a 3rd grade education system be reduced to a purely mechanical list of stuff?
Let's try applying your example of arithmetic to reading. What are the goals? To get children to read individual words? Or to get children to read a sentence, and obtain meaning from it? If it's just to read words, do you get those words from a defined vocab? Should they all be real words, or do you include nonsense words too?
While schools may not have stopped teaching students to read they have cut out other parts of the curriculum to focus on what's being tested.
And there's a risk of a cut off point - there's a bunch of children who will read, and there's a few children who are struggling to read. Do you spend extra time and money on the few strugglers (some of whom are going to fail whatever you do), or do you concentrate on the majority (and get most of them through the test and thus look good)?
Yes, I would hope that a multi-million dollar enterprise can clearly define their goals.
What are the goals? To get children to read individual words? Or to get children to read a sentence, and obtain meaning from it?
I don't know off the top of my head whether the latter should be learned by 3rd grade. Ultimately setting the goals of our educational system is up to the various bureaucrats in the school system.
However, regardless of what the goal is, you still haven't given an way to game the system apart from "teach kids to read [words/sentences]".
While schools may not have stopped teaching students to read they have cut out other parts of the curriculum to focus on what's being tested.
Indeed - if the school system is not achieving their primary goals, they should cut secondary goals and focus on the primary ones. That's a good thing.
Do you spend extra time and money on the few strugglers (some of whom are going to fail whatever you do), or do you concentrate on the majority (and get most of them through the test and thus look good)?
This depends on what the goals of the school system are. If you want to ignore the strugglers, set the goal to be maximizing this function:
student_scores.max()
student_scores.min()
student_scores.mean()
log(student_scores).mean()
Sure I have. You ignore everything that is not tested. This gets you children that pass the tests. But it ignores all the other work that schools should be doing, and it reduces education to the worst, least inspiring, mechanical drudge work.
> I don't know off the top of my head whether the latter should be learned by 3rd grade. Ultimately setting the goals of our educational system is up to the various bureaucrats in the school system.
They can't agree. That's why I listed nonsense words in one of the requirements. There's an argument about whether phonics or whole-word approaches are better, even though we have good research showing that phonics is better. And so when you look at phonics methods (include nonsense words in the tests), you get disagreement in the phonics camp, and you also have all the non-phonics people piling on.
But this should be easy to discover, right? We have millions of children learning to read each year. We randomise them, we set up a control group, we give other groups different methods. Then we test. (assuming we can get agreement on what and how to test.)
I can't imagine any politician agreeing to that.
Ignore non-goals and focus on goals?
But it ignores all the other work that schools should be doing...
Such as?
But this should be easy to discover, right?
No. Choosing your goals is about subjective value choices. If nonsense words are intrinsically valuable, they should be included, otherwise they should be excluded.
You are conflating the setting of goals with the method used to achieve them. If phonics is superior (I agree with you that it is), it will achieve higher scores. If teaching children nonsense words helps them understand real words, then teachers wishing to maximize their score will teach them.
All we really need is to believe that it's a better approximation than the alternative, which is currently something along the lines of 0.5 x Principal's Opinion + 0.5 x Union Seniority (at best).
My main point is that I think it's better to frame this as "this is a better approximation" than as "reading is the goal". That's all.
cheating (see Atlanta Journal-Constitution series on this)
encouraging absenteeism amongst low performers
damaging the test materials of low performing students to get their scores excluded
getting low performers into IEPs so they don't test, take the alternate or receive accommodations
sabotage other teachers in various ways
vie for the most favorable students by pulling strings with the administration
I'm sure if there's actual motivation one might find even more.
Treating damaged test materials/absent students as a score of 0 is the simplest way to prevent hacking the student body.
vie for the most favorable students by pulling strings with the administration
The most favorable students are those who will perform better than their statistical predictor. How do teachers know who those will be?
Another way to manipulate the test results is to manipulate which students are in your class or your school.
The point is, people are endlessly creative in subverting rules to their own benefit - so they conform to the letter of the rules but not the spirit.
Consider also the "work to rule" technique used by unions as a bargaining tactic. It's as simple as the workers literally adhering to their job descriptions. It doesn't work out well for the company.
The fact that the current system has a bunch of cheaters is not an argument against more carefully and objectively measuring the current system. What next - bankers sometimes engage in rogue trading, so we should reduce monitoring of their behavior?
Another way to manipulate the test results is to manipulate which students are in your class or your school.
This is very difficult with VAM, since the goal is to increase (actual score - statistically predicted score). You need to reliably identify students who will do better than their statistical predictor.
I.e., you need to discover students who will improve drastically this year and then pack your student body with them.
I had a high school teacher who explicitly wanted to be popular by assigning no work. The students did like him, but they did not respect him (and laughed at him behind his back). Those of us who wanted to go on to college also knew he was shortchanging us.
While the author is correct in that averaging hides the variance of data, he does not try to answer the question of "Can a trend be found from noisy data?" or show other accepted statistical methods that show no correlation in the data.
Instead the author dismisses the report from Gates through name calling: "It seems like the point of this ‘research’ is to simply ‘prove’ that Gates was right about what he expected to be true. "
This is supposed to be evidence that the correlation is weak??
The blog link to this thread is a propaganda site. It conspicuously doesn't reference any of the original materials that it is attempting (weakly) to refute.
There was an article on HN last week that referenced a more source material article about the project.
Here's the link to the METProject's web site: http://www.metproject.org/
Note that this isn't strictly equivalent for several reasons. The game above the error is part of the generating process instead of the measurement process. Also the distribution is district rather than continuous. The key commonality is that a weak predictor at an individual level can be a meaningful predictor at a group level.
Imagine if one group had 55% of its dice labeled up to 110 and 45% up to 100, and the other group had the opposite. So if all you use for hiring is the group membership, you're going to very arbitrarily miss out on a lot of 110s and hire a lot of 100s. And now also imagine that this process of identifying who is in what group is ridiculously expensive.
http://news.ycombinator.com/item?id=4728123
and I'm glad to see that so many participants, from the founder on to the newest member, enjoy thinking about and checking facts on education issues.
There is other research on the issue of teacher quality and value-added measures to assess teacher quality,
http://hanushek.stanford.edu/publications/teacher-quality
and I encourage Hacker News participants to dig through some of the other publications to figure out what the best available current data show.
The other comments here suggest that the blog author of the submitted blog post is overstating his own case at least as much as he accuses the Gates Foundation supporters of overstating their case. If people who oppose the kind of teacher ratings proposed by the Gates Foundation really had courage of their convictions, they might try to let learners have full power to shop for teachers, on the grounds that teacher quality measures miss many dimensions of teacher performance that are important to individual learners. In fact, very few states even have as easy and time-proven a policy reform as statewide open enrollment (which my state has had for almost a generation now)
http://education.state.mn.us/MDE/StuSuc/EnrollChoice/index.h...
and no state yet lets learners shop for schools on an equal per-capita funding basis whether the school is state-operated or not (as the Netherlands has for more than a century). Teachers will have plenty of deserved professional respect and regard from their clients if their clients have power to shop. But if policy proposals for learner power to shop are rejected, hand in hand with rejecting proposals for evaluating teachers for their effectiveness in publicly subsidized instruction, I wonder what the true agenda is. What assurance do we (we parents, we taxpayers, we members of the general public) have that all schools are staffed with the best available teachers unless someone is checking on how the teachers are doing their work?
Even here on the allegedly rational Hacker News I've seen item after item sink without trace, whereas others with significantly fewer details and significantly worse evidence get voted up and read by thousands.
Alex Bellos recently passed on to me what his editor told him. It doesn't matter how good your article is, if no one reads it, or if no one gets to the point, it's no use and a waste of your time.
I'd like to think "Scholarship and evidence don't need hyperbolic headlines." Sadly, I think that scholarship and evidence say otherwise.
And really you would be amazed at the titles (and abstracts) of papers in journals of Pure Mathematics. Really. I've seen thing you people would't believe.
Can't? Won't? Then be quiet.
http://en.wikipedia.org/wiki/Value-added_modeling
It's not something that "Bill Gates" or anyone else made up yesterday. It's a state-of-the-art approach to a difficult problem. If you have a better way, let's hear it. Otherwise, be quiet.
In science, falsifying a theory does not require that you replace it with a better theory.
Having a bad teacher to me doesn't matter when any kid can turn on his phone or laptop and watch free MIT lectures, free Stanford lectures, free Khanacademy lectures, free Android and iOS developer courses all over youtube, ect.
Gates obsession with teacher metrics would be better spent on building an entirely free university with worldwide standardized credits, so somebody in Pakistan can have their education credentials accepted should they start working overseas instead of being forced to drive a cab around Montreal while spending thousands to retake what he/she already learned.
Alternative self-study education is small compensation for having your child literally waste a school year in a teacher's class who has no clue how to keep control of the kids or desire to go over the material in class vs assigning everything for homework. In my daughter's case, her fourth grade teacher was too busy using that vaunted Internet access to check her facebook page.
Also, your "doesn't matter" argument would hold a lot more water if I didn't have to pay significant taxes to support such a lame public education system.