ORIGINAL RESEARCH article
Developing and Validating an Item Bank for Alcohol Use Disorder Screening in the Chinese Population by Using the Computerized Adaptive Testing
- School of Psychology, Jiangxi Normal University, Nanchang, China
Objective: To detect the individual’s severity of alcohol use disorder (AUD) in an effective and accurate manner, this study aimed to build an item bank for AUD screening and derive the computerized adaptive testing (CAT) version of AUD (CAT-AUD).
Methods: The initial CAT-AUD item bank was selected from the Chinese version of the questionnaires related to AUD according to the DSM-5 criteria. Then 915 valid Chinese samples, covering the healthy individuals and the AUD high-risk individuals, completed the initial CAT-AUD item bank. By testing the unidimensionality, test fit, item fit, discrimination parameter and differential item functioning of the initial item bank, the final CAT-AUD item bank confirming to the requirements of the item response theory (IRT) were obtained. Subsequently, the CAT-AUD simulation study based on the real data of the final item bank conducted to detect characteristics, reliability, validity, and predictive utility (sensitivity and specificity) of CAT-AUD.
Results: The CAT-AUD item bank meeting the IRT psychometric measurement requirements could be well geared into the graded response model. The Pearson’s correlation between the estimated theta via CAT-AUD and the estimated theta via the full-length item bank reached 0.95, and the criterion-related validity was 0.63. CAT-AUD can provide highly reliable test results for subjects whose theta above −0.8 with an average of 16 items. Besides, the predictive utility of CAT-AUD was better than AUDIT and AUDIT-C.
Conclusion: In brief, the CAT-AUD developed in this study can effectively screen the AUD high-risk group and accurately measure the AUD severity of individuals.
As one of the world’s commonest mental disorders, alcohol use disorder (AUD) is a problematic pattern of alcohol use which could lead to clinically significant impairment or distress (American Psychiatric Association, 2013). In 2016, the harmful drinking pattern had resulted in an estimated 3 million deaths worldwide, which accounted for 5.3% of all deaths (World Health Organization, 2019). This disorder is also correlated with a wide range of physical and mental disorders (Cargiulo, 2007; Hasin et al., 2007; Rehm, 2011). Individuals with AUD are not competent in fulfilling their primary role at the workplace, school or home, while their social relationship and the financial situation will be adversely affected (Rehm et al., 2009; American Psychiatric Association, 2013). Besides, AUD high-risk individuals are inclined to cause social harm such as vehicle accidents and violent crime (World Health Organization, 2014).
In China, public health has also been endangered by AUD. Surveys of several provinces and cities showed that AUD had jumped to the forefront among mental diseases and became a serious public health problem in China (Shi et al., 2005; Pan et al., 2006; Li et al., 2008; Ruan et al., 2010). To be specific, there are some differences in the probability of suffering AUD among different demographic groups. In terms of gender, Zhang et al. (2010) found that the rate of suffering AUD in adults was 10.92% for males and 0.17% for females in Shandong Province. The study by Chang et al. (2016) also showed a statistically significant difference in suffering AUD among genders (χ2 = 125.598, p < 0.01). In urban and rural areas, the survey results of adults from Heilongjiang Province showed that the harmful drinking rate of the urban population was 5.20%, and that of the rural population was 9.63%, which was higher in rural areas than in urban areas (χ2 = 22.27, p < 0.05).
For the purpose of reducing the rate of prevalence and mortality caused by AUD, detections should be used in the general population to screen out the AUD high-risk group. Currently, numerous questionnaires such as the Alcohol Use Disorder Identification Test (AUDIT; Babor and Grant, 1989), the classical Michigan Alcoholism Screening Test (MAST; Selzer, 1975) and the Alcohol Dependence Scale (Skinner, 1981) have been developed to measure the specific symptoms rather than the whole symptoms of AUD under the framework of classical test theory (CTT). Nevertheless, several inherent defects were showed in these questionnaires. First, current questionnaires were not developed in accordance with the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM–5; American Psychiatric Association, 2013) criteria. Thus, some symptoms of AUD in DSM-5 could not be mirrored integrally by the existing questionnaires which were not keeping abreast with times and practice, and then the information offered by questionnaires would remain unreliable. Second, based on the principle of CTT, participants would respond to the same items which are equally weighted regardless of their level of some trait, and the summed total scores under the normal condition were the measurement results which would be representative of individuals’ level of some trait. As a matter of fact, each item has a different weight for AUD high-risk individuals of different levels of severity. Some researchers reported that AUD is a heterogeneous disorder and upheld for personalized treatment of AUD (Linden-Carmichael et al., 2018). Also, Weiss (2004) stated that some items in fixed-length instruments may lead to error since measurement precision declined at both the high and low levels of the trait. On the whole, these CTT based questionnaires for AUD are not comprehensive or adaptive sufficiently and should be required further improvement.
Traditional CTT had become inadequate for the demand for measurement accuracy and efficiency. Under the stimulation of such need, item response theory (IRT; Embretson and Reise, 2013) was developed. The characteristic of IRT is to describe in the form of probability function how the item response result is affected by the combination of the ability level of the subject and the item characteristic, which entitles IRT with many advantages. First, it has parametric invariance, that is, the population parameter estimated from different samples of the same population is constant. Secondly, the subjects’ parameter estimated by IRT and the measuring scale of the item difficulty parameter is unified (Baker, 2001). In this way, items whose difficulty parameter is suitable for the subjects’ ability level can be selected to make the estimates of individuals’ latent trait more efficient and accurate. Furthermore, IRT can accurately estimate the measurement error of each item and the total test for subjects with different trait levels (Kim, 2012) while the traditional CTT which calculated the standard error of measurement (SEM) for a group of scores could just offer the total reliability regardless of subjects’ trait levels. In recent years, CTT was applied to conditional standard error of measurement (CSEM), which could also make CTT provide high measurement reliability for groups with different ability levels (Culpepper, 2013).
The emergence and growth of computerized adaptive testing (CAT) were due to the creation of IRT and the inexpensive accessibility to high-speed desktop computers (Gershon, 2005). The core of CAT is to maximize the precision by only administering the items suitable to individuals’ latent trait (van der Linden and Glas, 2000, van der Linden and Glas, 2010; Wainer et al., 2000). In accordance with IRT (Embretson and Reise, 2013), the computer flexibly administers suitable items that have been calibrated for individuals by specific algorithms, so people are allowed to administer fewer items when the accuracy is ensured to be as much as possible (Hambleton et al., 1992; van der Linden and Hambleton, 1997). The adaptive algorithm guarantees the test accuracy of the subjects and makes the subjects with different latent traits answer questions with different parameters as few as possible. With such an adaptive ability in CAT, the shortcomings of current questionnaires can be compensated. To accurately measure an individual’s level of AUD, selecting the CAT version may be prioritized.
This study aimed to develop a CAT version of AUD (CAT-AUD) based on DSM-5 as a screening instrument to distinguish the healthy group and the AUD high-risk group in Chinese adult population, which is expected to be a comprehensive, effective, and highly accurate for the AUD assessment. Based on the guidance of the AUD criteria of DSM-5, items that measure at least one symptom of AUD in DSM-5 were selected from five famous AUD scales to constructed the initial CAT-AUD item bank. After that, various statistical analyses including unidimensionality test, monotonicity test, local independence test, the level of item-fit test, the item discrimination and differential item function (DIF) test based on IRT were conducted to the initial item bank to develop the final item bank. Finally, a CAT-AUD simulation was implemented to investigate the reliability, validity and prediction utility (sensitivity and specificity) of the CAT-AUD.
Materials and Methods
One thousand three hundred and twenty Chinese participants aged beyond 18 years from Henan, Jiangxi, Shandong, Hunan, Guangxi, Fujian Provinces were recruited for this survey. Before conducting the questionnaire, each participant was informed of the individual privacy protection principle and then they volunteered to participate in the survey. The informed consent of all participants was obtained in accordance with the declaration of Helsinki. The survey included the basic demographic information (gender and region), the AUD item bank (including three pairs of lie detection items), and questions used as the exclusion criteria. In order to obtain real and effective response data, we screened the original questionnaires in advance. If participants satisfy one of the exclusion criteria below, they would not be covered in this study: (1) those patients with psychosis illness; (2) those having received psychiatric medication in a week; (3) those abusing or relying on other substances other than tobacco; (4) those with impaired heart, liver and kidney function or other serious physical diseases not resulting from alcohol. Besides, if a participant had at least one pair of identical response data on the lie detection item, his or her response data would be excluded.
One thousand three hundred and twenty questionnaires were distributed and 1128 completed questionnaires were collected, with a recovery rate of 85.5%. Among them, 52 questionnaires contained missing demographic information, 73 questionnaires failed to meet the lie detection test criteria, 88 questionnaires met at least one of the exclusion criteria, then the number of valid questionnaires was 915, suggesting that the valid rate was 81.1%. Table 1 contains the demographic and clinical characteristics of the valid sample. Among these participants, 735 were males (80.33%), with the average age of 37.58 (SD = 14.28), and 180 were females (19.67%), with the average age of 33.95 (SD = 11.38). There were 420 subjects in urban areas (45.90%), aged 35.83 on average (SD = 11.40), and 495 subjects in rural areas (54.10%), aged 36.21 on average (SD = 11.51).
All participants took part in the survey above as well as took part in the questionnaire of the MAST. In accordance with the research of Selzer (1975), subjects with a total score of 6 or above were classified as the AUD high-risk group, while those with a score of <6 were classified as the healthy group. These two groups were used to test the validity of the CAT-AUD which would be constructed subsequently.
The authors searched domestic and foreign publications with words such as the Chinese version scale of “AUD,” “AUD scale,” and “AUD questionnaire.” In order to find the target questionnaires among the numerous questionnaires, we selected the questionnaires related to the AUD according to some certain criteria: (1) whether there is a validated Chinese version of the questionnaire; (2) whether the items in the questionnaire measure one of the AUD symptoms in DSM-5; (3) whether the reliability and validity are high enough: (a) the test-retest reliability or the Cronbach’s alpha of a questionnaire is higher than 0.8 (Wu, 2010); (b) the criterion-related validity of the questionnaire is higher than 0.4 (Streiner and Norman, 2008); (c) the scale content validity (S-CVI) is higher than 0.8 (Davis, 1992); (4) the high citation frequency of a questionnaire; (5) the number of items of a questionnaire should not be too small. For the original questionnaire without a Chinese version, the process of Chineseization is as follows: after obtaining the approval of the author of the original questionnaire, the Chinese version of the questionnaire was then translated using the Brislin’s classical back translation model to form the Chinese version of the questionnaire. Moreover, to compare the reliability and validity of different questionnaires for the sample in this study horizontally, the Crombach’s alpha and criterion-validity values of the Chinese version of questionnaires above were calculated and displayed.
The Chinese version of AUD scales employed in this study exhibited three major purposes. Five scales including the Alcohol Dependence Scale (ADS; Skinner, 1981), the Munich Alcoholism Test (MALT; Speckens et al., 1991), the Severity Alcohol Dependence Questionnaire (SADQ; Stockwell et al., 1983; Cheng et al., 2009), the Obsessive Compulsive Drinking Scale (OCDS; Anton et al., 1995), and the Approach and Avoidance of Alcohol Questionnaire (AAAQ; McEvoy et al., 2004) were used to build a comprehensive CAT-AUD item bank containing all AUD symptoms of DSM-5; the classical Michigan Alcoholism Screening Test (MAST; Selzer, 1975; Hsueh et al., 2014) was performed as a criterion scale to investigate the validity of the CAT-AUD; the Alcohol Use Disorder Identification Test (AUDIT; Babor and Grant, 1989; Zhang et al., 2017) and the AUDIT consumption questions (AUDIT-C; Bush et al., 1998) were applied for the comparison of the performance of the CAT-AUD.
The Alcohol Dependence Scale
The ADS was developed using the alcohol dependence factor correlated with chronic social debilitation and psychopathology (Skinner, 1981) to ascertain patients’ symptoms of alcohol. The ADS covered 29 items scored with 0–1, 0–2, and 0–3, respectively. The original ADS scale was used, which exhibited the internal consistency reliability of 0.92 and the criterion-related validity of 0.69 with MAST (Skinner and Allen, 1982) and the mean and standard deviation (SD) of the sample ADS score were 23.1 and 11.3, respectively. In the present study, the Chinese version of the ADS depicted a Cronbach’s alpha of 0.85 and the criterion-related validity of 0.69 with MAST.
The Munich Alcoholism Test
The MALT refers to a screening instrument for alcoholism especially applicable to medical populations (Speckens et al., 1991). Such a scale is commonly used in many counties worldwide. This study chose the self-rating part of MALT (MALT-S) with 24 dichotomously scoring items. The scale exhibited split-half reliability of 0.94 and convergent validity of 0.85 (Feuerlein et al., 1979). In the present study, the Chinese version of the MALT depicted a Cronbach’s alpha of 0.86 and the criterion-related validity of 0.74 with MAST.
The Severity Alcohol Dependence Questionnaire
The SADQ covered 20 items all graded in four levels (Stockwell et al., 1983). The SADQ consisted of five dimensions: (a) physical symptoms of withdrawal (PHYS), (b) affective symptoms of withdrawal (AFF), (c) craving and withdrawal-relief drinking (NEED), (d) typical daily consumption (ALC), as well as (e) rapidity of reinstatement of symptoms after a period of abstinence (POSTAB) (Stockwell et al., 1979). Given the study of Stockwell et al. (1979), the questionnaire’s test-retest reliability was 0.85, and the criterion-related validity was 0.63. For the sample who rated 0 or 1, the mean and SD of the SADQ score were 28.8 and 11.3, respectively. For sample who rated 2, the mean and SD of the SADQ score were 50.6 and 12.2, respectively. In the present study, the Chinese version of the SADQ (Cheng et al., 2009) depicted a Cronbach’s alpha of 0.95 and the criterion-related validity of 0.61 with MAST.
The Obsessive Compulsive Drinking Scale
The OCDS, developed by Anton et al. (1995) who modified Modell’s Yale-Brown Obsessive Compulsive Scale (Y-BOCS), refers to a self-reported instrument that is capable of rating the level of drinking crave (Modell et al., 1992). It consists of 14 items, all of which were scored from 0 to 4. Among them, the item numbers of 1–6 acted as the degree of psychological craving for alcohol use testers, and the item numbers of 7–14 indicated the testers’ behavior and motivation for alcohol consumption. The test-retest reliability of the OCDS was 0.96, and the convergent validity was 0.83, which are both relatively high (Anton et al., 1995) when the mean and SD of the sample score were 22.5 and 7.5, respectively. In the present study, the Chinese version of the OCDS depicted a Cronbach’s alpha of 0.92 and the criterion-related validity of 0.68 with MAST.
The Approach and Avoidance of Alcohol Questionnaire
The AAAQ has been developed to assess inclinations to drink and not drink separately (McEvoy et al., 2004). Participants suggested that how firmly they agreed with each item on a nine-point Likert-type scale, from “not at all” (0) to “very strongly” (8). Existing studies suggested that the AAAQ had high internal consistency and convergent and predictive validity (McEvoy et al., 2004; Klein et al., 2007). According to McEvoy et al. (2004), the internal consistency of AAAQ was 0.62, and the scale interpreted a substantial proportion of the variance in drinking frequency (41–53%), drinking quantity (49–60%) as well as drinking problems (43%). For the inclined sample, the mean and SD of score were 4.16 and 2.42, respectively. For the obsessed sample, the mean and SD of score were 1.56 and 1.54, respectively. In the present study, the Chinese version of the AAAQ depicted a Cronbach’s alpha of 0.89 and the criterion-related validity of 0.46 with MAST.
The Michigan Alcoholism Screening Test
The MAST refers to a self-rating scale used in this study to screen high-risk individuals for alcohol disorders (Selzer, 1975). MAST primarily measures five aspects (namely, problems recognized by oneself or others, work and social problems, seeking help for drinking, marital and family problems, as well as physical health problems). Compared with the AUD symptoms of DSM-5, these aspects might be more general but almost consistent, so this scale was adopted as the criterion scale. In this study, the Chinese version of MAST (MAST-C) was employed to investigate the validity of the CAT-AUD (Hsueh et al., 2014). The test consisted of 24 items, and the item response options were rated dichotomously (yes/no). The internal consistency reliability and the scale-level content validity index (S-CVI) of the MAST-C were 0.89 and 0.92, respectively (Hsueh et al., 2014). In the present study, the Chinese version of the MAST (Hsueh et al., 2014) depicted a Cronbach’s alpha of 0.88.
The Alcohol Use Disorder Identification Test
For the validity comparison, the AUDIT and the AUDIT-C were employed in the study. The AUDIT was built by investigators from WHO collaborated on the six-nation project in the 1980s (Saunders, 1987; Babor and Grant, 1989; Babor et al., 1992; Saunders et al., 1993). The results of the study by Zhang et al. (2017) revealed that Cronbach’s alpha of the Chinese version was 0.782, and the split-half reliability was 0.711. The content validity of the item-level (I-CVI) ranged from 0.83 to 1.00, and the average content validity for the scale-level (S-CVI) was 0.99 (Zhang et al., 2017). In the present study, the Chinese version of the AUDIT (Zhang et al., 2017) depicted a Cronbach’s alpha of 0.90 and the Chinese version of the AUDIT-C depicted a Cronbach’s alpha of 0.77.
The Initial Item Bank for the CAT-AUD
For subsequent analysis, items which measure at least one criterion of the DSM-5 for AUD for the constructive purpose were introduced into the initial item bank. In order to screen out items which met the above standard, we invited 3 clinicians with curing AUD experience to evaluate whether the items measure at least one criterion of the DSM-5 for AUD. By comparing the judgment results of different clinicians, items that measured at least one criterion were kept in the initial CAT-AUD item bank. Table 2 presents the source of the selected items in the item bank. The R package mirt (Version 1.29; Chalmers, 2012) was performed for the following IRT analyses.
Construction of the Final Item Bank for the CAT-AUD
To select high-quality items from the initial bank under the framework of IRT, unidimensionality test, item-fit test, discrimination test, and differential item functioning (DIF) test were performed in this study based on the initial item bank. Only those items meeting the requirements of IRT psychometric measurement can be involved in the final item bank to build the CAT-AUD. In order to understand the steps of item bank generation more intuitively, we provided a figure depicting the steps of the psychometric measurement requirements of IRT for the item bank generation (see Figure 1).
Figure 1. The steps of the psychometric measurement requirements of IRT used for the item bank generation.
In IRT, unidimensionality implies that one major ability or trait should “explain” or “account” for the test performance of examinees (Hambleton and Swaminathan, 2013). The criteria for the data which might be reasonably well fit by a unidimensional model by exploratory factor analysis (EFA) included: (1) the ratio of the first eigenvalue to the second eigenvalue was above 3 (Lord, 1980; Hattie, 1985); (2) the percentage of variance interpreted by the first factor was more than 20% of the total variance (Reckase, 1979). Only two criteria were both satisfied at the same time, the data of the CAT-AUD item bank might fit the unidimensionality model well (Reckase, 1979). The software SPSS 21.0 was used for the analysis of unidimensionality.
Test Fit and IRT Model Selection
To assess the degree of fitness for the whole item bank, the graded response model (GRM; Samejima, 1969) and the generalized partial credit model (GPCM; Muraki, 1992) dealing with polytomous-scored items were used in this study. Given the three test-level model-fit indices: negative twice Log-likelihood (−2LL; Spiegelhalter et al., 1998), Akaike’s information criterion (AIC; Akaike, 1974), and Bayesian information criterion (BIC; Schwarz, 1978), the model representing the indices with the relatively smaller values were taken as the optimal model for the subsequent analysis.
In the test-fit test, the optimal model was selected and the overall degree of data fitting was suggested. However, this did not indicate that each item could fit the optimal model well. Accordingly, the degree of item-fit further should be explored. The S-χ2 was adopted to show the degree of the item-fit index (Orlando and Thissen, 2000, 2003). According to Flens et al. (2017), items with a p-value of S-χ2 below 0.001 were considered to be poor item-fit. A stricter standard was taken here, namely, if the item’s p < 0.01, it would be considered a low item-fit one and then deleted.
The high discrimination (slope) parameter of the IRT model refers to a desirable characteristic of the item quality (Masters, 1988). An item with relatively high discrimination is capable of ensuring that each subject’s latent trait level is separated by the effective area of the item bank. Given the study of Fliege et al. (2005), items whose discrimination values are <0.7 should be excluded from the item bank.
Differential Item Functioning
The DIF analysis was conducted to identify systematic errors attributed to different groups (Zumbo, 1999). If an item had DIF, the probability of selecting the same item type might be different for groups at the identical latent trait level. In current study, DIF for gender (male/female) and region (rural/non-rural) was investigated. The ordinal logistic regression (LR) (Crane et al., 2006) was used for the DIF analysis. The McFadden’s pseudo ΔR2 ≥ 0.2 (Flens et al., 2017) revealed a functional difference in the item, which should be removed. The R package lordif (Version 0.3–3; Choi et al., 2011) was used for the DIF test.
Through the above steps of the IRT analysis, the remaining items of the final item bank for the CAT-AUD would be employed to reassess the person ability parameter.
CAT-AUD Simulation Study
The simulation study based on real participants’ response data was presented to investigate the characteristic, the criterion-related validity, as well as the predictive utility (sensitivity and specificity) of the CAT-AUD, AUDIT, and AUDIT-C for comparison. Before the investigation, the authors made some specifications about the program of the CAT-AUD.
Initial Item Selection
In the CAT-AUD simulation program, items which were administered to participants were based on their responses to prior items. At the beginning of the test, the initial item of the CAT-AUD was randomly selected from the final item bank since there was no response information of participants (Magis and Barrada, 2017).
Item Selection Algorithm
After the initial item selected, the Fisher information was picked as the selection index for the subsequent item selection progress (Baker, 1994). The item with the maximum Fisher information (MFI) based on the participant’s current theta in the rest item bank would be administered by subjects (Baker, 1994).
Expected a posteriori method (EAP; Bock and Mislevy, 1982) is one of the Bayesian approaches to estimate the ability parameter. The prior distribution of ability is generally assumed to be a standard normal distribution. EAP is considered as an efficient estimation method of the person parameter. It can obtain the expected estimation of the person parameter by a direct calculation rather than an iteration.
In this study, the two commonly used CAT stopping rules (a fixed measurement of standard error and a fixed number of items) were integrated as the stopping rule of the CAT-AUD. The selection of standard error (SE) could ascertain the corresponding IRT reliability. As the subject answered a growing number of items, the measurement SE would be lower, and the reliability of the test would be higher. According to the existing studies (Wainer et al., 2000), when IRT reliability of rxx ≥ 0.9, the result of the test could be considered credible and accurate. When the mean and standard deviation (SD) of theta are fixed to 0 and 1, the relationship between the IRT reliability and SE of the test is expressed as (Samejima, 1994):
According to the formula above, when the reliability reaches 0.9, the corresponding SE is around 0.32. If the reliability of each CAT procedure is expected to be at least 0.9, the corresponding maximum SE set to 0.32 will be relatively reasonable and acceptable. The maximum length of the test was set as 35, if the subject in testing has not reached the standard of SE ≤ 0.32 after answering 35 items, the CAT program would stop administering any items to the subject.
The Characteristics of the CAT-AUD
To explore the characteristics of the CAT-AUD, frequency distributions of theta values estimated via the full-length item bank and via the CAT-AUD were formed and compared first. Subsequently, the number of items administered by subjects, and the marginal reliability, representing the average of the corresponding reliability values under each theta (Smits et al., 2011), were investigated completely.
The Validity and Predictive Utility (Sensitivity and Specificity) of the CAT-AUD, AUDIT, and AUDIT-C
To investigate the validity of the CAT-AUD, the two scales including the AUDIT and AUDIT-C were introduced for the transverse comparison. The criterion-related validity of the CAT-AUD, AUDIT, and AUDIT-C was examined by assessing the Pearson’s correlation between individuals’ estimated theta values by these three tests and individuals’ corresponding raw scores of the MAST.
For the assessment of the predictive utility, the area under (AUC) the receiver operating characteristic curve (ROC) was introduced (Smits et al., 2011). The AUC refers to the probability that a randomly chosen patient is rated or ranked as more likely to be diseased than a randomly chosen healthy individual (Hajian-Tilaki, 2013). Youden Index (YI; Schisterman et al., 2005) is another major summary statistic of the ROC used for the interpretation and assessment, defining the maximum potential effectiveness of a test. YI equals the sum of the sensitivity and the specificity minus 1. The sensitivity (Se) refers to the probability that an AUD high-risk individual will be diagnosed with the disease, and the specificity (Sp) represents the probability that a healthy individual will be diagnosed with non-disease. The larger the value of these two indicators (Se and Sp), the better the diagnostic effect will be. The cut-off point is used to determine whether the estimate results show a clinical risk of subjects. In CAT-AUD, if the subjects’ theta value is higher than the cut-off point, they will be at risk of AUD; if not, they will be relatively healthy. At the cut-off point, the corresponding value of YI is the largest among all subjects’ tests. The MAST scale acted as a categorical variable of AUD and theta values estimated by the CAT-AUD, AUDIT, and AUDIT-C as the continuous variable. Subsequently, the ROC result was analyzed and portrayed using SPSS 21.0 software when the stopping rule was SE ≤ 0.32.
Construction of the Item Bank for the CAT-AUD
According to the EFA results of the initial item bank, there were 5 items with factor loadings on the first factor <0.3 (Nunally, 1978). After these items were excluded, the unidimensionality test for the rest items was performed. The eigenvalue of the first component and the second component were 53.55 and 4.83, respectively. Their ratio of the first component and the second component was 11.09, which was far more than the criteria of 3 (Lord, 1980; Hattie, 1985). Moreover, the first factor interpreted about 41.84% of all the variance, which was more than the criteria of 20% (Reckase, 1979). The above results indicated the remained items could fit the unideminsional model well. And the ability parameters estimated based on the data were related to the first principal component.
Test Fit and IRT Model Selection
The commonly used test fit indicators (−2LL, AIC, BIC) of the GRM and the GPCM are listed in Table 3. It was found that the fitting degree of the GRM was better than that of the GPCM. Based on such results, the GRM was subsequently selected as the IRT model with the current data for the subsequent analysis.
Item-Fit, Discrimination and Differential Item Functioning
The result of item-fit suggested that item 65 and item 94 did not fit the GRM well. For discrimination parameters, there were 3 items (138, 142, 147) <0.7. In terms of DIF, there was no DIF item in area groups or gender groups.
After 10 items that did not meet the above requirements of psychometric measurement from the initial item bank were deleted, the final item bank remained 113 items. The number of items in the item bank measuring each symptom criteria of AUD defined in DSM-5 varied from 6 to 15 with an average of 11.27 (SD = 3.10), which indicated that the questionnaire based on the AUD item bank could exhaustively measure each diagnose criterion for AUD in DSM-5. And when the item bank was used to assemble the CAT-AUD, the probability that the CAT-AUD would have adequate content validity was high (Waltz et al., 2005).
The item psychometric information of the final item bank is listed in Table 2. As shown in Table 2, the slope value ranged from 0.93 to 4.07 with an average of 2.34 (SD = 0.80), suggesting that the quality of the final item bank was relatively high. The threshold of the final item bank ranged from −0.62 to 3.82, indicating that the item bank’s coverage of the theta value is skewed toward the high-risk group.
In Table 4, the descriptive statistics for the full-length item bank score distribution were shown. Evidently, the number of subjects with scores ranging from 0 to 49 is the largest, followed by those with scores ranging from 50 to 99, and those with the least number of subjects ranging from 250 to 299. In general, the standard deviation of each score range is close to each other. The standard deviation is the lowest for a sample whose score range is 200–249, and the standard deviation is the highest for a sample whose score range is 50–99.
The CAT-AUD Simulation Study
The Characteristics of the CAT-AUD
Figure 2 displays frequency distributions of theta values estimated via the full-length item bank and via the CAT-AUD at a given SE level (SE ≤ 0.32). The theta values(M = 0.06, SD = 1.002) estimated via the full-length item bank and the theta values(M = 0.05, SD = 0.987) estimated via the CAT-AUD were almost matched on frequency distributions, with a Pearson’s correlation of 0.95 (p < 0.01).
Figure 2. Frequency distributions of theta values estimated via the full-length item bank and the CAT-AUD at a given SE level (SE ≤ 0.32).
Figure 3 presents the information and its corresponding SE function of theta values estimated by the full-length item bank. Obviously, for those whose theta values are <−2, the information provided by the item bank is limited that it is almost impossible to provide accurate test results. For the subjects whose theta values ranged from −2 to −0.8, the test could provide certain amount of information. For the subject whose theta value is more than −0.8, the test could provide large amount of information and high reliability. With the increase of the theta value, the amount of information for item bank increases rapidly. When the theta value reaches about 1.5, the corresponding SE of the full-length test approaches 0. Therefore, for high-risk and clinical subjects, the item bank is very suitable and accurate for them, while for non-clinical subjects, the item bank can provide very little information.
Figure 3. Statistical information and standard error function of the theta values estimated via the full-length item bank.
In Figure 4, the information and its corresponding SE function of theta values estimated via the CAT-AUD was shown. As can be seen, when the SE value is <0.32, theta value in the graph would be roughly >−1. The subjects whose theta values are >−1 could benefit a lot from the CAT-AUD as the amount of information are relatively high. While for the subjects whose theta values are <−1, the current CAT may not provide much information for them.
Figure 4. Statistical information and standard error function of the theta values estimated via the CAT-AUD at a given SE level (SE ≤ 0.32).
In Figure 5, the minimum number of items to be administrated by participants was around 3. And the average number of items to be answered was 16.00 (SD = 12.26), which only took up 14.16% of the item bank length. For individuals whose theta values of the CAT-AUD were <−0.5, the number of item usage remained unchanged and were all set at a maximum of 35 items. For those with a value >−0.5, the number of item usage decreased as theta increased until the theta value reached around 1. Moreover, the marginal reliability of the CAT-AUD (SE ≤ 0.32) was 0.87, indicating that the CAT-AUD had high reliability in estimating the person parameter.
Figure 5. Number of items administered by individuals in the CAT-AUD at a given SE level (SE ≤ 0.32).
The Validity and Predictive Utility (Sensitivity and Specificity) of the CAT-AUD, AUDIT, and AUDIT-C
Table 5 reveals the criterion-related validity between the criterion scale (MAST) and three questionnaires (CAT-AUD, AUDIT, AUDIT-C). As can be seen from the result, the criterion-related validity between the MAST and the CAT-AUD was the highest among three criterion-related validity values, which is suggesting that the estimated theta by the CAT-AUD was consistent with the result of the MAST.
Table 5. The predictive utility (sensitivity and specificity) of the CAT-AUD, AUDIT, and AUDIT-C and the criterion-related validity.
For the predictive utility, the performance of the CAT-AUD remains the best among the three questionnaires. To be specific, the AUC value of the CAT-AUD in Table 5 equaled to 0.856, which was higher than those of the AUDIT and AUDIT-C. The situation that the cut-off value of the CAT-AUD was higher than that of the AUDIT and AUDIT-C mirrors that the CAT-AUD could distinguish AUD high-risk individuals with higher severity in AUD than the other two questionnaires. The sensitivity of the CAT-AUD was the highest while the specificity of AUDIT was the highest.
In the current research, the CAT-AUD was developed to screen the AUD high-risk group and measure individuals’ severity of AUD. For these ends, a high-quality item bank was established for the CAT-AUD under the guidance of IRT to support the operation of the CAT-AUD simulation study.
The result of the study suggested that the CAT-AUD could provide accurate, valid psychometric information for the AUD high-risk group while it can only provide the healthy group with a low amount of information. From the perspective of screening high-risk group, the CAT-AUD achieved its intended goal. Specifically, the full-length item bank questionnaire and the CAT-AUD could provide accurate results for the AUD high-risk group and provide relatively inaccurate results for the healthy group. For the AUD high-risk group, the questionnaire could provide much more information than the healthy group. For instance, when the theta value was >−0.8, the corresponding SE was about <0.32 in Figure 3. When the mean and SD of theta values are fixed to 0 and 1 (Samejima, 1994), according to the relationship of reliability and SE, the corresponding reliability was above 0.9, which confirms that the CAT-AUD prefers to the AUD high-risk group. This parallels to the research in which items did not appear capable of precisely measuring subjects with low transmissible liability index (TLI) values (Kirisci et al., 2012). Because the fixed-length and the fixed SE roles were set in advance, the CAT-AUD could provide the reliability of about 0.9 in each test for participants. Hence, the bias problem which happened in the full-length questionnaire was attenuated but not completely avoided in the CAT-AUD. This situation indicates that the CAT-AUD was almost equivalent to the full-length questionnaire in terms of the test accuracy. As for the test efficiency, the performance of the CAT-AUD was far more different from the full-length questionnaire. Participants who took part in the full-length questionnaire were requested to complete tedious 123 items. Nevertheless, unlike the full-length questionnaire, participants with various levels of AUD would just be provided a few or dozens of items that are suitable for their estimated theta level, and the test constructed by these items could excavate much valid IRT information for each individual. Accordingly, CAT cannot only provide accurate test results for AUD high-risk group but also greatly reduce the burden of all subjects (high-risk group and healthy group) in testing.
The advantages of the CAT-AUD could be also suggested from the perspective of validity. It can be obtained from Table 5 that the CAT-AUD was more consistent with the screening of the criterion scale, while the consistency of the AUDIT and AUDIT-C was weaker. Besides, the CAT-AUD also showed a higher predictive utility than the other two questionnaires. The AUC value was higher than that of the AUDIT and AUDIT-C, and its sensitivity is the highest among the three questionnaires. Accordingly, though the 10-item AUDIT and the 3-item AUDIT-C can noticeably save the test time for individuals, the test validity is not as good as that of the CAT-AUD. The CAT-AUD can narrow the length of the test based on ensuring the accuracy of the test, and the average number of items used is only 16. Therefore, it can be considered that the CAT-AUD can serve as a reliable and efficient test tool for research and screening.
Though the results of the CAT-AUD are promising, some limitations remained. First, the valid subjects from all over the country in this study were 915, which is only a moderate size sample. When the sample size was enlarged, the research conclusion of AUD might be modified. So the influence of sample size on test results remains to be further studied. Second, since the CAT-AUD cannot provide sufficient information for those participants with a low AUD level, items with different levels of threshold (especially those with low levels of a threshold) should be added. Only when the item bank covers the whole range of theta value, can the CAT program provide more suitable items for the subjects with different theta values and also assess the potential trait level of the subjects more accurately. Third, in a real scenario, the response of participants will be affected by numerous factors in that the simulation research with real data may differ from the real experiment in many aspects (Smits et al., 2011). According to the results of existing researches (Kocalevent et al., 2009), the performance of CAT simulation was consistent with real administration results, but it is uncertain that the same conclusion could be generalized to real working CAT-AUD system. Hence, the real CAT-AUD administration is worth applying in real life to explore the efficiency, accuracy and other characteristics of the CAT-AUD. Fourth, in the current study, a scale proven to have good reliability and validity rather than a professional clinical diagnosis from health care institutions was selected as the criterion in that way the AUD high-risk individual rather than the AUD patients would be screened out. As a consequence, the result of the CAT-AUD for individuals is highly correlated with the diagnostic result. Furthermore, though the item bank of this study complies with the standard of unidimensionality, it is a chance to consider conducting the above studies under the framework of Multi-dimensional IRT (MIRT) to report symptoms information of AUD defined in DSM-5, which may be a very noteworthy and exciting entry point for subsequent studies.
Data Availability Statement
The datasets generated for this study are available on request to the first author.
The studies involving human participants were reviewed and approved by the Research Center of Mental Health, Jiangxi Normal University. The patients/participants provided their written informed consent to participate in this study. Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this manuscript.
JL was responsible for the data processing and manuscript writing. DT and YC guided the process of the data processing and the manuscript writing. CX was responsible for the data collection. All authors contributed to the article and approved the submitted version.
This work was supported by the National Natural Science Foundation of China (31960186, 31760288, and 31660278).
Conflict of Interest
Anton, R. F., Moak, D. H., and Latham, P. (1995). The obsessive compulsive drinking scale: a self-rated instrument for the quantification of thoughts about alcohol and drinking behavior. Alcohol. Clin. Exp. Res. 19, 92–99. doi: 10.1111/j.1530-0277.1995.tb01475.x
Babor, T. F., and Grant, M. (1989). From clinical research to secondary prevention: international collaboration in the development of the alcohol disorders identification test (AUDIT). Alcohol Health Res. World 13, 371–375.
Bush, K., Kivlahan, D. R., McDonell, M. B., Fihn, S. D., and Bradley, K. A. (1998). The AUDIT alcohol consumption questions (AUDIT-C): an effective brief screening test for problem drinking. Arch. Inte. Med. 158, 1789–1795. doi: 10.1001/archinte.158.16.1789
Chang, G., Wang, P., Li, J., Xin, P., Wang, M., Li, J., et al. (2016). Investigation of drinking status in residents(15 years old) of urban and rural areas in Tianjin. Chin. J. Prevent. Control Chronic Dis. 24, 493–497.
Cheng, W. J., Huang, M. C., Huang, P. H., Gau, Y. F., and Chen, C. H. (2009). The Chinese version of the severity of alcohol dependence questionnaire: reliability and factor structure. Taiwanese J. Psychiatry 23, 159–166.
Choi, S. W., Gibbons, L. E., and Crane, P. K. (2011). lordif: an R Package for detecting differential item functioning using iterative hybrid ordinal logistic regression/item response theory and monte carlo simulations. J. Stat. Softw. 39, 1–30.
Crane, P. K., Gibbons, L. E., Jolley, L., and van Belle, G. (2006). Differential item functioning analysis with ordinal logistic regression techniques: DIFdetect and difwithpar. Med. Care 44(11 Suppl. 3), S115–S123. doi: 10.1097/01.mlr.0000245183.28384.ed
Culpepper, S. A. (2013). The reliability and precision of total scores and IRT estimates as a function of polytomous IRT parameters and latent trait distribution. Appl. Psychol. Meas. 37, 201–225. doi: 10.1177/0146621612470210
Flens, G., Smits, N., Terwee, C. B., Dekker, J., Huijbrechts, I., and de Beurs, E. (2017). Development of a computer adaptive test for depression based on the Dutch-Flemish version of the PROMIS Item Bank. Eval. Health Profess. 40, 79–105. doi: 10.1177/0163278716684168
Fliege, H., Becker, J., Walter, O. B., Bjorner, J. B., Klapp, B. F., and Rose, M. (2005). Development of a computer-adaptive test for depression (D-CAT). Qual. Life Res. 14:2277. doi: 10.1007/s11136-005-6651-9
Hasin, D. S., Stinson, F. S., Ogburn, E., and Grant, B. F. (2007). Prevalence, correlates, disability, and comorbidity of DSM-IV alcohol abuse and dependence in the united states: results from the national epidemiologic survey on alcohol and related conditions. Arch. Gen. Psychiatry 64, 830–842. doi: 10.1001/archpsyc.64.7.830
Hsueh, Y. J., Chu, H., Huang, C. C., Ou, K. L., Chen, C. H., and Chou, K. R. (2014). Psychometric properties of the Chinese version of the michigan alcoholism screening test (MAST-C) for patients with alcoholism. Perspect. Psychiatric Care 50, 83–92. doi: 10.1111/ppc.12022
Kirisci, L., Tarter, R., Reynolds, M., Ridenour, T., Stone, C., and Vanyukov, M. (2012). Computer adaptive testing of liability to addiction: Identifying individuals at risk. Drug Alcohol Depend. 123, S79–S86. doi: 10.1016/j.drugalcdep.2012.01.016
Klein, A. A., Stasiewicz, P. R., Koutsky, J. R., Bradizza, C. M., and Coffey, S. F. (2007). A psychometric evaluation of the approach and avoidance of alcohol questionnaire (aaaq) in alcohol dependent outpatients. J. Psychopathol. Behav. Assess. 29, 231–240. doi: 10.1007/s10862-007-9044-2
Kocalevent, R. D., Rose, M., Becker, J., Walter, O. B., Fliege, H., Bjorner, J. B., et al. (2009). An evaluation of patient-reported outcomes found computerized adaptive testing was efficient in assessing stress perception. J. Clin. Epidemiol. 62, 278–287. doi: 10.1016/j.jclinepi.2008.03.003
Li, K. Q., Cui, Z., Cui, L. J., Jiang, Q. P., Shi, G., Wu, H. R., et al. (2008). Epidemiological survey of mental disorders in the people aged 18 and older in Hebei Province. Asian J. Psychiatry 1, 51–55. doi: 10.1016/j.ajp.2008.09.010
Linden-Carmichael, A. N., Dziak, J. J., and Lanza, S. T. (2018). Dynamic features of problematic drinking: alcohol use disorder latent classes across ages 18–64. Alcohol Alcoholism 54, 97–103. doi: 10.1093/alcalc/agy074
McEvoy, P. M., Stritzke, W. G., French, D. J., Lang, A. R., and Ketterman, R. (2004). Comparison of three models of alcohol craving in young adults: a cross-validation. Addiction 99, 482–497. doi: 10.1111/j.1360-0443.2004.00714.x
Modell, J. G., Glaser, F. B., Cyr, L., and Mountz, J. M. (1992). Obsessive and compulsive characteristics of craving for alcohol in alcohol abuse and dependence. Alcohol. Clin. Exp. Res. 16, 272–274. doi: 10.1111/j.1530-0277.1992.tb01375.x
Orlando, M., and Thissen, D. (2003). Further investigation of the performance of s - x2: an item fit index for use with dichotomous item response theory models. Appl. Psychol. Meas. 27, 289–298. doi: 10.1177/0146621603027004004
Pan, G. W., Jiang, C., Yang, X. L., Na, J., Fu, Z. G., Yu, Q. F., et al. (2006). Epidemiological survey of mental disorders in urban and rural areas of Liaoning province. Chin. J. Public Health 22, 1505–1507.
Rehm, J., Mathers, C., Popova, S., Thavorncharoensap, M., Teerawattananon, Y., and Patra, J. (2009). Global burden of disease and injury and economic cost attributable to alcohol use and alcohol-use disorders. lancet 373, 2223–2233. doi: 10.1016/S0140-6736(09)60746-7
Saunders, J. B., Aasland, O. G., Babor, T. F., De la Fuente, J. R., and Grant, M. (1993). Development of the alcohol use disorder identification test (AUDIT): WHO collaborative project on early detection of persons with harmful alcohol consumption-II. Addiction 88, 791–804. doi: 10.1111/j.1360-0443.1993.tb02093.x
Schisterman, E. F., Perkins, N. J., Liu, A., and Bondell, H. (2005). Optimal cut-point and its corresponding Youden Index to discriminate individuals using pooled blood samples. Epidemiology 16, 73–81. doi: 10.1097/01.ede.0000147512.81966.ba
Shi, Q. C., Zhang, J. M., Xu, F. Z., Fei, L. P., Xu, Y., Fu, Y. L., et al. (2005). Epidemiological survey of mental illnesses in the people aged 15 and older in Zhejiang Province. Chin. J. Prev. Med. 39, 229–236.
Speckens, A. E. M., Heeren, T. J., and Rooijmans, H. G. M. (1991). Alcohol abuse among elderly patients in a general hospital as identified by the munich alcoholism test. Acta Psychiatr. Scand. 83, 460–462. doi: 10.1111/j.1600-0447.1991.tb05576.x
Spiegelhalter, D. J., Best, N. G., Carlin, B. P., and van der Linde, A. (1998). Bayesian Deviance, the Effective Number of Parameters, and the Comparison of Arbitrarily Complex Models. Technical Report, 98–009. Division of Biostatistics, University of Minnesota. Cambridge, UK: Medical Research Council Biostatistics Unit.
Stockwell, T., Hodgson, R., Edwards, G., Taylor, C., and Rankin, H. (1979). The development of a questionnaire to measure severity of alcohol dependence. Br. J. Addict. Alcohol Other Drugs 74, 79–87. doi: 10.1111/j.1360-0443.1979.tb02415.x
Stockwell, T., Murphy, D., and Hodgson, R. (1983). The severity of alcohol dependence questionnaire: its use, reliability and validity. Br. J. Addict. 78, 145–155. doi: 10.1111/j.1360-0443.1983.tb05502.x
Zhang, C., Yang, G. P., Li, Z., Li, X. N., Li, Y., Hu, J., et al. (2017). Reliability and validity of the Chinese version on alcohol use disorder Identification Test. Chin. J. Epidemiol. 38, 1064–1067. doi: 10.3760/cma.j.issn.0254-6450.2017.08.013
Zhang, J. X., Lu, C. H., Tang, J. S., Qiu, H. M., Liu, L. F., Wang, S. B., et al. (2010). Epidemiological survey of mental disorders in persons aged 18 years and older in shandong province. Chin. Ment. Health J. 24, 161–167.
Zumbo, B. D. (1999). A Handbook on the Theory and Methods of Differential Item Functioning (dif): Logistic Regression Modeling as a Unitary Framework for Binary and Likert-Type (ordinal) Item Scores. Ottawa: National Defense Headquarters.
Keywords: alcohol use disorder, item bank, item response theory, computerized adaptive testing, psychometric measurement
Citation: Lian J, Cai Y, Tu D and Xi C (2020) Developing and Validating an Item Bank for Alcohol Use Disorder Screening in the Chinese Population by Using the Computerized Adaptive Testing. Front. Psychol. 11:1652. doi: 10.3389/fpsyg.2020.01652
Received: 11 January 2020; Accepted: 17 June 2020;
Published: 08 July 2020.
Edited by:Caterina Primi, University of Florence, Italy
Reviewed by:Caterina Galandra, Istituti Clinici Scientifici Maugeri (ICS Maugeri), Italy
Mark D. Reckase, Michigan State University, United States
Copyright © 2020 Lian, Cai, Tu and Xi. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.