Problem gambling (PG) is a major public health concern, which is estimated to affect approximately 1.29% of the global population [Reference Gabellini, Lucchini and Gattoni1]. PG is characterized by problematic gambling behaviors such as escalating gambling frequency and expenditure, chasing losses, and lying about gambling, as well as other gambling-related harms such as health problems, guilt, and interpersonal conflict [2–Reference Langham, Thorne, Browne, Donaldson, Rose and Rockloff4]. Composite severity scores are widely used in screening and epidemiological research, but they offer limited guidance in terms of which specific symptoms carry the most weight in sustaining or escalating gambling problems. Psychometric network analysis provides a finer-grained, symptom-level framework well suited to this task [Reference Borsboom, Deserno, Rhemtulla, Epskamp, Fried and McNally5, Reference Briganti, Scutari, Epskamp, Borsboom, Hoekstra and Golino6].
In a network model, individual symptoms and related factors are represented as nodes and the statistical associations between them as edges, providing a map of the unique pairwise relationships between symptoms while controlling for all other symptoms in the system [Reference Borsboom, Deserno, Rhemtulla, Epskamp, Fried and McNally5]. Notably, this approach can provide key insights into the specific symptoms that are most influential in maintaining a condition [Reference Briganti, Scutari, Epskamp, Borsboom, Hoekstra and Golino6]. Cross-lagged network models extend this logic by estimating whether a given symptom at one time predicts changes in the same and other symptoms at a later time, thus moving from describing symptom co-occurrence to approximating symptom dynamics over time [Reference Wysocki, McCarthy, Van Bork, Cramer and Rhemtulla7].
A small but growing body of research has applied network analysis to PG and has examined PG symptoms alongside comorbid conditions and related factors such as gender, substance use, and emotion dysregulation [e.g., Reference Baggio, Gainsbury, Starcevic, Richard, Beck and Billieux8–Reference Peng, Hao, Zhang, Ma, Liu and Shen11]. PG symptoms appear intertwined with multiple game types among men such as poker, sports betting, and casino games, whereas they are linked primarily to electronic gaming machines among women [Reference Baggio, Gainsbury, Starcevic, Richard, Beck and Billieux8]. Among adolescent gamblers, stealing money to fund gambling was central, with strong ties to academic deterioration [Reference Oh, Joung and Baek10]. In a clinical sample with comorbid schizophrenia and gambling disorder, novelty seeking bridged psychosis symptoms and gambling severity [Reference Granero, Fernández-Aranda, Demetrovics, Lara-Huallipe, Morón-Fernández and Jiménez-Murcia12].
Network studies have also examined PG symptoms in isolation. In an undirected network of the criteria listed in the 5th edition of the Diagnostic and Statistical Manual of Mental Disorder (DSM-5; [2]), withdrawal (i.e., becoming restless or irritable when attempting to cut down or stop gambling) emerged as the most central symptom [Reference Lucas, Mora-Maltas, Granero, Demetrovics, Ciudad-Fernández and Nigro13]. In directed cross-lagged networks based on the widely used Problem Gambling Severity Index (PGSI; [Reference Ferris and Wynne3]), health problems due to gambling predicted subsequent PG symptoms [Reference Horváth, Paksi, Fernández-Aranda, Jiménez-Murcia and Demetrovics14]. Together, these findings underscore that which symptom emerges as most central may depend on the sample type, measurement instrument, and time frame under study. Accordingly, finding central symptoms that are consistent across samples serves as an important priority.
Understanding which PG symptoms are most central has practical implications for screening and early intervention: Symptoms that consistently occupy central positions across networks may serve as informative markers of emerging severity. Previous network studies on PG are limited by small sample sizes, scarcity of longitudinal investigations, and heterogeneity in terms of variables included. The present study addresses these limitations by applying undirected cross-sectional and directed cross-lagged psychometric network models to four large nationally representative datasets. We use the PGSI [Reference Ferris and Wynne3], which is well suited for network analysis because it includes items on both problematic gambling behaviors and negative consequences (harms). We also include gambling expenditure to account for overall gambling involvement while retaining model parsimony. Two research questions (RQs) guided the study:
RQ1: How do PG symptoms and gambling expenditure interact cross-sectionally and over time, and which symptoms emerge as the most central in these undirected and directed networks?
RQ2: How stable and generalizable are undirected psychometric networks, and how generalizable are directed psychometric networks of PG in representative population samples of gamblers?
Methods
Study design and procedure
The current study uses a mixed cross-sectional and longitudinal design based on four representative Norwegian general population surveys collected in 2013, 2015, 2019, and 2022 among individuals aged 16–74 [Reference Pallesen, Hanss, Mentzoni, Molde and Morken15–Reference Pallesen, Syvertsen, Kristensen, Mentzoni, Erevik and Morken20]. Each survey year included independent cross-sectional samples; the 2015 and 2022 surveys additionally re-invited respondents from 2013 and 2019, respectively, providing a longitudinal component. Participants were included in the present study if they reported gambling within the last 12 months. For the two longitudinal samples, we included participants reporting gambling in both years. See Tables 1 and 2 for descriptives on sample n, demographic variables, PG symptoms, and the sum of gambling expenditure. For Table 2, participants with missing data on PG symptoms were excluded for analytic reasons (see “Statistical analysis” below). See Supplementary Table A7 for details on participant flow and longitudinal attrition.
Demographic and problem gambling symptom distribution for cross-sectional samples
Table 1. Long description
The table compares four samples: 2013 (n = 6,034), 2015 (n = 3,232), 2019 (n = 5,830), and 2022 (n = 4,621).
Demographics:
* Gender, women: Percentages remain stable between 47 percent and 50 percent across all years.
* Age, years: Mean age ranges from 46 to 50, with a standard deviation of 15 to 16.
Problem Gambling Symptoms (n and percentage reporting ‘sometimes’ or higher):
* Loss of control: Fluctuates between 3 percent and 5 percent.
* Tolerance: Highest in 2019 (13 percent) and lowest in 2022 (10 percent).
* Chasing losses: Peaked in 2019 at 8 percent, compared to 5 percent in other years.
* Borrowed: Consistently low at 1 percent to 2 percent.
* Problem recognition: Ranges from 3 percent to 4 percent.
* Health problems: Ranges from 1 percent to 3 percent.
* Criticism: Ranges from 2 percent to 3 percent.
* Financial problems: Ranges from 1 percent to 2 percent.
* Guilt: Peaked in 2019 at 8 percent, with other years between 5 percent and 6 percent.
Financial Data:
* Gambling sum in N O K: Mean values show a slight downward trend from 5,133 in 2013 to 4,528 in 2022, with high standard deviations.
Abbreviations: N O K equals Norwegian Kroner (10 N O K is approximately 1 U S D).
Abbreviations: NOK = Norwegian Kroner (10 NOK ≈ 1 USD).
Note: Missing values on problem gambling symptoms: 2013 sample, n = 29; 2015 sample, n = 34; 2019 sample, n = 25; 2022 sample, n = 23.
a n (%).
b M (SD).
c n (%) reporting any intensity of PG symptom (1 = “sometimes” or higher).
Demographic and problem gambling symptom distribution for longitudinal samples
Table 2. Long description
The table compares two longitudinal samples: 2013 to 2015 (Wave 1 and Wave 2) and 2019 to 2022 (Wave 1 and Wave 2).
* Sample Sizes: n = 3,043 for the 2013 to 2015 wave and n = 2,749 for the 2019 to 2022 wave.
* Demographics: Women represent 50 percent (1,536) of the first sample and 47 percent (1,302) of the second. Mean age is 53 years for both samples.
* Problem Gambling Symptoms (n and percentage reporting ‘sometimes’ or higher):
- Loss of control: Stable at 3 percent across all waves.
- Tolerance: 11 percent in Wave 1 of both samples, dropping to 10 percent in Wave 2.
- Chasing losses: 3 percent in the 2013 to 2015 sample; 5 percent in 2019 to 2022 Wave 1 and 4 percent in Wave 2.
- Borrowed: 0 to 1 percent across all waves.
- Problem recognition: 2 percent in the first sample; 3 percent in the second sample.
- Health problems: 1 to 2 percent across all waves.
- Criticism: Stable at 2 percent across all waves.
- Financial problems: Stable at 1 percent across all waves.
- Guilt: 4 percent in the first sample; 6 percent in the second sample.
* Gambling sum in N O K: Mean values range from 4,812 to 5,539 across the waves.
Abbreviations: NOK = Norwegian Kroner (10 NOK ≈ 1 USD).
Note: Sample size for those reporting gambling both years before listwise deletion on problem gambling symptoms: 2013–2015, n = 3,090 (specific problem gambling symptom missingness, n = 11–36); 2019–2022, n = 2,778 (specific problem gambling symptom missingness, n = 5–17).
a n (%).
b M (SD).
c n (%) reporting any intensity of PG symptom (1 = “sometimes” or higher).
Preregistration
This study was preregistered (AsPredicted #151,908). Deviations: Undirected networks were estimated on continuous rank-transformed data (omitting cross-validation) and inclusion criteria were broadened to all who report gambling. For directed networks, the Network Comparison Test (NCT) was unavailable and bootstrap-based comparisons were not validated for this purpose.
Measures
Problem gambling
PG was measured with the PGSI [Reference Ferris and Wynne3]. The PGSI consists of nine items assessing problematic gambling behavior and gambling harms; each item is scored on a 4-point scale (0 = “never,” 1 = “sometimes,” 2 = “most of the time,” and 3 = “always”). The four problematic gambling behaviors are: Betting more than one can afford (loss of control), needing to gamble with larger amounts to get the same excitement (tolerance), going back another day to win money back (chasing losses), and borrowing money or selling possessions to fund gambling (borrowing). The five harms are: Feeling that one has a problem with gambling (problem recognition), gambling causes health problems (health problems), people criticizing one or saying that one has a problem with gambling (criticism), gambling causes financial problems (financial problems), and feeling guilty from gambling or consequences from gambling (guilt). PGSI Cronbach’s α were = 0.83–0.91 in study samples.
Gambling expenditure
Gambling expenditure was measured by self-report items on the amount bet in Norwegian kroner (NOK; 10 NOK ≈ 1 USD) during the last 12 months on individual game types (e.g., lottery games, scratch cards, sports betting, horse racing). Response options for each individual game were “nothing,” “1–1,000 NOK,” “1,001–5,000 NOK,” “5,001–10,000 NOK,” “10,001–25,000 NOK,” and “More than 25,000 NOK.” Continuous expenditure was computed by assigning midpoint values to each bracket (30,000 NOK for the top open-ended bracket), summed across all games, and inflation-adjusted to 2022 using rates from the Central Bank of Norway [21].
Statistical analysis
All statistical analyses were performed using R Statistical Software version 4.5.2 [22]. To investigate RQ1, we estimated undirected networks with a model search algorithm (“ggmModSelect”) as recommended by simulation-based guidelines for large sample sizes [Reference Isvoranu and Epskamp23] using the bootnet package version 1.6 [Reference Epskamp, Borsboom and Fried24]. The algorithm has high specificity at large sample sizes and first generates a sequence of 100 candidate structures with regularization, which are re-estimated without regularization using maximum likelihood. The best-fitting structure is then used as a starting point for a stepwise model search that iteratively adds and removes edges until the Bayesian Information Criterion (BIC) is optimized. We also applied rank transformation due to data skewness. All plots and strength centrality measures were drawn using qgraph version 1.9.8 [Reference Epskamp, Cramer, Waldorp, Schmittmann and Borsboom25].
To investigate RQ1, directed cross-lagged network model estimation followed the approach detailed by Wysocki et al. [Reference Wysocki, McCarthy, Van Bork, Cramer and Rhemtulla7], which involves estimating a series of regressions using penalized maximum likelihood and Lasso penalty on estimated regression coefficients with glmnet package 4.1–10 [Reference Friedman, Hastie and Tibshirani26]. The Lasso penalty results in regularization in which small regression coefficients are set to zero. The cross-lagged network models require complete data and assess how one PG symptom at time 1 (2013 or 2019) predicts variation in the same (autoregressive effect) and other (cross-lagged effect) PG symptoms at time 2 (2015 or 2022). Because this study focuses on temporal relationships between different PG symptoms, we excluded autoregressive effects when calculating centrality measures: In-strength centrality (the degree that variable is predicted by other variables) and out-strength centrality (the degree that variable predicts other variables).
To investigate RQ2, we assessed stability and accuracy of undirected networks through bootstrapping with bootnet [Reference Epskamp, Borsboom and Fried24]. Edge weight accuracy was examined with non-parametric bootstraps, and strength centrality stability with case-drop bootstraps (1,000 bootstrap samples; [Reference Epskamp, Borsboom and Fried24]), summarized as the correlation stability coefficient (CS-coefficient) recommended to be above .25 [Reference Epskamp, Borsboom and Fried24]. We also conducted measurements of generalizability, including NCT using the NetworkComparisonTest package version 2.2.2 and correlating networks’ edge presence (phi coefficient) and edge weights (Pearson correlation) between networks [Reference Fried, Epskamp, Veenman and Van Borkulo27, Reference van Borkulo, van Bork, Boschloo, Kossakowski, Tio and Schoevers28]. The undirected model for 2022 and the directed model for 2019–2022 served as the point of statistical comparisons. The NCT is only available for cross-sectional undirected networks, and a total of 1,000 iterations were used in line with common practice [Reference van Borkulo, van Bork, Boschloo, Kossakowski, Tio and Schoevers28]. To assess generalizability between the cross-lagged network models, we relied on correlating networks’ edge presence and edge weights.
Results
Cross-sectional network models
The cross-sectional network model for 2022 is visualized in Figure 1 and all cross-sectional network models are visualized together in Supplementary Figure A1. The edges in the networks are partial regression coefficients. The 2022 network had non-zero edge weights ranging from 0.04 (gambling expenditure – criticism) to 0.37 (financial problems – borrowed), the mean edge weight was 0.09, and network density was 0.69 (31/45). Figure 2 shows the strength centrality of each node. Financial problems had the highest strength centrality in the 2022 network (1.06), meaning that endorsement of this PG symptom was most strongly associated with concurrent endorsement of other PG symptoms.
Cross-sectional network model of problem gambling symptoms and gambling expenditure. Contr = Loss of control, Toler = Tolerance, Chase = Chasing losses, Borw = Borrowing, Prob = Problem recognition, HlthP = Health problems, Critz = Criticism, FinaP = Financial problems, Guilt = Guilt, GSum = Gambling expenditure.
Figure 1. Long description
The network diagram consists of ten circular nodes connected by blue lines of varying thickness.
At the top left is a green node labeled G Sum.
In the upper-middle and right sections are orange nodes labeled Prob, Guilt, Critz, Hlth P, and Fina P.
In the lower-left and bottom sections are red nodes labeled Toler, Chase, Contr, and Borw.
Strongest connections, indicated by the thickest blue lines, are visible between Guilt and Prob, Fina P and Borw, and Hlth P and Fina P. Thinner lines connect G Sum to Toler and Prob.
A legend on the right side defines the node colors and edge properties:
* Red dots represent Problematic gambling behaviors.
* Orange dots represent Negative consequences or harms.
* Green dots represent Gambling expenditure G Sum.
* Edge thickness represents the strength of partial correlation.
* Blue edges represent positive associations.
Strength centrality for cross-sectional network models of problem gambling symptoms. Contr = Loss of control, Toler = Tolerance, Chase = Chasing losses, Borw = Borrowing, Prob = Problem recognition, HlthP = Health problems, Critz = Criticism, FinaP = Financial problems, Guilt = Guilt, GSum = Gambling expenditure.
Figure 2. Long description
The graph plots Strength on the X-axis from 0.0 to 1.2 and symptoms on the Y-axis. Four colored lines represent different years: red (2013), green (2015), teal (2019), and purple (2022).
From top to bottom, the symptoms and their approximate strength values are:
* Toler: Values cluster between 0.6 and 0.8.
* Prob: A peak for all years, ranging from 0.95 to 1.1.
* HlthP: Values cluster between 0.85 and 1.0.
* Guilt: Values cluster between 0.85 and 1.05.
* GSum: A significant trough for all years, with values dropping sharply to between 0.25 and 0.35.
* FinaP: The highest peak for all years, with values reaching between 1.05 and 1.15.
* Critz: Values cluster between 0.75 and 0.85.
* Contr: Values vary between 0.65 (2022) and 0.85 (2013).
* Chase: Values cluster between 0.75 and 0.85.
* Borw: High variability, with 2015 at 0.6, 2019 at 0.75, 2013 at 0.85, and 2022 at 0.95.
The overall trend shows that Financial Problems (FinaP) and Problem Recognition (Prob) consistently have the highest strength centrality, while Gambling Expenditure (GSum) consistently has the lowest.
Borrowing (0.97) and problem recognition (0.96) also showed higher strength centrality compared to the other nodes in the network. Notably, borrowing was positively associated with financial problems, health problems, and criticism. Moreover, problem recognition was positively associated with guilt, criticism, and health problems.
The edges observed in the 2022 network were reliably retrieved in most bootstrap samples (Supplementary Figure A4), and strength centrality estimates were highly stable (CS-coefficient = .75; Supplementary Figure A5). The NCT global invariance tests were statistically insignificant when comparing the undirected 2022 network with the 2019 network (p = .65), the 2015 network (p = .16), and the 2013 network (p = .65), thus supporting generalizability across samples. Correlations with the 2022 network were φ = .79 and r = .85 for 2019, φ = .53 and r = .66 for 2015, and φ = .29 and r = .75 for 2013 (edge presence and edge weights, respectively). Figure 2 reveals that strength centrality estimates were highly similar for most nodes between the 2022, 2019, 2015, and 2013 undirected networks.
Cross-lagged network models
The cross-lagged network model for 2019–2022 is visualized in Figure 3. Both cross-lagged models are visualized together in Supplementary Figures A10 and A11 (Supplementary Material). Non-zero edges in the 2019–2022 network ranged from −0.07 (T1 criticism → T2 guilt) to 0.22 (T1 borrowing → T2 health problems). Mean edge weight was 0.06 and network density was 0.7 (70/100). Strength centrality measures are visualized in Figure 4 and show that borrowing has the strongest out-strength in the 2019–2022 network (0.70). Borrowing at time 1 was associated with higher chasing, financial problems, health problems, and guilt at time 2. Guilt had the strongest in-strength in the 2019–2022 network (0.66) and had non-zero associations with all other nodes. Among the relatively stronger associations, chasing and health problems at time 1 predicted guilt at time 2.
Cross-lagged network model of problem gambling symptoms (excluding autoregressive paths). Contr = Loss of control, Toler = Tolerance, Chase = Chasing losses, Borw = Borrowing, Prob = Problem recognition, HlthP = Health problems, Critz = Criticism, FinaP = Financial problems, Guilt = Guilt, GSum = Gambling expenditure.
Figure 3. Long description
A network diagram with ten circular nodes connected by directed arrows of varying thickness and color.
Node Distribution:
* G Sum (green) is at the top.
* Contr (red) and Prob (orange) are in the upper-middle section.
* Toler (red) is on the far left.
* Fina P (orange) is on the right.
* Chase (red) and Borw (red) are in the center.
* Guilt (orange) is on the lower left.
* Critz (orange) is on the lower right.
* Hlth P (orange) is at the bottom.
Connections:
* Thick blue arrows indicate strong positive temporal effects, notably from Borw to Hlth P, Borw to Fina P, and Chase to Borw.
* Thinner blue arrows connect most other nodes, such as Prob to G Sum and Contr to Toler.
* A few thin red arrows indicate negative effects, such as the path from Borw to Critz.
Legend:
* Red dots represent problematic gambling behaviors.
* Orange dots represent negative consequences or harms.
* Green dots represent gambling expenditure (G Sum).
* Arrow thickness and clarity represent the strength of the temporal effect.
* Blue arrows indicate a positive effect.
* Red arrows indicate a negative effect.
Strength centrality for cross-lagged network models of problem gambling symptoms. Contr = Loss of control, Toler = Tolerance, Chase = Chasing losses, Borw = Borrowing, Prob = Problem recognition, HlthP = Health problems, Critz = Criticism, FinaP = Financial problems, Guilt = Guilt, GSum = Gambling expenditure. Values on x-axis represent the sum of the partial correlations for a specific node when excluding a node’s predictive effect on itself (autoregressive effect).
Figure 4. Long description
The figure consists of two vertical panels sharing a common y-axis of ten gambling symptoms. The x-axis measures centrality values from 0.0 to 0.8. A legend on the right identifies red lines as the 2013-2015 network and teal lines as the 2019-2022 network.
Left Panel: OutStrength
* Toler: Red is at 0.22, Teal is at 0.18.
* Guilt: Red is at 0.43, Teal is at 0.07.
* Prob: Both converge at 0.35.
* Chase: Red is at 0.43, Teal is at 0.54.
* HlthP: Red is at 0.45, Teal is at 0.30.
* Contr: Red is at 0.31, Teal is at 0.38.
* GSum: Both converge at 0.05.
* Critz: Red is at 0.39, Teal is at 0.17.
* FinaP: Red is at 0.18, Teal is at 0.52.
* Borw: Red is at 0.58, Teal is at 0.70.
Right Panel: InStrength
* Toler: Red is at 0.88, Teal is at 0.47.
* Guilt: Red is at 0.40, Teal is at 0.66.
* Prob: Red is at 0.57, Teal is at 0.43.
* Chase: Red is at 0.45, Teal is at 0.37.
* HlthP: Red is at 0.23, Teal is at 0.43.
* Contr: Red is at 0.33, Teal is at 0.25.
* GSum: Red is at 0.25, Teal is at 0.18.
* Critz: Red is at 0.20, Teal is at 0.18.
* FinaP: Red is at 0.10, Teal is at 0.27.
* Borw: Red is at 0.05, Teal is at 0.09.
For generalizability, the correlation between 2019–2022 and 2013–2015 networks revealed that edge presence was φ = .08 (φ = .20 when excluding autoregressive effects) and the edge weight was = .74 (r = .27 when excluding autoregressive effects). The results indicate a low overlap in edge presence across the two cross-lagged networks and the similarity in edge weights was driven more by autoregressive paths compared to cross-lagged paths between these networks. Figure 4 shows a similar pattern of strength centrality estimates across nodes, with out-strength estimates differing relatively more between the directed networks compared to in-strength estimates.
Discussion
The present study used cross-sectional and cross-lagged psychometric network models to identify unique connections between PG symptoms and gambling expenditure, aiming to identify the most central nodes in PG symptomatology. Across both network types and samples from 2013 to 2022, a consistent pattern emerged: Finance-related symptoms were most strongly connected with and associated with other symptoms.
In the 2022 undirected network, financial problems due to gambling were the most central symptom. The central role of financial problems suggests that it can function as a threshold marker: When financial problems emerge in a general population context, they tend to co-occur with a broad range of other difficulties, which could signal a shift from lower-severity gambling involvement to a more pervasive symptom profile. Financial problems were also found to be the most central PG symptom in a previous study applying psychometric network analysis to the PGSI, although its centrality changed between two separate samples [Reference Horváth, Paksi, Fernández-Aranda, Jiménez-Murcia and Demetrovics14]. More broadly, previous investigations have identified financial problems among the most frequent gambling harms [Reference Langham, Thorne, Browne, Donaldson, Rose and Rockloff4, Reference Shannon, Anjoul and Blaszczynski29, Reference Syvertsen, Kristensen, Browne, Li and Pallesen30]. Financial problems from gambling can include losing savings, debt, bankruptcy, and difficulties affording basic expenses. Financial worries are in turn associated with elevated psychological distress [Reference Ryu and Fan31]. Furthermore, financial problems are among the most common reasons for help-seeking, which highlights the importance that individuals with PG place on it [Reference Loy, Grüne, Braun, Samuelsson and Kraus32].
In the 2019–2022 directed network, borrowing money or selling possessions to fund gambling emerged as the most central symptom by predicting increases in other PG symptoms at the subsequent time point. This suggests that borrowing is a predictive marker of broader PG symptom severity across different domains (finance, mental/physical health, emotional). The two time-point design cannot establish a dynamic or causal process linking borrowing to subsequent symptom escalation, although previous studies corroborate the association between borrowing and adverse outcomes. Individuals with PG who seek external funding from friends, family, or creditors when depleting their own financial resources expose themselves to additional interpersonal and legal risks [Reference Krishnan and Orford33–Reference Swanton and Gainsbury35]. This may include consumer credit use and debt, which among individuals with PG are linked to higher rates of comorbid mental health disorders, substance use, suicidality, and crime [Reference Riley, Harvey, Crisp, Battersby and Lawn34]. Affected others can conversely experience pressure to lend money and/or experience serious financial stress and conflict due to gambling debt [Reference Krishnan and Orford33, Reference Riley, Harvey, Crisp, Battersby and Lawn34]. Moreover, within the family context, this can result in family members returning to work or working overtime to relieve debt and ensure basic needs [Reference Riley, Harvey, Crisp, Battersby and Lawn34].
Feeling guilt from gambling or the consequences of gambling showed the highest in-strength centrality in the 2019–2022 directed network, meaning it was the symptom most strongly predicted by all other symptoms and gambling expenditure over time. While the present design cannot confirm that guilt was caused by the PG symptoms endorsed at time 1, the item content is still consistent with guilt as a downstream consequence of gambling. Guilt from gambling may not be driven by any single symptom but rather accumulates as a combined effect of engaging in multiple problematic gambling behaviors and experiencing multiple harms. Shame and guilt related to gambling are well-established barriers to help-seeking among individuals with PG [Reference Loy, Grüne, Braun, Samuelsson and Kraus32]. Relatedly, individuals with PG can develop self-stigma by internalizing beliefs such as being irresponsible, greedy, or lacking in self-constraint, which is associated with lower self-esteem and greater risk of relapse as well [Reference Hing, Nuske, Gainsbury and Russell36, Reference Quigley37]. Self-stigma is positively associated with severity of PG and is often coped with through social withdrawal and concealment of gambling [Reference Quigley37]. The effects of guilt, self-stigma, and resultant concealment may therefore explain why individuals with PG report low help-seeking despite experiencing concurrent financial problems that otherwise facilitate such help-seeking [Reference Loy, Grüne, Braun, Samuelsson and Kraus32].
The centrality of financial problems and borrowing may have implications for help services and harm-reduction policy. Frontline workers such as general practitioners and social workers encountering clients with financial difficulties should routinely assess gambling habits, as gambling-related financial harm may represent a tipping point for a broader symptom profile of PG. This is particularly relevant for early identification, given that only 1 in 25 with moderate-risk gambling and 1 in 5 with PG seek help [Reference Bijker, Booth, Merkouris, Dowling and Rodda38]. The value of intervening on borrowing to prevent long-term symptom escalation should be treated as a hypothesis for future intervention research. Relatedly, frontline workers can initiate brief motivational interventions [Reference Diskin and Hodgins39], refer to PG helplines [Reference Wall, Kristiansen, Molander, Forsström and Marionneau40], and encourage use of various responsible gambling measures [Reference Engebø, Torsheim and Pallesen41]. For individuals with more severe PG, the present findings suggest that there is complexity in addressing guilt from gambling. As guilt appears embedded in the broader symptom network, specialized treatment is likely needed to address it adequately. Nonetheless, frontline workers making referrals can prepare clients by explicitly addressing guilt and shame, while also informing about available debt counselling and relief programs [Reference Swanton and Gainsbury35]. Finally, at the policy level, the findings corroborate existing recommendations that governing bodies should prioritize measures that limit gambling-related financial harm [Reference Wardle, Degenhardt, Marionneau, Reith, Livingstone and Sparrow42]. Relevant Norwegian examples include the introduction of a national debt registry to prevent accumulation of unmanageable debt and mandatory loss limits enforced by state-owned gambling providers [Reference Pallesen, Mentzoni, Syvertsen, Kristensen, Erevik and Morken18, Reference Engebø, Torsheim and Pallesen41].
The secondary aim of the present study was to evaluate the accuracy, stability, and generalizability of PG network models. These measures were broadly supported for the undirected networks, in contrast to a previous psychometric network study of PGSI [Reference Horváth, Paksi, Fernández-Aranda, Jiménez-Murcia and Demetrovics14]. The options to evaluate accuracy, stability, and generalizability of the directed cross-lagged network models were more limited, as these models are novel [Reference Wysocki, McCarthy, Van Bork, Cramer and Rhemtulla7]. Correlation analyses and visual inspection indicated low generalizability, although differences in societal factors and/or methodological factors suggest that the results on generalizability for the directed networks should be interpreted as inconclusive and exploratory, as discussed in the following.
The Covid-19 pandemic occurred during the sampling frame for the 2019–2022 directed network when lockdown measures affected gambling availability. The Covid-19 pandemic represents a possible, albeit speculative, explanation for the low generalizability observed in the present study. In Norway, electronic gambling machines and bingo venues were temporarily closed, shifting gambling more online [Reference Pallesen, Mentzoni, Syvertsen, Kristensen, Erevik and Morken18]. Because cross-sectional and cross-lagged networks capture different aspects of symptom structure, pandemic-related disruptions may have altered symptom dynamics over time without substantially changing co-occurrence at any given measurement point. An Australian longitudinal study spanning 2019 to 2023 observed that gambling participation and PG levels shifted during the pandemic, but were highly similar at each endpoint [Reference Suomi, Kim, Hahn and Biddle43] – a pattern consistent with our finding that cross-sectional networks were similar while cross-lagged networks differed. The pandemic also intensified psychosocial risk factors like psychological distress and loneliness [e.g., Reference Hoffart, Johnson and Ebrahimi44, Reference Lu, Wang, Ebrahimi, Shen, Unnarsdóttir and Hauksdóttir45], which could represent one set of factors influencing the system dynamics of the directed networks in different ways [Reference Suomi, Kim, Hahn and Biddle43]. Alternatively, the differences between the directed networks may reflect methodological issues as the 2013–2015 network had a two-year follow-up while the 2019–2022 network had 3 years. Cross-lagged effects are sensitive to the measurement interval, as processes that emerge at a 2-year lag may subside or be overtaken by other processes at a 3-year lag [Reference Gauld, Grasman and Bailly46]. Due to the study designs, the substantive and methodological explanations cannot be fully disentangled in the present study.
The strengths of the present study include four large nationally representative samples covering nearly a decade, the inclusion of two directed networks addressing the lack of temporal models in PG network research, and formal evaluation of accuracy, stability, and generalizability. Data were collected via both digital and paper-based formats, mitigating selection bias related to participants’ digital experience. Limitations common to self-report apply, including inability to ascertain causality and susceptibility to social and recall bias. We observed low base rates of PG symptoms, which are common in PG population studies [Reference Gabellini, Lucchini and Gattoni1, Reference Horváth, Paksi, Fernández-Aranda, Jiménez-Murcia and Demetrovics14]. Potential threats to statistical power were addressed by modeling variables continuously rather than dichotomously, with bootstrapping procedures for edge accuracy and centrality stability for cross-sectional networks, and by leveraging larger sample sizes than previous comparable studies (especially for cross-lagged networks for PG [Reference Horváth, Paksi, Fernández-Aranda, Jiménez-Murcia and Demetrovics14]). Cross-sectional network comparisons also served as sensitivity analyses for these networks. However, comparisons for cross-lagged networks are less conclusive because of less-developed modeling techniques, and methodological and societal differences affecting the networks differently as discussed above. These limitations suggest some caution when interpreting these results. Finally, this study examined PG at the population level, which includes more individuals reporting no or milder symptoms and harms when compared to clinical populations comprising fewer individuals but more severe symptoms and harms [Reference Angus, Anjoul, Shannon and Blaszczynski47]. Accordingly, undirected and directed networks of PG will likely differ between population and clinical levels.
Conclusions
Financial problems and borrowing money to fund gambling were the most central symptoms in undirected and directed networks of PG, respectively. These finance-related symptoms may serve as candidates for screening and may, following further study, represent targets for early intervention and harm-reduction policy. Financial problems were associated with a broader PG symptom profile, and borrowing was associated with a broader profile of gambling harms across financial, emotional, and health domains over time, although the generalizability of borrowing’s role in directed networks was inconclusive in the present study. The routine assessment of gambling among individuals presenting financial problems and rapid response when borrowing first appears offer actionable points for frontline workers aiming to reduce gambling harm and facilitate timely referrals to helplines and specialized treatment.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1192/j.eurpsy.2026.12261.
Data availability statement
Anonymized data and R code are available at osf.io/w27u8.
Author contribution
A.S.: Conceptualization, methodology, data collection, data curation, formal analysis, writing – original draft/revision, and visualization. O.V.E.: Conceptualization and writing – review & editing. D.S.: Conceptualization and writing – review & editing. J.H.K.: Conceptualization, data collection, and writing – review & editing. S.P.: Conceptualization, data collection, writing – review & editing, and funding acquisition.
Financial support
This work was funded by the Norwegian Competence Center for Gambling and Gaming Research. Study data were obtained from general population surveys on gaming and gambling problems funded by the Norwegian Gaming Authority. No specific grant numbers were assigned.
Competing interests
None.
Ethical approval and informed consent
This study was approved by the Norwegian Centre for Research Data, project numbers 2013/120, 2019/698, 528056, and the Regional Committee for Medical and Health Related Research Ethics in Western Norway project number 460136. Participants provided informed consent.
Open access
Comments
No Comments have been published for this article.