Abstract
Recent legislative and technological changes have transformed the gambling landscape in the United States, underscoring the need to better understand both problem gambling and responsible gambling behaviors. The Positive Play framework conceptualizes responsible gambling from the player’s perspective, yet little is known about how these behaviors change over time. The present study examined longitudinal changes in Positive Play scores and problem gambling symptoms among a sample of adult gamblers (n = 2,942). Participants completed repeated assessments of the Positive Play Scale and the Problem Gambling Severity Index across five waves over approximately two years. Confirmatory factor analyses supported the original four-factor structure of the Positive Play Scale. Latent growth models indicated that scores across all Positive Play subscales increased over time, while problem gambling symptoms decreased. Notably, increases in gambling literacy scores were associated with greater reductions in problem gambling severity, whereas changes in other Positive Play domains were not. These findings suggest that gambling literacy, in particular, may be associated with reductions in gambling-related harm over time.
Keywords: Gambling, Problem gambling, Responsible gambling, Positive Play, Gambling literacy, Longitudinal analyses
Introduction
The gambling landscape of the United States (U.S.) has undergone significant changes over the past decade, largely driven by the expansion of online platforms and the legalization of sports betting. These developments have contributed to increased gambling participation (Ukhova & Volberg, 2025). The rise in sports betting underscores the growing need to understand both problem gambling and responsible gambling behaviors (Yeola et al., 2025). Responsible gambling (RG) encompasses policies, procedures, and individual behaviors intended to prevent or minimize gambling harm (Wood & Griffiths, 2015). One recent initiative in RG research is the Positive Play framework, which conceptualizes it from the player’s perspective (Wood et al., 2017). Positive Play is conceptualized as involving four purportedly distinct domains: personal responsibility, gambling literacy, pre-commitment, and honesty/control. Personal responsibility reflects the extent to which individuals accept accountability for the time and money they spend gambling. Gambling literacy assesses players’ understanding of gambling odds and the realistic chances of winning. Honesty/control evaluates players’ abilities regulate their gambling behavior and maintain transparency about their gambling activities. Pre-commitment captures the extent to which individuals plan how much time and money they intend to spend before engaging in gambling activities (Tabri et al., 2020).
Since its introduction, the Positive Play Scale (PPS) has been widely used to measure RG behaviors and beliefs. The original 33-item PPS, developed by Wood et al. (2017) and later refined through principal components analysis, established the multidimensional structure of the construct. Subsequent validation by Tabri et al. (2020) using exploratory structural equation modeling in a Canadian sample confirmed excellent model fit and supported the distinct yet correlated nature of these four domains.
Additional work has replicated and expanded upon the initial PPS findings. Using a Chinese sample, higher Positive Play scores were associated with fewer gambling disorder symptoms, using a validated two-factor behavioral and two-factor belief structure of the PPS (Tong et al., 2020). Similarly, higher PPS scores, particularly in pre-commitment and honesty/control, were associated with lower gambling harm across risk levels, even after accounting for problem gambling severity (Delfabbro et al., 2020). These behavioral subscales also moderated the relationship between gambling severity and harm, highlighting their potential protective role.
Applied research has also explored how Positive Play relates to gambling attitudes and behaviors. Among U.S. casino loyalty members, gambling literacy uniquely predicted attitudinal loyalty and higher PPS scores were associated with reduced monetary wagering (Hollingshead et al., 2023). Beyond behavioral outcomes, intervention-based work has examined the utility of the PPS in promoting RG. The use of personalized normative feedback through the Positive Play Quiz found that participants who received personalized normative feedback demonstrated greater insight into their gambling behaviors, increased intentions to gamble responsibly, and stronger self-efficacy to use harm-reduction tools such as deposit limits over a three-month follow up period (Tabri et al., 2024).
Recently, researchers have applied the Positive Play framework to the domain of sports gambling, which represents one of the fastest growing forms of gambling in the U.S. (Winters & Derevensky, 2020). King et al. (2024) found that individuals seeking treatment for gambling exhibited particularly low PPS scores, especially in personal responsibility and honesty/control. In addition, Connolly et al. (2024, 2026) observed that Positive Play behaviors varied across bet types, gambling locations, confidence, and website usage. For instance, live betting and complex wager types such as parlays and player props were associated with lower gambling literacy and personal responsibility. Moreover, higher self-reported sports knowledge was paradoxically associated with lower gambling literacy, suggesting a possible overconfidence effect. Similarly, individuals scoring low on all PPS subscales were more likely to be young, single males who primarily engaged in online gambling (Turner et al., 2025).
Despite its growing empirical support, the Positive Play framework has faced conceptual criticism. Some scholars have argued that Positive Play represents a relatively narrow conceptualization of RG and lacks the comprehensive scope of broader harm-reduction models (Shaffer et al., 2023). In response, Wood et al. (2023) clarified that Positive Play is not intended to replace systemic approaches but rather to complement public health models like the Reno Model (Blaszczynski et al., 2004) by fostering sustainable gambling practices at the individual level.
Although Positive Play has demonstrated links with both gambling behavior and harm, most research has been cross-sectional. As the U.S. gambling landscape continues to expand and change, there remains a pressing need to examine how RG behaviors evolve over time and how they relate to problem gambling symptoms. The present study longitudinally assesses the relationship between RG, using the PPS, and problem gambling severity symptoms over a two-year period in a general sample of the American population. This approach allows for a more comprehensive understanding of how RG behaviors develop and how they may relate to gambling-related problems over time.
Present Study
The primary aim of this study was to examine trajectories of the four Positive Play subscales – personal responsibility, gambling literacy, pre-commitment, and honesty/control – across multiple time points. In addition, we aimed to explore how these trajectories correspond to changes in problem gambling symptoms. Although the study was largely exploratory, we generally expected that increases in Positive Play behaviors and beliefs would be associated with reductions in problem gambling symptoms over time. More precisely, we expected a negative association between the trajectories of Positive Play subscales and problem gambling symptoms. This work also aligns with a broader resurgence of interest in RG initiatives, particularly those that emphasize player-centered approaches (Shaffer et al., 2026). By focusing on longitudinal patterns of Positive Play, this study contributes to ongoing efforts to identify and develop evidence-based strategies for fostering and maintaining sustainable RG practices. This approach allows for a more comprehensive understanding of how RG behaviors develop over time and how they relate to problem gambling symptoms.
Method
To compile a sample that broadly matched the demographics of the U.S., we recruited participants via YouGov America (N = 4,363). At the time of our original survey, we collected both a Census-matched sample of American adults (n=2,806; Mage=48.9, SD=17.2; 1365[48.6%] male; response rate=87.6%), stratified and matched to U.S. norms for age, gender, education, census region, and race/ethnicity as of the 2019 American Community Survey, as well as a specific oversample of sports-gambling adults (n=1,557; Mage=41.7, SD=15.3; 1043[67%] men; response rate=78.7%). These samples were combined for analysis. Survey results are reported in accordance with American Association for Public Opinion Research guidelines for nonprobability internet panels (American Association for Public Opinion Research, 2016).
Participants and Procedure
Participants were initially recruited between late March 2022 through early April 2022. Subsequent to baseline measures, participants were invited to participate in follow-up surveys every six months, beginning in Fall of 2022 (n=2,808; average interval=194.1 days, SD=5.7), continuing through Spring (n=2,279; average interval=180.9 days, SD=4.2) and Fall (n=1,734; average interval=193.1 days, SD=4.8) of 2023, and concluding in Spring of 2024 (n=1,570; average interval=182.3 days, SD=5.4) for a total of five waves of data collection. The average total length of the study was slightly over two years (average interval=754.5 days, SD=6.8 days). Participants were invited at each follow-up time point, regardless of previous wave completion (i.e., participants could miss a wave and still be invited to all subsequent waves).
For the current study, we restricted analyses to those who had reported any gambling at baseline. Across time points, 2,945 gamblers completed the first wave of data collection; 1,949 completed the second wave, 1,688 completed the third wave, 1,355 completed the fourth wave, and 924 completed all five waves. All analyses were run using this subsample of 2,945 gamblers, of whom 2,942 provided complete baseline Positive Play Scale data and comprise the final analytic sample (demographics reported in Supplemental Table 1).
Measures
Demographics such as gender, age, ethnicity, marital status, education, and employment status were reported at baseline (see Supplemental Table 1). At each point, we assessed gambling symptoms and RG behaviors using the following measures.
Problem Gambling Symptoms
Participants who reported gambling during the interval preceding each wave completed the Problem Gambling Severity Index (Ferris & Wynne, 2001).This 9-item scale assesses the frequency of problem gambling behaviors and consequences (e.g., “Have you felt that you might have a problem with gambling?”). Responses were recorded on a 4-point Likert scale ranging from 0 (never) to 3 (almost always). Scores were summed to yield a total index of problem gambling severity (see Table 2 for reports on internal consistency).
Table 2.
Descriptive Statistics, Reliabilities, and Within-Construct Correlations Across Waves
| Measure | M | SD | α | ω | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|---|---|---|---|
| PGSI | |||||||||
| Wave 1 | 3.77 | 6.03 | .88 | .89 | — | ||||
| Wave 2 | 2.42 | 5.12 | .89 | .90 | .81 | — | |||
| Wave 3 | 1.94 | 4.50 | .90 | .91 | .79 | .77 | — | ||
| Wave 4 | 1.57 | 3.92 | .91 | .91 | .76 | .76 | .75 | — | |
| Wave 5 | 1.34 | 3.68 | .92 | .92 | .72 | .68 | .73 | .75 | — |
| Personal Responsibility | |||||||||
| Wave 1 | 6.14 | 1.07 | .85 | .85 | — | ||||
| Wave 2 | 6.24 | 1.04 | .85 | .85 | .59 | — | |||
| Wave 3 | 6.37 | 0.93 | .84 | .84 | .57 | .55 | — | ||
| Wave 4 | 6.36 | 0.98 | .85 | .85 | .50 | .47 | .51 | — | |
| Wave 5 | 6.46 | 0.83 | .81 | .81 | .49 | .51 | .51 | .52 | — |
| Honesty and Control | |||||||||
| Wave 1 | 5.74 | 1.64 | .84 | .85 | — | ||||
| Wave 2 | 5.96 | 1.57 | .83 | .84 | .53 | — | |||
| Wave 3 | 6.06 | 1.49 | .82 | .83 | .54 | .50 | — | ||
| Wave 4 | 6.07 | 1.54 | .81 | .82 | .45 | .53 | .52 | — | |
| Wave 5 | 6.18 | 1.43 | .80 | .81 | .55 | .45 | .52 | .49 | — |
| Gambling Literacy | |||||||||
| Wave 1 | 5.21 | 1.52 | .75 | .76 | — | ||||
| Wave 2 | 5.67 | 1.27 | .72 | .74 | .71 | — | |||
| Wave 3 | 5.80 | 1.28 | .76 | .78 | .73 | .71 | — | ||
| Wave 4 | 5.92 | 1.22 | .73 | .76 | .68 | .68 | .72 | — | |
| Wave 5 | 6.02 | 1.14 | .75 | .77 | .70 | .67 | .70 | .74 | — |
| Pre-Commitment | |||||||||
| Wave 1 | 5.64 | 1.37 | .83 | .83 | — | ||||
| Wave 2 | 5.78 | 1.55 | .88 | .88 | .52 | — | |||
| Wave 3 | 5.89 | 1.49 | .88 | .88 | .51 | .47 | — | ||
| Wave 4 | 5.91 | 1.57 | .91 | .91 | .45 | .52 | .46 | — | |
| Wave 5 | 6.01 | 1.42 | .87 | .87 | .50 | .44 | .47 | .42 | — |
Note. All correlations are Pearson’s r and significant at p < .001. Means and standard deviations reflect participants with complete data for each wave. Cronbach’s α and omega total (ω) represent internal consistency reliability.
Positive Play Scale (PPS)
We assessed RG behaviors using the 14-item Positive Play Scale (Wood et al., 2017) which includes four subscales: personal responsibility (four items), gambling literacy (three items), pre-commitment (four items), and honesty/control (three items). The personal responsibility and gambling literacy items were anchored at 1 (strongly disagree) to 7 (strongly agree) while the pre-commitment and honesty and control items were anchored at 1 (never) and 7 (always). Two items in the Gambling Literacy subscale (“My chances of winning get better after I have lost” and “If I gamble more often, it will help me win more than I lose”) were both reverse coded, such that a higher score on the Gambling Literacy subscale indicated higher Positive Play in this domain. Subscale scores were calculated as the average of their constituent items (see Table 2 for reports on internal consistency). All participants who endorsed past six-month gambling frequency received this measure at each wave.
Data Analysis
Statistical Analysis
For all measures, we computed both descriptive statistics and measures of internal consistency (Cronbach’s α and McDonald’s ω) using the psych package in R Studio (Revelle, 2026). Within and across time points, we calculated Pearson correlations with Holm adjusted test statistics for all key variables. The Holm-adjustment is a sequentially rejective version of the simple Bonferroni correction for multiple comparisons and strongly controls the family-wise error rate at a designated alpha.
Measurement Structural and Confirmatory Factor Analysis
Prior research examining the structure of the PPS has used exploratory structural equation modeling (Tabri et al., 2020) and Confirmatory Factor Analysis (CFA; Tong et al., 2020). However, the CFA tested two alternative two-factor solutions corresponding to the Positive Play Beliefs and Positive Play Behaviors. In the current study, we began by testing the factorial structure of the PPS using CFA in our sample. We evaluated four potential models: the original four factor model (Wood et al., 2017), the two-factor model (Tong et al., 2020), and two versions of a three-factor model, one in which we combined behaviors into a single index while keeping beliefs as distinct categories and one model in which beliefs were combined into a single latent variable and behaviors were kept distinct.
Model fit was evaluated using multiple fit indices, including the chi-square goodness-of-fit test, the Comparative Fit Index, the Tucker-Lewis Index, the Root Mean Square Error of Approximation with 90% confidence interval, and the Standardized Root Mean Square Residual. In accordance with standard thresholds (Hu & Bentler, 1999), CFI and TLI values ≥.95, SRMR values ≤.08, and RMSEA values ≤.06 were utilized as indicators of good model fit, though RMSEA values up to .08 were considered acceptable. To evaluate competing, non-nested measurement models, the Akaike Information Criterion and Bayesian Information Criterion were examined, with lower values indicating superior fit.
Longitudinal Measurement Invariance
Because our primary hypotheses concern change over time, it is essential to establish that the PPS subscales measure the same constructs consistently across waves before interpreting latent growth parameters. We therefore tested longitudinal measurement invariance for each of the four Positive Play subscales across all five waves. For each subscale, we sequentially tested configural (equal factor structure), metric (equal factor loadings), and scalar (equal item intercepts) invariance. Model comparisons were evaluated using established fit-based criteria (ΔCFI ≥ −.010 and ΔRMSEA ≤ .015; Chen, 2007; Cheung & Rensvold, 2002) rather than chi-square difference tests, given the latter’s sensitivity to sample size.
Latent Growth Curve Models
To examine change in problem gambling and Positive Play constructs over time, we estimated a series of latent growth curve models (LGCs) using the lavaan (Rosseel, 2012), miTools (Lumley, 2019), and semTools (Jorgensen et al., 2026) packages. Missing data were handled using Full Information Maximum Likelihood methods. For LGCs, we used robust maximum likelihood estimation (MLR), which provides standard errors and a chi square test statistics that are robust to non-normality and non-independence of observations, as well as multiple fit indices, including Robust Comparative Fit Index (CFI), Robust Tucker-Lewis Indices (TLI), Robust Root Mean Square Error of Approximation (RMSEA), and Standardized Root Mean Squared Residual (SRMR). Following initial CFA, five univariate LGCs were estimated for each construct (PGSI and the four Positive Play subscales), followed by four parallel process LGCs evaluating the association between longitudinal changes in PGSI and each Positive Play Subscale.
Results
Confirmatory Factor Analysis
At the outset of this work, we sought to examine the structure of the scale in a confirmatory framework. Therefore, to validate competing theoretical conceptualizations of the PPS, a series of CFAs were conducted comparing four alternative measurement models (see above). As shown in Table 1, the original four-factor model demonstrated superior fit, with all global fit indices meeting or exceeding standard thresholds for acceptable to good psychometric fit. Furthermore, all individual items loaded strongly and significantly onto their respective latent factors within this four-factor model (see Supplemental Table 2 for the complete list of standardized factor loadings). In contrast, the alternative specifications failed to meet these baseline criteria.
Table 1. Confirmatory Factor Analyses Models.
Model Fit Indices for the four-factor, three-factor behaviors combined, three-factor beliefs combined, and two-factor models.
| χ2 | df | CFI | TLI | RMSEA [90% CI] | SRMR | AIC | BIC | ΔAIC | ΔBIC | |
|---|---|---|---|---|---|---|---|---|---|---|
| Four-factor | 929.33 | 71 | .961 | .951 | .068 [.064, .072] | .056 | 114,713.84 | 114,913.23 | - | - |
| Three-factor (behaviors combined) | 1490.44 | 74 | .936 | .922 | .086 [.082, .090] | .056 | 115,268.94 | 115,450.74 | 555.10 | 537.51 |
| Three-factor (beliefs combined) | 2007.41 | 74 | .913 | .893 | .100 [.096, .104] | .062 | 115,785.91 | 115,967.71 | 1072.07 | 1054.48 |
| Two-factor | 2555.41 | 76 | .889 | .867 | .112 [.108, .116] | .061 | 116329.92 | 116,499.98 | 1616.08 | 1586.75 |
Lower values of AIC and BIC indicate better relative model fit. ΔAIC and ΔBIC values are reported relative to the four-factor model.
Specifically, the three-factor model combining pre-commitment and honesty/control into a single factor exhibited poorer, inadequate fit, while the alternative three-factor model combining gambling literacy and personal responsibility also exhibited unacceptable fit. Finally, the two-factor model demonstrated the poorest overall fit, failing to meet most established psychometric thresholds (see Table 1).
Model comparison statistics further favored the four-factor model. Because a decrease in information criteria indicates a relatively better-fitting and more parsimonious model, the substantial differences in AIC and BIC values (Δ > 10) provided overwhelming statistical evidence in favor of the original four-factor structure over all alternative configurations. As such, the four-factor model was retained for all subsequent analyses (see Table 1).
Longitudinal Measurement Invariance
Personal Responsibility, Honesty and Control, and Pre-Commitment each demonstrated full metric and scalar invariance across all five waves, indicating that these constructs were measured consistently over time. Gambling Literacy showed full configural invariance and partial metric/scalar invariance: initial testing revealed non-invariance driven by a single item (Item G), which was worded affirmatively at Wave 1 but negatively at Waves 2 through 5. After releasing the equality constraints on this item’s loading and intercept at Wave 1, the subscale demonstrated excellent partial metric and scalar invariance; a supplementary analysis restricted to Waves 2 through 5 confirmed full invariance across the remaining four waves. Full model comparison statistics for all four subscales are reported in Supplementary Table 5. Because full or partial scalar invariance was established for every subscale, the observed changes in Positive Play scores over time can be interpreted as reflecting genuine change in the underlying constructs rather than shifts in item functioning.
Descriptive Statistics
Descriptive statistics and correlations for all observed variables are presented in Table 2. Correlations among baseline measures were in the expected directions, with higher scores on Positive Play subscales generally associated with lower PGSI scores. All measures were significantly correlated both within and across time points.
Single Latent Growth Curve Models
Single latent growth curve (LGC) models were estimated for each of the five key constructs (the four Positive Play subscales and PGSI) to examine change over time. All single LGC models demonstrated excellent fit (CFI and TLI > .98, RMSEA < .05; see Table 3). PGSI scores declined significantly over time, and under our primary specification the slope variance was positive, admissible, and statistically significant (slope Var = 0.244, SE = .113, p = .030; intercept Var = 25.491, SE = 1.296, p < .001), indicating significant individual variability in trajectories of change. As detailed below and in the Supplementary Materials, this estimate was replicated under multiple imputation and was attenuated (similar in magnitude but no longer significant) in a more conservative active-gambler sensitivity analysis. Each Positive Play subscale demonstrated a significant positive slope, indicating increases in responsible-gambling behaviors and beliefs over time. Slope variance was significant for gambling literacy (slope = .103, SE = .011, p < .001; slope Var = 0.011, SE = .004, p = .015; intercept Var = 1.265, SE = .051, p < .001), but not for personal responsibility (slope = .032, SE = .009, p < .001; slope Var = −0.005, SE = .004, p = .225; intercept Var = 0.532, SE = .038, p < .001) or honesty/control (slope = .065, SE = .013, p < .001; slope Var = 0.010, SE = .008, p = .196; intercept Var = 1.223, SE = .074, p < .001). For pre-commitment, preliminary estimation produced a negative (Heywood) slope-variance estimate, so this parameter was fixed to zero in the final model (slope = .068, SE = .011, p < .001; intercept Var = 0.885, SE = .041, p < .001). All slope means were significantly different from zero and in the hypothesized directions, indicating a general trend of decreasing gambling severity alongside increasing Positive Play behaviors and attitudes.
Table 3. Single Growth Curve Models.
Intercept, Slope, and Model Fit statistics for unconditional single process latent growth curves.
| Mean | SE | p-value | Variance | SE | p-value | ||
|---|---|---|---|---|---|---|---|
|
| |||||||
| PGSI | Intercept | 3.718 | 0.152 | <.001 | 25.491 | 1.296 | <.001 |
| Slope | −0.328 | 0.042 | <.001 | 0.244 | 0.113 | .030 | |
| Gambling Literacy | Intercept | 5.337 | 0.036 | <.001 | 1.265 | 0.051 | <.001 |
| Slope | 0.103 | 0.011 | <.001 | 0.011 | 0.004 | .015 | |
| Personal Responsibility | Intercept | 6.176 | 0.027 | <.001 | 0.532 | 0.038 | <.001 |
| Slope | 0.032 | 0.009 | <.001 | −0.005 | 0.004 | .225 | |
| Pre-Commitment | Intercept | 5.680 | 0.036 | <.001 | 0.885 | 0.041 | <.001 |
| Slope | 0.068 | 0.011 | <.001 | Fixed to 0 | - | - | |
| Honesty and Control | Intercept | 5.795 | 0.042 | <.001 | 1.223 | 0.074 | <.001 |
| Slope | 0.065 | 0.013 | <.001 | 0.010 | 0.008 | .196 | |
|
| |||||||
| Model Fit Statistics | |||||||
|
| |||||||
| CFI | TLI | RMSEA | SRMR | ||||
| PGSI | 1.000 | 1.005 | 0.000 | 0.019 | |||
| Gambling Literacy | 0.990 | 0.987 | 0.052 | 0.033 | |||
| Personal Responsibility | 0.996 | 0.996 | 0.028 | 0.024 | |||
| Pre-commitment | 1.000 | 1.000 | 0.000 | 0.021 | |||
| Honesty and Control | 0.997 | 0.997 | 0.022 | 0.018 | |||
Models were estimated using latent growth curve modeling with robust maximum likelihood estimation and full-information maximum likelihood methods for missing data. SE = Standard Error; PGSI = Problem Gambling Severity Index, CFI = Comparative Fit Index, TLI = Tucker-Lewis Index, RMSEA = Root Mean Square Error of Approximation, SRMR = Standardized Root Mean Square Residual
Parallel Process Latent Growth Models
To evaluate whether longitudinal changes in Positive Play constructs were associated with changes in problem gambling severity, four parallel process latent growth models were estimated. Each model included a simultaneous estimation of PGSI intercept and slope alongside one Positive Play construct, controlling for age and gender. Model fit was consistently strong, with robust CFI and TLI values above .99 and robust RMSEA values below .04 across all four models (see Tables 4–7).
Table 4. PGSI GL.
Intercept, Slope, and Model Fit statistics for parallel process latent growth curves.
| Mean | SE | p-value | Variance | SE | p-value | |
|---|---|---|---|---|---|---|
|
| ||||||
| PGSI Intercept | 3.717 | 0.152 | <.001 | 25.591 | 1.288 | <.001 |
| PGSI Slope | −0.330 | 0.041 | <.001 | 0.247 | 0.112 | .027 |
| Gambling Literacy Intercept | 5.336 | 0.037 | <.001 | 1.309 | 0.052 | <.001 |
| Gambling Literacy Slope | 0.103 | 0.011 | <.001 | 0.012 | 0.004 | .005 |
|
| ||||||
| Slope/Intercept Correlations | ||||||
|
| ||||||
| PGSI Intercept | PGSI Slope | Gambling Literacy Intercept | ||||
| PGSI Intercept | 1 | |||||
| PGSI Slope | r=−.620, p=.001 | 1 | ||||
| Gambling Literacy Intercept | r=−.683, p<.001 | r=.552, p<.001 | 1 | |||
| Gambling Literacy Slope | r=.409, p<.001 | r=−.787, p<.001 | r=−.322, p=.002 | |||
|
| ||||||
| Model Fit: CFI= .993, TLI=.991, RMSEA= .038, SRMR= .040 | ||||||
Models were estimated using latent growth curve modeling with robust maximum likelihood estimation and full-information maximum likelihood methods for missing data. SE = Standard Error; PGSI = Problem Gambling Severity Index, CFI = Comparative Fit Index, TLI = Tucker-Lewis Index, RMSEA = Root Mean Square Error of Approximation, SRMR = Standardized Root Mean Square Residual
Table 7.
PGSI Honesty and Control
| Mean | SE | p-value | Variance | SE | p-value | |||
|---|---|---|---|---|---|---|---|---|
|
| ||||||||
| PGSI Intercept | 3.719 | 0.152 | <.001 | 25.486 | 1.294 | <.001 | ||
| PGSI Slope | −0.325 | 0.041 | <.001 | 0.245 | 0.113 | .030 | ||
| HC Intercept | 5.805 | 0.041 | <.001 | 1.194 | 0.072 | <.001 | ||
| HC Slope | 0.062 | 0.013 | <.001 | 0.010 | 0.008 | .173 | ||
|
| ||||||||
| Slope/Intercept Correlations | ||||||||
| PGSI Intercept | PGSI Slope | HC Intercept | ||||||
| PGSI Intercept | 1 | |||||||
| PGSI Slope | r=−.621, p=.001 | 1 | ||||||
| HC Intercept | r=−.572, p<.001 | r=.397, p=.001 | 1 | |||||
| HC Slope | r=−.168, p=.140 | r=−.389 p=.094 | r=−.191, p=.277 | |||||
|
| ||||||||
| Model Fit: CFI= 1.000, TLI=1.001, RMSEA= .000, SRMR= .019 | ||||||||
Models were estimated using latent growth curve modeling with robust maximum likelihood estimation and full-information maximum likelihood methods for missing data. SE = Standard Error; PGSI = Problem Gambling Severity Index, CFI = Comparative Fit Index, TLI = Tucker-Lewis Index, RMSEA = Root Mean Square Error of Approximation, SRMR = Standardized Root Mean Square Residual.
In the model examining gambling literacy and PGSI, the slopes were significantly negatively correlated (r = −.787, p < .001), indicating that greater increases in gambling literacy over time were associated with steeper declines in problem gambling severity. This association was robust across estimation approaches (pooled multiple imputation r = −.72) and, in a conservative sensitivity analysis restricted to active gamblers, was similar in magnitude or stronger (r = −.92; see Supplementary Materials). This association also held when Gambling Literacy was recomputed excluding the single item whose wording direction reversed between Wave 1 and Waves 2–5 (Item G), ruling out this item as the source of the association (r = −.718, p = .011; see Supplemental Table 8). Baseline levels were also strongly correlated, with a significant negative correlation between intercepts (r = −.683, p < .001; see Table 4).
In the model examining personal responsibility and PGSI, intercepts were significantly negatively correlated (r = −.572, p < .001), suggesting that individuals with higher baseline personal responsibility reported lower baseline gambling problems. However, the slope-slope correlation was not significant (r = −.159, p = .511), indicating that increases in personal responsibility over time were not associated with concurrent decreases in PGSI (see Table 5).
Table 5. PGSI PR.
Intercept, Slope, and Model Fit statistics for parallel process latent growth curves.
| Mean | SE | p-value | Variance | SE | p-value | |
|---|---|---|---|---|---|---|
|
| ||||||
| PGSI Intercept | 3.723 | 0.152 | <.001 | 25.648 | 1.290 | <.001 |
| PGSI Slope | −0.328 | 0.043 | <.001 | 0.256 | 0.113 | .024 |
| Personal Responsibility Intercept | 6.181 | 0.027 | <.001 | 0.535 | 0.037 | <.001 |
| Personal Responsibility Slope | 0.031 | 0.009 | <.001 | −0.005 | 0.004 | .270 |
|
| ||||||
| Slope/Intercept Correlations | ||||||
| PGSI Intercept | PGSI Slope | Personal Responsibility Intercept | ||||
| PGSI Intercept | 1 | |||||
| PGSI Slope | r=−.613, p=.002 | 1 | ||||
| Personal Responsibility Intercept | r=−.572, p<.001 | r=.511, p<.001 | 1 | |||
| Personal Responsibility Slope | r=−.201, p=.069 | r=−.159, p=.511 | r=.212, p=.247 | |||
|
| ||||||
| Model Fit: CFI= 1.000, TLI=1.002, RMSEA= .000, SRMR= .026 | ||||||
Models were estimated using latent growth curve modeling with robust maximum likelihood estimation and full-information maximum likelihood methods for missing data. SE = Standard Error; PGSI = Problem Gambling Severity Index, CFI = Comparative Fit Index, TLI = Tucker-Lewis Index, RMSEA = Root Mean Square Error of Approximation, SRMR = Standardized Root Mean Square Residual
Similarly, in the pre-commitment-PGSI model, baseline levels were negatively associated (r = −.593, p < .001; see Table 6). Because preliminary estimation produced a negative (Heywood) slope variance for pre-commitment, this parameter was fixed to zero in the final model; as a result, individual differences in the rate of change for pre-commitment could not be meaningfully estimated or correlated with change in PGSI. The honesty/control–PGSI model showed a similar pattern for intercepts, which were significantly correlated (r = −.572, p < .001), but no significant correlation was observed between the slopes (r = −.389, p = .094; see Table 7).
Table 6.
PGSI Pre-commitment
| Mean | SE | p-value | Variance | SE | p-value | |
|---|---|---|---|---|---|---|
|
| ||||||
| PGSI Intercept | 3.716 | 0.152 | <.001 | 25.500 | 1.297 | <.001 |
| PGSI Slope | −0.320 | 0.041 | <.001 | 0.254 | 0.114 | .026 |
| Pre-Commitment Intercept | 5.688 | 0.036 | <.001 | 0.859 | 0.036 | <.001 |
| Pre-Commitment Slope | 0.066 | 0.011 | <.001 | 0 (fixed) | - | - |
|
| ||||||
| Slope/Intercept Correlations | ||||||
| PGSI Intercept | PGSI Slope | Pre-Commitment Intercept | ||||
| PGSI Intercept | 1 | |||||
| PGSI Slope | r=−.622, p=.001 | 1 | ||||
| Pre-Commitment Intercept | r=−.593, p<.001 | r=.297, p<.001 | 1 | |||
| Pre-Commitment Slope | - | - | - | |||
|
| ||||||
| Model Fit: CFI= .997, TLI=.996, RMSEA= .022, SRMR= .032 | ||||||
Models were estimated using latent growth curve modeling with robust maximum likelihood estimation and full-information maximum likelihood methods for missing data. SE = Standard Error; PGSI = Problem Gambling Severity Index, CFI = Comparative Fit Index, TLI = Tucker-Lewis Index, RMSEA = Root Mean Square Error of Approximation, SRMR = Standardized Root Mean Square Residual. Pre-Commitment slope variance was fixed to zero after preliminary estimation produced a negative (Heywood) variance; associated covariances were constrained accordingly.
Discussion
The present study represents the first known longitudinal investigation of Positive Play behaviors and their relationship with problem gambling symptom severity over time in a census-matched U.S. sample. Overall, these findings provide encouraging evidence that increases in at least one element of RG, gambling literacy, were associated with decreases in problem gambling symptoms over time. Given the observational, correlational design of this study these associations should not be interpreted as evidence of a causal effect.
Consistent with prior validation work (Tabri et al., 2020), our CFAs supported the original four factor structure of the Positive Play Scale (Wood et al., 2017) over all alternative models. In short, the measure demonstrated acceptable structural integrity within this large, longitudinal dataset when the four-factor model was retained. As hypothesized, PGSI scores declined significantly across the study period, indicating a general reduction in gambling-related symptoms among participants. Concurrently, all four Positive Play subscales showed significant growth, suggesting that RG beliefs and behaviors tend to strengthen over time. This simultaneous trend suggests it may be worthwhile to explore interventions or public health efforts that foster greater RG engagement for consumers, particularly those emphasizing both cognitive and behavioral dimensions of gambling behavior.
Although overall trends pointed toward decreasing gambling problems and increasing responsible play, the trajectories of individual Positive Play domains differed in their association with problem gambling severity. Notably, gambling literacy emerged as the only robust associate of reduced problem gambling severity. The strong negative correlation between changes in gambling literacy and PGSI scores indicates that faster gains in gambling literacy paralleled faster declines in problem gambling severity. This aligns with previous cross-sectional work (Connolly et al., 2024; 2026) and suggests gambling literacy co-occurs with reduced problem gambling severity over time.
In contrast, while personal responsibility, pre-commitment, and honesty/control all increased over time on average, individual differences in their rates of change were not associated with concurrent change in PGSI (for pre-commitment, slope variance was fixed to zero after preliminary estimation produced a negative estimate, precluding a slope-slope correlation). In each case, baseline levels of these constructs were linked with lower initial problem gambling, but changes over time did not significantly predict concurrent changes in gambling severity. This suggests that although these domains show upward trends over the long term, they may function as steady, enduring orientations toward responsible gambling rather than dynamic, volatile behaviors that mirror short-term changes in gambling severity. These results echo prior observation that certain RG constructs may be trait-like and less sensitive to interventions targeting behavior change (Gainsbury et al., 2014; Tabri et al., 2020).
Taken together, these findings expand the growing body of literature on Positive Play as a framework for understanding individual-level RG behaviors (Wood et al., 2017). However, while previous studies have primarily examined cross-sectional associations between Positive Play and problem gambling symptoms, the longitudinal design of the present work offers new insight into how these relationships evolve over time. In particular, the results highlight gambling literacy as a construct worth further attention in efforts to reduce problem gambling severity among at-risk groups, relative to personal responsibility, honesty/control, and pre-commitment. Given the correlational nature of these findings, however, we frame this as a promising target for future intervention research rather than a demonstrated causal mechanism. Importantly, while these consumer-side factors are worth investigating, individual-level behaviors represent only one component of a holistic responsible gambling framework, which must operate alongside systemic industry practices and environmental regulations.
From a policy perspective, these findings support gambling literacy may be a promising target for future intervention efforts, perhaps analogous to public health initiatives focused on financial or digital literacy. Such programs may be especially effective if tailored to specific subgroups (e.g., young adults, sports bettors) or embedded directly into online gambling environments through in-app tools, prompts, or feedback mechanisms.
Limitations and Future Directions
Several limitations warrant attention. First, while the longitudinal design strengthens causal inference, the study remains observational, and unmeasured confounding variables cannot be ruled out. A further limitation is that the observed changes may partly reflect methodological artifact rather than genuine psychological change. Repeated assessment can itself alter self-reported behavior (French & Sutton, 2010; Miles et al., 2020) and with five waves of assessment, we cannot rule this out. Regression to the mean is also plausible given the sample’s elevated baseline PGSI (M = 3.77). Socially desirable responding is a further documented source of bias in gambling self-report, affecting both the accuracy of retrospective recall (Goldstein et al., 2017) and the magnitude of change captured across repeated assessment (Schell et al., 2021); relatedly, participants whose gambling worsens may be disproportionately likely to drop out, inflating apparent improvement among completers. We note that these are general mechanisms that should act similarly across all four Positive Play subscales, yet only Gambling Literacy showed a reliable association with PGSI change, a pattern not fully predicted by a generic artifact account, though it does not rule one out. To probe this directly, we conducted two additional analyses reported in the Supplementary Materials. First, a decomposition of the PGSI decline showed that the proportion of non-gambling (zero) intervals was stable across follow-up (approximately 18–22%) rather than rising, and that mean PGSI continued to decline substantially among active gamblers alone, indicating that the observed reduction reflects predominantly genuine severity change rather than reduced gambling frequency (see Supplemental Table 7). Second, a sensitivity analysis restricted to active gamblers (treating non-gambling intervals as missing rather than zero) reproduced the gambling-literacy–PGSI slope association, which cannot be an artifact of the structural-zero coding because it holds where no such zeros are present (see Supplemental Table 6). Future work using behavioral or unobtrusive indicators of gambling activity would help disentangle true change from measurement artifact.
Second, participant attrition was nontrivial across waves, with approximately 31% completing all five time points, potentially introducing bias. We addressed this by comparing completers and non-completers on baseline characteristics, modeling predictors of non-completion, and re-estimating all growth models via multiple imputation as a robustness check (Supplementary Tables 3 and 4); missingness was related to observed baseline variables, consistent with a missing-at-random mechanism, though we cannot rule out that unobserved factors such as continued gambling involvement also contributed. Future work should examine retention strategies and further sensitivity analyses.
Moreover, the lack of significant slope correlations between some Positive Play subscales and PGSI suggests the need for more granular measurement. It is possible that specific behaviors within broader constructs of pre-commitment or honesty/control are more predictive of gambling severity than aggregate scores suggest. Future research could employ ecological momentary assessment or passive digital tracking to capture within-person fluctuations in gambling behavior more accurately.
Finally, although the current sample was large and combined a census-matched general-population sample with an oversample of sports gamblers, further research is needed to examine how these trajectories differ across demographic groups, gambling modalities, and clinical versus non-clinical populations. Tailoring interventions based on these subgroup differences may enhance the precision and effectiveness of RG initiatives.
Conclusion
Together, these findings indicate that although all four Positive Play constructs improved over time, only gambling literacy demonstrated a significant longitudinal association with reductions in problem gambling severity. Specifically, increased gambling literacy was associated with both lower baseline PGSI scores and steeper declines in gambling problems across time. In contrast, although higher baseline levels of personal responsibility, pre-commitment, and honesty/control were linked with fewer gambling problems at baseline, growth in these constructs was not associated with changes in gambling-related severity. These results suggest that gambling literacy, of all the Positive Play constructs, may be a promising target for future research into problem gambling prevention and treatment.
Supplementary Material
Footnotes
Generative AI Statement: During the revision process, the authors used Claude Opus 4.8 to annotate and troubleshoot statistical coding and used for editing and clarity. The author(s) reviewed and edited the output as needed and take full responsibility for the content of the published article.
Data Availability Statement:
The data that supports the findings of this study will be made publicly available in an open-access repository following publication. A link to the dataset will be provided upon acceptance.
References
- American Association for Public Opinion Research. (2016). Standard Definitions: Final Dispositions of Case Codes and Outcome Rates for Surveys. [https://aapor.org/standards-and-ethics/standard-definitions/](https://aapor.org/standards-and-ethics/standard-definitions/)
- Blaszczynski A, Ladouceur R, & Shaffer HJ. (2004). A science-based framework for responsible gambling: The Reno model. Journal of Gambling Studies, 20(3), 301–317. 10.1023/B:JOGS.0000040281.49444.e2 [DOI] [PubMed] [Google Scholar]
- Chen FF. (2007). Sensitivity of Goodness of Fit Indexes to Lack of Measurement Invariance. Structural Equation Modeling: A Multidisciplinary Journal, 14(3), 464–504. 10.1080/10705510701301834 [DOI] [Google Scholar]
- Cheung GW, & Rensvold RB. (2002). Evaluating Goodness-of-Fit Indexes for Testing Measurement Invariance. Structural Equation Modeling: A Multidisciplinary Journal, 9(2), 233–255. 10.1207/S15328007SEM0902_5 [DOI] [Google Scholar]
- Connolly AJ, Faruk R, Grant Weinandy JT, Stall LM, Kraus SW, & Grubbs JB. (2026). Exploring the relationship between Positive Play and perceived knowledge in traditional sports gambling. International Gambling Studies, 0(0), 1–20. 10.1080/14459795.2026.2697457 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Connolly A, Stall L, Floyd C, Kraus S, & Grubbs J. (2024). Exploring the Differences in Positive Play among Various Sports Wagering Behaviors. Journal of Gambling Studies, 1–24. 10.1007/s10899-024-10304-8 [DOI] [PubMed] [Google Scholar]
- Delfabbro P, King DL, & Georgiou N. (2020). Positive play and its relationship with gambling harms and benefits. Journal of Behavioral Addictions, 9(2), 363–370. 10.1556/2006.2020.00041 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ferris J, & Wynne H. (2001). THE CANADIAN PROBLEM GAMBLING INDEX: FINAL REPORT. [Google Scholar]
- French DP, & Sutton S. (2010). Reactivity of measurement in health psychology: How much of a problem is it? What can be done about it? British Journal of Health Psychology, 15(3), 453–468. 10.1348/135910710X492341 [DOI] [PubMed] [Google Scholar]
- Gainsbury SM, Suhonen N, & Saastamoinen J. (2014). Chasing losses in online poker and casino games: Characteristics and game play of Internet gamblers at risk of disordered gambling. Psychiatry Research, 217(3), 220–225. 10.1016/j.psychres.2014.03.033 [DOI] [PubMed] [Google Scholar]
- Goldstein A, Vilhena-Churchill N, Munroe M, Stewart S, Flett G, & Hoaken P. (2017). Understanding the Effects of Social Desirability on Gambling Self-Reports. International Journal of Mental Health and Addiction, 15. 10.1007/s11469-016-9668-0 [DOI] [Google Scholar]
- Hollingshead SJ, Davis CG, & Wohl MJA. (2023). The customer-brand relationship in the gambling industry: Positive play predicts attitudinal and behavioral loyalty. International Gambling Studies, 23(1), 118–138. 10.1080/14459795.2022.2086992 [DOI] [Google Scholar]
- Hu L, & Bentler PM. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling: A Multidisciplinary Journal, 6(1), 1–55. 10.1080/10705519909540118 [DOI] [Google Scholar]
- Jorgensen TD, Pornprasertmanit S, Schoemann AM, Rosseel Y, Miller P, Quick C, Garnier-Villarreal M, Selig J, Boulton A, Preacher K, Coffman D, Rhemtulla M, Robitzsch A, Enders C, Arslan R, Clinton B, Panko P, Merkle E, Chesnut S, … Vanbrabant L. (2026). semTools: Useful Tools for Structural Equation Modeling (Version 0.5-8) [Computer software]. https://cran.r-project.org/web/packages/semTools/index.html [Google Scholar]
- King A, Kraus SW, Wong-Padoongpatt G, & Grubbs JB. (2024). Can Positive Play Deficits Explain the Associations Between Posttraumatic Stress Symptoms, Gambling Motives, and Problem Gambling? Results of a National U.S. Sample. Journal of Gambling Studies, 40(4), 2093–2117. 10.1007/s10899-024-10348-w [DOI] [PubMed] [Google Scholar]
- Lumley T. (2019). mitools: Tools for Multiple Imputation of Missing Data (Version 2.4) [Computer software]. https://cran.r-project.org/web/packages/mitools/index.html
- Miles LM, Rodrigues AM, Sniehotta FF, & French DP. (2020). Asking questions changes health-related behavior: An updated systematic review and meta-analysis. Journal of Clinical Epidemiology, 123, 59–68. 10.1016/j.jclinepi.2020.03.014 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Revelle W. (2026). psych: Procedures for Psychological, Psychometric, and Personality Research (Version 2.6.1) [Computer software]. https://cran.r-project.org/web/packages/psych/index.html [Google Scholar]
- Rosseel Y. (2012). lavaan: An R Package for Structural Equation Modeling. Journal of Statistical Software, 48, 1–36. 10.18637/jss.v048.i02 [DOI] [Google Scholar]
- Schell C, Godinho A, & Cunningham JA. (2021). Examining Change in Self-Reported Gambling Measures Over Time as Related to Socially Desirable Responding Bias. Journal of Gambling Studies, 37(3), 1043–1054. 10.1007/s10899-020-09970-1 [DOI] [PubMed] [Google Scholar]
- Shaffer HJ, Ladouceur R, & Blaszczynski A. (2023). A Comment: Positive Play is a Subset of Responsible Gambling. Journal of Gambling Studies, 39(2), 1019–1025. 10.1007/s10899-023-10204-3 [DOI] [PubMed] [Google Scholar]
- Shaffer HJ, Ladouceur R, & Blaszczynski A. (2026). Twenty Years of Responsible Gambling: The Science-Based Glass is Half Full. International Journal of Mental Health and Addiction, 24(1), 21–23. 10.1007/s11469-024-01398-7 [DOI] [Google Scholar]
- Tabri N, Wood RTA, Philander K, & Wohl MJA. (2020). An examination of the validity and reliability of the Positive Play Scale: Findings from a Canadian national study. International Gambling Studies, 20(2), 282–295. 10.1080/14459795.2020.1732442 [DOI] [Google Scholar]
- Tabri N, Wood RTA, & Wohl MJA. (n.d.). Effects of personalized and normative feedback via the Positive Play Quiz on responsible gambling intention, self-efficacy and behavior: A randomized controlled trial. 10.1111/add.16722 [DOI] [PMC free article] [PubMed]
- Tong KK, Chen JH, & Wu AMS. (2020). Validation and Application of the Positive Play Scale Adapted for Chinese Gamblers: Its Relation to Disordered Gambling and Gambling Attitudes. Frontiers in Psychology, 11, 263. 10.3389/fpsyg.2020.00263 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Turner NE, AlMakadma Y, & Christensen DR. (2025). The Profiles of People who Do not Engage in Positive Play while Gambling. Journal of Gambling Studies. 10.1007/s10899-025-10384-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ukhova D, & Volberg R. (2025). Gambling Harms in the Era of Legal Gambling Expansion-A Growing Health Problem. JAMA Internal Medicine, 185(4), 389–390. 10.1001/jamainternmed.2024.8189 [DOI] [PubMed] [Google Scholar]
- Winters KC, & Derevensky JL. (2020). A review of sports wagering: Prevalence, characteristics of sports bettors, and association with problem gambling. Journal of Gambling Issues, 43, 102–127. psyh (2020-60635-006). [Google Scholar]
- Wood RTA, & Griffiths MD. (2015). Understanding Positive Play: An Exploration of Playing Experiences and Responsible Gambling Practices. Journal of Gambling Studies, 31(4), 1715–1734. 10.1007/s10899-014-9489-7 [DOI] [PubMed] [Google Scholar]
- Wood RTA, Wohl MJA, Tabri N, & Philander K. (2017). Measuring Responsible Gambling amongst Players: Development of the Positive Play Scale. Frontiers in Psychology, 8, 227. 10.3389/fpsyg.2017.00227 [DOI] [PMC free article] [PubMed] [Google Scholar]
- Wood RTA, Wohl MJA, Tabri N, & Philander K. (2023). Responsible Gambling as an Evolving Concept and the Benefits of a Positive Play Approach: A Reply to Shaffer et al. Journal of Gambling Studies. 10.1007/s10899-023-10245-8 [DOI] [PubMed] [Google Scholar]
- Yeola A, Allen MR, Desai N, Poliak A, Yang KH, Smith DM, & Ayers JW. (2025). Growing Health Concern Regarding Gambling Addiction in the Age of Sportsbooks. JAMA Internal Medicine, 185(4), 382–389. 10.1001/jamainternmed.2024.8193 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The data that supports the findings of this study will be made publicly available in an open-access repository following publication. A link to the dataset will be provided upon acceptance.
