Assessing gambling disorder using frequency‐ and time‐based response options: A Rasch analysis of the gambling disorder identification test

Abstract

Objectives

The Gambling Disorder Identification Test (GDIT) is a recently developed self‐report measure. The GDIT includes items with multiple response options that are either based on frequency or time, and item response theory evaluations of these could yield vital knowledge on its measurement performance.

Methods

The GDIT was evaluated using Rasch analysis in a study involving 597 Swedish gamblers.

Results

In a three‐dimensional Rasch model, the item response difficulty range extended from −1.88 to 4.06 and increased with higher time‐ and frequency‐based responses. Differential item functioning showed that some GDIT items displayed age and gender‐related differences. Additionally, person‐separation reliability indicated the GDIT could reliably be divided into three to four diagnostic levels.

Conclusions

The frequency‐ and time‐based item response options of the GDIT offer excellent measurement, allowing for elaborate assessment across both lower and higher gambling severity. The GDIT can be used to detect DSM‐5 Gambling Disorder, thereby holding significance from both epidemiological and clinical standpoints. Notably, the 3‐item GDIT Gambling Behavior subscale also shows potential as a brief screening tool for identifying at‐risk gambling behavior.

1. INTRODUCTION

Gambling Disorder (GD; American Psychiatric Association, 2013) is diagnosed based on a pattern of repeated problematic gambling behavior, leading to negative consequences for individuals, their families and society as a whole. With the introduction of the 5th edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM‐5), the diagnostic criteria for GD were revised, with one of the 10 original criteria removed (committing illegal acts to finance gambling). Compared to the fourth edition of the DSM (DSM‐IV) (American Psychiatric Association, 1994), GD was also reclassified from an impulse control disorder to an addictive disorder in the DSM‐5, alongside Alcohol Use Disorder and Substance Use Disorder (American Psychiatric Association, 2013). It was further suggested that the GD diagnosis comprise three levels of symptom severity: meeting 4 or 5 of the 9 criteria GD (mild GD), meeting 6 or 7 of the 9 criteria (moderate GD) or meeting 8 or 9 of the 9 criteria (severe GD).

Despite these important advances, little is known about the GD DSM‐5 criteria from epidemiological or clinical perspectives, including current prevalence estimates of GD in the general population and among treatment‐seeking patients. This could be, at least in part, explained by measurement factors, such as issues related to item content validity, or conceptualization of public health‐based measurement frameworks such as problem gambling (Dowling et al., 2019; Walker et al., 2006). Although GD according to DSM‐5 can be assessed using self‐reported criteria (Molander et al., 2023; Stinchfield et al., 2016), a recent systematic review of existing gambling measures concluded that there is a general lack of diagnostic accuracy in relation to DSM‐5‐based interviews (Otto et al., 2020). One challenge associated with DSM‐5 has been the efficient identification of individuals with different levels of GD symptom severity (Molander, Wennberg, & Berman, 2021). Expanding gambling measurement beyond public health‐based conceptualizations to include DSM‐5 GD would enable more precise measurement, providing advantages from primary, secondary, and tertiary prevention perspectives. Such an expansion would result in better differentiation between recreational, at‐risk, and diagnostic gambling levels, offering improved opportunities to estimate the number of individuals with GD in need of treatment within healthcare systems.

In response to this challenge, the Gambling Disorder Identification Test (GDIT; www.gditscale.com) was developed as a gambling self‐report measure analogous to the Alcohol Use Disorders Identification Test (AUDIT; Saunders et al., 1993) and the Drug Use Disorders Identification Test (DUDIT; Berman et al., 2005), with a particular emphasis on establishing cut‐off scores in relation to DSM‐5 GD. The development process of the GDIT included a comprehensive analysis of the item content of 47 existing gambling measures (Molander et al., 2019), as well as the selection of GDIT items and response categories in an international expert Delphi study and consensus meeting, which included feedback from individuals with lived experience of gambling harm (Molander, Wennberg, & Berman, 2021). The final GDIT comprises 14 items across three theoretical domains (i.e., sub‐scales; Gambling Behaviors, Gambling Symptoms, and Negative Consequences). In a validation study (Molander, Wennberg, & Berman, 2021), 603 gamblers from diverse populations completed self‐report measures, a GDIT retest (n = 499), and diagnostic semi‐structured interviews assessing GD according to DSM‐5 (n = 203). The GDIT demonstrated excellent internal consistency reliability (α = 0.94) and test–retest reliability (6–16 days, intraclass correlation coefficient = 0.93). Regarding convergent and divergent validity, the GDIT exhibited strong positive correlations with other gambling measures and with gambling debts, along with weaker positive and negative correlations with measures assessing other psychiatric conditions and quality of life, respectively. Confirmatory factor analysis identified three domains, supporting the theoretical GDIT sub‐scales. Finally, GDIT cut‐off scores were established in relation to semi‐structured DSM‐5 diagnostic interviews assessing GD, utilizing receiver operating characteristic (ROC) curves.

Similar to the AUDIT and the DUDIT, the GDIT includes items with multiple response options that are either based on frequency or time. For example, the first item of the GDIT employs frequency‐based response options: “How often do you gamble”: (0) “Never”, (1) “Monthly or less”, (2) “2–4 times a month”, (3) “2–3 times a week”, (4): “4 or more times a week”, (5) “Daily”, (6) “Several times a day”. In contrast, the second item employs time‐based response options: “How much time do you spend gambling on a typical day?”: (0) “No time”, (1) “Less than an hour”, (2) “1–2 hours”, (3) “3–4 hours”, (4) “5–6 hours”, (5) “7–9 hours”, (6) “10–24 hours”. It has been argued that such multiple response options based on actual frequencies and time periods facilitate clearer measurement procedures than more general response options based on frequency (e.g., “almost always”, “sometimes”, “never” etc.) or time (e.g., “most of the time”, “some of the time”, “none of the time”) (Molander, Volberg, et al., 2021). However, psychometric evaluations of the validity of this argument have not yet been performed.

To date, the psychometric properties of the GDIT have been investigated using methods from classical test theory (Molander, Wennberg, & Berman, 2021). The performance of the GDIT items has not yet been examined using methods within item response theory, such as Rasch analysis. Briefly, classical test theory evaluates psychometric performance of one or more samples on an entire measure, typically a total score (Bond & Fox, 2007). As such, classical test theory assumes that all items make equal contributions to the performance of the measure. An alternative psychometric approach is item response theory. As in classical test theory, there is an assumption of a latent trait which is defined within the model, which is estimated in relation to both individual participants (i.e., person abilities) and responses to specific items (item difficulties) across a continuum (see e.g., Bond & Fox, 2007, for an in‐depth description of differences between classical test theory and item response theory). Rasch analysis is an item response theory method (Gorin et al., 2008), which holds several advantages from a psychometric perspective. First, a Rasch analysis estimates specific items or item response options across a difficulty or severity continuum (Wind & Hua, 2021). This analysis involves an expression of the proportion of respondents selecting each response alternative, which is important information both from research and clinical perspectives. Second, a Rasch analysis can be used to test how many levels (or strata) a measure can be reliably divided into, which is important in relation to the intended use and cut‐off scoring of the measure, such as diagnostic cut‐offs. Third, a Rasch analysis could ideally establish whether a measure is on a data level that permits parametric analyses (Bond & Fox, 2007). If the data align with the Rasch model, it is assumed to be at least on an interval level and allow for parametric analyses, which is important for the use of the measure within research contexts. Rasch analysis thus offers a highly useful evaluation of measurement performance from a psychometric perspective, and several such studies have been conducted on gambling measures and diagnostic interviews (Cowlishaw et al., 2019; Miller et al., 2013; Molander et al., 2023; Molander & Wennberg, 2022; Molde et al., 2010). Rasch evaluation of the GDIT could yield vital knowledge on its measurement performance, but such investigation has been lacking to date.

The aim of the current study was therefore to complement our earlier psychometric analyses of the GDIT according to classical test theory (Molander, Wennberg, & Berman, 2021) by evaluating the Rasch properties of the GDIT. Specifically, the aims were to: (1) explore the degree to which there is equal discrimination across GDIT items and estimate the difficulty of each GDIT item, including multiple response options, across a severity continuum, using Rasch modelling; (2) investigate whether there are gender and age variations in GDIT item difficulty; and (3) identify the number of diagnostic levels (or strata) into which GDIT scores can reliably be divided.

2. METHOD

2.1. Participants and procedure

The current study analyzed data from a previous psychometric gambling study (see Molander et al., 2023; Molander & Wennberg, 2022; Molander, Wennberg, & Berman, 2021). In that study, Swedish gamblers (N = 603) were recruited from a non‐treatment seeking social media sample (n = 292) and from help‐seeking samples, including gambling self‐help groups (n = 47), treatment‐seeking samples within healthcare (n = 79), and support‐seeking samples (n = 185) from stodlinjen.se, the Swedish national gambling helpline. Participants were administered an online survey that included the GDIT and several additional measures. In the current study, Rasch analyses were conducted on the total sample (N = 597; six participants were removed from the total sample which either endorsed no or all GDIT items, as Rasch modelling requires variability in responses). See Table 1 for participant characteristics.

TABLE 1.

Participant characteristics.

Sample	Total (N = 597)
Demographic characteristics
Age M (SD)	33 (12.1)
Sex (%)
Men	73.7%
Women	25.3%
Not stated	1.0%
Source of income (%)
Employed	64.3%
Studies	20.8%
Other a	14.9%
Highest level of education (%)
University	34.7%
High school	52.4%
Junior high school	10.6%
Civil status (%)
Cohabiting	52.3%
Children	41.2%
GDIT Gambling diagnostics (%)
Recreational gambling	51.8%
Problem gambling	7.0%
Gambling Disorder, any	41.2%
Mild	6.4%
Moderate	5.2%
Severe	29.6%
Gambling characteristics (%)
Gambling debts	36.7%
Gambling types b
Casino online	51.3%
Casino land‐based	11.9%
Sport games online	43.9%
Sport games venue	13.6%
Poker online	16.8%
Poker club	5.7%
EGM	7.2%
Number games	8.9%
Lotteries	26.5%
Horse betting	17.4%
Bingo	10.4%
Other	10.7%

Abbreviations: EGM, electronic gambling machines; GDIT, Gambling Disorder Identification Test (Molander, Wennberg, & Berman, 2021).

^a This category included several sources for example, unemployment insurance, income support, sickness compensation, sickness benefit, and pension.

^b Participants were able to report multiple gambling types.

2.2. The GDIT

The GDIT (www.gditscale.com) is a recently developed gambling self‐report measure, which includes 14 items across three sub‐scales: Gambling Behaviors (GDIT‐B; GDIT_1–3), Gambling Symptoms (GDIT_4–10), and Negative Consequences (GDIT_11–14) (see Table 2). The GDIT scoring index, with a total score of a maximum of 62 points, is classified into: Non‐gambling (Total score = 0), Recreational gambling (Total score <15), Problem gambling (Total score 15–19), and Any GD (Total score ≥20), with Any GD further categorized into Mild GD (Total score 20–24), Moderate GD (Total score 25–29), and Severe GD (Total score ≥30) (Molander, Wennberg, & Berman, 2021). In addition to the 14 items, the GDIT assesses gambling expenditure and involvement in a range of gambling types in an appendix. See Table 2 for an overview of GDIT items, response options and scoring. The GDIT scores for the current sample were most often classified in the Recreational gambling (51.8%) and GD (41.2%) categories, with a smaller proportion of participants classified in the Problem gambling category (7.0%) (see Table 1).

TABLE 2.

GDIT items and scoring.

The gambling disorder identification test (GDIT)		Response options	Scoring	Sub‐scales
GDIT1	How often do you gamble?	“Never”, “Monthly or less”, “2–4 times a month”, “2–3 times a week”, “4 or more times a week”, “Daily”, “Several times a day”	0–6	Gambling behaviors
GDIT2	How much time do you spend gambling on a typical day?	“No time”, “Less than an hour”, “1–2 hours”, “3–4 hours”, “5–6 hours”, “7–9 hours”, “10–24 hours”	0–6	Gambling behaviors
GDIT3	How much time do you spend thinking about gambling on a typical day?	“No time”, “Less than an hour”, “1–2 hours”, “3–4 hours”, “5–6 hours”, “7–9 hours”, “10–24 hours”	0–6	Gambling behaviors
GDIT4	How often have you tried to control, cut down or stop your gambling, in the past 12 months?	“Never”, “Less often than monthly”, “Monthly”, “Weekly”, “Daily or almost daily”	0–4	Gambling symptoms
GDIT5	How often have you gambled to win back money you lost on gambling, in the past 12 months?	“Never”, “Less often than monthly”, “Monthly”, “Weekly”, “Daily or almost daily”	0–6	Gambling symptoms
GDIT6	How often, in the past 12 months, have you gambled more than you planned (more occasions, longer time or larger sums)?	“Never”, “Less often than monthly”, “Monthly”, “Weekly”, “Daily or almost daily”	0–6	Gambling symptoms
GDIT7	How often have you lied to others about your gambling, in the past 12 months?	“Never”, “Less often than monthly”, “Monthly”, “Weekly”, “Daily or almost daily”	0–6	Gambling symptoms
GDIT8	How often have you borrowed money or sold something to obtain money for gambling, in the past 12 months?	“Never”, “Less often than monthly”, “Monthly”, “Weekly”, “Daily or almost daily”	0–6	Gambling symptoms
GDIT9	How often have you gambled as a way of escaping problems or relieving negative feelings, in the past 12 months?	“Never”, “Less often than monthly”, “Monthly”, “Weekly”, “Daily or almost daily”	0–6	Gambling symptoms
GDIT10	How often have you gambled with larger sums to get the same feeling of excitement as before, in the past 12 months?	“Never”, “Less often than monthly”, “Monthly”, “Weekly”, “Daily or almost daily”	0–6	Gambling symptoms
GDIT11	Have you or anyone close to you experienced financial problems due to your gambling?	“No”, “Yes, but not in the past year”, “Yes, in the past year”	0, 2, 4	Negative consequences
GDIT12	Has your gambling worsened your mental health?	“No”, “Yes, but not in the past year”, “Yes, in the past year”	0, 2, 4	Negative consequences
GDIT13	Have you experienced serious problems in any important relationship because of your gambling?	“No”, “Yes, but not in the past year”, “Yes, in the past year”	0, 2, 4	Negative consequences
GDIT14	Have you experienced serious problems at work or in school because of your gambling?	“No”, “Yes, but not in the past year”, “Yes, in the past year”	0, 2, 4	Negative consequences

Abbreviation: GDIT, The Gambling Disorder Identification Test (Molander, Wennberg, & Berman, 2021).

2.3. Statistical analysis

A statistical analysis plan was published a priori (see https://aspredicted.org/qi873.pdf and https://archive.org/details/osf‐registrations‐cfq82‐v1). Analyses were performed in R Studio (R Core Team, 2018), using the following key packages: dplyr, psych, lavaan, eRm and TAM.

2.3.1. Dimensionality

Although Rasch models can be fitted to address multidimensionality, the standard Rasch model assumes that a single latent variable explains most of the variation in item responses (i.e., unidimensionality of a measure) (Katz et al., 2021; Wind & Hua, 2021). To test this assumption, a standard GDIT Rasch model was fitted first. An analysis of variance was performed, which revealed that the variance in responses on the primary latent variable (52%) was above the critical value of 20%, thereby indicating unidimensionality (Reckase, 1979; Wind & Hua, 2021). However, a principal components analysis of standardized residual correlations showed that two eigen values were above the critical value of 2.00 (2.78, 2.15), supporting a three‐dimensional structure (Wind & Hua, 2021). Although the GDIT was developed to encompass items across three specific sub‐scales, the main intended use of the instrument is screening and assessment via interpretation of the GDIT total score. In the current study, the decision was therefore made to fit and present Rasch analyses both as a three‐dimensional model (Gambling Behaviors GDIT_1–3, Gambling Symptoms GDIT_4–10, and Negative Consequences GDIT_11–14), and a unidimensional model.

2.4. Rasch modelling

The Rasch properties of GDIT items and multiple frequency‐ and time‐based response options were tested using two partial credit Rasch models, one for the three‐dimensional model, and the second for the unidimensional model. A partial credit Rasch model is suitable for measures with multiple item response options which vary across items (Katz et al., 2021), as they do in the GDIT. In order to maintain equidistant response options throughout the scale when estimating the partial credit model, the scoring for GDIT_11–14 was recoded (0 to 0, 2 to 1, 4 to 2).

2.4.1. Item discrimination

A Rasch analysis is based on an assumption of equal discrimination across items, which can be tested by item fit measures (Linacre, 2002, 2003; Miller et al., 2013; Wind & Hua, 2021). To test the degree to which there is equal discrimination across GDIT items, outlier‐sensitive fit (outfit) and inlier‐sensitive fit (infit) was estimated for each of the four GDIT Rasch models conducted in this study. In Rasch modeling, the outfit measure is sensitive to outliers, thereby indicating, for example, that high severity individuals endorse severe, moderate as well as mild severity items. The infit measure estimates the degree of fit between items and individual participants, thereby indicating, for example, to what extent low severity individuals endorse low severity items (Linacre, 2002; Miller et al., 2013; Molde et al., 2010). Infit and outfit item estimates between 0.50 and 1.50 indicate equal discrimination across items and can be considered productive for measurement. Values of <0.50 indicate overfit and values >1.50 indicate underfit (Linacre, 2002).

2.4.2. Item and multiple response option difficulties

The difficulty of each GDIT item and multiple response option was estimated across a severity continuum in both a three‐dimensional and a unidimensional model (see Tables 3 and 4, respectively).

TABLE 3.

Three‐dimensional Rasch analysis of GDIT item response options (N = 597).

Difficulty (SE)	GDIT multiple response options a						Outfit MNSQ	Infit MNSQ
	1	2	3	4	5	6
GDIT1	−1.88 (0.14)	−0.35 (0.11)	0.77 (0.12)	1.19 (0.14)	1.05 (0.15)	1.70 (0.18)	0 | 4 | 2	0 | 6 | 0
GDIT2	−1.76 (0.13)	0.19 (0.11)	0.72 (0.12)	1.76 (0.15)	2.43 (0.20)	2.61 (0.29)	0 | 3 | 3	0 | 6 | 0
GDIT3	−0.62 (0.11)	0.84 (0.12)	1.06 (0.14)	1.36 (0.16)	2.06 (0.18)	1.30 (0.20)	1 | 5 | 0	0 | 6 | 0
GDIT4	−0.29 (0.13)	0.63 (0.14)	1.87 (0.15)	1.86 (0.15)	‐	‐	0 | 0 | 4	0 | 3 | 1
GDIT5	−0.69 (0.12)	1.21 (0.14)	1.47 (0.15)	2.25 (0.16)	‐	‐	0 | 2 | 2	0 | 4 | 0
GDIT6	−1.20 (0.13)	1.13 (0.14)	1.50 (0.14)	2.62 (0.16)	‐	‐	2 | 2 | 0	0 | 4 | 0
GDIT7	0.52 (0.13)	1.58 (0.15)	1.66 (0.15)	2.14 (0.16)	‐	‐	0 | 4 | 0	0 | 4 | 0
GDIT8	1.41 (0.12)	1.76 (0.14)	3.63 (0.20)	4.06 (0.31)	‐	‐	1 | 3 | 0	0 | 4 | 0
GDIT9	0.43 (0.13)	1.12 (0.14)	2.06 (0.15)	2.35 (0.17)	‐	‐	0 | 3 | 1	0 | 4 | 0
GDIT10	0.59 (0.12)	1.81 (0.14)	2.48 (0.16)	3.31 (0.22)	‐	‐	2 | 2 | 0	0 | 4 | 0
GDIT11	1.23 (0.12)	1.55 (0.14)	‐	‐	‐	‐	0 | 1 | 1	0 | 2 | 0
GDIT12	0.56 (0.12)	0.64 (0.12)	‐	‐	‐	‐	0 | 2 | 0	0 | 2 | 0
GDIT13	1.56 (0.12)	1.25 (0.14)	‐	‐	‐	‐	0 | 2 | 0	0 | 2 | 0
GDIT14	2.87 (0.14)	2.14 (0.18)	‐	‐	‐	‐	0 | 1 | 1	0 | 2 | 0
						Underfit >1.50	14	1
						Overfit <0.50	6	0

Abbreviations: GDIT, The Gambling Disorder Identification Test (Molander, Wennberg, & Berman, 2021); MNSQ, Mean‐square; SE, Standard error.

^a GDIT uses the following item response options: GDIT_item1, Never (not estimated), Monthly or less, 2–4 times a month, 2–3 times a week, 4 or more times a week, Daily and Several times a day; GDIT_items2–3, No time (not estimated), Less than an hour, 1–2 hours, 3–4 hours, 5–6 hours, 7–9 hours, and 10–24 hours; GDIT_items4–10, Never (not estimated), Less often than monthly, Monthly, Weekly, and Daily or almost daily; GDIT_item11–14, No (not estimated), Yes, but not in the past year, and Yes, in the past year.

TABLE 4.

Unidimensional Rasch analysis of GDIT response options (N = 597).

Difficulty (SE)	GDIT multiple response options a						Outfit MNSQ	Infit MNSQ
	1	2	3	4	5	6
GDIT1	−1.80 (0.14)	−0.31 (0.11)	0.82 (0.12)	1.22 (0.14)	1.00 (0.14)	1.53 (0.17)	0 | 0 | 5	0 | 5 | 1
GDIT2	−1.68 (0.13)	0.23 (0.11)	0.74 (0.12)	1.70 (0.14)	2.25 (0.20)	2.33 (0.28)	0 | 0 | 6	0 | 6 | 0
GDIT3	−0.57 (0.11)	0.88 (0.12)	1.09 (0.14)	1.33 (0.15)	1.94 (0.17)	1.08 (0.19)	0 | 3 | 3	0 | 6 | 0
GDIT4	0.15 (0.11)	0.35 (0.12)	1.18 (0.13)	0.84 (0.14)	‐	‐	0 | 1 | 3	0 | 6 | 0
GDIT5	−0.22 (0.11)	0.90 (0.12)	0.75 (0.13)	1.20 (0.14)	‐	‐	1 | 3 | 0	0 | 4 | 0
GDIT6	−0.61 (0.11)	0.88 (0.12)	0.77 (0.13)	1.52 (0.15)	‐	‐	3 | 1 | 0	0 | 4 | 0
GDIT7	0.70 (0.11)	1.06 (0.13)	0.86 (0.14)	1.06 (0.15)	‐	‐	1 | 3 | 0	0 | 4 | 0
GDIT8	1.29 (0.11)	0.91 (0.13)	2.46 (0.18)	2.51 (0.29)	‐	‐	1 | 3 | 0	0 | 4 | 0
GDIT9	0.65 (0.11)	0.63 (0.12)	1.24 (0.13)	1.22 (0.15)	‐	‐	0 | 4 | 0	0 | 4 | 0
GDIT10	0.68 (0.11)	1.13 (0.12)	1.49 (0.15)	1.99 (0.20)	‐	‐	2 | 2 | 0	0 | 4 | 0
GDIT11	1.16 (0.10)	0.89 (0.12)	‐	‐	‐	‐	0 | 1 | 1	0 | 2 | 0
GDIT12	0.69 (0.10)	0.16 (0.11)	‐	‐	‐	‐	0 | 2 | 0	0 | 2 | 0
GDIT13	1.47 (0.11)	0.60 (0.12)	‐	‐	‐	‐	0 | 2 | 0	0 | 2 | 0
GDIT14	2.53 (0.12)	1.22 (0.16)	‐	‐	‐	‐	0 | 2 | 0	0 | 2 | 0
						Underfit >1.50	18	1
						Overfit <0.50	8	0

Abbreviations: GDIT, The Gambling Disorder Identification Test (Molander, Wennberg, & Berman, 2021); MNSQ, Mean‐square; SE, Standard error.

^a GDIT uses the following item response options: GDIT_item1, Never (not estimated), Monthly or less, 2–4 times a month, 2–3 times a week, 4 or more times a week, Daily and Several times a day; GDIT_items2–3, No time (not estimated), Less than an hour, 1–2 hours, 3–4 hours, 5–6 hours, 7–9 hours, and 10–24 hours; GDIT_items4–10, Never (not estimated), Less often than monthly, Monthly, Weekly, and Daily or almost daily; GDIT_item11–14, No (not estimated), Yes, but not in the past year, and Yes, in the past year.

2.4.3. Gender and age variations in item difficulty

Differential item functioning was analyzed to investigate whether differences in GDIT item difficulty estimates could be found between men and women (n = 440 and n = 151, respectively; six participants did not identity as either men or women), as well as between younger and older gamblers (≤30 vs. >30 years; n = 302 and n = 295, respectively, based on the median age). The differential item functioning of the GDIT was estimated using a standard unidimensional Rasch model with dichotomized responses.

2.4.4. Number of diagnostic strata

Rasch person‐separation reliability can be used to test into how many diagnostic levels (or strata) a measure reliably could be divided. A Rasch person‐separation reliability estimate of >0.70 corresponds to two strata, an estimate of >0.80 corresponds to three strata, and estimates of >0.90 and >0.94 indicate four and five strata, respectively (Wright & Masters, 1982). In a multidimensional Rasch model, person‐separation reliability is estimated per dimension. Given that the aim of this study related to an overall estimate of diagnostic levels/strata, person‐separation reliability was only analyzed for the unidimensional GDIT Rasch model.

3. RESULTS

3.1. Rasch modelling

Two Rasch analyses estimated the item discrimination and item difficulty of each of the GDIT's frequency‐ or time‐based multiple item response options across a severity continuum, in three‐dimensional and unidimensional models (see Tables 3 and 4, respectively).

3.1.1. Item discrimination

Several GDIT response options showed underfit or overfit in both the three‐dimensional and unidimensional Rasch models, particularly in the unidimensional model (Tables 3 and 4).

3.1.2. Item and response difficulties

The three‐dimensional model (Table 3) showed a GDIT item response difficulty range of −1.88 to 4.06. The item response alternative with the lowest difficulty was GDIT₁ (gambling monthly or less) and that with the highest difficulty was GDIT₈ (having borrowed money/sold something to gamble, daily or almost daily). Similarly, the unidimensional yielded a GDIT item response difficulty range of −1.80 to 2.53. The item response alternative with the lowest difficulty was GDIT₁ (gambling monthly or less), while the highest difficulty was shown in GDIT₁₄ (gambling‐related school/work problems, but not in the past year). In general, Rasch difficulty estimates in both models consistently increased in relation to the elevated time‐ and frequency‐based response options of the GDIT. An exception to this was that several items in the GDIT Negative Consequences subscale (GDIT_10–14) showed lower difficulty estimates for the response option “Yes, in the past year” compared to the option “Yes, but not in the past year”.

3.1.3. Gender and age‐related variations in item difficulty

Differential item functioning was analyzed to test whether differences in GDIT item difficulty estimates could be found between men and women, as well as younger and older gamblers (≤30 vs. >30 years). Regarding gender, three items – GDIT₈ (borrowed/sold something to gamble), GDIT₉ (escape gambling), and GDIT₁₂ (worse mental health due to gambling) – showed significantly higher difficulty for men. Regarding age, significant differences in difficulty occurred for nine of the 14 GDIT items. Specifically, item difficulty was significantly higher for older gamblers on the GDIT₁ (gambling frequency) and GDIT₂ (gambling duration) items. Conversely, GDIT₅ (chasing losses), GDIT₆ (gambled more than intended), GDIT₈ (borrowed/sold something to gamble), GDIT₉ (escape gambling), GDIT₁₁ (gambling‐related financial problems), GDIT₁₂ (worse mental health due to gambling), and GDIT₁₃ (gambling‐related relationship problems) showed significantly higher item difficulty for younger gamblers.

3.1.4. Number of diagnostic strata

Person‐separation reliability was estimated to test the number of diagnostic levels (or strata) using an unidimensional Rasch model. The result indicated that the GDIT reliably could be divided into three to four strata (person‐separation reliability 0.89; Wright & Masters, 1982).

4. DISCUSSION

The aim of this study was to test item discrimination and item and response option difficulties of the GDIT via Rasch analyses, as well to assess as the number of reliable diagnostic strata and demographic variations in item difficulty. We view the GDIT as a three‐dimensional measure, based on a previous confirmatory factor analysis (Molander, Wennberg, & Berman, 2021), as well as the Rasch item and dimensionality fit measures in the current study. The intended use of the GDIT however, involves screening and assessment via interpretation of the GDIT total score. Also, most previous Rasch studies on other gambling measures have been conducted using unidimensional models (Cowlishaw et al., 2019; Miller et al., 2013; Molander et al., 2023; Molander & Wennberg, 2022; Molde et al., 2010). In the current study, a unidimensional Rasch model was therefore presented alongside a three‐dimensional model, for transparency and to facilitate comparison.

4.1. Item discrimination

In terms of item discrimination, the GDIT generally followed the Rasch model. Based on this, it is concluded that GDIT can be used in parametric analyses (Bond & Fox, 2007), and to monitor severity across a continuum, ranging from low‐risk gambling levels to high diagnostic GD levels. However, some minor discrepancies are noted. Several item responses showed under and overfit in both the three‐dimensional and unidimensional models. In such partial credit Rasch models, each item response option is modeled in relation to the assumption of equal discrimination, thereby creating a more demanding test of model fit. A possible explanation for the lesser item response fit for these models might be that the frequency‐ and time‐based response options of the GDIT were primarily developed to maximize theoretical and face validity, rather than to reflect a strict interval (see Molander, Volberg, et al., 2021).

4.2. Item and response difficulties for the GDIT

The GDIT showed item response difficulty ranges of −1.88 to 4.06 and −1.80 to 2.53, in the three‐ and unidimensional models, respectively. These GDIT item difficulty ranges were larger or comparable to Rasch analyses of other gambling self‐report measures (Miller et al., 2013; Molander & Wennberg, 2022; Molde et al., 2010) and to GD criteria assessed via diagnostic interviews (Molander et al., 2023; see Table 5). Item difficulty range is important for the ability of a measure to cover aspects of a theoretical construct across a sufficiently wide continuum. For example, the Problem Gambling Severity Index (PGSI; Ferris & Wynne, 2001) was developed as a low gambling‐severity measure for use in the general population but has shown weakness in assessing such low problem severity levels in Rasch studies, which is a limitation shared by most brief gambling screens (Cowlishaw et al., 2019; Miller et al., 2013; Molander & Wennberg, 2022). In the current study, the GDIT showed low item difficulties, in particular the Gambling Behaviors subscale (GDIT_1‐3). These three GDIT items targets frequencies and times of gambling behavior, a measurement feature which has been recommended (Walker et al., 2006), but has been lacking in most gambling self‐report measures (Molander et al., 2019). The Gambling Behaviors subscale has shown good internal consistency reliability (α = 0.85) and test–retest reliability (6–16 days, intraclass correlation coefficient = 0.80) (Molander, Wennberg, & Berman, 2021). As the Gambling Behaviors subscale only includes three items with a low severity threshold, it has the potential to be used a brief measure to detect at‐risk gambling levels in the general population.

TABLE 5.

Rasch studies on gambling instruments.

Rasch studies and instruments	References indicating the source of development aims and intended use	Difficulty range of items/response options	Number of reliable diagnostic strata
Current study
GDIT, three‐dimensional	“[…] a measure that could identify and assess fulfilment of the revised diagnostic criteria for GD, for example, levels of symptom severity” Molander, Wennberg, and Berman (2021)	−1.88 to 4.06 (5.94)	‐
GDIT, unidimensional	(See above)	−1.80 to 2.53 (4.33)	3–4 (0.89)
Molander and Wennberg (2022)
PPGM (unidimensional)	“Utility in both population surveys and clinical settings”	0.47 to 5.11 (4.64)	2 (0.78)
	“[Assessment of] all potential harms deriving from gambling“
	Williams and Volberg (2013)
NODS (unidimensional)	“[A measure] designed to be more demanding and restrictive in assessing problematic behaviors than (…) other screens based on the DSM – IV criteria” Gerstein et al. (1999)	1.50 to 4.61 (3.11)	1–2 (0.50)
PGSI (unidimensional)	“[A measure for] use in general population surveys, one that reflected a more holistic view of gambling, and included more indicators of social context.” Ferris and Wynne (2001)	1.30 to 2.85 (1.55)	3 (0.82)
Molander et al. (2023)
SCI‐GD interview (unidimensional)	“[…] a clinician‐administered structured diagnostic interview which has been used as a gold standard assessment procedure in gambling research” Molander et al. (2023)	−0.50 to 1.28 (1.78)	2 (0.70)
Miller et al. (2013)
PGSI (unidimensional)	(See above)	−0.85 to 1.17 (2.02)	‐
Molde (2010)
NODS (unidimensional)	(See above)	−2.20 to 1.77 (3.97)	‐

Abbreviations: GDIT, The Gambling Disorder Identification Test (Molander, Wennberg, & Berman, 2021); NODS, The NORC Diagnostic Screen for Gambling Problems (Gerstein et al., 1999), 30 days version; PGSI, The Problem Gambling Severity Index (Ferris & Wynne, 2001); PPGM, The Problem and Pathological Gambling Measure (PPGM; Williams & Volberg, 2013); SCI‐GD, The Structured Clinical Interview for Gambling Disorder (Grant et al., 2004).

Overall, the difficulty estimates increased continuously in relation to the elevated time‐ and frequency‐based response options of the GDIT. However, some exceptions are noted. Both GDIT₁₃ (gambling‐related relationship problems) and GDIT₁₄ (gambling‐related work or school problems) showed a higher difficulty for the alternative “Yes, but not in the past year” compared to “Yes, in the past year” (see Table 4). A possible explanation for these findings might be that individuals with previous experience of gambling‐related problems in relationships or work/school might reflect a group with higher symptom severity. Few gambling measures include frequency‐ and time‐based item responses (Molander et al., 2019), and previous Rasch studies specifically targeting item responses are scarce. However, such analyses might be important from a measurement perspective. In the current study, each GDIT item showed a specific item response difficulty range. Taken together, these results indicate that the frequency‐ and time‐based item responses of the GDIT might facilitate clearer measurement, including assessment of both lower and higher gambling severity. According to our knowledge, GDIT is the only self‐report measure which can be used as a proxy to detect the revised DSM‐5 criteria for GD, including diagnostic severity (Molander, Wennberg, & Berman, 2021). Such assessment of GD is crucial from epidemiological, prevention and clinical perspectives, for establishment of diagnostic prevalence in the general population or and among treatment‐seeking patients.

4.3. Gender and age variations in item difficulties

Differential item functioning showed differences in item difficulties in relation to age and gender. A lifespan perspective might explain these differences, whereby younger gamblers might engage in a higher intensity gambling behavior, while older gamblers might face the aftermath of such gambling (i.e., gambling symptoms and negative consequences) to a greater degree. In terms of gender differences, studies have suggested that it is more common for women than men to gamble in order to escape or reduce symptoms of anxiety, depression or stress (i.e., emotionally vulnerable gamblers; see for example Nower et al., 2021). Gender‐based income differences might also be related to differences in use of strategies to acquire money for continuous gambling.

4.4. Number of diagnostic strata

In terms of diagnostic outcomes, person‐separation reliability indicated that the GDIT could reliably be divided into three to four strata. This adds to previous findings, where one to three strata have been reported for DSM‐IV based gambling measures (Molander & Wennberg, 2022) and two strata for DSM‐5 GD criteria assessed via diagnostic interviews (Molander et al., 2023; see Table 5). The proposed number of GD symptom severity levels (no, mild, moderate or severe GD) is less studied, but has been subject to critique. For instance, in a psychometric study, Grant et al. (2017) reported that several measures of psychopathology and gambling symptoms showed similar estimates between moderate and severe GD patients. Given that the GD DSM‐5 criteria were revised relatively recently, more research is needed to validate the symptom severity levels from an empirical perspective. Until such potential diagnostic revisions are made, we choose to retain the scoring index of the GDIT, as it was derived in relation to diagnostic interviews based on the DSM‐5 (Molander, Wennberg, & Berman, 2021).

4.5. Implications

The study has three major practical implications. The first is that the Rasch analysis of GDIT multiple responses showed a good fit for item difficulty measures that generally follow response progression for each item, meaning the total GDIT score can be used as a continuous variable that indicates GD severity. One practical implication of this finding is that the GDIT can be applied as a continuous variable in research studies or clinical contexts in which categorization of respondents by cut‐off‐based level of severity is not appropriate. A second implication is that there is not necessarily any need to dichotomize GDIT severity outcomes, which would require non‐parametric analyses; the continuous total score can instead be used in more robust parametric analyses. Thirdly, in treatment contexts, the GDIT can be used to detect and assess GD, including symptom severity. An additional finding of high practical value is that the GDIT time‐ and frequency‐based responses can provide precise measures of gambling patterns, both as snapshots and in time‐series measures. In the latter case, risk levels can be easily and robustly measured across time in groups exposed to various types of interventions. In future GDIT validation research, analysis of the GDIT income variables may also provide an additional measure of severity that could be very helpful in clinical and research settings (for example, in assessing the percentage of available monthly income spent on gambling).

This was the first study to evaluate the psychometric properties of the GDIT using Rasch methods. The study yielded several important findings in relation to measurement of problematic gambling behavior that have implications for the diagnostic assessment of GD. The study included past‐year gamblers from both low severity (general population) and high severity (help‐seeking) samples, yielding a satisfactorily wide range of difficulty item ratings. Also, the study encompassed uni‐ and three‐dimensional modeling, including testing of multiple response options, an approach that reaches beyond the standard Rasch model, to assess GDIT measurement properties.

In terms of limitations, the sample size was relatively small compared to other gambling Rasch studies with a range of N = 1305–25000 (see Cowlishaw et al., 2019; Miller et al., 2013; Molde et al., 2010). As noted, a partial credit Rasch model was used here, despite the fact that the multiple response options in the GDIT were not designed to constitute strict frequency‐ and time‐based intervals. Other item response theory methods in which equal discrimination is not assumed could have been used to assess item response theory; nonetheless, a decision was made to use the Rasch model to facilitate comparisons of results with other studies. Future psychometric studies could investigate GDIT responses using alternative item response theory methods, investigate item response theory adaptive testing to reduce item burden, or corroborate findings among larger and/or international gambling samples. From an epidemiological perspective, establishment of specific frequency‐ and time‐based cut‐offs for gambling behavior (measured by the GDIT Gambling Behaviors Subscale) in relation to symptom severity and low‐risk levels, is also of interest. From a clinical and research perspective, establishment of DSM‐5 GD prevalence rates among clinical and non‐clinical populations is a prioritized endeavor.

5. CONCLUSION

Rasch analysis of the GDIT in this study's sample of recreational and help‐seeking gamblers has clarified that the GDIT consists of three‐dimensional strata and that multiple response items generally discriminate equally between respondents. Differences in item difficulty by age and gender suggest alignment with a lifespan perspective. Our findings support using the GDIT to estimate DSM‐5 prevalence of GD in epidemiological and clinical samples, and we would welcome multi‐national collaboration in this effort.

AUTHOR CONTRIBUTIONS

Olof Molander: Conceptualization; writing ‐ review & editing; writing ‐ original draft; investigation; validation; methodology; formal analysis; data curation; visualization; software; project administration. Peter Wennberg: Conceptualization; writing ‐ review & editing; investigation; methodology. Nicki A. Dowling: Conceptualization; investigation; writing ‐ review & editing; methodology. Anne H. Berman: Conceptualization; investigation; methodology; writing ‐ review & editing.

CONFLICT OF INTEREST STATEMENT

The authors identify no conflicts of interest. From January 2013 until December 2022, a period that coincided with this study, author AHB was a board member of the independent research council funded by the state‐owned gambling company Svenska Spel AB. The council funds gambling related studies, but SvenskaSpel AB has no influence whatsoever on decisions to grant funds to researchers. Author OM has received grants from this research council for separate independent studies, unrelated to the work reported in this article. In the last 3 years, ND has received research and consultancy funding from multiple sources, including via hypothecated taxes from gambling revenue. ND has received research funding from the Victorian Responsible Gambling Foundation, New South Wales Office of Responsible Gambling, Tasmanian Department of Treasury and Finance, Gambling Research Australia, Swedish Gambling Research Council, Health Research Council of New Zealand, and New Zealand Ministry of Health. She has been the recipient of a Deakin University Faculty of Health Mid‐Career Fellowship. She has not knowingly received research or consultancy funding from the gambling, tobacco, or alcohol industries or any industry‐sponsored organization.

ETHICS STATEMENT

This study was approved by the Regional Ethics Board of Stockholm, Sweden (ref. no. 2017/1479‐31). All participants provided informed consent for participation and publication.

Molander, O. , Wennberg, P. , Dowling, N. A. , & Berman, A. H. (2024). Assessing gambling disorder using frequency‐ and time‐based response options: A Rasch analysis of the gambling disorder identification test. International Journal of Methods in Psychiatric Research, e2018. 10.1002/mpr.2018

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.