Tuesday, September 22, 2009

What is "The Gold Standard"?

Did you hear about the big report that came out this week? You know, the one that "shows" that NYC charter schools are better than traditional non-charter public schools? It has gotten a ton of attention, probably because it uses "'the gold standard' method[ology]." The report is not subtle about this. It is right there in the very first sentence of the executive summary, "The distinctive feature of this study is that charter schools' effects on achievement are estimated by the best available, "gold standard" method: lotteries." It even uses the term "gold standard" four more times throughout the report.

Everyone wants to follow The Gold Standard -- or at least be able to say that they do. Of course! I mean, who wouldn't? But I do not think that we actually have a gold standard in education research. In fact, I am quite sure that we do not, and appropriating biomedical research's gold standard does not make it appropriate for us.

However, if we are going to borrow their standard, can we not at least get it right?

The biomedical standard uses double-blind experimental studies with random assignment. That means that some research participants get the experimental treatment and some get a placebo, and both are assigned randomly. It also means that neither the researchers nor the participants know who is getting which treatment. After all, expectations are important, and the mind can set us up for all kinds of things.


One of the latest ideas about The Gold Standard in educational research concerns charter schools.

We all want to know whether charter schools are better than traditional non-charter public schools. On one level, we certainly do want to know about individual schools. But on the policy level, we want to know about the average charter school, because we want to figure out if "charterness" helps a school be better. If it does, then we want more charter schools. If it does not, then we want fewer or none. And if we cannot be sure, we want to keep checking.

Let me say this quite clearly: Some charter schools are better than most non-charter public schools, and some are worse. And some non-charter public schools are better than most charters, and some are worse.

The Gold Standard crowd have a favorite method for comparing charter schools to non-charter public schools, one of which they are quite proud, but one that is so full of problems that I am shocked that they keep using it.

They rightly want to control for self-selection bias among charter school students. We know that children and families that apply to charter schools are different from those who do not, even if we do not know what all those differences are. This seems like the perfect time to do a randomized assignment, because that is the best method to make sure that these differences cancel out between groups. Luckily, we have some randomized assignments. Oversubscribed charter schools are virtually always supposed to accept students using a random lottery. This allows researchers to compare the outcomes of those were were randomly accepted, and those who were not.

Sounds good, right?

Well, it does sound good. But serious issues remain. Some are more obvious than others, and some are correctable by those interested in getting the correct answer, rather than the one that fits their pre-ordained conclusions.

Issue #1: No Placebo

Biomedical research does not just include randomization of treatment. It also is at least single blind. If some patients know that they are getting the new treatment, they might react differently. The mind is a powerful thing. They might be more diligent. Perhaps do their rehab exercises more often. Maybe pay more attention to diet. Who knows? And those who know that they are getting the new treatment might not lose hope.

If a student and/or his family do not get into the school of his/her/their choice, how might they react? I know from my own experience teaching that students who get their choice of schools take a bit more ownership. If they get their second choice, or last choice, or somehow do not get their choice, that's a big hurdle for their teachers and parents to overcome. If parents do not get their choice of schools for their students, are they going to be as supportive of their child's teachers? Of the assignments? Are they going to have the same kind of faith in their child's school? I think that the answer is really quite obvious.

The problem with these studies is that the students and families who "lose" these lotteries are no longer like the students and families who "win" these lotteries. There simply is no basis for thinking that their views of their schools are like those of the lottery "winners." In fact, one could quite simply argue that this method of analysis ensures that the "winning" charter school students are being compared to students who did not want to go to the schools they attend.

Obviously, that's not an even comparison.

Issue #2: Peer Effects

We all know that peer effects matter. Research, experience and common sense back this up. If you put a student in a class with a "better" group of peers, s/he will do better than s/he would have done in a class with a "worse" group of peers. The other kids all do their homework, or their parents were more likely to read to them, or they are somehow smarter, or harder working, or bring more cultural capital to school with them, or however else you think "better students" might be defined.

We also know that charter school students are not, as a group, like non-charter school students. That is how the Gold Standard crowd justifies their approach here. So: trying to control for applicant differences, but not controlling for ongoing peer effects? I don't know if that is just lazy or actually dishonest. The importance of peer effects is so well recognized that I tend to think it is the latter, especially because techniques to control for them are so well established.

Of course, there is another way to look at this. From a personal level, if you have a child, you don't care about controlling for peer effects. Actually, you want the effects to be left in so that you can take advantage of them. If charter schools have "better" students, that's a reason to send your own child to a charter school. However, if this analysis is done for policy purposes, to influence policy-makers, then peer effects do matter. If you are thinking about all students, not just the select few who can get into the "better" school, you need to control for peer effects.

Issue #3: Selection Bias on the School Level

The goal of this lottery-based study design is to avoid self-selection bias in the data. However, those who use it do not acknowledge the additional selection problems they create.

The most important problem is that not all charter schools are oversubscribed, so not all charter schools can be included in these studies. This wouldn't be a problem if we had good reason to believe that a random selection of charter schools were included, but that is obviously not the case. Clearly, the "better" charter schools are far, far, far more likely to be oversubscribed than the "worse" charter schools. This biases the sample rather severely towards better charter schools.

Unfortunately, the sample bias problem doesn't stop there.

A really strong traditional non-charter public school is not going to lose a lot of students to a simply above-average charter school. In order to be oversubscribed, a significant number of students and/or families have got to believe that the charter school option is superior to the non-charter public school option, which suggests a level of dissatisfaction with the local traditional public schools. This biases the sample towards inferior non-charter schools.

Issue #4: Generalizability

The hardest thing in educational research -- and perhaps research overall -- is to be able to generalize one's results to the broader population or wider world. And yet, that is usually the end goal of policy-oriented research.

These kinds of lottery-based studies only include the kinds of students and families that apply to charter schools in the first place. Even if the previous issues could be corrected, how can one know that other sorts of students and families would see the same benefits? The fact is that different populations might benefit less or more from going to a charter school. It is simply impossible to know from this kind of study. Of course, if you are only concerned about benefitting the kids of families who already opt for charter schools, then this is not a problem. But if you aim to help a broader population than that, you need a better methodology.

These generalizability concerns also apply to schools. Oversubscribed charter schools might well be better than average non-charter public schools, and I do not really question whether they are better than their local traditional alternatives. But on a policy level, we need to be concerned with charters more generally than that. If we raise or lift caps on charter schools, or approve new charter schools, we have to expect an average charter school to result, not an exceptional one. But these studies really tell us nothing about the majority of charter schools that are not oversubscribed. Nor do they tell us anything about the relative quality of non-charter public schools that lack charter school alternatives.


I understand the desire to find a Gold Standard for educational research. But simply grabbing that label because a methodology has some resemblance to biomedical research is not good enough, despite what Prof. Caroline Hoxby may claim. Moreover, the popular press really must do a better job of examining these claims critically, rather than cheerleading for them like this.

This, of course, means that researchers, journalists and the rest of us must be sure to take a more thoughtful stance than has become our habit.