{"id":"855549a3-2fd2-4d97-a7d3-417a230bf2bf","arxiv_id":"1908.02614","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Dynamic social network features predict depression and anxiety more accurately than static network or non-network features on the same smartphone data.","lead":"This paper tests whether social interaction data that changes over time, modeled as a dynamic network, predicts depression and anxiety better than the same data collapsed into a static network. The authors report that dynamic network features beat static and non-network baselines on data from a college smartphone study.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported adjusted p-values are impossible under the described Wilcoxon signed-rank test with five paired runs, undermining the central significance claim.","rationale":"The reader's verdict identified the significance testing as a major problem in the rationale but selected the SMS-proxy assumption as the weakest assumption. The SMS-proxy issue is a concern about generalization of the mechanism, not about the internal validity of the reported statistical claim. The Wilcoxon signed-rank problem is more load-bearing: it directly invalidates the paper's central quantitative claim as stated. The described test cannot yield adjusted p-values below 0.05 with five paired runs and multiple comparisons. This is not a matter of disagreement with consensus or a minor methodological choice; it is a mathematical impossibility. The claim that 'all three dynamic network feature-based classification models are significantly more accurate' is therefore unsupported, and the paper's conclusion that dynamic network data have more predictive power than static or non-network equivalents is not established. While the underlying empirical pattern might be real, the paper as written does not provide the evidence needed to sustain it. A conditional verdict could be appropriate if the authors supplied the missing raw data and re-analysis, but as submitted the central claim is not credible. Hence, REJECT is the appropriate verdict for this preprint in its current form.","tokens_in":12389,"tokens_out":3153,"duration_ms":33727,"concrete_test":"Obtain the raw per-run performance values (e.g., precision for each of the five runs for each of the eight models) for depression and anxiety. Recompute the pairwise Wilcoxon signed-rank tests exactly as described using the five paired differences, and then apply the same FDR adjustment across the reported number of comparisons. If the minimum possible adjusted p-value across all comparisons exceeds 0.05, as it must with n=5 and the stated number of tests, Section 3.3's significance claim is false. Alternatively, ask the authors to provide the exact test variant and per-run data; re-run the analysis with a paired t-test or bootstrap over the five runs to check whether any dynamic-versus-static difference survives without relying on the impossible Wilcoxon p-values.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim in Section 3.3—that all three dynamic network feature-based models are significantly more accurate (adjusted p-value<0.05) than static, DMF, non-network, and random models—rests entirely on the statistical test described in Section 2.2e. The authors state: 'When comparing performance of any two predictive models in terms of their five paired runs of 5-fold cross-validation, we evaluate the statistical significance of their performance difference by using the Wilcoxon signed-rank test.' With only five paired runs, the Wilcoxon signed-rank test has n=5 differences. The minimum achievable two-sided exact p-value is 0.0625 (1/16), which already exceeds the 0.05 threshold before any multiple-testing correction. Even for a one-sided test, the minimum exact p-value is 0.03125, but after Benjamini-Hochberg adjustment across the multiple pairwise comparisons (each dynamic model is compared against at least six other models: two static, DMF, raw SMS, random, plus possibly other dynamic models), the smallest possible adjusted p-value is at least 0.03125 × m, where m ≥ 6, giving >0.05. Therefore, the reported 'adjusted p-value<0.05' cannot be produced by the procedure described. This is an internal mathematical inconsistency, not a matter of external validity. If the authors used a normal approximation or a different test, they did not disclose it, and the claim as stated is unsupported. The paper's headline conclusion—that dynamic network features have superior predictive power—depends on this significance claim, so the finding is not currently established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper uses NetHealth smartphone SMS logs (31 weekly snapshots, August 2015-August 2016) for 576 iPhone users, of whom 274 have depression/anxiety survey labels, to ask whether dynamic social network features are more predictive of mental health than static network or non-network alternatives. The authors perform three tasks: comparing centrality magnitudes and fluctuations between depressed/anxious and non-depressed/non-anxious individuals; clustering individuals by their evolving centrality profiles and testing enrichment in mental health outcomes; and training logistic regression classifiers on three dynamic features (dynamic centralities, dynamic graphlet degree vectors, graphlet orbit transitions), two static features (static centralities, static GDV), raw SMS counts, a recommender-system baseline (DMF), and a random-guess model. The central claim, stated in Section 3.3 and the abstract, is that all three dynamic-network feature models are significantly more accurate (adjusted p-value < 0.05) than the static, DMF, non-network, and random models for both depression and anxiety across precision, recall, F1, and accuracy, based on 5-fold cross-validation repeated five times.","tokens_in":12648,"tokens_out":4628,"duration_ms":99839,"significance":"If the central claim were fully supported, the paper would make a useful contribution by showing, on the same data and with the same classifier, that temporal network features add predictive value over static and non-network features for mental health outcomes. The paper has real strengths: the comparison is carefully framed so that dynamic and static features are extracted from the same underlying SMS data; the exploratory Tasks 1 and 2 provide supporting evidence that network position is associated with depression and anxiety; and the dynamic graphlet and GoT features are nontrivial, complementary, and of current methodological interest. The supplementary material is promised and the study design is in principle reproducible. However, the headline significance claim rests on a statistical test that, as described in Section 2.2e, cannot produce the reported adjusted p-values. This is an internal mathematical inconsistency, not merely an external-validity concern, and it directly affects the paper's central conclusion.","major_comments":[{"comment":"The Wilcoxon signed-rank test as described cannot yield adjusted p-values below 0.05 for the claims made. With five paired runs of 5-fold cross-validation, the sample size for the signed-rank test is n = 5. The smallest exact two-sided p-value is 1/16 = 0.0625, already above 0.05 before any multiple-testing correction. Even using a one-sided test, the smallest exact p-value is 1/32 = 0.03125, and after Benjamini-Hochberg adjustment across the at least six pairwise comparisons each dynamic model is said to win (two static models, DMF, raw SMS, random, plus comparisons among dynamic models), the smallest achievable adjusted p-value is at least 0.03125 x 6 = 0.1875. Unless the authors used a different procedure (e.g., a normal approximation, a different test, or unadjusted p-values), the statement in Section 3.3 that all three dynamic models are significantly more accurate with adjusted p-value < 0.05 is unsupported. The paper should state exactly which test, which number of comparisons, and which p-value adjustment was used, and it should report the pairwise p-values and effect sizes. If the described five-run protocol is retained, the central claim cannot be made as stated; more cross-validation repeats or an alternative analysis are needed.","section":"Section 2.2e / Section 3.3"},{"comment":"The manuscript does not report the actual adjusted p-values for any pairwise model comparison. The statement 'significantly more accurate (adjusted p-value<0.05)' is asserted globally for all three dynamic models, both outcomes, and all four evaluation measures, but no table or figure shows the comparison counts, the raw p-values, or which comparisons were included in the multiple-testing correction. Without these details, the central comparison cannot be verified or reproduced. Please provide a complete table of pairwise p-values and adjusted p-values, or otherwise make the full comparison results available.","section":"Section 2.2e / Section 3.3"},{"comment":"The temporal relationship between the SMS-derived network features and the survey-derived mental health labels is not specified. The study period runs from August 2015 to August 2016, and the features are computed over the 31 weekly snapshots of that same period, but it is not stated when the depression and anxiety surveys were administered relative to the SMS data. If the labels were collected during or at the end of the same interval, the model performs concurrent classification rather than prospective prediction. This affects the interpretation of 'predict' in the title and abstract, and it should be clarified or the language should be tempered accordingly.","section":"Section 2.1 / Section 4"}],"minor_comments":[{"comment":"The choice of the 'best' pre- versus post-PCA version of each feature appears to be made based on performance in the same five cross-validation runs used for evaluation. If the PCA dimensionality or the pre/post selection is chosen using test-fold performance, this can introduce selection bias. Please describe the PCA dimensionality selection procedure and clarify whether it was nested inside the cross-validation loop.","section":"Section 2.2d / Section 3.3"},{"comment":"The statement that the three dynamic models 'perform similarly, with none of them having perfect performance' is not a statistical comparison. The paper should either test differences among the dynamic models or explicitly refrain from interpreting their relative performance.","section":"Section 3.3"},{"comment":"The claim of being 'the first' to develop predictive mental-health models from dynamic social network data should be qualified in light of the concurrent work [26] acknowledged in Section 1. The current wording is stronger than the body of the paper supports.","section":"Abstract / Introduction, Table 1"},{"comment":"The paper equates SMS logs with 'social interactions' and interprets the Task 1 results as validating smartphone data as a proxy for real-world friendships. This is an external-validity assumption, not established by the data. A brief limitation statement and, if feasible, a sensitivity analysis using another communication channel (e.g., call logs) would strengthen the paper.","section":"Section 2.1 / Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The statistical issue in Section 2.2e is the main barrier to acceptance. If the authors cannot produce valid adjusted p-values under the described protocol, the headline conclusion in Section 3.3 must be weakened or the experiments must be rerun with a sufficient number of cross-validation repeats. The paper is otherwise within scope and the study design is thoughtful, so I would encourage a revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the one thing to know: this is a useful paper with a real empirical question—do temporal network features improve mental health prediction over static network features—but the central statistical claim is not supported as described, because the test they report cannot produce the p-values they claim. That doesn't mean the result is wrong; it means the evidence as written doesn't establish it.\n\nWhat's new: first to build predictive models for depression and anxiety from dynamic social networks and to compare dynamic vs static network features on the same data. That comparison is a fair and useful contribution. They use three dynamic features (centrality trajectories, dynamic GDV, graphlet orbit transitions) against static counterparts and a non-network SMS feature, all under the same logistic regression. The exploratory analyses showing depressed/anxious individuals occupy more peripheral and more fluctuating positions are sensible and consistent with prior work. The writing is clear and the feature construction is reproducible in principle.\n\nWhere the soft spots are:\n1. The significance testing. In Section 2.2e, they compare models via Wilcoxon signed-rank test on five paired runs of 5-fold CV. With n=5, the minimum two-sided exact p-value is 0.0625, above 0.05 before any multiple-testing correction. Even a one-sided test gives a minimum of 0.03125, and after Benjamini-Hochberg correction across at least six pairwise comparisons per dynamic model, the smallest adjusted p-value is >0.05. So the reported adjusted p-value<0.05 cannot come from the described procedure. This is an internal inconsistency, not a matter of external validity. The authors need to disclose the exact test, report raw per-run values, or use a valid resampling approach with more repeats. It's fixable, but it's load-bearing.\n2. The SMS proxy assumption. They treat SMS logs as social interactions and invoke social support/isolation as the mechanism. If texting doesn't track real relationships, the generalization to actual social connectedness is weak. They cite existing survey-based correlations, which softens it, but it's a limitation they should state explicitly.\n3. No code or data. The supplementary has details, but no public implementation. Given the statistical question, reproducibility would help materially.\n\nThe citation pattern is fine; self-citations to prior NetHealth work are appropriate as baselines, and they acknowledge concurrent work [26] even though it came to their attention late.\n\nBottom line: the paper is worth engaging and deserves a serious referee, but the significance claim needs to be fixed before I'd trust the headline. The core comparison is plausible and the dynamic features likely do help, but 'likely' is not 'significantly' without valid statistics. I'd ask for a revision with corrected testing and more transparency.","headline":"Dynamic network features likely improve mental health prediction, but the reported Wilcoxon p-values are impossible with five paired runs, so the central claim needs a statistical fix before it is credible.","tokens_in":13222,"tokens_out":3495,"would_cite":false,"duration_ms":31971,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Weekly changes in a person's texting network predict depression and anxiety more accurately than static network features, non-network phone use, a recommender-system baseline, or random guessing, on the same college dataset.","keywords":["dynamic social networks","mental health prediction","depression","anxiety","graphlet degree vectors","graphlet orbit transitions","smartphone sensing","logistic regression"],"falsifier":"Shuffle the order of the 31 weekly snapshots separately for each individual and rerun the dynamic-centrality classifier; if prediction accuracy does not drop, temporal ordering is not what carries the signal, and the static network contains the same information.","tokens_in":12176,"feed_emoji":"🧠","tokens_out":10295,"duration_ms":98594,"temperature":0.7,"pith_summary":"This paper asks whether the timing and ordering of a person's social interactions carry mental-health information that a static snapshot of the same social network misses. Using a year of weekly SMS interaction snapshots from 576 college students, the authors build three families of dynamic network features—evolving centrality ranks, dynamic graphlet degree vectors, and graphlet orbit transitions—and train a logistic-regression classifier for depression and anxiety on each. They report that every dynamic-feature model significantly beats its static-network counterparts, a raw SMS-volume model, a recommender-system baseline, and random guessing, at adjusted p<0.05 across precision, recall, F1, and accuracy. If that holds, passive phone metadata could screen for depression and anxiety without a single survey response.","feed_headline":"Dynamic social networks predict mental health better than static ones","feed_subtitle":"A smartphone study of 576 students found that week-to-week shifts in texting ties outperform static friend counts in predicting depression…","key_machinery":"The load-bearing object is the dynamic social network, built as 31 weekly snapshots from SMS logs, with an edge between two people in a week if they exchanged at least one text. Three node-level feature families carry the argument: dynamic centrality, eight centrality measures converted to ranks in each snapshot and concatenated into a 248-dimensional vector; dynamic graphlet degree vectors, which count how often a node takes part in each small connected temporal subgraph type; and graphlet orbit transitions, which count how often a static graphlet type around a node changes into another type between consecutive weekly snapshots. The static comparison uses the same centrality measures collapsed over all weeks and static graphlet degree vectors, and the fairness argument is that each dynamic feature has a direct static counterpart, except graphlet orbit transitions, which have none. Logistic regression on PCA-reduced features is the shared classifier.","core_discovery":"The paper's central claim is that dynamic social network data are more predictive of an individual's depression and anxiety status than static social network data, non-network phone-use data, or a recommender-system baseline, on the same data and under the same classifier. For each of three dynamic feature sets—evolving centrality ranks, dynamic graphlet degree vectors, and graphlet orbit transitions—a logistic regression model significantly outperforms its static counterparts (static centralities and static graphlet degree vectors), the DMF model, a raw SMS-volume model, and random guessing, with adjusted p<0.05 for both depression and anxiety and for all four evaluation measures (precision, recall, F1, accuracy). The dynamic models perform similarly to one another rather than any one dominating, which the paper interprets as likely complementarity. Two preliminary analyses show that depressed and anxious individuals occupy more peripheral network positions and have positions that fluctuate more over time.","pith_inferences":["Editorial extension: the dynamic-centrality feature has 248 dimensions while its static counterpart has 8, so a static feature expanded to comparable dimensionality before PCA would isolate whether the gain comes from temporality or from richer descriptive capacity.","Editorial extension: since only SMS channels were used, testing the same models on call logs, co-location, or messaging-app metadata would show whether the dynamic-network signal is specific to texting or generalizes to other interaction channels.","Editorial extension: the observational design cannot separate selection from influence—people who become depressed may withdraw from texting, or withdrawal may precede depression—so panel-style temporal modeling is needed to test which direction dominates."],"forward_implications":["Mental-health screening could run on passive smartphone metadata, with risk scores updated week by week as a person's network position shifts.","Because the three dynamic feature families reach similar accuracy with different errors, an ensemble of all three is the natural route to further gains.","Depression and anxiety are marked not only by having fewer contacts but by more volatile contact patterns; models that ignore time collapse this distinction.","Static-network recommender-system models for mental health should be re-built on temporal network data to recover information lost by aggregation."],"supporting_citations":[{"why":"supplies the longitudinal smartphone data-collection protocol that produced the SMS logs.","marker":"[15]"},{"why":"documents the wearable/smartphone data collection that defined the study cohort.","marker":"[39]"},{"why":"defines dynamic graphlets and dynamic graphlet degree vectors, one of the three dynamic feature families.","marker":"[22]"},{"why":"defines graphlet orbit transitions, the temporal transition feature used as the third dynamic class.","marker":"[5]"},{"why":"defines graphlet degree signatures used for the static graphlet baseline.","marker":"[33]"},{"why":"provides the DMF recommender-system model used as a static-network baseline.","marker":"[13]"},{"why":"is the authors' earlier heterogeneous-network mental-health prediction study whose static framework the dynamic features are benchmarked against.","marker":"[28]"}],"fun_headline_variants":["Week-to-week texting ties forecast depression and anxiety","Shift in texting habits predicts mental health better than friend counts","Dynamic social ties beat static ones in predicting depression","Evolving networks outperform static in mental health prediction","Texting tie changes predict depression more accurately than static data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that SMS logs reliably measure real-world social relationships, so that a person's texting centrality and its instability reflect genuine social connectedness and isolation; if texting volume is dominated by logistics or platform habits, the predictive signal would not transfer to the social-support mechanism the paper invokes.","fun_headline_variants_meta":{"raw":{"variants":["Week-to-week texting ties forecast depression and anxiety","Shift in texting habits predicts mental health better than friend counts","Dynamic social ties beat static ones in predicting depression","Evolving networks outperform static in mental health prediction","Texting tie changes predict depression more accurately than static data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000567,"raw_usage":{"total_tokens":2643,"prompt_tokens":860,"completion_tokens":1783,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":1708}},"tokens_in":476,"tokens_out":1783,"duration_ms":12193,"temperature":1.0,"reasoning_tokens":1708,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:59:17.200177+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Shuffle the order of the 31 weekly snapshots separately for each individual and rerun the dynamic-centrality classifier; if prediction accuracy does not drop, temporal ordering is not what carries the signal, and the static network contains the same information.","supporting_citations":[{"cited_title":"Faust, R","cited_arxiv_id":null,"evidence_quote":"supplies the longitudinal smartphone data-collection protocol that produced the SMS logs."},{"cited_title":"Purta, S","cited_arxiv_id":null,"evidence_quote":"documents the wearable/smartphone data collection that defined the study cohort."},{"cited_title":"Hulovatyy, H","cited_arxiv_id":null,"evidence_quote":"defines dynamic graphlets and dynamic graphlet degree vectors, one of the three dynamic feature families."},{"cited_title":"Apar´ ıcio, P","cited_arxiv_id":null,"evidence_quote":"defines graphlet orbit transitions, the temporal transition feature used as the third dynamic class."},{"cited_title":"Milenkovi´ c and N","cited_arxiv_id":null,"evidence_quote":"defines graphlet degree signatures used for the static graphlet baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the DMF recommender-system model used as a static-network baseline."},{"cited_title":"Heterogeneous network approach to predict individuals' mental health","cited_arxiv_id":"1906.04346","evidence_quote":"is the authors' earlier heterogeneous-network mental-health prediction study whose static framework the dynamic features are benchmarked against."}],"review_version":1}