Teaching is a lot like programming in this respect -- if you try to enforce metrics of "performance", you usually end up with optimization for those metrics and dumber, less-educated students. Similarly, students tend to optimize too, just from a different standpoint. Examples of this:
If you rate teachers based on the grades their students get, they will usually award their students higher grades for the same (or worse) performance. This isn't just a joke; many, many professors have reported their deans bearing down on them for awarding too many bad grades, often to students who are so incompetent that they cannot write basic English sentences or perform long division. The solution -- not remedial courses, or letting them drop out (that would hurt financial aid income) -- but letting them pass even though they lack the basic skills to learn in that class, let alone pass. This also leads to the student mentality that they are being "given" grades (and thus can beg for higher grades) as opposed to earning them through their own effort in the class.
If you rate teachers based on the number of their students pass a standardized test, the teachers will focus solely on the contents of that standardized test and put a huge amount of effort into educating the few students who Don't Get It, at the cost of all the students that do. I had this experience in primary and secondary school, where teachers lamented that they had to go over material for the Standards of Learning tests constantly and repeatedly, dramatically lowering the amount they actually had time to teach. The end result is a quick dropping of everyone to the lowest common denominator -- which often just ends up shifting the entire bell curve to a lower point.
If you rate teachers based on student evaluations, your results are nearly useless. Most students don't bother to write in much detail, since they're just glad to be done with the final exam (I did this too!)-- and often they use it for "revenge" on professors they thought were too hard. The teachers who get the best evaluations are often (though not always) the teachers who gave the easiest work and grades, and whose students turn out to be the most woefully unprepared for what comes next. Just like customers of your software are not always good judges of what they really want, students are often terrible judges of what good education actually is.
As far as I can see, nobody has come up with a magical metric for rating teachers. If anything, most of the metrics out there are outright counterproductive -- and you can see it in schools across America, where teachers are being constantly urged to inflate grades for their 'customers', overlook plagiarism, ignore the most talented students in favor of the least talented ones, and help drive down the quality of education for the next generation.
>As far as I can see, nobody has come up with a magical metric for rating teachers. If anything, most of the metrics out there are outright counterproductive -- and you can see it in schools across America, where teachers are being constantly urged to inflate grades for their 'customers', overlook plagiarism, ignore the most talented students in favor of the least talented ones, and help drive down the quality of education for the next generation.
You can do things to improve the quality of data - use multiple metrics, don't announce them up front and certainly don't standardize - so that optimizing for a specific test doesn't justify the cost, also change testing as needed/with dynamic feedback and use the results to get insight in to the problem. Computers can only help with this sort of thing. Ultimately it's about defining the objective of education, what environment students are supposed to be prepared by it and exposing them to that environment, just like you want to test the system in production as soon as you can, benchmarks are always synthetic and can't replace "real world" testing.
Data won't capture everything and could in fact be 100% misleading - that's the nature of imperfect knowledge - but if used correctly it can help provide insight in to situation. Just don't conflate benchmarks and reality, use it as a tool for improved feedback/analysis. Of course this approach is absolutely incompatible with bureaucratic systems. Heck the entire education system is based around optimizing for arbitrary/artificial evaluation metrics (all that crap about higher education and signaling, that starts with primary education).
As for favoring least talented over most talented, that's just different objectives, I don't think it changes with more quantifiable information about students performance, do you want to raise the bottom at the expense of the top.
If you rate teachers based on the number of their students pass a standardized test, the teachers will focus solely on the contents of that standardized test and put a huge amount of effort into educating the few students who Don't Get It, at the cost of all the students that do.
Your issue here is not with standardized tests. Your issue is with the goals of the school system.
I.e., you want schools to maximize mean(performance_data), but they are actually trying to maximize len([p for p in performance_data if p >= pmin]).
Gathering good data is not the problem, it just reveals the problem.
Of course, the real problem is crappy metrics. And this is why "teach to the test" is given such a bad rap.
If the tests were actually well-written, testing all sorts of critical reasoning and problem-solving abilities, then teaching-to-the-test would be the correct strategy, and metrics would be tremendously useful.
It makes me think somewhat of exams used for foreign students of English. The American TOEFL exam is extremely limited, and test prep can be extremely effective, without actually improving your day-to-day English. The British IELTS exam is extremely wide-ranging, and test prep is basically worthless, because it really does a good job of testing your English. To improve your test scores, you truly do have to improve your real-world English.
I don't see a problem with effective metrics; they might help identify where teachers might need help.
The problem, is that usually nobody really does proper statistical analysis. The metrics I've seen, are usually like "mean student grade performance". In which case, what you'd want to do is have a really small class of smart kids. (Ergo, kudos to the teacher who finds a way to dump the idiots.)
I do think, however, you could do polling of teachers by asking the students, teachers, and other staff. In most cases you could easily have enough coworkers to figure out who's respected and who needs help.
The author is making a straw man argument, though it does not seem to be intentional (edit: I am referring to Rubinstein's original analysis). The author points out that using data for a single class and single year is not a reliable indicator (second graph). That is uncontested and exactly why ratings are based on 3 years of data and multiple classes where possible. Even this cannot provide a single, accurate percentile value. That is why confidence intervals are used and are rather prominently displayed in all the graphs; e.g. <http://www.nytimes.com/schoolbook/school/656-ps-009-teunis-g...;
Given the example above, there is one teacher who has is 50th percentile for career math, but the confidence interval (CI) indicates that this may really be anywhere from well below average to well above average. In contrast there is another teacher with a value of 3 and the highest bounds of the CI still place them well below average. Conversely there is a teacher that is 90th pctl and even the lowest bounds of the the confidence interval still makes this a highly effective teacher.
In sum, the data make evident that a few teachers are fairly unambiguously ineffective (at least with regards to test results), a few are unambiguously effective, but for the majority the most that can be said is that they are neither the best nor the worst, but somewhere in the middle.
So if you reserve yourself to making conclusions merited by the data, then there is some value to this (i.e. identifying the extremely ineffective and effective). As it stands now, the teacher with a 3 may very well have tenure and be paid considerably more that the teacher with a 90 (given that pay is a function largely of seniority and education, and nearly all teachers have tenure). That is what this exercise is meant to address.
So you definitely cannot assign a fine-grained single value to every teacher (and the authors analyses speak to that issue), but there does not mean you cannot make any conclusions at all. If folks make conclusions that exceed what is supported by the data that is a fault of the analyst or perhaps a function of poor communication or visualization, not evidence that the data is junk.
That said, the test could be improved (and supposedly are being improved) and the very least the measurements place more focus on the subjects being tested.
And the teacher effectiveness data is only 40% of the new approach to evaluating teachers in NY; the other 60% including peer evaluation. It may be useful if that became part of the public data so a more balanced picture is available; but it is unclear if that data would be public.
You're right, I did take the route of attacking single-metric incentive systems despite the fact that they're not 100% single-metric. I wanted to get a point across that I think still needs getting across.
I'm not arguing against the use of data in helping teachers understand their effectiveness for one second, and I mention in the article that the data becomes more correlated and useful as more years are included.
"If folks make conclusions that exceed what is supported by the data that is a fault of the analyst or perhaps a function of poor communication or visualization, not evidence that the data is junk."
...completely agree. If my article seemed to argue that "metrics are dangerous" instead of "metrics are dangerous if you choose to publish them publicly while simultaneously using them as a significant component in compensation calculations," then I missed my intended point.
I was speaking to Rubinstein's underlying article. Sorry for the ambiguity. I strongly agree with you regarding the importance of data being used to empower relevant stakeholders. This issue has been on my mind a lot lately. I would even take it a step further: the priority should be empowering the individuals closest to the data and then working outward. So first priority is making the data empowering for students, e.g. so they have the access and tools to be more reflective about their individual results, learning practices, strengths, and so forth. Depending on age, parents would be here as well. Then teachers. And last should be using the data to empower administrators or bureaucrats. What we have seen is precisely the opposite - the people at greatest distance from the activities generating the data, i.e. the actual learning activities, are most empowered by it. This is reflective of big data business models in general. And I think that is at the heart of where the use of data in education has gone most astray. And this issue is relevant not only to government, but edutech companies as well. Are they maximizing empowering themselves with the data they are collecting to the detriment of empowering students, etc.? Who owns the insights extracted from the data? Are students free to extract their own data and take it elsewhere? etc.
It seems Khan Academy has taken this student-first approach as well and has put it into practice. I would be interested in hearing more about there philosophies, practices, or intents with respect to these other dimensions of educational data.
We have "focus on the student" hanging all over our office (which is pretty tricky b/c we change offices all the time these days for reasons I won't get into :) ). It's easy to see how other agendas could creep into educational organizations (especially if you're worried about profit), but for us right now...and hopefully forever...the student is our clear priority.
So if you reserve yourself to making conclusions merited by the data, then there is some value to this (i.e. identifying the extremely ineffective and effective).
Yes. Anyone who cares about the education of children will welcome any further improvements in evaluating teachers, but there is already actionable information available today. Bill Gates makes the good point that the best system of evaluating teachers should lead to interventions to help each teacher do better. But the existing performance gap, which may also be shown by which teachers other teachers who have children choose for their own children, could guide a system of encouraging the worst performers to seek other occupations
Current accountability reform places too much faith in the power of incentives and hence many metrics (not only teacher effectiveness and not only the NYC accountability system) are not directed to offering any sort of guidance towards improving schools. The NYC Progress Reports, which give schools grades, are another prime example of this. Personally, I hope that the next stage in accountability will see a shift to making the data more actionable from the perspective of practitioners and it is something I have been encouraging for a few years. I think part of the obstacle, though perhaps not the primary one, is the criticism that measurements are not accurate. So given the option of re-thinking how to make actionable reports or measurements vs. just improving what exists (via extending and revising statistical methods, refining business rules, improving data quality, and including additional data sources) the latter is winning out even in areas where there may be diminishing returns to doing this.
On the other hand, one could argue that the issue of how to improve the schools is addressed by other projects or other areas in the system, and it sufficient for these tools just to make clear and accurate evaluations. We are dealing with a big problem, both in terms of depth and significance. So there is room to approach this from many angles.
Regarding the issue of making the data public (i.e. Gates' Op Ed), there are competing values at play. Yes, it will be misused and some folks will be unfairly hurt. I would rather have data public (not just the teacher effectiveness data), and allow everyone the option to create there own diverse conclusions, perform their own analyses (like the one linked to), and as a society we can become more sophisticated and knowledgeable about these issues. The alternative is that the only people that will has access to this data is bureaucrats with a specific policy agenda and a handful of academics. Broader society, especially parents, has a right to weigh-in on these issues and be a part of this discussion and they cannot do this effectively without open data. To reiterate, I don't want technocrats driving education (that is part of the broader problem with current reform); hopefully open data such as this can limit that.
The solution is to do a better job of disclosing the information and providing the necessary context rather than just hide it. NYTimes did a pretty sweet job in this respect . I like where they are going with schoolbook.
I would like to see the media and public consistently push government for better open data. Right now that is not happening, and agencies are not going to make big changes in this area in the absence of that pressure.
In particular, the "same grade, different subjects" plot looks like a good correlation. Some noise, but otherwise exactly what you'd expect. Anyone know where to get the raw data so I can do more careful plots (e.g., a density plot instead of a scatter plot [1])?
Note that noise simply means you need more than a single year's data to make a decision.
[1] In a scatterplot, if two points overlap (which they likely do in this data), the apparent density is understated.
Any field that produces only one metric will find people optimizing on it. Eventually the metric loses it's value. This is true in teaching, programming, and most everything. Life's not so one dimensional.
14 comments
[ 4.9 ms ] story [ 47.2 ms ] threadIf you rate teachers based on the grades their students get, they will usually award their students higher grades for the same (or worse) performance. This isn't just a joke; many, many professors have reported their deans bearing down on them for awarding too many bad grades, often to students who are so incompetent that they cannot write basic English sentences or perform long division. The solution -- not remedial courses, or letting them drop out (that would hurt financial aid income) -- but letting them pass even though they lack the basic skills to learn in that class, let alone pass. This also leads to the student mentality that they are being "given" grades (and thus can beg for higher grades) as opposed to earning them through their own effort in the class.
If you rate teachers based on the number of their students pass a standardized test, the teachers will focus solely on the contents of that standardized test and put a huge amount of effort into educating the few students who Don't Get It, at the cost of all the students that do. I had this experience in primary and secondary school, where teachers lamented that they had to go over material for the Standards of Learning tests constantly and repeatedly, dramatically lowering the amount they actually had time to teach. The end result is a quick dropping of everyone to the lowest common denominator -- which often just ends up shifting the entire bell curve to a lower point.
If you rate teachers based on student evaluations, your results are nearly useless. Most students don't bother to write in much detail, since they're just glad to be done with the final exam (I did this too!)-- and often they use it for "revenge" on professors they thought were too hard. The teachers who get the best evaluations are often (though not always) the teachers who gave the easiest work and grades, and whose students turn out to be the most woefully unprepared for what comes next. Just like customers of your software are not always good judges of what they really want, students are often terrible judges of what good education actually is.
As far as I can see, nobody has come up with a magical metric for rating teachers. If anything, most of the metrics out there are outright counterproductive -- and you can see it in schools across America, where teachers are being constantly urged to inflate grades for their 'customers', overlook plagiarism, ignore the most talented students in favor of the least talented ones, and help drive down the quality of education for the next generation.
You can do things to improve the quality of data - use multiple metrics, don't announce them up front and certainly don't standardize - so that optimizing for a specific test doesn't justify the cost, also change testing as needed/with dynamic feedback and use the results to get insight in to the problem. Computers can only help with this sort of thing. Ultimately it's about defining the objective of education, what environment students are supposed to be prepared by it and exposing them to that environment, just like you want to test the system in production as soon as you can, benchmarks are always synthetic and can't replace "real world" testing.
Data won't capture everything and could in fact be 100% misleading - that's the nature of imperfect knowledge - but if used correctly it can help provide insight in to situation. Just don't conflate benchmarks and reality, use it as a tool for improved feedback/analysis. Of course this approach is absolutely incompatible with bureaucratic systems. Heck the entire education system is based around optimizing for arbitrary/artificial evaluation metrics (all that crap about higher education and signaling, that starts with primary education).
As for favoring least talented over most talented, that's just different objectives, I don't think it changes with more quantifiable information about students performance, do you want to raise the bottom at the expense of the top.
Your issue here is not with standardized tests. Your issue is with the goals of the school system.
I.e., you want schools to maximize mean(performance_data), but they are actually trying to maximize len([p for p in performance_data if p >= pmin]).
Gathering good data is not the problem, it just reveals the problem.
If the tests were actually well-written, testing all sorts of critical reasoning and problem-solving abilities, then teaching-to-the-test would be the correct strategy, and metrics would be tremendously useful.
It makes me think somewhat of exams used for foreign students of English. The American TOEFL exam is extremely limited, and test prep can be extremely effective, without actually improving your day-to-day English. The British IELTS exam is extremely wide-ranging, and test prep is basically worthless, because it really does a good job of testing your English. To improve your test scores, you truly do have to improve your real-world English.
The problem, is that usually nobody really does proper statistical analysis. The metrics I've seen, are usually like "mean student grade performance". In which case, what you'd want to do is have a really small class of smart kids. (Ergo, kudos to the teacher who finds a way to dump the idiots.)
I do think, however, you could do polling of teachers by asking the students, teachers, and other staff. In most cases you could easily have enough coworkers to figure out who's respected and who needs help.
Given the example above, there is one teacher who has is 50th percentile for career math, but the confidence interval (CI) indicates that this may really be anywhere from well below average to well above average. In contrast there is another teacher with a value of 3 and the highest bounds of the CI still place them well below average. Conversely there is a teacher that is 90th pctl and even the lowest bounds of the the confidence interval still makes this a highly effective teacher.
In sum, the data make evident that a few teachers are fairly unambiguously ineffective (at least with regards to test results), a few are unambiguously effective, but for the majority the most that can be said is that they are neither the best nor the worst, but somewhere in the middle.
So if you reserve yourself to making conclusions merited by the data, then there is some value to this (i.e. identifying the extremely ineffective and effective). As it stands now, the teacher with a 3 may very well have tenure and be paid considerably more that the teacher with a 90 (given that pay is a function largely of seniority and education, and nearly all teachers have tenure). That is what this exercise is meant to address.
So you definitely cannot assign a fine-grained single value to every teacher (and the authors analyses speak to that issue), but there does not mean you cannot make any conclusions at all. If folks make conclusions that exceed what is supported by the data that is a fault of the analyst or perhaps a function of poor communication or visualization, not evidence that the data is junk.
That said, the test could be improved (and supposedly are being improved) and the very least the measurements place more focus on the subjects being tested.
And the teacher effectiveness data is only 40% of the new approach to evaluating teachers in NY; the other 60% including peer evaluation. It may be useful if that became part of the public data so a more balanced picture is available; but it is unclear if that data would be public.
I'm not arguing against the use of data in helping teachers understand their effectiveness for one second, and I mention in the article that the data becomes more correlated and useful as more years are included.
"If folks make conclusions that exceed what is supported by the data that is a fault of the analyst or perhaps a function of poor communication or visualization, not evidence that the data is junk."
...completely agree. If my article seemed to argue that "metrics are dangerous" instead of "metrics are dangerous if you choose to publish them publicly while simultaneously using them as a significant component in compensation calculations," then I missed my intended point.
It seems Khan Academy has taken this student-first approach as well and has put it into practice. I would be interested in hearing more about there philosophies, practices, or intents with respect to these other dimensions of educational data.
We have "focus on the student" hanging all over our office (which is pretty tricky b/c we change offices all the time these days for reasons I won't get into :) ). It's easy to see how other agendas could creep into educational organizations (especially if you're worried about profit), but for us right now...and hopefully forever...the student is our clear priority.
Yes. Anyone who cares about the education of children will welcome any further improvements in evaluating teachers, but there is already actionable information available today. Bill Gates makes the good point that the best system of evaluating teachers should lead to interventions to help each teacher do better. But the existing performance gap, which may also be shown by which teachers other teachers who have children choose for their own children, could guide a system of encouraging the worst performers to seek other occupations
http://hanushek.stanford.edu/sites/default/files/publication...
while rewarding teachers who teach effectively and are willing to take on challenging classrooms with students who have the greatest need.
http://hanushek.stanford.edu/sites/default/files/publication...
On the other hand, one could argue that the issue of how to improve the schools is addressed by other projects or other areas in the system, and it sufficient for these tools just to make clear and accurate evaluations. We are dealing with a big problem, both in terms of depth and significance. So there is room to approach this from many angles.
Regarding the issue of making the data public (i.e. Gates' Op Ed), there are competing values at play. Yes, it will be misused and some folks will be unfairly hurt. I would rather have data public (not just the teacher effectiveness data), and allow everyone the option to create there own diverse conclusions, perform their own analyses (like the one linked to), and as a society we can become more sophisticated and knowledgeable about these issues. The alternative is that the only people that will has access to this data is bureaucrats with a specific policy agenda and a handful of academics. Broader society, especially parents, has a right to weigh-in on these issues and be a part of this discussion and they cannot do this effectively without open data. To reiterate, I don't want technocrats driving education (that is part of the broader problem with current reform); hopefully open data such as this can limit that.
The solution is to do a better job of disclosing the information and providing the necessary context rather than just hide it. NYTimes did a pretty sweet job in this respect . I like where they are going with schoolbook.
I would like to see the media and public consistently push government for better open data. Right now that is not happening, and agencies are not going to make big changes in this area in the absence of that pressure.
http://garyrubinstein.teachforus.org/2012/02/26/analyzing-re...
http://garyrubinstein.teachforus.org/2012/02/28/analyzing-re...
In particular, the "same grade, different subjects" plot looks like a good correlation. Some noise, but otherwise exactly what you'd expect. Anyone know where to get the raw data so I can do more careful plots (e.g., a density plot instead of a scatter plot [1])?
Note that noise simply means you need more than a single year's data to make a decision.
[1] In a scatterplot, if two points overlap (which they likely do in this data), the apparent density is understated.
[edit: found the data here: http://www.ny1.com/content/top_stories/156599/now-available-... ]