Eleven independent samples drawn a year apart, across a decade containing a pandemic, the rise of TikTok, and a full turnover in what the phrase "social media" denotes. The signal appears every year, same direction, similar magnitude.

Having read the methods I want to give this more credit than I did on first pass. The authors benchmarked against a majority-class baseline, a demographics-only logistic regression, and a demographics-only forest, so the incremental contribution of media use is actually isolated rather than assumed. They ran three separate importance metrics because Gini alone is known to favor high-cardinality predictors. They ran the COVID sensitivity three ways, removing 2020 and 2021 outright, adding a pandemic indicator, and testing an interaction, and none of it moved the result. They restricted features to inter-correlations below 0.3. And they trained on 2014 to 2018 and tested on 2019 to 2024, where the model held at AUROC 0.68 to 0.72.

That last one is the check that separates a finding from a fit, and most papers in this area do not bother.

Table 1 reports covariate-adjusted pooled odds ratios per minute per day for all six media categories. Converted to a per-hour basis, they run: gaming 19.7 percent higher odds, internet 17.6 percent, social media 16.9 percent, television 15.5 percent, email 14.1 percent, news 3.7 percent. Every category except news was significant in all eleven years after Bonferroni correction.

The section is headed "SMU is the most important predictor of SRD among multiple forms of media usage." The discussion states that "social media use should not be collapsed into a generic screen-time measure when studying depression-related outcomes." On the odds-ratio scale the paper reports, social media ranks third, behind gaming and internet use, and sits inside a hand's width of television and email. The confidence intervals are so tight and so close together that the ordering is real, but the ordering does not put social media on top.

The forest's feature-importance ranking is where social media does well, and the authors themselves note the obvious problem with that: "Given one must access the internet for social media, these metrics may be related and thus similarly ranked." They wrote the caveat and then wrote the headline anyway.

This one is not the fault of the press. This is the paper.

A SHAP dependence plot shows how a feature's contribution to a model's output varies across the range of that feature. Where the curve crosses zero, the variable stops pushing predictions down and starts pushing them up. In these models the crossing sat near 150 on the derived exposure scale.

Three separate problems stack here. The units are summed hourly probability sliders, not measured time. The zero crossing is a property of the model and the distribution it was fit to, so moving the sample moves the crossing. And the authors say so plainly: "The 150-min value should therefore be treated as a hypothesis for future longitudinal or experimental studies." Every outlet that ran the number dropped the sentence that followed it.

The tell was always the roundness. Real thresholds in medicine are ugly numbers. 150 is what a number looks like after it has been rounded toward memorability.

Eleven repeated cross sections. Different people each year. Nobody was followed. The paper says "repeated cross-sectional" in the second sentence of its abstract and returns to it in the limitations.

"Researchers Studied 183,000 Adults for 11 Years" describes a cohort study nobody ran. The authors are careful about this and lay out the reverse-causation case themselves, noting that depressed people may seek online escape, emotional support, or distraction, or may turn to digital interaction as offline social contact gets harder. That paragraph exists. It did not make the coverage.

There is also a selection problem the paper gestures at without naming. It concedes that "the commercial recruitment channels may have its own sampling/selection bias," but does not connect that to the specific fact that one of the three recruitment channels was social media platforms. Recruiting part of your sample through the exposure you are measuring inflates the high-exposure tail, which is exactly the tail the 150-minute crossing lives in.

Table 2 reports the numbers nobody quoted. Sensitivity ran from 0.395 to 0.534. Precision ran from 0.272 to 0.393. In the temporal validation on held-out years, sensitivity fell to 0.392 to 0.431.

The model misses roughly half to sixty percent of the people who report depression, and between 61 and 73 percent of the people it flags do not. The paper's closing line offers "a scalable predictive framework for identifying individuals at elevated risk." A framework with those operating characteristics identifies a population, not an individual. AUROC around 0.70 is correctly described in the paper as moderate discrimination, and the authors do not oversell it. The closing sentence does.

The abstract reports accuracies between 0.71 and 0.75 and AUROCs between 0.65 and 0.73. Table 2 reports accuracies from 0.723 to 0.771 and AUROCs from 0.661 to 0.739. Both cannot be right.

Social media's rank among the predictors is given four different ways. It was "the fourth most important independent variable for predicting SRD for 2024 and the prior six years." It was "the top-ranked feature each year" from 2019 onward. It "came in second with internet use and age range in 3rd and 4th." It was "ninth in importance by SHAP in 2014 and ranged between 3rd and 4th for the last 6 years." Personal income, meanwhile, is the strongest predictor across all three importance metrics, which is a finding about poverty that the paper mentions and then walks past.

The forest is specified with 200 estimators, then with a default of 20,000, then described as aggregating across 100 trees. Cross-validation is called twenty-fold in one paragraph and specified as ten splits with twenty repeats in another. The sample total is 183,000 in the abstract, 183,300 in the methods, and 183,400 in the results. One limitation reads, in full, "there may be unmeasured confounding factors from the fact the Figure was always surveyed at the same time each year (January)."

This is an Article in Press, an uncorrected proof, and some of that is copyediting that a production pass will catch. "The Figure" plainly used to be a word about people. But an abstract that disagrees with its own results table is not a typo, and four incompatible statements about the study's headline variable are not either.