Network structure of gambling-related cognitions and risk behaviors among sports bettors in the Democratic Republic of the Congo: a cross-sectional EBICglasso network analysis (N = 2261) Authors: Degani Banzulu Bomba, Mechac Koffi Nkazi, Kevy Nguya Makonda, Christian Badimanya Albert, Merci Lusala Lutsasa, Jacques Benangindu Kikumpa Published: 2026 ABSTRACT AbstractBackground and aims: Sports betting is expanding rapidly in sub-Saharan Africa, yet the network structure of gambling-related cognitions and risk indicators remains understudied. This study analyzed these cognitions and their financial risk, psychological, and social correlates among DRC sports bettors.Methods: A cross-sectional sample of 2261 sports bettors (85.9% male; mean age = 25.59 years, SD = 5.79) was recruited at Kinshasa betting venues (March–May 2026). The Gambling-Related Cognitions Scale (GRCS; α = 0.924) was administered alongside financial, psychological, and social measures. Network structure was estimated using thresholded EBICglasso (γ = 0.5). Bridge centrality and sex-based invariance were assessed via NetworkComparisonTest (100,000 permutations).Results: The thresholded network retained 150 edges (density = 0.302). Guilt and G1 were the most central nodes. Bridge analysis identified guilt and social pressure as bridge nodes. Seven items showed zero bridge strength. Item-level network structure showed no global sex difference (M = 0.195, p = .121), whereas a 14-node sensitivity analysis detected an omnibus structural difference (M = 0.217, p = .022) not localized to any edge. After Bonferroni and FDR corrections, applied separately to 496 edge-level and 90 centrality-level tests, one edge (G1–G15), but no centrality index, differed by sex.Conclusions: This is among the first network analyses of gambling-related cognitions in sub-Saharan Africa, with Gambling Expectancies as the most connected cognitive domain, G1 as a central hub, and guilt and social pressure as bridge nodes. These findings extend the Interaction of Person-Affect-Cognition-Execution (I-PACE) model and inform intervention research.Keywords: Gambling-related cognitions, Network analysis, Sports betting, Sub-Saharan Africa, Bridge centrality, Behavioral addiction Introduction Sports betting has emerged as one of the fastest-growing forms of gambling in sub-Saharan Africa, with the Democratic Republic of the Congo (DRC) witnessing rapid proliferation of betting venues and mobile platforms, driven by smartphone penetration and youth unemployment (Bitanihirwe et al., 2022). In a six-country survey of 3879 youth aged 17–35, 54% had gambled, with estimates ranging from 42% (Ghana) to 76% (Kenya) (Ssewanyana & Bitanihirwe, 2018); lifetime sports-betting participation in East Africa has ranged from 32.3% (Kenya) to 73% (Uganda) (Bitanihirwe et al., 2022). Gambling disorder prevalence estimates vary considerably by diagnostic threshold, from approximately 0.2–0.3% for clinically significant cases under DSM-5 (American Psychiatric Association, 2013) to several percent under broader definitions. The cognitive mechanisms underlying gambling disorder in sub-Saharan Africa remain largely unexamined, and no network-based study of gambling cognitions has yet been conducted in this rapidly growing betting market. In Kinshasa, young adults often wager daily under strong peer pressure (Ssewanyana & Bitanihirwe, 2018); guilt and anxiety, commonly reported, may paradoxically reinforce betting, perhaps when losses are seen as threatening familial obligations. The Interaction of Person-Affect-Cognition-Execution (I-PACE) model (Brand et al., 2016, 2019) posits that predisposing cognitive factors interact with affective states, executive functioning, and situational cues to maintain addictive cycles. The Gambling-Related Cognitions Scale (GRCS; Raylu & Oei, 2004) operationalizes these cognitions across five dimensions: Gambling Expectancies, Illusion of Control, Predictive Control, Inability to Stop, and Interpretive Bias. Traditional variable-centered approaches treat these dimensions as homogeneous blocks, obscuring individual items' differential roles and their connections to behavioral and emotional outcomes. Gambling-related cognitions merit attention for three reasons: they are among the most consistent correlates of gambling severity; they are modifiable, and thus the operative target of cognitive-behavioral treatment and public education; and they can be addressed through brief, low-cost, non-specialist-delivered formats realistic in settings without addiction treatment infrastructure — making a cognition-centred model the appropriate entry point here. Network analysis offers a complementary paradigm — increasingly applied to behavioral addictions research (Fried et al., 2017; Zarate et al., 2022) — in which constructs are modeled as interacting nodes rather than reflective indicators of a latent factor (Borsboom et al., 2021; Borsboom & Cramer, 2013). Bridge nodes, at the intersection of distinct communities, serve as conduits for cross-domain activation (Jones et al., 2021). The approach offers what a latent-variable model does not: mutually interacting rather than exchangeable indicators, differential item positions, bridge analysis across a priori domains, and testable intervention priorities — though node position in a cross-sectional network remains descriptive and does not license causal inference about intervening on that node (Burger et al., 2023). The present study addresses this gap by examining the GRCS network in a context where limited regulatory frameworks, intense social pressure, and absent treatment infrastructure represent amplifiers largely absent from Western samples. This study pursued three objectives: (1) to estimate the network structure of gambling-related cognitions, financial risk behaviors, psychological impact, and social influences using EBICglasso; (2) to identify central and bridge nodes; and (3) to test sex-based invariance. We hypothesized that (H1) Inability to Stop and Predictive Control items would show the highest centrality; (H2) nodes with high bridge centrality would connect the GRCS community to financial, psychological, and social communities; and (H3) global network structure would be invariant across sex. This study examined 2261 bettors in a non-WEIRD, sub-Saharan African setting. Methods Participants and procedure This cross-sectional study recruited participants face-to-face at sports betting venues in Kinshasa. Trained research assistants systematically approached every third patron, explained study objectives, and obtained written informed consent. Inclusion criteria were: (a) active sports bettor status (at least one bet placed in the preceding 30 days); (b) age 18 years or older; and (c) written informed consent. Following ethical approval on March 23, 2026, data collection occurred between late March and May 2026. Games of chance and prediction contests are legally regulated in the DRC under a State licensing regime administered through the Société Nationale de Loterie (SONAL); licensed operators run street-level shops, kiosks, and mobile platforms in Kinshasa. Persons younger than 18 are prohibited from admission to gambling venues (Arrêté ministériel n° 041/MJS/CAB/2100/01/2011). The questionnaire was interviewer-administered, read aloud individually and privately rather than self-administered, to avoid literacy-related selection bias. Research assistants were physicians in postgraduate psychiatry specialty training (Diplôme d'Études Spécialisées, DES) at a university-affiliated neuropsychiatric center in Kinshasa, trained in consent procedures, neutral administration, and confidentiality. French, using the published GRCS adaptation (Grall-Bronnec et al., 2012), was the primary language; a standardized French–Lingala oral guide, reviewed by bilingual team members, was available for clarification but did not undergo independent back-translation and is not a validated Lingala version of the GRCS. The final sample comprised 2261 participants (men: n = 1943, 85.9%; women: n = 318, 14.1%); see Table 1 and Section 3.1 for demographic and gambling-history characteristics. Table 1Descriptive Statistics for GRCS Subscales, Financial Risk Behaviors, Psychological Impact, and Social Influences (N = 2261).VariableTotal (n = 2261)Men (n = 1943)Women (n = 318)pAge, years25.59 (5.79)25.79 (5.99)24.35 (4.15)<0.001Age at gambling onset, years20.61 (5.06)20.57 (5.22)20.86 (3.92)0.235 GRCS subscales Gambling Expectancies (GE)17.36 (5.24)17.24 (5.16)18.05 (5.69)0.018 Illusion of Control (IC)12.82 (5.47)12.74 (5.35)13.35 (6.12)0.096 Predictive Control (PC)26.57 (6.31)26.53 (6.13)26.83 (7.32)0.485 Inability to Stop (IS)21.32 (7.43)21.46 (7.26)20.47 (8.36)0.046 Interpretive Bias (IB)18.43 (4.78)18.50 (4.69)18.01 (5.27)0.123 GRCS Total96.49 (24.03)96.46 (23.41)96.70 (27.52)0.885 Financial risk behaviors (% endorsing) Borrowing money to gamble55.5%57.3%44.0%<0.001 Diverting funds77.8%80.0%63.8%<0.001 Lying about losses41.8%43.5%31.4%<0.001 Psychological impact (% endorsing) Anxiety91.5%91.9%88.7%0.071 Escapism67.7%69.3%57.5%<0.001 Guilt77.7%78.7%71.4%0.005 Social influences (% endorsing) Social pressure57.7%60.4%41.2%<0.001 Friends' influence0.775None (0)12.8%12.7%13.5%A few (1)60.8%60.6%62.3%Most (2)20.7%20.9%19.5%All (3)5.7%5.8%4.7% Online forums78.5%81.2%61.9%<0.001 Gambling history (% endorsing / M, SD) Daily gambling68.4%70.5%56.0%<0.001 Self-imposed spending limit58.2%59.9%48.1%<0.001 Betting frequency55.8% “occasional”57.6% “occasional”44.3% “occasional”<0.001 Source of gambling funds50.7% “Other” (modal)52.9% “others”37.4% “others”<0.001 Method of financing44.1% “Income” (modal)46.3% “income”30.8% “income”<0.001Note. M = mean; SD = standard deviation. For binary variables, mean columns report prevalence (% endorsing). GRCS = Gambling-Related Cognitions Scale. Internal consistency (Cronbach's alpha): GE = 0.749; IC = 0.744; PC = 0.743; IS = 0.881; IB = 0.744; Total = 0.924. Sex differences were evaluated using Welch's t-tests for continuous variables and chi-square tests for categorical variables. Reported p-values are unadjusted descriptive comparisons. Daily gambling status and betting frequency were assessed as separate items with different reference framings (a direct yes/no question versus a categorical frequency scale) and are not mutually exclusive by design; the two indicators should not be read as contradictory. Percentages shown are for the most frequently reported category; full distributions are available from the corresponding author upon request. Possible ranges: Gambling Expectancies 4–28; Illusion of Control 4–28; Predictive Control 6–42; Inability to Stop 5–35; Interpretive Bias 4–28; GRCS Total 23–161. Descriptive Statistics for GRCS Subscales, Financial Risk Behaviors, Psychological Impact, and Social Influences (N = 2261). Note. M = mean; SD = standard deviation. For binary variables, mean columns report prevalence (% endorsing). GRCS = Gambling-Related Cognitions Scale. Internal consistency (Cronbach's alpha): GE = 0.749; IC = 0.744; PC = 0.743; IS = 0.881; IB = 0.744; Total = 0.924. Sex differences were evaluated using Welch's t-tests for continuous variables and chi-square tests for categorical variables. Reported p-values are unadjusted descriptive comparisons. Daily gambling status and betting frequency were assessed as separate items with different reference framings (a direct yes/no question versus a categorical frequency scale) and are not mutually exclusive by design; the two indicators should not be read as contradictory. Percentages shown are for the most frequently reported category; full distributions are available from the corresponding author upon request. Possible ranges: Gambling Expectancies 4–28; Illusion of Control 4–28; Predictive Control 6–42; Inability to Stop 5–35; Interpretive Bias 4–28; GRCS Total 23–161. Measures Gambling-related cognitions scale (GRCS) The GRCS (Raylu & Oei, 2004; French version: Grall-Bronnec et al., 2012) is a 23-item measure rated on a 7-point Likert scale. It comprises five subscales: Gambling Expectancies (GE), Illusion of Control (IC), Predictive Control (PC), Inability to Stop (IS), and Interpretive Bias (IB). Higher scores indicate stronger gambling-related cognitions. Non-verbatim content descriptors for all items are provided in Supplementary Table S3. Internal consistency was excellent for the total scale (Cronbach's α = 0.924). Subscale alphas were as follows: GE = 0.749, IC = 0.744, PC = 0.743, IS = 0.881, and IB = 0.744. Financial risk behaviors Three binary items (0 = absent, 1 = present) measured borrowing money to gamble, diverting funds toward gambling, and lying about gambling losses. Psychological impact Three binary items assessed gambling-related anxiety, escapism, and guilt — affective variables consistent with pathways models of emotional vulnerability (Blaszczynski & Nower, 2002). Social Influences Three items captured social pressure to bet (binary), friends' influence (ordinal, 0–3), and online forum participation (binary). Gambling History Additional self-reported indicators captured betting frequency, daily gambling status, possession of a self-imposed spending limit, and the source and method of gambling finance. Statistical analysis All analyses were conducted in R version 4.6.0, α = 0.05. The EBICglasso algorithm (Epskamp & Fried, 2018) was used via bootnet (v1.5), with cor_auto for mixed Likert/binary data (Epskamp et al., 2012; polychoric/tetrachoric correlations via lavaan, Rosseel, 2012) and tuning parameter γ = 0.5. A significance threshold controlled for inflated edge weights at N = 2261 (Isvoranu & Epskamp, 2023). Both thresholded (main) and unthresholded (Supplementary Fig. S1) networks are reported. As a sensitivity analysis, a complementary 14-node network (5 GRCS subscales +3 financial +3 psychological +3 social indicators) was estimated identically. A confirmatory factor analysis (WLSMV) of the five-factor GRCS structure was conducted using lavaan, with item allocation verified against Raylu and Oei (2004) as itemized by Kale and Dubelaar (2013). — Unthresholded Network Fig. S1 presents the EBICglasso network without significance threshold (289 edges, density = 0.583). Provided for transparency; the thresholded network (Fig. 1) is recommended for interpretation. Strength, closeness, and betweenness centrality were computed via qgraph (Epskamp et al., 2012). Stability was evaluated using case-dropping bootstrap (n = 1000; Epskamp et al., 2018), yielding centrality stability (CS) coefficients for each index. CS coefficients below 0.50 indicate the corresponding index should be interpreted with caution (Epskamp et al., 2018). Case-dropping bootstrap stability was also evaluated separately for men and women at both the item level and the 14-node subscale level. Bridge strength and bridge betweenness were computed via networktools (Jones et al., 2021) across four a priori communities: GRCS items (23 nodes), financial, psychological, and social indicators (3 nodes each). Sex-based invariance was evaluated via NetworkComparisonTest (NCT; van Borkulo et al., 2023; 100,000 permutations, selected to support Bonferroni correction across the resulting tests), testing global structure (M) and strength (S) invariance. Edge-level and node-centrality p-values were corrected separately using Benjamini-Hochberg FDR and Bonferroni procedures. Results Descriptive statistics The sample was predominantly male (n = 1943, 85.9%), with mean age 25.59 years (SD = 5.79; range: 18–68) and mean gambling onset 20.61 years (SD = 5.06). GRCS total scores ranged from 23 to 161 (Table 1). Predictive Control and Inability to Stop had the highest subscale means; Illusion of Control the lowest. Financial risk behaviors were highly prevalent: diverting funds (77.8%), borrowing money (55.5%), and lying about losses (41.8%). Psychological impact was widespread: anxiety (91.5%), guilt (77.7%), escapism (67.7%). Online forum use was reported by 78.5% and social pressure by 57.7%. Table 1 is stratified by sex. Men were slightly older (25.79 vs. 24.35 years; t(559.6) = 5.35, p < .001) but did not differ in gambling onset age (p = .235) or GRCS total (p = .885). Women scored higher on Gambling Expectancies (18.05 vs. 17.24, p = .018) and men marginally higher on Inability to Stop (21.46 vs. 20.47, p = .046); no other subscale differed (all p > .09), and none would survive multiple-testing correction. Daily gambling was more common among men (70.5% vs. 56.0%; p < .001), and betting frequency, spending-limit behavior, source of funds, and financing method also differed significantly by sex (all p < .001). Network structure Network density and structure The thresholded EBICglasso network (Fig. 1) contained 150 edges (density = 0.302). The unthresholded network produced 289 edges (density = 0.583; Supplementary Fig. S1). Visual inspection indicated a densely interconnected GRCS community, with guilt, social pressure, anxiety, and diverting funds emerging as cross-community connectors (see Section 3.4 for zero-bridge nodes). Thresholded EBICglasso network (N = 2261; 150 edges, density = 0.302). Node colors: blue = GRCS items (G1–G23; see Supplementary Table S3 for item content descriptors); red = financial risk behaviors (Borr = borrowing; Divt = diverting funds; Lie = lying about losses); orange = psychological impact (Anxt = anxiety; Esc = escapism; Glt = guilt); teal = social influences (SPrs = social pressure; FrInf = friends' influence; Frum = online forums). Edge thickness reflects absolute edge weight. Unthresholded network in Supplementary Fig. S1. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Thresholded EBICglasso network (N = 2261; 150 edges, density = 0.302). Node colors: blue = GRCS items (G1–G23; see Supplementary Table S3 for item content descriptors); red = financial risk behaviors (Borr = borrowing; Divt = diverting funds; Lie = lying about losses); orange = psychological impact (Anxt = anxiety; Esc = escapism; Glt = guilt); teal = social influences (SPrs = social pressure; FrInf = friends' influence; Frum = online forums). Edge thickness reflects absolute edge weight. Unthresholded network in Supplementary Fig. S1. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Confirmatory factor analysis of the GRCS A confirmatory factor analysis of the canonical five-factor GRCS structure (item allocation per Raylu & Oei, 2004, as itemized by Kale & Dubelaar, 2013) showed inadequate fit in this sample: χ2(220) = 7976.18, χ2/df ≈ 36.3, CFI = 0.846, TLI = 0.823, RMSEA = 0.125 [90% CI: 0.123, 0.127], SRMR = 0.074. Given this inadequate fit, the item-level network is retained as the primary analytic unit, complemented by a 14-node subscale-level network reported as a sensitivity analysis in Section 3.5 (see Supplementary Fig. S3 for the corresponding network visualization). Thresholded 14-node subscale-level network (N = 2261; 50 edges, density = 0.549), estimated using EBICglasso with the same specification as the item-level network (Fig. 1). Node colors indicate the four theoretical communities: blue = GRCS subscales (GE, IC, PC, IS, IB); red = financial risk behaviors; orange = psychological impact; green = social influences. This sensitivity analysis complements Table 4. Centrality indices Centrality stability analysis yielded CS coefficients of 0.672 (strength), 0.594 (closeness), and 0.361 (betweenness); given that betweenness fell below the preferred threshold, it is reported descriptively (Fig. 2, Table 2). Guilt had the highest strength (z = 2.790); G1 had the highest closeness (z = 2.332) and ranked second in strength (z = 2.180). H1 was not fully supported: several IS items (G2, G7, G17, G21) ranked among the most central nodes, whereas no PC item reached comparable centrality; guilt and anxiety also ranked highly, unpredicted by H1. Centrality indices (strength, closeness, betweenness; standardized z-scores) for all 32 nodes ordered by strength. CS: strength = 0.672; closeness = 0.594; betweenness = 0.361. Table 2Centrality Indices for All Network Nodes (Standardized z-Scores).NodeLabelCommunityStrengthClosenessBetweennessGuiltGltPsycho2.790*2.103*2.946*G1G1GRCS2.180*2.332*2.528*G2G2GRCS1.378*1.854*1.970*AnxietyAnxtPsycho1.181*1.132*−0.539Social pressureSPrsSocial1.113*0.8930.576Diverting fundsDivtFinancial1.0360.2660.902*G21G21GRCS0.6140.4590.298G18G18GRCS0.5950.1120.484G20G20GRCS0.5580.6610.855G17G17GRCS0.3780.436−0.121G7G7GRCS0.2881.229*0.902*G15G15GRCS0.1840.229−0.446G6G6GRCS0.1760.2960.809G11G11GRCS0.0490.1790.065G12G12GRCS−0.130−0.047−0.492G8G8GRCS−0.218−0.488−0.539G22G22GRCS−0.274−0.502−0.725Borrowing moneyBorrFinancial−0.307−0.635−0.818Online forumsFrumSocial−0.389−0.456−0.678G14G14GRCS−0.430−0.347−0.678G10G10GRCS−0.529−0.030−0.028G13G13GRCS−0.531−0.654−0.353G3G3GRCS−0.578−0.077−0.771G16G16GRCS−0.617−0.265−0.446G19G19GRCS−0.669−0.598−0.632EscapismEscPsycho−0.761−0.472−0.864G9G9GRCS−0.880−1.428−0.725G5G5GRCS−0.907−1.053−0.306Friends' influenceFrInfSocial−1.230−0.487−0.725G4G4GRCS−1.306−1.441−0.864G23G23GRCS−1.312−1.847−0.725Lying about lossesLieFinancial−1.453−1.354−0.864Note. Values are standardized z-scores. * = top five per index. CS: Strength = 0.672, Closeness = 0.594, Betweenness = 0.361 (below preferred threshold — interpret with caution). GRCS = gambling cognition items; Financial = financial risk behaviors; Psycho = psychological impact; Social = social influences. Centrality Indices for All Network Nodes (Standardized z-Scores). Note. Values are standardized z-scores. * = top five per index. CS: Strength = 0.672, Closeness = 0.594, Betweenness = 0.361 (below preferred threshold — interpret with caution). GRCS = gambling cognition items; Financial = financial risk behaviors; Psycho = psychological impact; Social = social influences. Bridge centrality Guilt showed the highest bridge strength (BS = 1.9624) and bridge betweenness (BB = 51), linking the psychological community to GRCS and social-influences items (Fig. 3, Table 3). Social pressure ranked second (BS = 1.5957, BB = 17). G1 showed the highest bridge betweenness (BB = 51, tied with guilt) and third-highest bridge strength (BS = 1.1285). Anxiety and diverting funds completed the top five (values in Table 3). Consistent with H2, several high-bridge-centrality nodes connected the GRCS community with financial, psychological, and social indicators. Seven GRCS items (G4, G5, G9, G13, G16, G19, G23) showed zero bridge strength (Table 3). Bridge centrality indices (bridge strength and bridge betweenness) for all 32 nodes across four theoretical communities, ordered by bridge strength. Top five bridge nodes identified in Table 3. Table 3Bridge Centrality Indices for All Network Nodes, Ordered by Bridge Strength.NodeLabelCommunityBridge StrengthBridge BetweennessGuiltGltPsycho1.9624*51*Social pressureSPrsSocial1.5957*17G1G1GRCS1.1285*51*AnxietyAnxtPsycho1.0709*34*Diverting fundsDivtFinancial1.0507*48*Online forumsFrumSocial0.70383EscapismEscPsycho0.56537G2G2GRCS0.496938*Friends' influenceFrInfSocial0.45015Borrowing moneyBorrFinancial0.40241G11G11GRCS0.397518G3G3GRCS0.29704G15G15GRCS0.20038G17G17GRCS0.165415G14G14GRCS0.15253G6G6GRCS0.148915G7G7GRCS0.145915G20G20GRCS0.129815G12G12GRCS0.12642G18G18GRCS0.11791Lying about lossesLieFinancial0.11630G10G10GRCS0.10798G21G21GRCS0.103712G8G8GRCS0.072012G22G22GRCS0.06050G4 (zero bridge)G4GRCS0.00000G5 (zero bridge)G5GRCS0.00000G9 (zero bridge)G9GRCS0.00000G13 (zero bridge)G13GRCS0.00000G16 (zero bridge)G16GRCS0.00000G19 (zero bridge)G19GRCS0.00000G23 (zero bridge)G23GRCS0.00000Note. Bridge strength and bridge betweenness computed via networktools across four communities: GRCS (n = 23), Financial (n = 3), Psychological impact (n = 3), Social influences (n = 3). * = top five per index. Zero bridge strength = no cross-community edge was retained for this node under the present network specification. G1 = “Gambling makes me happier” (Gambling Expectancies item). Bridge Centrality Indices for All Network Nodes, Ordered by Bridge Strength. Note. Bridge strength and bridge betweenness computed via networktools across four communities: GRCS (n = 23), Financial (n = 3), Psychological impact (n = 3), Social influences (n = 3). * = top five per index. Zero bridge strength = no cross-community edge was retained for this node under the present network specification. G1 = “Gambling makes me happier” (Gambling Expectancies item). Sex-based network comparison Separate item-level networks for men (n = 1943; 132 edges; density = 0.266) and women (n = 318; 50 edges; density = 0.101) showed no significant global structural (M = 0.195, p = .121) or strength (S = 1.313, p = .778) difference. Bootstrap analysis showed the women's item-level network stability could not be estimated (non-positive-definite matrices under all configurations), precluding confident interpretation of local differences. As a sensitivity analysis, the 14-node subscale-level network was also compared by sex: structural invariance was rejected (M = 0.217, p = .022) while strength invariance held (S = 0.910, p = .449); bootstrap stability was again poor for women (CS[strength] = 0.129, below the 0.25 threshold; Epskamp et al., 2018). The NCT was re-run with 100,000 permutations to support a valid Bonferroni threshold across the 496 item-level edges. After Bonferroni and FDR correction, one item-level edge remained significant (G1–G15, linking a Gambling Expectancies item to an Interpretive Bias item; p_uncorrected <0.0001, p_Bonferroni = p_FDR = 0.005); zero edges remained significant at the subscale level (0/91). Given this combination of non-estimable or below-threshold stability with a single, narrow item-level exception and no corrected subscale-level edges, sex-based local differences are treated as exploratory rather than substantive. Full centrality and bridge centrality indices for the 14-node subscale-level network are reported in Table 4. Full details appear in Supplementary Tables S1–S2 and Fig. S2. Table 4Centrality and Bridge Centrality Indices for the 14-Node Subscale-Level Network (Sensitivity Analysis).NodeLabelCommunityStrength (z)Bridge StrengthBridge BetweennessGambling ExpectanciesGEGRCS1.220.8195Illusion of ControlICGRCS−0.400.3930Predictive ControlPCGRCS−0.130.0372Inability to StopISGRCS0.100.4457Interpretive BiasIBGRCS−0.100.25911Borrowing moneyBorrFinancial−0.330.3630Diverting fundsDivtFinancial1.070.95023Lying about lossesLieFinancial−1.680.0870AnxietyAnxtPsychological0.440.6177EscapismEscPsychological−0.920.5191GuiltGltPsychological2.021.39314Social pressureSPrsSocial0.461.1569Friends' influenceFrInfSocial−1.310.4650Online forumsFrumSocial−0.460.6481Note. Strength values are standardized z-scores; bridge strength and bridge betweenness are reported in their original metrics. Indices were calculated for the full-sample 14-node sensitivity network (N = 2261), comprising five GRCS subscales and nine financial, psychological, and social indicators, estimated via EBICglasso across four theoretical communities. These indices are descriptive and do not represent tests of sex differences. Centrality and Bridge Centrality Indices for the 14-Node Subscale-Level Network (Sensitivity Analysis). Note. Strength values are standardized z-scores; bridge strength and bridge betweenness are reported in their original metrics. Indices were calculated for the full-sample 14-node sensitivity network (N = 2261), comprising five GRCS subscales and nine financial, psychological, and social indicators, estimated via EBICglasso across four theoretical communities. These indices are descriptive and do not represent tests of sex differences. Discussion This study examined the network structure of gambling-related cognitions among 2261 sports bettors in Kinshasa, DRC — among the first such analyses in sub-Saharan Africa. Four key findings emerged: (1) a moderately sparse network (density = 0.302) with clear community organization; (2) guilt as the most central node; (3) guilt and social pressure as the leading bridge nodes; (4) no item-level difference in network structure or strength, apart from one edge (G1–G15) surviving multiple-testing correction, alongside a significant omnibus structural difference at the subscale level not localized to any individual edge. Centrality: the unexpected prominence of guilt and G1 H1 was only partially supported: several IS items (G2, G7, G17, G21) ranked among the most central nodes, whereas no PC item reached comparable centrality. Guilt showed the highest strength and the second-highest closeness, while G1 (“Gambling makes me happier”; Kale & Dubelaar, 2013) — cited as a single illustrative item, consistent with the non-verbatim policy for the full scale (Section 2.2.1) — showed the highest closeness and the second-highest strength. Guilt's high endorsement (77.7%) may reflect the sample's high level of gambling involvement and prevalence of self-reported financial and psychological risk indicators. Its centrality suggests affective consequences of gambling are at least as interconnected as cognitive distortions, challenging the unidirectional cognitive-to-affective pathway of classical models (Ladouceur & Walker, 1996) and supporting a more symmetric affect-cognition role in I-PACE (Brand et al., 2019). One possible explanation is that guilt may be especially salient where gambling losses affect both individual and family responsibilities. Network analyses of gambling cognitions in Western samples remain scarce, precluding systematic cross-cultural comparison; future studies should test whether affective nodes occupy more peripheral positions there, using continuous rather than binary guilt measures. Bridge nodes: guilt and social pressure as transdiagnostic connectors Guilt was the dominant bridge node (BS = 1.9624, BB = 51), consistent with the I-PACE feedback loop linking negative affect to compensatory cognitions in the addictive cycle (Brand et al., 2019), bridging cognitive distortions, financial behaviors, and social influences as a transdiagnostic correlate (Fried et al., 2017). Social pressure ranked second (BS = 1.5957): in the DRC, where betting venues may function as social hubs where peer norms favor risk-taking, social pressure showed cross-community associations with cognitive and behavioral indicators, highlighting the inadequacy of cognition-only interventions. G1's centrality may reflect a broadly relevant gambling expectancy, whereas the prominence of guilt and social pressure may be particularly relevant in the present sociocultural context; cross-cultural studies are needed to test this interpretation. Seven GRCS items (G4, G5, G9, G13, G16, G19, G23) show zero bridge strength, with no retained cross-community edges under the present specification; this does not imply a distinct empirical cluster. Item codes are retained only where they add information the domain level cannot (G1; the seven zero-bridge items). We note a methodological circularity: the item-level specification is partly motivated by the confirmatory factor model's inadequate subscale-level fit (Section 3.2.2), yet interpretations above still draw on subscale labels (GE, IC, PC, IS, IB) as organizing categories — a pragmatic but unresolved tension common in network psychometrics. Sex-based network invariance Results diverged by level of specification (Section 3.5): the item-level NCT did not detect a global difference in network structure (M = 0.195, p = .121) or strength (S = 1.313, p = .778), apart from one edge (G1–G15, linking a Gambling Expectancies item to an Interpretive Bias item) surviving correction. At the subscale level, the omnibus structure test was significant (M = 0.217, p = .022), although no individual edge remained significant after correction for the 91 local comparisons; given the poor stability of the women's network (CS[strength] = 0.129), this omnibus result should be considered exploratory. Taken together, these results indicate that H3 was only partially supported: sex-related differences in this network appear specification-dependent rather than demonstrating either robust invariance or a clearly localized mechanism; the sample imbalance (86% male) further constrains power for women and warrants replication in balanced samples before any clinical interpretation. Theoretical and clinical contributions We propose two candidate bridge processes that may complement existing I-PACE formulations: guilt, consistent with a possible affective-amplification role in the cognitive-behavioral feedback loop, and social pressure, consistent with a possible contextual-activation role. Mapped onto the model (Brand et al., 2016, 2019): Gambling Expectancies and Illusion of Control reflect the cognitive-bias component; guilt and anxiety, affective responses to losses; and social pressure, situational triggering cues. Financial risk behaviors represent behavioral correlates rather than a direct measure of the Execution component, which in I-PACE denotes executive functions (e.g., inhibitory control, decision-making) not directly assessed here; accordingly, this study addresses selected I-PACE processes without operationalizing the full model. Guilt's bridge position supports the updated I-PACE assumption of reciprocal affect-cognition loops. Future I-PACE formulations for sports betting in LMIC contexts should incorporate these constructs. Clinically, three candidate priorities emerge for low-resource settings, framed as targets for future evaluation rather than demonstrated effects. First, guilt and social pressure as entry points: brief psychosocial approaches to guilt, combined with peer resistance training, may merit testing given their bridge status; both constructs may also warrant evaluation as candidate indicators for future screening or triage tools, although the present single-item measures are not validated screening instruments. Second, Gambling Expectancies, the most connected domain, as a target for public education addressing affective framing rather than probability-based messaging alone. Third, the seven zero-bridge items appear, on present evidence, lower-priority targets; whether intervening on them would alter the network cannot be determined from this cross-sectional design. Limitations Several limitations apply. First, cross-sectional design precludes causal inference. Second, convenience sampling may overrepresent frequent or highly engaged bettors, limiting generalizability to occasional, online-only, or community bettors. Third, the sex imbalance (86% male) constrains NCT power for women. Fourth, guilt and psychological variables used single binary items combined with Likert-type GRCS items; despite cor_auto's appropriate handling, item format differences may influence edge magnitudes, possibly affecting guilt's centrality. Fifth, all variables were self-reported via interviewer administration (adopted to reduce literacy-related bias), which does not permit independent verification of disclosures and may have introduced social-desirability or interviewer effects. Sixth, betweenness CS = 0.361 fell below the preferred threshold; findings are exploratory. Seventh, single-city data (Kinshasa) limit generalizability within the DRC. Eighth, the GRCS's French adaptation (Grall-Bronnec et al., 2012) has not been formally validated in a Congolese context, consistent with the inadequate CFA fit reported (Section 3.2.2); the Lingala oral guide was reviewed by bilingual staff but not independently back-translated. Ninth, the amount of money wagered or lost was not collected; financial risk is therefore captured only through the binary indicators described in Section 2.2.2, without quantifying the magnitude of exposure. Tenth, data collection spanned multiple Kinshasa betting venues (∼50 participants per venue); venue counts and refusal rates were not systematically tracked, precluding a formal response-rate estimate. Conclusions This study provides among the first network analyses mapping gambling-related cognitions alongside financial risk, psychological impact, and social influences in sub-Saharan Africa. Among 2261 sports bettors in Kinshasa, guilt emerged as a central bridge node, while social pressure also showed high bridge centrality, with Gambling Expectancies as the most connected cognitive domain and G1 as a central hub. The item-level omnibus test showed no global sex difference; after correction for multiple local tests, a single edge remained significant and warrants replication in balanced samples; a subscale-level sensitivity analysis detected a significant global structural difference not localized to any individual edge. These findings extend the I-PACE model by identifying guilt and social pressure as bridge mechanisms not previously modeled. Rather than a clean separation between a purely cognitive core and affective bridges, the results point to affective content — G1's expectancy of feeling happier when gambling, and guilt — occupying central positions both within and between communities, suggesting that affect and cognition are closely intertwined at the network's core. Where sports betting outpaces clinical and regulatory infrastructure, this network map provides hypotheses for the development and prospective evaluation of culturally adapted prevention and intervention strategies. The following are the supplementary data related to this article. — NCT Permutation Distributions Fig. S2 presents permutation distributions for the two NCT omnibus tests (M and S statistics) with observed statistics as vertical lines. Men: n = 1943; Women: n = 318; permutations:100,000. Neither statistic reached significance (M: p = .121; S: p = .778). Supplementary material 1 Supplementary material containing Figures S1–S3 and Tables S1–S3. Supplementary Data 1 Complete edge-level NetworkComparisonTest results for all 496 tested item-level edges (100,000 permutations), including raw, Bonferroni-corrected, and FDR-corrected p-values. Test statistic E = absolute difference in edge weight between men (n = 1,943) and women (n = 318). Supplementary Data 2 Complete centrality-level Network Comparison Test results for all 90 testable indices (strength, closeness, betweenness) across 32 nodes (100,000 permutations), including raw, Bonferroni-corrected, and FDR-corrected p-values. Men: n = 1,943; Women: n = 318. Declaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process During the preparation of this work, the authors used Claude (Anthropic) in order to assist with manuscript drafting. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. CRediT authorship contribution statement Degani Banzulu Bomba: Writing – review & editing, Writing – original draft, Supervision, Project administration, Methodology, Formal analysis, Conceptualization. Mechac Koffi Nkazi: Writing – review & editing, Investigation, Data curation. Kevy Nguya Makonda: Writing – review & editing, Investigation, Data curation. Christian Badimanya Albert: Writing – review & editing, Investigation, Data curation. Merci Lusala Lutsasa: Writing – review & editing, Investigation, Data curation. Jacques Benangindu Kikumpa: Writing – review & editing, Validation, Supervision. Ethics declaration Informed consent and patient details Written informed consent to take part in the study and to publish the article has been obtained from all participants or their legal representatives. The privacy rights of participants have been observed. Studies in Human This study was conducted in accordance with the Declaration of Helsinki. Ethical approval was granted by the Ethics Committee of the École de Santé Publique, Université de Kinshasa (reference ESP/CE/60B/2026) on March 23, 2026. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements The authors thank all participants and data collection teams. Contributor Information Degani Banzulu Bomba, Email: banzulu.bomba@unikin.ac.cd. Jacques Benangindu Kikumpa, Email: sj.benangindu@unikin.ac.cd. Data availability Data are available upon reasonable request to the corresponding author. References Associated Data Supplementary Materials Supplementary material 1 Supplementary material containing Figures S1–S3 and Tables S1–S3. Supplementary Data 1 Complete edge-level NetworkComparisonTest results for all 496 tested item-level edges (100,000 permutations), including raw, Bonferroni-corrected, and FDR-corrected p-values. Test statistic E = absolute difference in edge weight between men (n = 1,943) and women (n = 318). Supplementary Data 2 Complete centrality-level Network Comparison Test results for all 90 testable indices (strength, closeness, betweenness) across 32 nodes (100,000 permutations), including raw, Bonferroni-corrected, and FDR-corrected p-values. Men: n = 1,943; Women: n = 318. Data Availability Statement Data are available upon reasonable request to the corresponding author.