Evaluation of the continuum of gambling problems using the DSM-IV

Transcription

1 METHODS AND TECHNIQUES doi: /j x Evaluation of the continuum of gambling problems using the DSM-IV David R. Strong 1 & Christopher W. Kahler 2 Butler Hospital, Brown Medical School 1 and Center for Alcohol and Addictions Studies, Brown University, Providence, RI, USA 2 ABSTRACT Aims To assess the measurement properties of the 10 symptoms of pathological gambling defined by the Diagnostic and Statistical Manual of Mental Disorders, fourth edition (DSM-IV) in a general population survey, including examination of unidimensionality and relative severity of the symptoms and their typical patterning. Design We conducted a Rasch model item response analysis of data from the National Epidemiologic Survey on Alcohol and Related Conditions (NESARC). Setting The NESARC surveyed a nationally representative sample from the United States. Participants Participants were individuals who reported gambling more than five times in a single year. Measurements The NESARC survey employed the Alcohol Use Disorder and Associated Disability Interview Schedule-DSM-IV version (AUDADIS-IV). Findings and conclusions DSM-IV problem gambling symptoms were strongly unidimensional and maintained a reliable ordering across a broad range of the problem gambling continuum, indicating that the 10 symptoms can be used to create an additive index of problem severity. The DSM-IV-based symptom index appeared to have sufficient reliability to separate life-time pathological gamblers from other gamblers using the current diagnostic threshold of five or more symptoms. However, the DSM-IV symptom index did not have sufficient reliability to separate further among groups of gamblers who reported fewer than five symptoms. After equating for the level of gambling problem severity, the likelihood of reporting certain symptoms was biased across groups characterized by their age and gender. Implications for understanding the construct of gambling problems severity and use of symptom counts as a continuous index are discussed. Keywords DSM-IV, gambling, item response theory, pathological gambling, Rasch. Correspondence to: David R. Strong, Brown Medical School, Butler Hospital, 345 Blackstone Blvd, Providence, RI 02906, USA. david_strong@brown.edu Submitted 18 July 2006; initial review completed 6 November 2006; final version accepted 8 December 2006 INTRODUCTION The syndrome of pathological gambling (PG) has been defined using the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) [1] as a persistent pattern of recurring maladaptive gambling behavior as evidenced by the presence of five (or more) of 10 specified symptoms. These symptoms of PG include increased preoccupation with gambling, needing to gamble with increasing amounts of money, loss of control, irritability when reducing gambling, gambling to escape problems or relieve negative moods, chasing losses, lying to others about gambling and relying on others for money to relieve gambling-related financial consequences, as well as relationship, job and legal difficulties. While the DSM-IV currently defines PG as present or absent, arguments continue over whether there are discrete shifts in the distribution of PG symptoms or if the pathological gambling syndrome could be conceptualized more continuously [2 8]. Evaluating the effectiveness of a diagnostic threshold with a specific number of PG symptoms requires an increased understanding of how individual symptoms are related to levels of gambling problem severity, whether the endorsement of additional symptoms is related to measurable increments in problem severity and whether symptom reports are influenced by demographic characteristics. In a recent investigation of the frequency of individual symptoms of PG, Toce-Gerstein, Gerstein & Volberg [8] examined the report of the DSM-IV symptoms of PG among individuals from two stratified communitybased surveys. The authors examined the proportion of

2 714 David R. Strong & Christopher W. Kahler gamblers reporting each DSM-IV PG symptom within groups of gamblers rank-ordered by the total number of PG symptoms they endorsed. By examining symptom prevalence as a function of the total number of symptoms, we can begin to understand which symptoms are associated with high or low levels of gambling problem severity. Reports of chasing losses characterized the lowest levels of gambling problem severity (one to two symptoms) followed by gambling to escape problems (three to four symptoms), loss of control and jeopardizing relationships (five to seven symptoms) and finally committing illegal acts to support gambling (eight to 10 symptoms). Orford, Sproston & Erens [9] observed a very similar ordering in the frequency of responses to PG symptoms in a community sample from the British Gambling Prevalence Study. These studies support the potential to generate a continuous graded index from individual symptoms of PG, and provide some initial information on the kinds of symptoms that would be expected within specified regions of a latent continuum of gambling problem severity. However, these estimates do not provide information about how precisely these regions can be specified or whether levels or thresholds can be established. Additionally, prevalence estimates cannot characterize whether important subgroups conform to this expected rank ordering. A significant challenge in assessing an index of gambling problem severity comes from the desire to ensure that the severity of DSM-IV symptoms and the index s relationship to the underlying construct is not affected by respondent characteristics. Summed scores should maintain relationships to levels of gambling problem severity, regardless of who is being assessed. For example, gender, some racial, and age groups have been shown to be at increased risk for PG and thus would be expected to produce higher symptom counts [10]. While these findings point to probable group differences in diagnosable levels of gambling problem severity (e.g. PG), they do not address whether individual symptom counts reflect similar levels of gambling problem severity across these demographic groups. Understanding the degree of bias in an index of PG will help to ensure that inferences about the relative severity of individual symptoms will generalize appropriately across demographic groups. There are several psychometric approaches for determining the severity of each DSM-IV symptom of gambling problems, whether symptom counts can be used as a continuous index and whether individual symptoms are biased for different demographic groups. These techniques include the comparison of the prevalence of each symptom across demographic groups, comparison of separate estimates of the correlation of individual symptoms with total symptom counts within each group, or the comparison of the factor structure of symptoms across groups [11]. However, techniques based upon prevalence or correlations are limited, because they are linked directly to the level of gambling problems among respondents. Prevalence estimates would be expected to change across samples composed of individuals that differ in their average level of gambling problem severity (e.g. from population surveys to clinical samples). This sample-based instability is particularly problematic when evaluating whether symptoms are reported similarly across important demographic subgroups. When using prevalence estimates alone, interpretation of any demographic differences in the prevalence of symptom endorsement would be confounded, as it would be unclear if the demographic groups differed in the level of gambling problems or in the way group members interpret and respond to particular symptoms. Over the last 40 years, efforts to overcome limitations of examining only prevalence estimates and intercorrelations have led psychometric researchers in educational psychology to gravitate to model-based psychometric methods based on item response theory (cf. [12]). Item response models (e.g. [13,14]) provide significant advantages over prevalence-based estimates in exploring the relationship between items and the underlying latent construct across groups. Of the item response models, the Rasch model [14] is distinguished by its focus on a measurement model that defines the relative position of each symptom on a common latent continuum (i.e. interval-level properties) and determines the most probable pattern of symptoms that leads to each possible total score (cf. [15]). In order for one symptom to be characterized as more severe than another symptom, the model presumes that individuals who endorsed the more severe symptom (i.e. lower prevalence symptom) have a high probability of endorsing the less severe symptom(s). This study sought to replicate and extend existing studies that suggest that PG symptoms can be graded by their severity to form a continuous index of gambling problem severity. We apply the Rasch model to pathological gambling symptom questions collected from a large representative sample as part of the US National Epidemiologic Survey on Alcohol and Related Conditions (NESARC). The primary aims of this study are (1) to assess the unidimensionality of the construct underlying symptoms of pathological gambling; (2) to determine the relative severity and typical patterning of individual DSM-IV pathological gambling symptoms; (3) to generate an estimate of the precision of the DSM-IV based index across levels of gambling problems; and (4) to assess whether the severity of DSM-IV pathological gambling symptoms and typical symptom patterns are influenced by demographic characteristics including age, sex, race and income level.

3 Gambling problem continuum 715 Table 1 Item parameter estimates, fitting statistics and symptom bias estimates for the DSM-IV symptoms in ascending order of severity obtained from respondents who reported gambling more than five times in a single year. DSM-IV criterion Symptom Item parameters Fit statistics Symptom bias % Severity SE In-fit Out-fit Gender Age Race Income 1 Is preoccupied with gambling 12.2% Needs to gamble with increasing amounts 6.4% of money to achieve the desired excitement 5 Gambles to escape problems or relieve 6.0% * dysphoric mood 6 After losing money, often returns another 4.4% * day to get even ( chasing ) 7 Lies to conceal the extent of involvement 3.3% with gambling 3 Repeated unsuccessful efforts to control, 2.9% cut back, or stop gambling 10 Relies on others money to relieve a 1.3% desperate financial situation 4 Is restless or irritable when cutting down 1.2% or stopping gambling 9 Jeopardized or lost relationship, job, or 1.0% career opportunity 8 Has committed illegal acts to finance gambling 0.04% Gender bias reflects the difference between symptom severity estimates from women compared to men, age bias reflects the difference between symptom severity estimates from younger gamblers (age < 25) compared to older gamblers, race bias reflects the difference between symptom severity estimates from African Americans compared to Caucasians, and income bias reflects the difference between symptom severity estimate from lower incomes (< $20 000) compared to higher incomes. Positive and negative logit values indicate the magnitude and direction of differences between symptom severity estimates relative to the comparison group. *P < (Bonferroni adjustment = 0.05/40). METHOD Data collection The NESARC is a survey of a nationally representative sample from the United States. Methods for obtaining the sample have been described in detail in previous studies [16,17]. The survey was conducted using interviews from adults aged 18 years and older. Although data can be weighted to reflect accurately the design of the NESARC survey these weights are not used in the current analyses, as our purpose was not to make generalizations about the US population but rather to examine the relationships of symptoms to a latent continuum and whether these relationships differed across subgroups. The complete section on pathological gambling assessing the life-time occurrence of DSM-IV based symptoms of PG was administered to individuals who responded affirmatively to the question Have you ever gambled at least 5 times in any one year of your life? (n = ). The sample of gamblers was 46.1% women (n = 5146) and 14.2% Hispanic (n = 1582). Racially, the sample of gamblers was 76.8% white (n = 8560) and 20.1% black (n = 2242). The mean age of gamblers was (SD = 18.51). Income levels were as follows: 24.1% < $20 000, 21.7% between $ and $34 999, 33.1% between $ and $ and 21.1% $ Symptoms of pathological gambling The NESARC survey was developed to assess DSM-IV criteria for PG. Fifteen dichotomous items were administered that allowed the 10 DSM-IV criteria to be coded as present or absent. The full survey is available at niaaa.census.gov/index.html. One of the 10 DSM-IV criteria was assessed with three NESARC questions, two DSM-IV criteria were assessed with two NESARC questions, seven of the DSM-IV criteria were assessed with one NESARC question and an additional item inquiring about chasing wins was added to the NESARC interview. Of those who reported ever gambling more than five times in a single year (n = 11,153), 19.6% (n = 2186) reported at least one PG symptom. The prevalence rates of the DSM-IV criteria among life-time gamblers are listed in Table 1. We also assessed the DSM-IV 12-month clustering criterion using three questions that asked whether multiple symptoms occurred within a 12-month period. Participants answering positive to any of these questions

4 716 David R. Strong & Christopher W. Kahler or reporting five or more symptoms in the past year were coded as positive for the clustering criterion. Analyses We used the Rasch model to evaluate the measurement properties of the 10-symptom index of gambling problem severity. The Rasch model places symptoms and people on the same metric so that the level of gambling problem severity associated with a symptom and the level of gambling problem severity of a given person can be compared directly. Thus, if we know the relative severity of symptoms, we can understand more clearly how severe an individual s gambling problem must be before a particular symptom is likely to be present. The analysis also provides information about the degree to which each individual s response pattern conforms to or fits with the expected pattern of symptoms. The model assumes that endorsement of more severe symptoms presumes a high likelihood of also endorsing less severe symptoms. Individual symptoms that do not meet this expectation are identified as misfitting the model. By modeling the latent continuum of gambling severity, we can also estimate how precisely the DSM-IV index rank-orders individuals across levels of the gambling problem continuum. Unidimensionality and local independence The first assumption of the Rasch model is that responses to symptom queries are a function of individual variation along a single underlying dimension of gambling severity. Secondly, responses to a given symptom should be independent from responses to other symptoms (e.g. locally independent). Previous factor analyses of DSM-IV PG symptoms support a strong general gambling problem severity factor underlying responses [8,9,18]. None the less, before interpreting the fit of the current data to the Rasch model, we first conducted an exploratory principal factors analysis to ensure unidimensionality and then conducted principal components analysis of residual variance after the Rasch model was fitted to ensure local independence [19]. We restricted factor analysis to those individuals who met the life-time gambling criteria (n = ). Rasch modeling further limits analysis to individuals who display variability in their symptom profiles. Therefore, the Rasch model fit and analysis of residual variance of remaining factors exclude those individuals who either do not endorse any symptoms (n = 8967) or endorse all symptoms (n = 6). For the analysis of residuals after fitting the Rasch model, additional factors accounting for > 1.5 units of variance were considered significant [19,20]. Two c 2 ratios are used typically to determine how well the data fit a Rasch model, the in-fit and out-fit statistics [21]. The out-fit statistic is based on conventional sum of squared standardized residuals generated by the difference between what would be expected by the model and what was observed in the actual data. The in-fit statistic incorporates information about the severity of symptoms and respondents level of gambling problem severity. When patterns of responding by individuals or patterns in the ordering of symptom severity fit a Rasch model, values of observed minus expected scores (in-fit and outfit) will fall within an acceptable range of [22]. In-fit and out-fit statistics are sensitive to unexpected variation in response patterns (e.g. a person with little gambling problem severity endorses a severe symptom) or when items are not discriminating equally across levels of gambling problem severity (e.g. unequal slopes across item response functions; in the present sample, in-fit values correlated with discrimination parameters that were obtained by fitting a two-parameter logistic model to the data) [22]. However, in-fit statistics are generally preferred, as they are weighted locally and are thus less susceptible to outlier influences [22]. While other item response models are available for this type of analysis (e.g. [13]), we chose the Rasch model for its ability to provide an independent estimation of symptom severities that is not affected by differences in levels of gambling problem severity across demographic subsamples (cf. [23]). This comparison ensures that equal comparisons of levels of gambling problem severity can be made across demographic groups. The estimation of differential item functioning (DIF) involves comparing analyses conducted separately within each group [11]. If the symptoms behave similarly across groups, then symptom severity parameters estimated independently in different samples will fall within an acceptable range of agreement (e.g. 95% CI). While there are many means to assess DIF (cf. [11]) comparison of Rasch severity estimates and standard errors from separate subgroups has been shown to be a reliable means of detecting DIF and is comparable to other DIF indices such as the Mantel Haenszel test [24]. In this study, we compared women to men, younger gamblers (age < 24) to older gamblers (age 25), African American gamblers to Caucasian gamblers and lower-income gamblers (income < $20 000) to other gamblers (income $20 000). Given the large sample, we considered differences in item severities greater than logits to be a clinically meaningful difference (cf. [21]); this difference is equal to about half of the sample standard deviation of item severities, i.e. a medium effect size. We used BIGSTEPS [25] for all item response model analyses. Clustering criterion The clustering criterion for pathological gambling can be endorsed only if the respondent endorses at least five

5 Gambling problem continuum 717 Figure 1 Plot of item characteristic curves for the 10 DSM-IV symptoms of gambling problem severity. This figure shows the relationship between the latent level of gambling problem severity and the probability of endorsing a symptom.the ICC are labeled with numbers that correspond to the DSM-IV symptoms listed in Table 1 symptoms. Therefore it is not an independent symptom compared to the other items examined. None the less, we wanted to model the relation of the clustering criterion to the other symptoms to determine how much higher of a bar this criterion introduces into the diagnostic criteria set. Therefore, after our primary analyses, we fitted a Rasch model including clustering in the item set. RESULTS Unidimensionality and local independence Maximum likelihood factor analysis of tetrachoric correlations revealed a strong single factor among respondents meeting the life-time gambling criteria. The first factor was significantly larger than the second, accounting for 68.9% and 4.9% of the common variance, respectively. Loadings on the first factor ranged from 0.66 to Principal components analysis of residual variance after fitting the Rasch model among only individuals who contributed to variability among the 10 DSM-IV symptoms (n = 2180) again supported the unidimensionality and local independence of this set, with residual factors accounting for only 1.41 and 1.22 units of variance. Analysis of residuals within gender, age, African American race and economic status subgroups were consistent with the overall analysis. Symptom-level estimates of problem severity Table 1 lists symptom-level estimates of the severity of the 10 DSM-IV PG symptoms listed in ascending order using a logit scale. The PG symptoms (see Table 1) fitted a unidimensional Rasch model well, with in-fit values ranging from and out-fit values from In addition, the Rasch reliability of symptom severity estimates was very high with an estimate of 0.99, suggesting the high likelihood of obtaining the same ordering of symptom severity with other samples. Figure 1 shows the item characteristic curves (ICC) for the 10 symptoms. Examination of these curves provides a visual representation of the relationship between the increased likelihood of a symptom (e.g. the y axis) and individual levels of gambling problem severity (e.g. the x axis). For example, symptom (1), Is preoccupied with gambling, is most helpful in separating individuals with the lowest levels of gambling problem severity while symptom (8), illegal acts to finance gambling, is most helpful only at the highest levels of gambling problem severity. Symptoms (2), Needs to gamble with increasing amounts of money and (5), Gambles to escape problems or relieve dysphoric mood, are observed at very similar points on the latent continuum and thus would be expected to be observed at very similar levels of problem severity. When modeled in the context of the primary 10 DSM-IV symptoms, chasing wins also fitted the Rasch model well (in-fit = 0.95), was associated with lower range of gambling problems (prevalence = 5.28%; severity =-0.90, SE = 0.05) and was similar in severity to chasing losses. The Rasch model can inform us about the odds of different symptoms being endorsed by individuals from

6 718 David R. Strong & Christopher W. Kahler non-pathological (no symptoms), those with subclinical symptoms (one to four symptoms) and pathological gamblers (more than four symptoms). For example, the average Rasch severity in the current sample was (SD = 0.00) among non-pathological gamblers (n = 8967), (SD = 0.82) among those with subclinical symptoms (n = 1991) and 0.87 (SD = 1.10) among those who met diagnostic criteria (n = 195). As individuals and item severity levels are scaled on the same metric, we can estimate the likelihood of observing a symptom if we know the distance between the item severity and the person severity. The likelihood of observing a symptom is < 50%, as the person severity decreases below the item severity, and > 50% as the person severity increases above the item severity. For example, the observed prevalence of symptom 1 is 58% among those with subclinical symptoms who, as a group, have an average level of gambling severity of -2.57, slightly higher than the item severity of Among pathological gamblers (PG), those who met diagnostic criteria, we do not expect to observe prevalence below 50% until symptom 10, where the item severity estimate of 1.08 exceeds the average person severity of In this way, we can characterize the types of symptoms to be expected for gamblers above and below a diagnostic threshold. To examine the clustering criterion, we fitted a Rasch model with the 10 life-time symptoms along with the clustering criterion. With a severity estimate of 0.89 (SE = 0.09) and a low model fit statistic (in-fit = 0.47), the criterion discriminated participants strongly in the region of the continuum in which the six least severe DSM-IV criteria (1, 2, 5, 6, 7, 3) would be likely to be expressed, but other more severe indicators of gambling problems would be unlikely. Measurement precision of the DSM-IV symptom index We can examine the precision of the Rasch-modeled DSM-IV index by estimating the standard error of measurement across all assessed levels of gambling problem severity (SE range ), with lower values indicating more precise measurement. The Rasch-modeled DSM-IV index provides the most information or is most reliable (SE < 1.0) in rank-ordering individuals with midrange levels of gambling problem severity and is least effective (SE > 1.0) with individuals with either low (less than three symptoms) or high-range (more than eight symptoms) levels of gambling problem severity. Using methods described by Wright [26], we can use the estimated measurement precision (i.e. SE) of DSM-IV PG symptoms to evaluate how reliably this set of symptoms can distinguish whether an individual belongs above or below a given location on the latent continuum of gambling problem severity. We estimated that the first reliable threshold (using the joint standard error for placing someone above or below a given point on the latent continuum) occurs when four or more symptoms are endorsed. This finding suggests that the least severe symptoms gather within a similar level of gambling problems. There is a relatively large gap along increasing levels of gambling problem severity after symptom 3 (see Fig. 1) before symptoms such as (10), relies on others money to relieve a desperate financial situation, (4), is restless or irritable when cutting down or stopping gambling and (9), jeopardized or lost relationship, job, or career opportunity become likely ( logit units higher than symptom 3). These symptoms mark a final reliable threshold beyond which the most severe symptom (8), illegal activities, becomes likely. The gaps in severity estimates between some items indicate that there are regions along the continuum in which a marked increase in gambling problem severity is required before additional DSM-IV symptoms are observed. Demographic influence on symptom endorsement Plots of symptom-severity estimates obtained in separate analysis of demographic subgroups are displayed in Fig. 2. If symptom severity is estimated similarly, plots of symptom severity estimates should fall on a 45-degree angle within bounds of the joint standard errors from each analysis [i.e. the 95% confidence interval (CI)]. Women reported (5), gambling as a way of escaping from problems at lower levels of gambling problem severity than men. For example, the prevalence of endorsing symptom 5, gambling to escape problems among subclinical and PG was 38.6% and 79.2% for women and 19.2% and 61.0% for men, respectively. Younger gamblers were more likely to report chasing losses (symptom 6) at lower levels of gambling problem severity than older gamblers. The prevalence of endorsing symptom 6 among subclinical and PG was 27.1% and 80.9% for younger gamblers and 15.0% and 78.7% for older gamblers, respectively. Table 1 lists the differences between severity estimates in logit units. The observed differences fell outside 95% CI and remained significant statistically using a Bonferroni-adjusted alpha level. DISCUSSION This study employed a Rasch latent trait measurement model to understand more clearly the symptoms that characterize a continuum of gambling problem severity in a large community sample obtained from the NESARC. We evaluated the degree to which the 10 DSM-IV PG criteria indexed a single unidimensional construct, estimated the relative severity of each symptom and evaluated whether symptoms were biased among demographic

7 Gambling problem continuum 719 Figure 2 Plot of item severity estimates and 95% confidence interval from differential item function analyses by gender, age, race and socio-economic status groupings based on age, gender, race and economic level. The symptoms of gambling problem severity fitted a unidimensional Rasch model well. With fitting to the Rasch model we were able to describe the symptoms that characterize various levels of gambling problem severity and predict the pattern of symptoms that led to each score on the 10-item index. After equating for the level of gambling problem severity, the likelihood of reporting specific symptoms was different for women and younger gamblers. Implications for understanding the construct of gambling problems severity and use of symptom counts as a continuous index are discussed below. Estimates of the rank-order of individual PG symptoms from the current study largely mirrored those of previous community-based samples that relied upon prevalence estimates alone (cf. [8,9]). However, by placing symptom prevalence estimates within a latent measurement model, we can evaluate more clearly assumptions that are inherent when using symptom counts as an index of gambling problem severity. With fitting to the Rasch model, we now (a) know that adding DSM-IV PG symptoms together can result in a total score that reflects variation on a primary construct of gambling problem severity, (b) have estimates of the relative position of symptoms on this continuum and (c) know that certain symptom patterns are likely to characterize given levels of severity. The Rasch model provides a context for each symptom, and estimates from the model (e.g. individual levels of gambling problems) do not depend on all the symptoms. This model-based approach allows researchers to place symptoms on a common metric that can be used to examine alternative symptoms for placement in relation to the DSM-IV symptoms without losing the original frame of reference or established thresholds for diagnosis that the DSM-IV criteria provide. Further, this frame of reference for the items is established independently from the level of gambling problems among the respondents. For example, the Rasch estimate of the precision of the DSM-IV index from the current study mirrored closely results from a large clinical sample of treatment-seeking gamblers where the prevalence of DSM-IV symptoms was significantly higher, ranging from 46% to 89% [27]. Rasch analyses across previous clinical samples [27] and the current community study suggest that when

8 720 David R. Strong & Christopher W. Kahler using the existing DSM-IV symptoms, a reliable threshold occurs when gamblers report more than four symptoms. These studies did not suggest that the current set of DSM-IV symptoms can reliably classify further among gamblers with fewer than four symptoms. Symptoms reflecting preoccupation with reliving past experiences, spending increased time gambling and using gambling as a means of lifting a bad mood are linked consistently to levels of gambling problem severity that are significantly below the DSM-IV diagnostic threshold for PG. The level of gambling problem severity associated with a transition above a DSM-IV diagnostic threshold of PG is marked by symptoms including concealing the extent of gambling behavior and unsuccessful efforts to control gambling. Committing illegal acts and jeopardizing relationships and/or job opportunities mark the highest levels of gambling problem severity, well above a DSM-IV diagnostic threshold. Unlike previous studies, chasing losses and feeling restless or irritable when cutting down were ranked as more severe symptoms in the current sample than in previous US community samples that indexed DSM-IV PG symptoms with a survey used by the National Opinion Research Center (NODS [28]). These discrepancies could be related to the wording of these questions in these two surveys. For example, the NESARC question about chasing inquires whether the individual had to gamble as soon as possible after losing..., whereas the NODS question inquires whether the individual would ever return another day to get even after a loss. The difference in emphasis on the immediate need to chase losses may explain why this symptom is ranked as more severe in the NESARC data. Ultimately, each criterion must be assessed with a question or multiple questions and differences in wording may play a substantial role in the relative ordering of symptoms observed. The Rasch model allows for examination of these severity estimates within a common frame of reference. The Rasch approach also provided evidence for significant bias in symptom severity estimates across demographic groupings based on age and gender. The most significant differences were found among women. Women with the same level of gambling problem severity as men were more likely to report gambling to lift a bad mood and gambling to forget problems. Additionally, younger gamblers reported chasing losses at lower levels of gambling problem severity than older gamblers. The two symptoms that displayed bias were located at levels of gambling problem severity that were similar to other non-biased items. This decreases the chance that systematic bias in rank-ordering individuals would dramatically affect estimates of an individual s level of gambling problem severity. However, these biases were large and future research may reveal the source of these qualitative differences among gamblers. Despite the large sample used in the current study, many of the symptoms of PG were endorsed infrequently. This sparse distribution of symptom reports may limit significantly the ability to define patterns of responses among individuals at the highest levels of gambling problem severity within this sample. Despite this limitation, the Rasch estimates of the measurement precision of the DSM-IV index were almost identical to estimates in a large sample of predominantly pathological gamblers [27]. For future studies, targeted sampling strategies that increase the prevalence of pathological gamblers with high levels of problem severity would be valuable. The use of a sample from the United States also limits our ability to assess cultural factors that may influence selfdescriptions. Another limitation of the current study was the lack of detailed information regarding gambling histories that might be associated with levels of gambling severity or the timing of symptom onsets that may help anchor a pattern of progression towards more severe symptoms. This cross-sectional study relied on life-time retrospective reports and although an ordered pattern of symptom responses was observed, any attribution of developmental progression ultimately requires longitudinal evaluation. We view the modeling of the continuum of gambling problems as a first step in anchoring the symptoms generated by the extant DSM-IV conceptual model. Given that the DSM-IV provides a widely utilized referent, alternative models can be proposed and fitted with previous knowledge of the structure of the current DSM-IV conceptualization. It should be noted that the DSM-IV criteria themselves are abstract constructs which must be assessed ultimately with measures, be they questionnaires, structured interviews or biological markers. The criteria, in and of themselves, do not have fixed severities. Rather, the items for each symptom can have estimated severity levels that can vary greatly based on changes in method. Alternative models can be developed using the existing conceptualization as a frame of reference to support the benefits of departure from this established standard. Acknowledgements The National Epidemiologic Survey on Alcohol and Related Conditions (NESARC) was conducted and funded by the National Institute on Alcohol Abuse and Alcoholism, with supplemental support form the National Institute on Drug Abuse. We appreciate research assistance from Chelsea Spector, Jessica Kaplan and Amy Cameron. References 1. American Psychiatric Association Diagnostic and Statistical Manual of Mental Disorders: Text Revision. Arlington, VA: American Psychiatric Publishing; 2000.

9 Gambling problem continuum Blaszczynski A., McConaghy N. Anxiety and/or depression in the pathogenesis of addictive gambling. Int J Addict 1989; 24: Blaszczynski A., Nower L. A pathways model of problem and pathological gambling. Addiction 2002; 97: National Research Council Pathological Gambling: a critical review. Washington, DC: National Academy Press; Orford J. The fascination of psychometrics: commentary on Gerstein et al. (2003). Addiction 2003; 98: Shaffer H. J., Hall M. N. Estimating the prevalence of adolescent gambling: a quantitative synthesis and guide toward standard gambling nomenclature. J Gambling Stud 1996; 12: Slutske W. S., Eisen S., True W. R., Lyons M. J., Goldberg J., Tsuang M. Common genetic vulnerability for pathological gambling and alcohol dependence in men. Arch Gen Psychiatry 2000; 57: Toce-Gerstein M., Gerstein D. R., Volberg R. A. A hierarchy of gambling disorders in the community. Addiction 2003; 98: Orford J., Sproston K., Erens B. SOGS and DSM-IV in the British gambling prevalence survey: reliability and factor structure. Int Gambling Stud 2003; 3: Petry N. M., Stinson F. S., Grant B. F. Comorbidity of DSM-IV pathological gambling and other psychiatric disorders: results from the National Epidemiologic Survey on Alcohol and Related Conditions. J Clin Psychiatry 2005; 66: Holland P. W., Wainer H. Differential Item Functioning. Hillsdale, NJ: Erlbaum; Lord F. Applications of Item Response Theory to Practical Testing Problems. Hillsdale, NJ: Erlbaum; Birnbaum A. Some latent trait models and their use in inferring an examinee s ability. In: Novick F. M., editor. Statistical Theories on Mental Test Scores. Reading, PA: Addison-Wesley; 1968, pp Rasch G. Probabilistic Models for Some Intelligence and Attainment Tests. Copenhagen: Danish Institute for Educational Research; Guttman L. The basis for scalogram analysis. In: Stouffer S. A., editor. Measurement and Prediction. New York: Wiley; 1950, pp Grant B. F., Moore T. C., Shepard J. Source and Accuracy Statement: Wave 1. National Epidemiological Survey on Alcohol and Related Conditions (NESARC). Bethesda, MD: National Institute on Alcohol Abuse and Alcoholism; Grant B. F., Stinson F. S., Dawson D. A., Chou S. P., Dufour M. C., Compton W. et al. Prevalence and co-occurrence of substance use disorders and independent mood and anxiety disorders: results from the National Epidemiologic Survey on Alcohol and Related Conditions. Arch Gen Psychiatry 2004; 61: Stinchfield R. Reliability, validity, and classification accuracy of a measure of DSM-IV diagnostic criteria for pathological gambling. Am J Psychiatry 2003; 160: Smith R. M., Miao C. Y. Assessing unidimensionality for Rasch measurement. In: Wilson M., editor. Objective Measurement:Theory Into Practice. Norwood, NJ: Ablex; 1994, pp Linacre J. M. Detecting multidimensionality: which residual data-type works best? J Outcome Measure 1998; 3: Wright B. D., Masters G. N. Rating scale analysis. Chicago: MESA Press; Bond T., Fox C. Applying the Rasch model: Fundamental measurement in the human sciences. Mahway, NJ: Erlbaum; Andrich D. Rasch Models for Measurement. Newbury Park, CA: Sage; Schulz E. M. DIF detection: Rasch versus Mantel Haenszel. Rasch Measure Transact 1990; 4: Linacre J. M., Wright B. D. A Users Guide to BIGSTEPS: Rasch model computer programs. Chicago, IL: MESA Press; Wright B. D. Separation, reliability and skewed distributions. Rasch Measurement Transactions 2001; 14: Strong D. R., Lesieur H. R., Breen R. B., Stinchfield R., Lejuez C. W. Using a Rasch model to examine the utility of the South Oaks Gambling Screen across clinical and community samples. Addict Behav 2004; 29: Gerstein D., Murphy S., Toce M., Hoffmann J., Palmer A., Johnson R. et al. Gambling Impact and Behavior Study. Report to the National Gambling Impact Study Commission. Chicago, IL: National Opinion Research Center; 1999.