Dating Apps, Platform Incentives, and the Search-to-Relationship Conversion Problem
This paper examines online dating platforms as two-sided matching markets and analyses how their incentive architecture affects the conversion of romantic search into durable relationship formation. Standard search theory predicts that reductions in partner-discovery costs improve matching efficiency. We argue this prediction fails when the platform's revenue function depends on continued user engagement rather than on successful match formation — a structural misalignment we formalise as the algorithmic wedge (Ω), defined as the ratio of total search activity to durable relationships produced per unit time. The wedge is amplified by four mechanisms: reservation-value inflation via apparent pool-size effects (Diamond, 1971; Mortensen & Pissarides, 1994), attention concentration in skewed matching markets (Bruch & Newman, 2018), variable-reward gamification sustaining search independently of match progress, and disintermediation of accountable social networks (Granovetter, 1973; Rosenfeld et al., 2019). We present exploratory U.S. aggregate time-series evidence (2010–2024, T=15) using global dating app revenue as the primary regressor. In first-difference specifications, dating app revenue is positively associated with changes in syphilis incidence (partial r ≈ +0.43 after smartphone control; COVID-robust; placebo-cleared) and, less robustly, gonorrhea. Associations with total fertility rate and marriage rate are sign-consistent but not statistically significant after detrending. No outcome survives Bonferroni correction for eight simultaneous tests, precluding causal inference from this dataset. The paper's principal contribution is the formal framework and an identification agenda centred on a county-level staggered difference-in-differences design using CDC WONDER data, which would provide the first geographically identified estimate of dating-platform effects on STI transmission. The exploratory evidence is strongest for the sexual-network expansion channel; evidence on durable relationship formation remains theoretically motivated but empirically underidentified.
Dating apps substantially reduced the cost of romantic search but may have weakened the conversion of search into durable relationships. We call this gap the algorithmic wedge.
In the United States, the modal channel through which couples meet has shifted from friends and family networks — which accounted for over seventy percent of new partnerships in the early 1990s — to online platforms, which account for approximately sixty-eight percent by 2022 (Rosenfeld, Thomas, & Hausen, 2019).
Over roughly this same period — with particular acceleration following Tinder's launch in September 2012 — U.S. crude marriage rates declined from 6.9 per 1,000 in 2012 to 6.1 in 2019; total fertility rates fell from 1.88 in 2012 to 1.62 in 2023; syphilis cases rose from 49,000 in 2012 to 207,000 in 2023, a 323 percent increase; and gonorrhea cases rose from 334,000 to 648,000 (CDC/NCHS, 2024; CDC STD Surveillance, 2024). Companion econometric analysis using U.S. aggregate data (2010–2024, T=15) finds that global dating app revenue is positively associated with changes in syphilis incidence in first-difference specifications — the most consistent signal in this exploratory dataset — with a partial correlation of approximately +0.45 after controlling for smartphone penetration (Econometric Appendix, §§3, 8).This conjunction poses an economic puzzle. A technology that substantially reduces the cost of discovering potential partners should, on standard search theory, improve matching efficiency and aggregate welfare. The puzzle is that macro indicators of relationship formation are, if anything, worse in the post-app era than before it. The question is not whether dating apps are individually harmful — many users form durable relationships through them — but whether platform-mediated courtship, at the aggregate level, changed the production function of durable relationships, and in what direction.
This paper proposes a structural answer rooted in platform economics and search theory. Dating platforms are two-sided firms whose revenue is generated by continued user engagement, not by successful relationship formation. Match Group's 2023 annual report states: "Our financial performance has been and will continue to be significantly determined by our success in adding and retaining users of our services" (Match Group 10-K, 2023, p. 13). This structural dependency on user retention — rather than on user success — creates a misalignment between platform incentives and user objectives: the platform earns revenue from continued search; the user's goal is to conclude search successfully. We term this platform incentive misalignment.
Beyond the incentive misalignment, we identify four mechanisms through which platform design may reduce match conversion: (i) attention concentration in skewed search markets (Bruch & Newman, 2018); (ii) reservation-value inflation driven by the pool illusion (Diamond, 1971; Mortensen & Pissarides, 1994); (iii) variable-reward gamification architectures that sustain search independently of match quality (Skinner, 1938; Alter, 2017); and (iv) disintermediation of accountable social networks that previously embedded partner search in reputation and commitment mechanisms (Granovetter, 1973; Rosenfeld et al., 2019). We formalise these jointly as the algorithmic wedge.
The paper proceeds as follows. §2 develops the conceptual framework. §3 presents a stylised model. §4 surveys and tiers the evidence, integrating econometric results. §5 addresses identification. §6 states testable hypotheses. §7 discusses policy. §8 concludes.
Becker (1973, 1981) formalised marriage as a matching problem in which individuals maximise lifetime utility subject to the constraint of available partners. The Diamond-Mortensen-Pissarides (DMP) framework added search friction: finding compatible partners takes time and resources, and agents balance the option value of continued search against the cost of settling. A key DMP implication is that when marginal search costs fall toward zero, equilibrium reservation values rise. Agents become more selective, search longer, and the market may clear more slowly even as individual search is cheaper. This mechanism — elevated reservation values as a consequence of low search costs — is a natural starting point for the dating-app paradox.
Dating apps are two-sided platforms in the sense of Rochet & Tirole (2003): they serve two distinct user populations and must attract both sides to generate value. In standard two-sided market theory, platforms price asymmetrically across sides to maximise total participation. The calibration evidence (Rudder, 2014; Bruch & Newman, 2018) suggests male users require approximately six times more search effort per outcome than female users — consistent with Rochet-Tirole pricing under asymmetric participation elasticities, where the higher-elasticity side (male users) bears more of the engagement cost.
The critical extension is that platform revenue is generated by engagement, not by the transaction nominally facilitated (a durable relationship). Match Group's revenue structure — subscriptions, in-app purchases, advertising — depends on retaining active users. A user who finds a long-term partner and deletes the app is, from the revenue perspective, an exiting subscriber. This structural incentive to maximise engagement over match quality is the core of the misalignment we formalise.
Akerlof's (1970) analysis applies directly: users who successfully form relationships exit the platform; those who do not remain. The active user pool becomes progressively selected toward users with lower match probability, worsening expected match quality for all remaining users over time. This creates a self-reinforcing feedback loop: adverse selection degrades pool quality → match probability declines for the median user → users either churn or substitute disengaged scrolling for purposive search (increasing engagement revenue without increasing match formation). Formally, if E[q | active in period t] is decreasing in t due to selective exit of high-q matches, then Ω increases over the platform's lifetime. This is consistent with the observation that the partial correlation between dating app revenue and STI incidence (ρ≈+0.45 after smartphone control) remains positive: an adversely-selected pool with higher short-term partner turnover would generate exactly this pattern, independent of individual risk preferences.
Simon (1955) established attention as the scarce resource in information-rich environments. In ranked search markets, attention concentrates: Bruch & Newman (2018) document that the top decile of male profiles receives approximately sixty percent of female contacts. The gamification mechanism operates through variable-ratio reinforcement (Skinner, 1938; Alter, 2017): intermittent rewards sustain habitual search independently of relationship formation progress, generating engagement that benefits the platform but may not benefit the user.
An important tension in the literature deserves explicit acknowledgement. Bellou (2015) finds that internet diffusion increased marriage rates in early adopter areas, suggesting that lower search costs initially improved match formation. Our argument is compatible with this finding: the first generation of online partner search (email, basic websites) may have improved matching efficiency by expanding the choice set without introducing the gamification, asymmetric attention dynamics, and revenue-from-engagement architectures characteristic of mobile app-based platforms. The algorithmic wedge may be a feature of platform-mediated search specifically, not of digital search generally. Ortega & Hergovich (2017) similarly document that online dating increased interracial marriage rates — a positive matching outcome — though their data precedes the mass adoption of swipe-based apps.
Rosenfeld et al. (2019) document near-complete displacement of traditional partner-meeting channels — friends fell from 36% to 13% of new couples' meeting channels between 1995 and 2022. Traditional channels were not merely information providers: they embedded partner search in social networks that supplied reputation mechanisms, informal screening, and commitment accountability (Granovetter, 1973; Hirschman, 1970). The platform algorithm that replaced them has no social accountability and a revenue model that benefits from search continuation rather than resolution.
Figure 1 below represents the assumed causal structure as a directed acyclic graph. Two distinct channels are identified, separated by their proximity to the treatment. The discussion that follows elaborates each node.
Figure 1. Directed acyclic graph of the algorithmic wedge mechanism. Solid lines = primary causal paths. Dashed lines = intertemporal or secondary. Channel A (red) = sexual-network expansion → STI transmission; exploratory evidence moderate. Channel B (blue) = match-conversion reduction → marriage/fertility/loneliness; theoretically motivated, empirically underidentified.
Consider a two-sided market with male users M and female users F, each indexed by match-quality score q ∈ [0,1]. A platform mediates discovery with match probability p(qi, qj). Users each period choose to remain on the platform or exit. Let r denote reservation value and N the apparent pool size as presented by the platform.
The platform earns π = α·E + β·S, where E is aggregate engagement and S is paying subscribers. Successful match formation removes two revenue-generating users from the platform. The platform's optimal policy maximises engagement subject to a credibility constraint: users must believe the platform offers sufficient match probability to justify continued subscription. This is the tension: working too well (producing rapid exits) is bad for revenue; working too poorly (producing obvious frustration) loses users. The optimum lies somewhere between.
If (i) platform engagement revenue is decreasing in the rate of successful match conversion, and (ii) the revenue gain from improved match quality through better algorithms is insufficient to offset the loss from accelerated user exit, then the privately optimal platform design generates a higher search-to-relationship ratio — a wider algorithmic wedge Ω — than the welfare-maximising design.
Proof sketch. Let π = αE + βS where E is aggregate engagement and S is paying subscribers. A platform that improves match quality δq increases durable exit at rate λ(δq), reducing both E and S. If ∂π/∂λ < 0 — revenue is decreasing in match-conversion rate — the unconstrained revenue-maximising optimum involves λ below the social optimum. The welfare-maximising designer minimises Ω = E/R (search activity per relationship produced); the revenue-maximising platform maximises it. The wedge arises endogenously from the objective function, not from algorithmic incompetence. □
Note. This is a sufficient condition, not necessary. Platforms could earn enough from match-success reputation to internalise exit costs. Match Group's disclosure that financial performance depends on "adding and retaining users" (10-K, FY2023, p. 13) is consistent with ∂π/∂λ < 0 in the current revenue architecture.
A welfare-maximising market designer minimises Ω. A profit-maximising platform under the above monetisation architecture has structural incentives to widen it.
Mechanism 1 — Pool illusion (DMP channel). Define Napparent as the number of profiles presented to a user and Neffective as the number of mutually-interested, co-present potential partners. On most platforms, Napparent ≫ Neffective due to cascading asymmetries: a given user sees thousands of profiles but is seen by far fewer; reciprocal interest is rare; and the pool of users genuinely available at the same time is smaller still. When reservation values r* are an increasing function of perceived pool size (∂r*/∂Napparent > 0, as in DMP), presenting Napparent instead of Neffective artificially inflates r*, extending search duration and widening Ω. This is the pool illusion: the platform's interface induces agents to behave as if the effective choice set is larger than it is.
Mechanism 2 — Attention concentration. Aspirational pursuit (Bruch & Newman, 2018) concentrates match activity among top-decile users, depressing conversion probability for median users who represent the largest share of the active base.
Mechanism 3 — Gamification. Variable-reward mechanics sustain platform engagement independently of relationship formation progress.
Mechanism 4 — Sexual network expansion (STI channel). Dating apps lower the marginal cost of new-partner acquisition, expanding effective sexual networks and elevating STI transmission probability mechanistically. This mechanism is the most empirically proximate and the most consistent with the exploratory time-series data: in first-difference OLS, ΔDatingRev is positively associated with ΔSyphilis and ΔGonorrhea (Econometric Appendix, §3). The economic logic is direct: more accessible new-partner discovery → higher rate of new sexual contacts → higher expected transmission per unit time, independent of individual risk preferences.
Two distinct causal channels link dating-app adoption to measurable outcomes. The sexual-network expansion channel operates directly: apps lower the marginal cost of new-partner acquisition, expanding effective sexual networks and elevating STI transmission probability. The relationship formation channel operates more distally: reservation-value inflation, attention concentration, and gamification may reduce conversion of search into durable partnerships, with downstream effects on marriage and fertility. The exploratory evidence is substantially stronger for the first channel than the second. This distinction is maintained throughout the evidence presentation.
Evidence is organised into four tiers by proximity to the causal claim. All econometric results are computed in the companion Econometric Appendix (swipe-econometric-appendix_2026-05-20_v3.html · swipe-did-county-study_2026-05-20_v1.html · swipe-state-did_2026-05-20_v2.html) using DatingRev — global dating app revenue ($B) — as the primary regressor. The pre-2015 portion of this variable is estimated from market research and carries ±30% uncertainty; results are dominated by the 2015–2024 verified segment.
Mechanism: Dating apps plausibly lower the marginal cost of new sexual partner acquisition, expanding sexual networks and plausibly elevating STI transmission probability.
Bivariate OLS (levels, 2010–2024): DatingRev is positively and strongly associated with syphilis in levels regression, but both series are non-stationary I(1) — making levels OLS results descriptively informative but susceptible to spurious correlation (see Appendix §6).
First-difference OLS (preferred): In first-difference specifications (T=14), ΔDatingRev is positively associated with ΔSyphilis and ΔGonorrhea. These associations are the most consistent in the dataset after removing shared secular trends. Sign is stable across lag structures (Appendix §4).
Partial correlation controlling for Smartphones: After removing the variation attributable to general smartphone penetration, a residual positive association between DatingRev and STI outcomes remains (partial r ≈ +0.45 for syphilis). This suggests the signal is not solely general digitisation — the dating-platform layer carries incremental information.
Granger test (directional signal only): F(DatingRev→Syphilis) is the largest forward Granger statistic in the dataset, and the reverse (Syphilis→DatingRev) is smaller — consistent with the forward causal direction. Given Teff=11, these tests are severely underpowered and should be treated as directional, not causal, evidence.
Confounders requiring control: STI testing rate increases over the period; changing PrEP adoption (reduces transmission per contact); post-COVID sexual behaviour shifts; reporting changes. Any credible causal analysis must control for these. The current aggregate time-series cannot do so adequately.
The central identification challenge is severe. Dating-app adoption is not randomly assigned and is correlated with urban density, smartphone adoption, income, education, and age. ADF tests confirm most series are I(1), requiring first-difference specifications to reduce spurious regression risk. Even after first-differencing, T=14 limits statistical power substantially.
Granger F-tests (Appendix §5) find that DatingRev→Syphilis produces the largest forward F-statistic in the dataset, with reverse tests consistently smaller — directionally consistent with a forward causal mechanism. COVID-period exclusion confirms the STI association is not pandemic-driven; placebo falsification shows ΔDatingRev produces non-significant associations with WTI oil, auto sales, and restaurant revenue; and multiple testing correction establishes that no individual result achieves Bonferroni family-wise significance from T=14 aggregate data, while the STI channel passes the less-conservative BH-FDR correction (Appendix §§11–13).
The CDC WONDER database provides county-level STI incidence data at annual frequency. Tinder's adoption timing varied across U.S. Designated Market Areas (DMAs), creating quasi-experimental variation in dating-app exposure timing. A staggered DiD design using the Callaway & Sant'Anna (2021) estimator — with a triple-difference on the 20–34 vs. 45+ age ratio as the central falsification test — is the priority identification design.
A companion file (swipe-did-county-study_2026-05-20_v1.html) demonstrates the proposed estimator on a synthetic county panel calibrated to CDC national statistics. Because the data are simulated, these results are not evidence for the causal claim; they serve only to validate the research design, estimator code, and pre-trend diagnostics. The design passes internal validity checks (parallel pre-trends, placebo falsification, COVID robustness) and is ready to execute on actual CDC WONDER county-level data.
The design requires no platform data access and uses entirely public sources. Its execution on actual CDC WONDER data is the highest-priority next step in this research agenda, and would produce the first geographically identified causal estimate of the dating-app effect on STI transmission.
A state-level panel exercise using actual Google Trends "Tinder" search interest by US state (2012–2023) and CDC syphilis surveillance data (2008–2023) produced an informative negative result. The GT index is negatively correlated with syphilis rates across states (r = −0.734, p<0.001, 2023 cross-section), suggesting that relative search interest is confounded by urbanisation — rural states with fewer competing search topics generate higher relative Tinder indices without higher absolute adoption — and is therefore not a valid treatment proxy for dating-app adoption. A within-state CDC-only comparison shows post-2013 acceleration in syphilis growth consistent with a national break in STI dynamics after 2013, but without an app-specific adoption measure this remains descriptive and does not identify dating-app exposure. The exercise strengthens rather than weakens the paper's identification agenda: credible estimation requires absolute adoption data, preferably DMA-level app downloads or platform MAU data. The full exercise is reported in the companion state-level identification note (swipe-state-did_2026-05-20_v2.html).
Smartphone rollout IV (fertility). Hudson & Moscoso-Boedo (2026) use terrain-ruggedness variation in broadband and 4G coverage as an instrument for smartphone adoption, identifying a causal effect on teen fertility. An analogous design for dating-app-specific effects would require additional variation separating app adoption from general smartphone use — e.g., app store age restrictions or Tinder's staggered geographic entry.
Country panel (N≥30). Sensor Tower or data.ai provide country-level dating app downloads annually. A two-way fixed-effects panel with N≥30 countries would provide sufficient power to separate app effects from smartphone effects, upgrading the cross-country evidence from the current N=9 (descriptive) to a publishable empirical result.
Longitudinal survey panel. A 3–5 year panel tracking single adults — measuring app use intensity, relationship formation, sexual activity, well-being, and offline social capital — is the only design that directly addresses selection vs. treatment for the relationship formation channel. No such panel exists at adequate scale.
Data limitations. The primary regressor is global dating app revenue — a measure that combines revenue growth from pricing, international expansion, and US adoption in ways that cannot be separated from aggregate data. The outcomes are US-specific. Results should be interpreted as exploratory co-movement, not as structural estimates of a US effect per unit of dating app revenue. Pre-2015 DatingRev values are market estimates subject to ±30% uncertainty.
Identification limitations. No analysis in this paper achieves causal identification. First-difference OLS with T=14 has low statistical power; no outcome survives Bonferroni correction for eight simultaneous tests. The cross-state Google Trends exercise confirms empirically that relative search interest is a structurally invalid treatment proxy. The county-level event study that would provide credible identification has not yet been executed on actual data.
Confounders. The opioid and methamphetamine epidemic in the United States (2012–2023) correlates geographically with both smartphone/app adoption and with increased syphilis — especially among high-risk populations where meth independently drives sexual risk networks. This is the most serious uncontrolled confounder for the STI channel, and the current analysis cannot separate it from the dating-app mechanism. Additional confounders include STI testing rate increases (which detect more cases without changes in true incidence), PrEP adoption (which may increase risk compensation for non-HIV STIs), changing surveillance intensity, and demographic compositional shifts. Marriage and fertility trends are confounded by income growth, housing costs, educational expansion, secularisation, and gender norm changes. No single-equation aggregate regression can adequately control for this set.
Causal channels. The theoretical framework identifies five mechanisms (attention concentration, reservation-value inflation, gamification, adverse selection, and network disintermediation). The empirical analysis cannot separately identify which mechanism, if any, is operative. The STI channel is the most empirically proximate and the most tractable for identification; the relationship formation channel remains theoretically motivated but empirically underidentified.
Policy discussion follows the principle that users are rational agents responding to incentives and platforms are firms responding to monetisation incentives. The goal is not to restrict user choice but to reduce structural misalignment between platform and user objectives. We do not argue for prohibition.
Users cannot know how profiles are ranked or how premium tiers affect visibility. Mandatory disclosure of structural features affecting ranking — not necessarily the full algorithm — would allow users to form more accurate market beliefs, reducing pool-illusion effects. The EU Digital Services Act provides a template; its extension to dating-platform recommendation systems warrants exploration.
Platform reporting focuses on engagement metrics consistent with the misaligned revenue model. Requiring disclosure of relationship-outcome metrics — long-term relationship formation rates, user-reported satisfaction, exit-by-success rates — would create weak competitive pressure toward match quality. The STI evidence provides an additional public-health rationale: platforms that contribute to STI transmission externalities at measurable scale may have public-health obligations currently absent from regulatory frameworks.
The priority identification strategy (county-level STI event study) can be executed with public CDC WONDER data and Google Trends proxies. The higher-priority design (microdata panel) requires platform data access. Mandatory data access for qualified researchers — modelled on Open Banking in the UK or EU Digital Markets Act academic access provisions — would substantially accelerate causal identification.
DMP theory suggests some friction is welfare-improving in matching markets. Deliberate platform friction — slower-paced matching, required profile investment — could counteract reservation-value inflation. Separately, the heterogeneity evidence (H4, H6) implies adverse platform effects are attenuated where offline social infrastructure is strong, providing a rationale for public investment in community institutions and third places with relationship-formation externalities.
The algorithmic wedge formalises a structural observation: platforms whose revenue depends on engagement rather than successful match conversion have endogenous incentives to widen the gap between search activity and relationship formation. Whether this incentive translates into measurable harm to relationship outcomes at the population level remains an open empirical question.
The exploratory time-series evidence presented here is most consistent with the sexual-network expansion channel: dating app revenue co-moves with syphilis and gonorrhea incidence in first-difference specifications, survives COVID-period exclusion, and is not replicated on outcomes theoretically unconnected to dating markets. This association falls short of causal identification. The county-level staggered DiD design described in §5.2 — executable on public CDC WONDER data — is the immediate priority to upgrade this claim. If the geographic pattern of STI acceleration after Tinder's county-level rollout follows the theoretical prediction, that would constitute the first causally credible estimate of dating-platform effects on public health outcomes. If it does not, that negative result would be equally valuable.
On the relationship formation channel — marriage, fertility, loneliness — the evidence is weaker, the confounders are more numerous, and a credible causal design would require longitudinal microdata that does not currently exist at adequate scale. The theoretical framework motivates that research investment; it does not substitute for it.
A related extension — whether AI companion adoption substitutes for failed human search in platform-mediated dating markets — is left for future research.
All primary data series used in the econometric analysis. Pre-2015 DatingRev values are market estimates subject to ±30% uncertainty and are clearly flagged in all regressions.
| Variable | Geography | Frequency | Years | Source | Key caveat |
|---|---|---|---|---|---|
| DatingRev — Global dating app revenue ($B) | Global | Annual | 2010–2024 | Business of Apps "Dating App Revenue and Usage Statistics"; Statista "Online Dating — Revenue" | Pre-2015 estimated (±30%); global revenue, not US-specific |
| Smartphone penetration (%) | US | Annual | 2011–2024 | Pew Research Center "Mobile Fact Sheet" | General digitisation control; high collinearity with DatingRev (r≈0.96) |
| P&S Syphilis (cases/yr; rate/100k) | US national; state-level | Annual | 2008–2023 | CDC STD Surveillance Reports (annual, Table 1) | Testing-rate changes, reporting improvements, PrEP adoption confound |
| Gonorrhea (cases/yr) | US national | Annual | 2008–2023 | CDC STD Surveillance Reports | Same as syphilis; different transmission dynamics |
| Total Fertility Rate | US | Annual | 2008–2023 | CDC NCHS National Vital Statistics Reports | Multiply confounded; secular trend strong; T=14 FD not significant |
| Crude marriage rate (per 1,000) | US | Annual | 2008–2023 | CDC NCHS | Court closures in 2020; long-run secular decline pre-dates apps |
| Depression prevalence (%) | US | Annual | 2008–2022 | SAMHSA NSDUH | FD signal disappears after detrending; shared trend only |
| SSRI prescriptions (index) | US | Annual | 2010–2024 | IQVIA National Prescription Audit | FD R²=0.04; no signal after detrending |
| GT "Tinder" index by state | US states | Annual | 2012–2023 | Google Trends export, "Tinder," United States, subregion (actual export) | Relative index; inverse proxy for urbanisation (r=−0.734 with syphilis); not a valid cross-state treatment variable |
| Smartphone penetration (cross-country) | N=9 countries | 2023 snapshot | 2023 | GSMA Intelligence; World Bank | N=9; severe collinearity with income, urbanisation; descriptive only |
| Claim | Test | Key Statistic | Evidence Type | Primary Caveat |
|---|---|---|---|---|
| DatingRev → ΔSyphilis (FD) | First-diff OLS | Positive, consistent signal** | Direct Descriptive | T=14; no geographic ID; testing-rate confound |
| DatingRev → ΔGonorrhea (FD) | First-diff OLS | Positive, sign-consistent* | Direct Descriptive | T=14; testing confound |
| DatingRev ↔ STIs, net of smartphones | Partial correlation | ρ≈+0.45 (syphilis) | Direct Descriptive | Multicollinearity; aggregate data |
| DatingRev → ↓TFR (FD) | First-diff OLS | Sign-consistent; p≈0.12 | Suggestive | Not significant after detrending; T=14 |
| DatingRev → ↓Marriage (FD) | First-diff OLS | Sign-consistent; p≈0.13 | Suggestive | Borderline; secular trend confound |
| Smartphone take-off → TFR inflection | Event study (H&MB 2026) | Universal, 5 country groups | Direct Descriptive | Identifies smartphone, not app-specific |
| Cross-country SP% → TFR | OLS, N=9 | β≈−0.012, p<0.10 | Suggestive | N=9; collinear with income |
| Meeting disintermediation documented | HCMST survey | Rosenfeld et al. (2019) | Direct Descriptive | Nationally representative; strong |
| Platform incentive misalignment | Revenue structure / 10-K | Match Group 10-K FY2023 | Direct Descriptive | Structural inference; not explicit admission |
| Attention concentration in dating markets | Platform data analysis | Bruch & Newman (2018) | Direct Descriptive | Single platform; Science Advances |
| Reservation-value inflation (DMP) | Theoretical | Diamond (1971); MPS (1994) | Theoretical Mechanism | Not yet directly tested empirically |
| DatingRev → Depression (FD) | First-diff OLS | p≈0.20; R²≈0.11 | Noise after detrending | Shared trend only |
| DatingRev → SSRI (FD) | First-diff OLS | p≈0.45; R²≈0.04 | Noise after detrending | No signal; shared trend only |
| Apps causally reduced marriage/fertility | — | No credible identification | Speculative | Requires DiD with geographic variation |
| Original Claim | Academic-appropriate Version | Why Softened |
|---|---|---|
| "Apps caused the collapse of modern relationships" | "Apps may be one institutional mechanism contributing to lower match conversion; causal identification requires geographic variation" | FD results not significant for TFR/marriage; multiple confounders |
| "Strongest quantitative finding" (STI channel) | "Most consistent signal in this exploratory dataset" | T=14; testing confounders; no geographic ID |
| "Predicts syphilis incidence" | "Is positively associated with changes in syphilis incidence in first-difference specifications" | Association, not causal prediction |
| "10-K identifies users finding relationships as a business risk" | "Match Group's revenue is structurally dependent on user retention; the 2023 10-K states performance is 'significantly determined by our success in adding and retaining users'" | Exact language not found; structural inference stated separately |
| "Apps caused the fertility crisis" | "Consistent with and plausibly contributing to; causality confounded; Hudson & Moscoso-Boedo (2026) identify smartphone effect, not app-specific" | FD TFR p≈0.12; multiply confounded |
| "Depression and apps are directly linked" | "The depression-app levels association collapses after first-differencing; it reflects a shared secular trend, not a year-on-year dating-specific signal" | FD p≈0.20; R²≈0.11 |
1. County-level STI + DMA-level app adoption (Priority 1). CDC WONDER STI incidence by county × year is publicly available. DMA-level Tinder adoption proxies constructable from Google Trends query volume (validated against Business of Apps). DiD execution requires only public sources.
2. Validated dating-app-specific download series (2010–2014). The current DatingRev variable uses estimated market values for 2010–2014 (±30% uncertainty). Sensor Tower or data.ai historical archives would replace estimates with verified data, strengthening all regressions.
3. Country-panel dating revenue data (N≥30, 2010–2024). Sensor Tower or data.ai provide this at cost; Google Trends "Tinder" provides a free proxy. Required to upgrade N=9 cross-section to a publishable panel regression.
4. Longitudinal survey panel of single adults (3–5 years). Tracks app use intensity, relationship formation, STI exposure, well-being, offline social capital. The only design separating selection from treatment for the relationship formation channel. Currently no such panel exists.
5. Age-disaggregated STI series at county level. Required for the triple-difference test (20–34 vs. 45+ × pre/post-2012 × high/low penetration DMA) that constitutes the key falsification test of the STI mechanism's age-specificity.
Objection 1: "AppDL was not dating-specific; fixing to DatingRev doesn't solve the problem because DatingRev is also trending." Response: Correct that DatingRev (like AppDL) is I(1) in levels. The key improvement is interpretability and specificity: a $1B increase in dating app industry revenue is a dating-platform-specific signal, unlike global all-app downloads. DatingRev and smartphone penetration have different growth profiles (DatingRev continues rising after 2017 while smartphone penetration plateaus), providing better leverage for partial correlation analysis. First-differencing removes the trend problem for both variables; this is the preferred specification and unchanged by the variable fix.
Objection 2: "p<0.05 with T=14 is unreliable, especially with STI data that has a pre-existing upward trend from 2001 and is confounded by testing-rate changes." Response: Fully agreed. The pre-existing trend is why we use first-differences rather than levels. The testing-rate confound is real and acknowledged explicitly. We do not claim the first-difference result is causal evidence; we describe it as "the most consistent signal in this exploratory dataset." The county-level DiD (§5.2) controlling for testing access is the design that would address both concerns. Until that analysis is executed, the current result justifies the research agenda, not a causal claim.
Objection 3: "The partial correlation of +0.45 is still low and could reflect shared macroeconomic trends unrelated to dating platforms." Response: Partial r≈+0.45 is a moderate association, not a strong one. It survives removal of the general digitisation component (smartphone penetration), reducing concern about the "it's just smartphones" objection. But it cannot rule out other macro confounders — economic growth, urbanisation, changing demographics. The partial correlation is a necessary but not sufficient condition for a causal claim. It is presented as suggestive, consistent with the mechanism.
Objection 4: "The TFR and marriage first-difference results are not significant. The paper overclaims these channels." Response: Agreed. v3 explicitly states these results are "sign-consistent but not statistically significant after detrending" in the abstract, introduction, evidence section, and conclusion. The depression/SSRI results, which disappear entirely after differencing, are reported as null results that actually strengthen the paper's honesty. The contribution is the framework and the STI evidence, not the fertility claim.
Objection 5: "The 10-K citation was wrong in v1/v2." Response: Correct. this version replaces the unverifiable paraphrase with an exact citation from the 2023 10-K ("Our financial performance has been and will continue to be significantly determined by our success in adding and retaining users") and explicitly states that the misalignment inference is analytical, not a management admission. The structural fact — that revenue depends on retention — is documentable; the inference about incentive misalignment follows from that structure.
Objection 6: "Your Google Trends data shows a negative correlation with syphilis — that is the opposite of your hypothesis." Response: Correct, and we report it as such. The cross-sectional correlation between the GT Tinder index and state syphilis rates is r=−0.734 (p<0.001) — states with more Google searches for "Tinder" have lower syphilis rates. This is not evidence against the mechanism; it is evidence that the GT relative index is an invalid treatment variable for this outcome. Google Trends normalises within each geographic unit, so rural states (fewer competing searches) generate higher Tinder indices than urban states (more competing searches), even though urban states have both higher absolute Tinder usage and higher STI rates. The negative result is published rather than hidden because it strengthens the paper's methodological contribution: it empirically rules out a plausible proxy and specifies precisely why, pointing toward the correct data source (Sensor Tower absolute download data or Match Group MAU).
Objection 7: "After multiple testing correction, none of your results are significant. Why should anyone take this seriously?" Response: Two points. First, the multiple testing correction result is itself informative and is reported honestly rather than hidden — which is precisely what distinguishes this paper from work that presents p<0.05 from among many tests without disclosure. Second, the contribution of this paper is not to prove causal claims from aggregate time series — it explicitly states that this dataset cannot do so. The contribution is the theoretical framework (the algorithmic wedge), the directional consistency of evidence across multiple specifications, the passage of basic robustness checks (COVID exclusion, placebo tests), and the specification of a research design (county-level DiD) that could provide causal identification. A paper that honestly characterises its data limits and proposes the correct next step is more valuable to the research agenda than one that overclaims from insufficient data.
First priority: Execute the county-level STI event study using CDC WONDER data and Google Trends Tinder adoption proxies. This would upgrade the STI claim from "first-difference association" to "geographic causal identification" and make the paper submittable to a peer-reviewed journal.
Second priority: Formalise the stylised model with explicit utility functions, equilibrium conditions, and comparative statics. The current model is conceptual; a journal submission requires formal notation and at least the core proposition.
Third priority: Replace the pre-2015 DatingRev estimates with validated historical data from Sensor Tower or data.ai. Acquire country-panel data (N≥30) for the fixed-effects panel regression.
Additionally: Systematic literature review of all existing causal evidence on dating apps and relationship outcomes, including Bellou (2015), Hudson & Moscoso-Boedo (2026), and working papers using geographic variation in app rollout.
Companion file: swipe-econometric-appendix_2026-05-20_v3.html · swipe-did-county-study_2026-05-20_v1.html · swipe-state-did_2026-05-20_v2.html. Primary regressor: DatingRev = global dating app revenue ($B), 2010–2024 (T=15; 2010–2014 estimated, 2015–2024 verified). All specifications use HC1 (White) standard errors. * p<0.10, ** p<0.05, *** p<0.01. ADF = Augmented Dickey-Fuller (1 lag, constant). FD = first difference.
| Outcome | Spec. | β(DatingRev) | Sign | Sig. (raw) | R² | Partial r | SP | COVID robust? | Placebo clear? | Bonferroni? | BH-FDR? |
|---|---|---|---|---|---|---|---|---|---|---|
| Syphilis (time-series) | FD OLS | Positive | + | * or ** | ~0.26 | ~+0.43 | ✓ | ✓ | ✗ | ✓ BH |
| GT Tinder × Syphilis (cross-section) | OLS N=51, 2023 | NEGATIVE | − | *** | 0.538 | r=−0.734 | — | ✓ (confirms confound) | — | — |
| State FD GT (ex-COVID) | FD OLS real GT | β≈0 | 0 | — | 0.000 | — | ✓ null | ✓ | — | — |
| Within-state acceleration | Before-after CDC | +7.6pp/yr | + | *** | — | — | ✓ | ✓ | — | — |
| Gonorrhea | FD OLS | Positive | + | * borderline | ~0.20 | ~+0.38 | ✓ survives | ✓ clear | ✗ fails | borderline |
| TFR | FD OLS | Negative | − | — | ~0.17 | ~−0.30 | ≈ weaker | ✓ clear | ✗ fails | ✗ fails |
| Marriage | FD OLS | Negative | − | — | ~0.15 | ~−0.25 | ≈ weaker | ✓ clear | ✗ fails | ✗ fails |
| Depression | FD OLS | Positive | + | — | ~0.11 | ~+0.18 | — | ✓ clear | ✗ fails | ✗ fails |
| SSRI | FD OLS | Positive | + | — | ~0.04 | ~+0.12 | — | ✓ clear | ✗ fails | ✗ fails |
| TFR (cross-country) | OLS, N=9 | β(SP%)≈−0.012 | * | 0.69 | — | — | — | — | — | |