Clarifying the picture over teen mental health
A graph I published last year included an error. This piece corrects the error, walks through the full recalculation, and then describes some surprising trends.
I got an email from David Stein recently suggesting that the first graph from my post last year with Taylor Barkley titled, “Thoughts on what the CDC YRBS data means for social media, teens, and mental health” was likely calculated wrong. He then published a Substack piece explaining all of the details.
It is a serious graph because it shows the suicide rate for teen boys and girls aged 15-19. So, I pulled all the data again and discovered my mistake. I accidentally pulled a segment of the data for 13-19 year olds instead of simply 15-19 year olds for 1999 to 2020.
It was not my intention to include this broader group in the graph, but as even Stein expressed, “this is an understandable mistake that anyone could make.”
While the change in the data doesn’t fundamentally change my analysis either then or now, it means the chart should look like this:
For transparency's sake, I thought I would share the full recalculation, walk through the framing of trends and cycles we suggested in the first post, and then describe some surprising trends, in my estimation.
Teen suicide is a very serious public health problem. While my mistake did change some aspects of the graph, the big-picture story laid out in the original post is still the same. Teen suicides rose into the 1990s, saw a decline into the 2000s, and rose again in recent years. Understanding what’s driving these decade-long changes is critical to know how to turn things around.
The calculation
As Stein notes in the Substack post,
Rinehart had to make three separate inquiries corresponding to ICD 8 (1968-78), ICD 9 (1979-1998), and ICD 10 classifications.
The International Classification of Diseases (ICD) is the international standard for comparable statistics on causes of mortality and morbidity. ICD-8, ICD-9, and ICD-10 are successive iterations of the standard.
Pulling each one of these segments is how you construct a graph that includes all available suicide data from the 1960s. ICD-8 matches with ICD-9, so you use E950-E959 for both. Under ICD-10, the term previously known as Suicide was renamed Intentional self-harm, and it was reassigned to a new category X60–X84 in addition to Y87.0.
This is the step of the process where I made a mistake. The original graph selected for X60–X84 in 13-19 year olds when it should have been 15-19 year olds in both X60–X84 and Y87.0.
It remains a crude rate and not an age-adjusted rate because I wasn't comparing geographic areas but the entire United States against itself over time. But this could be an incorrect assumption.
Altogether then, the data was constructed via segments from
1968-1978, saved here;
1979-1998, saved here; and
1999-2020, saved here.
Results were grouped by Age Group, Year, and then Gender. This updated Google Sheet contains the original as well as updated data.
After pulling the data several times, I confirmed that the old graph uses data for 13-19-year-olds instead of 15-19-year-olds from 1999 to 2020.
As Stein rightly point out,
Neither Rinehart nor Barkley are psychologists or public health experts, and the graph was published in a mere blog post, not in a peer-reviewed journal.
Indeed, this is correct! I am a trained economist, and the graph was from a blog post we published that was meant to start exploring this topic. I never intended it to be the absolute word on the subject, but was trying to begin to build a body of work and better understand the dynamics of play. Part of that post also teased a framing of the issue as trends and cycles. More on that later.
The original looks like this:
Instead, It should look like this:
This also comports with the Twenge post here cited by Stein.
Stacking the new and the old data:
What’s changed between the older, incorrect graph, and the updated version:
The total suicide rate reached an absolute peak in 2017, although as of 2020, it is still below all-time highs.
The total rate, boys rate, and girls rate are higher.
What hasn’t changed between the two graphs:
The early 1990s was a relative highpoint in teen suicides.
The mid to late 2000s were a relative lowpoint in teen suicides.
Suicide rates are again rising.
Girls suicide rates are the highest they have ever been, this was the case with the previous graph as well.
Trends and cycles
The dominant framing in the original post contrasted trends with cycles. Here is what we said,
Is what we are seeing a trend or a cycle?
Trends are different from cycles. Cycles revert towards the mean. They fluctuate around an average point over years. Trends, on the other hand, are something new. They mark a change that is out of the normal cycle.
While I understand there can be confusion over it, the binary framing of a trend versus a cycle highlights one way to think through policy questions. I was attempting to colloquially explain the process of isolating the true effect from everything else. Is what we are seeing a trend, or something else, a cycle?
But I can understand why this framing of trend and cycle is confusing. Indeed, even in that second paragraph, I define cycles as both reverting towards the mean, which is a strictly defined financial term for an asset’s price to converge over time, as well as fluctuating towards an average point, which can suggest a moving average process or even variance.
To add to this, cycles can sometimes be confused with cyclical components which has strict definitions. Sometimes, cycles are even used to describe seasonality. Cycles, as a term, is intensely polysemic.
If I were to rewrite the original piece, I would focus on the binary of trends versus everything else, and just do away with all of the baggage of cycles. Of course, the trend that is driving the conversation is how technology is contributing to mental health changes, holding all else constant.
But not everything is constant.
The school season
One of the more surprising parts of diving into the data was realizing the seasonality in teen suicides. Teen suicides tend to parallel the school year with peaks in April and October, and a significant drop in July. But there have always been deep questions about this relationship because spring peaks are common for adults as well.
COVID seems to have confirmed the association between school and suicide.
COVID acted like a natural experiment. Schools were unexpectedly closed when they should been opened. And the closures happened just when the suicide peak should have occurred but didn’t.
In compiling emergency department visits and hospitalizations for suicidal ideation and suicide attempts, Kim, Krause & Lane (2023) found, “the presence of seasonal patterns and an observed unexpected decrease in suicidality among children and adolescents after COVID-19–related school closures in March 2020.” Their work underscores the association between suicidality and the school calendar, which is displayed in the graph from their report below.
Notwithstanding, the rise in suicide rates is troubling. Grasping the underlying factors behind the decade-long changes is essential for devising effective interventions.