upvote
That ≈0.01 SD effect is actually 0.0087 +/- 0.0043 (standard error), giving a 95% confidence interval of roughly 0.0003 to 0.0171 standard deviations.

So the effect could be tiny.

Worse, the report says it excludes observations outside the 0.5th and 99.5th percentiles of the 2021–22 fall-to-spring score-change distribution. I'm not sure how this is defensible. Either you trust your measurement instrument or you don't.

If you're using a difference-in-differences design, excluding observations can change what you're measuring, particularly if the exclusions differ between treatment and control groups or over time. Without trimming those observations, the estimated effect drops from 0.0087 SD to 0.0058 SD (and is no longer statistically significant).

The authors preregistered their study, but there's nothing in the report indicating that the trimming of 'extreme' values was part of that preregistration...

reply
Sounds like they noticed a few extreme outliers were pulling the difference down and hacked it out. Methodologically that’s really sus to do this after the study is done. But generally I think it’s okay to exclude extreme outliers - we generally assume Gaussian distributions, but the real world happens and messes things up. Having a few kids that are 5 or 6 sigma out pulling your averages - I can believe the rules are different for them. But this should have been decided ahead of time.
reply
I emailed the authors and received a reply. That part of the study was not preregistered.

My message to them:

  Dear Dr. Burland and Professor Kane,
  I was reading your November 2026 report, Phone-Free Schools and Achievement Growth, and saw that the study was preregistered on the Open Science Framework. I haven’t been able to locate the registration.
  Could you please share a link to the preregistration or let me know where I can find it? I’m particularly interested in checking whether the exclusion of extreme fall-to-spring score changes was specified in the original analysis plan.
  Thank you for your time and for making the research available.
  Best regards
Their reply:

  Rahim:  That part of the study was not pre-registered.   We have used the NWEA data from earlier years, but had not encountered such extreme outliers.  The magnitude of gains or losses we dropped were clearly errors—multiple standard deviation gains in a single year, also inconsistent with the midyear measurement.   I’m ccing our colleague, Samantha Cullum, who can provide the link to the pre-registration.   Tom Kane
reply
Aside:

Wow! I love this.

I love that you bothered to email them with a thought out question.

I love that they bothered to respond with a thought out answer and they took you quite seriously.

I love that you bothered to tell us about the exchange.

Thank you so much for all of this. This is what we should all strive to do with our time and as a community.

reply
Alright mate
reply
Yeah! Fuck follow-through!
reply
Phenomenal. This is why I come to Hacker News. Thanks so much.
reply
Respect! Good science
reply
Interesting that their take is that the outliers represent measurement errors. Plausible and reasonable. But the fact that it changes the conclusion of the study deserves skepticism.
reply
The outliers would have made the claimed effect even stronger. They were trying not to overhype their findings.
reply
This is incorrect. Please see my earlier comment.
reply
Contrast that with all the anecdotes in this thread from parents about how great banning phones is, I'm left with two possible conclusions:

- Trust the anecdotes and this effect can't be fully captured by the study

- Trust the study and assume the parents are projecting their desire for banning phones to be a good thing.

I'm leaning towards #2, since I also have a strong desire to ban phones around my kids (they're 3 and 1(x2) so yes, around them not for them). I'd personally be really prone to confirmation bias myself around phone bans.

reply
When the anecdotes and the data point in different directions, trust the anecdotes; it usually means you are measuring the wrong thing.

-Jeff Bezos, paraphrased

reply
What does this mean for what is arguably the most famous data-anecdote gap: crime rates? People have been thinking crime has been skyrocketing for decades (well before the pandemic), but statistics show the opposite?
reply
People - the "individuals" - really dislike being victims of crime. There's also the broken windows theory: a crime, especially when it goes unpunished, sends a signal to every actual or potential criminal: "Here, you can do as you please."

So, yes, I almost agree with the statement "People have been thinking crime has been skyrocketing." IMO, we're never necessarily far from a situation in which widespread impunity could cause crime to escalate rapidly.

reply
Very simple: the news is not anecdotal; the news is editorial. If you were to talk to people about their own direct experiences with crime, I am confident that the aggregate anecdotes would look far more like the crime stats. Instead, you have in news media an industry that chooses what to cover, and it never covers: "this part of town had no crime today," which is an anecdote itself. Thus, because the news is biased (towards something rather than nothing) in what it reports, and because most people believe the news as an accurate representation of the world, this malforms their view of the world.
reply
the tests they are using are measuring something. they just aren't measuring what everyone wants to use these tests for (admissions, explaining differences in race and gender, closing schools, picking the best schools, how to assign scholarships / give kids IEPs, the impact of cellphone bans, etc.)
reply
Individual test students are going to be hard to predict anyway. Though I think this is compared to the sample standard deviation, and for most students their improvement over time is going to be smaller than the difference between two random students.
reply
Don't forget that scores are normalized for many tests. It's not a linear scale from "nothing correct is a zero, everything correct is a 10". Rather, the average number of correct answers (often weighted, but ok) is a pass. This will mess up any analysis.
reply
The test they're using is the NWEA MAP Growth https://www.nwea.org/map-growth/ which doesn't have a concept of pass or fail. It doesn't even have a fixed set of questions, but adapts the difficulty dynamically based on earlier answers to get a more accurate measurement.
reply
Yea, this study seems pretty bad
reply
deleted
reply