The two items that inflate every social-media addiction score
- Cross-sectional school survey in Spain: 3,463 adolescents recruited across 16 schools in Valencia and Madrid between September 2023 and April 2024, 2,761 retained after exclusions (mean age 14.80 years, SD 1.91; 49.0% girls, 49.0% boys). Confirmatory factor analysis pitted the standard one-factor structure of the BSMAS and the SMD against two-factor structures that separate core criteria from peripheral ones.
- Pulling salience and tolerance out as a separate factor fit better on both instruments. On the six-item BSMAS, CFI rose from 0.961 to 0.990 and RMSEA fell from 0.077 to 0.041 (Δχ² = 113.37, Δdf = 1, p < 0.001). On the nine-item SMD the gain was real but modest: CFI 0.917 to 0.930, RMSEA 0.082 to 0.077 (Δχ² = 73.82, Δdf = 1, p < 0.001), still short of conventional thresholds.
- The core factor (mood modification, relapse, withdrawal, conflict) tracked distress hard. Standardised paths to depression were β = 0.909 on the BSMAS and β = 0.798 on the SMD; to anxiety β = 0.746 and β = 0.751; to life satisfaction β = -0.589 and β = -0.531; to self-esteem β = -0.544 and β = -0.456. All p < 0.001. These are standardised paths between latent variables with both factors in the model; the raw correlation between the core BSMAS sum score and depression is .51.
- With the core factor in the same model, the salience-tolerance factor pointed the other way: depression β = -0.429 (BSMAS), anxiety β = -0.306 (BSMAS) and β = -0.328 (SMD), life satisfaction β = 0.290 and β = 0.224, all p < 0.001. These are partial coefficients from a two-predictor model, not raw associations: in the supplementary correlation table the peripheral factor correlates positively with depression (.37 and .38), anxiety (.36 and .35) and loneliness (.37 and .39), and with the core factor at .74 and .75. The survey is cross-sectional, so this describes who scores what at one moment, not what produces what.
Most published prevalence figures for social-media addiction rest on a total score from one of the two most widely used questionnaires. This study took both apart in 2,761 Spanish adolescents and found that the total is adding up two things that move in opposite directions.
Where the score comes apart
The BSMAS and the SMD are both built on Griffiths' components model, which imports six features of substance dependence into behaviour: salience, tolerance, mood modification, relapse, withdrawal and conflict. Both instruments are scored as a single sum, which presumes a single underlying dimension. The authors tested that presumption directly. A two-factor model placing salience and tolerance on their own outperformed the one-factor model on both scales, and the improvement on the BSMAS was large: CFI 0.961 to 0.990, TLI 0.935 to 0.982, RMSEA 0.077 to 0.041, with AIC dropping from 53,301 to 53,151. On the SMD the same split helped less (CFI 0.930, RMSEA 0.077). The authors call that the best-fitting of the three SMD models they tested. Calling it a misfit is my reading rather than their wording, though the indices are theirs and they sit under the conventional 0.95 mark for CFI and TLI. One more hedge belongs here: the paper's component table does not treat the split as clean, classing mood modification as core or peripheral, and one of the tested models put it on the peripheral side.
The structural models are where the clinical content sits. Each outcome was regressed on both factors at once. The core factor behaved exactly as a disorder construct should: depression β = 0.909 on the BSMAS and β = 0.798 on the SMD, anxiety β = 0.746 and β = 0.751, loneliness β = 0.756 and β = 0.592, with life satisfaction and self-esteem negative throughout. Adolescents who endorse using social media to shift their mood, failing to cut down, feeling bad when they cannot use it, and losing sleep, school or relationships to it are not a statistical artefact. They are the signal.
Salience and tolerance are the noise. Once the core factor was held constant, higher scores on the two items covering preoccupation and escalating desire went with less depression, less anxiety, less loneliness and higher life satisfaction. That is a partial effect and not a raw one, and the supplementary table shows exactly how much the distinction matters. On raw scores the peripheral factor runs with distress rather than against it: .37 with depression, .36 with anxiety and .37 with loneliness on the BSMAS, and .38, .35 and .39 on the SMD. The negative signs appear only after the core factor enters the model, and the two factors correlate .74 on the BSMAS and .75 on the SMD. This is suppression under collinearity, and it licenses one reading and not another: among adolescents at the same level of core symptoms, the ones who are simply very involved are doing better. Involvement is not protective. The authors' own component table makes the point from the other end. Salience and tolerance map onto DSM-5 internet gaming disorder criteria 1 and 3, but against ICD-11 gaming disorder they are listed as having no direct analogue. ICD-11 asks about impaired control, displacement of other activities, and continuation despite harm. It never asks how much you think about the thing.
What to do with a score in the room
Stop reporting the sum. On the BSMAS, two of six items are engagement indicators; on the SMD, two of nine. A 15-year-old can reach a substantial total on preoccupation and escalating desire alone while denying every core criterion, and that adolescent is not the one these data single out. Look at which items were endorsed, and record them. The authors go further and recommend instruments built on ICD-11 rather than on substance-dependence criteria, naming the ACSID-11, which asks about impaired control and functional impairment rather than about preoccupation. They also cite a deliberately satirical study in which 69% of participants met "addiction" criteria for spending time offline with friends, which is the reductio the components model invites.
The mirror-image error is just as costly. Nothing here says that problematic social-media use is a moral panic or that it dissolves under proper criteria. It says the opposite about the core items: their association with depression was the strongest path in either model. An adolescent who sits below your cut-off but endorses conflict, withdrawal and failed attempts to cut down deserves a formulation, not reassurance that the score was fine. Cut-offs on these scales were calibrated on totals that include the two items now shown to run the other way, so a total near threshold tells you very little about which side of it the young person is actually on.
Practically: when a screening total comes back high, go back through the questionnaire item by item with the patient and separate the four core endorsements from the two engagement ones before you write anything in the notes.
A high total on a social-media addiction scale tells you how much of the scale a young person endorsed, not which part of it, and only one part of it travels with depression and anxiety.
This is a cross-sectional school survey of 2,761 adolescents in two Spanish cities, so it establishes no direction of effect and does not generalise to adults or to clinical samples. The negative coefficients for salience and tolerance are partial estimates from models containing a core factor with which they correlate .74 and .75, and the paper's own supplementary correlation table shows the raw associations running the other way, positively with depression, anxiety and loneliness; what the model demonstrates is suppression under collinearity, not a protective effect of involvement. The analysed sample rests on random-forest imputation covering the 15.86% of participants with incomplete data, which the authors report as not missing completely at random and with an out-of-bag error they themselves call not optimal, so they repeated the analyses by listwise deletion on 2,323 cases and report the same conclusions. The peripheral factor rests on only two items, no ICD-11-based instrument was administered for comparison, and for the SMD the two-factor model, though a significant improvement, still fell short of conventional fit thresholds.