Sex differences in the latent class structure of alcohol use disorder: Does (dis)aggregation of indicators matter?

Introduction

Among the research and practice communities there are disagreements as to how many diagnostic categories are needed to fully characterize the true nature of some psychological disorders, or whether a dimensional structure is a more valid representation of the data (Hasin et al, 2013). This is an important consideration in the determination of the public health burden of a disorder (Agrawal, Heath, & Lynskey, 2011). Alcohol Use Disorder (AUD), specifically, has moved from a three-categorical structure in the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (no diagnosis, alcohol abuse, and alcohol dependence; DSM-IV) to a four-categorical, severity-graded structure in the DSM-Fifth Edition (no diagnosis, mild, moderate, and severe AUD; DSM-5, American Psychiatric Association, 2013).

Further complicating matters, population subgroups can have alternative presentations of AUD. Many researchers have noted sex differences in the presentation of AUD, in addition to differential social consequences, comorbidity, and biological course, resulting in difficulties in AUD assessment (Babor et al, 1992; Brady & Randall, 1999; Bucholz et al, 1996; Saha, Chou, & Grant, 2006; Winokur, Rimmer, & Reich, 1971; Schuckit et al, 1969). Nevertheless, researchers attempting to model the diagnostic structure of AUD frequently combine men and women and model diagnostic groups with latent class analysis (LCA). There have been several class structures represented in the literature (see Table 1). Some studies support the DSM-5 severity-graded class structure (Beseler et al, 2012; Chung et al, 2001; Sacco et al, 2009), others support a 4-6 class structure of groups which differ in the symptoms or criteria they tend to endorse (Kendler et al, 1998; Moss et al, 2007; Rist et al, 2009; Smith & Shevlin, 2008), while other research indicates that AUD lies on a latent dimension of severity (Heath et al, 1994). Very few studies include parallel analyses for men and women (Bucholz et al, 1996), and some examine only men (Heath et al, 1994; Kendler et al, 1998).

Table 1.

LCAs of AUD

Study	Year	# Classes	Input Variables	Sample	Gender
Bucholz et al	1996	4	27 Symptoms of AUD	Relatives of alcoholics	Separated
Beseler et al	2012	3	AUD diagnostic criteria	Undergraduates	Combined
Chassin et al	2004	3	Alcohol and drug use	COAs1, matched non-COAs	Combined
Chung et al	2001	3	AUD diagnostic criteria	Adolescent clinical	Combined
Heath et al	1994	5	Symptoms of alcoholism	Twins, Australia	Men Only
Kendler et al	1998	5	Reasons for temperance board registration, age at onset	Twins, clinical	Men Only
Moss et al	2007	5	AUD diagnostic criteria and lifetime psychiatric diagnoses	Clinical	Combined
Reboussin et al	2006	2	Drinking behaviors and alcohol-related problems	Underage, community	Combined
Rist et al	2009	4	Alcohol consumption	Clinical	Combined
Sacco et al	2009	3	Alcohol consumption and AUD diagnostic indicators	Older individuals, NESARC2	Combined
Smith et al	2008	6	Alcohol consumption and related problems	Population representative, UK	Combined
Whitesell et al	2006	4(lifetime) 3(past year)	Alcohol and drug use	Native American and Population- Representative	Combined

¹Children of Alcoholics

²National Epidemiologic Survey of Alcohol and Related Conditions Table 2

The LCAs of AUD conducted to date have also varied in that they use different sets of observed indicators of the latent classes. In some studies, symptom items, meant to describe the criteria in more layman’s terms, are used as input variables (Bucholz et al, 1996; Heath et al, 1994), others use an aggregated distillation of the symptoms into each of the diagnostic criteria (Beseler et al, 2012; Chung et al, 2001; Moss et al, 2007; Sacco et al, 2009), while other studies have used more inclusive sets of indicators including (noncriteria) drinking consequences, and various measures of alcohol consumption (Chassin et al, 2004; Smith et al, 2008; Whitesell et al, 2006). Considering the sensitivity of LCA to the number and quality of indicators (Nylund, Asparouhov, & Muthen, 2007; Raftery & Dean, 2006; Swanson, Lindenberg, Bauer & Crosby, 2012), differential data structure may be partially explaining differences in class structure of AUD reported in the literature. To clarify these issues, this paper examines (1) the latent class structure of AUD, (2) sex differences in this structure, and (3) and aggregating data from item-level symptoms to diagnostic criteria impacts the analysis, and whether this differs between men and women.

Reduction of the symptom items into criteria has been largely regarded as an arbitrary decision. Although the third column in Table 1 demonstrates that class structure of AUD has varied based on whether symptoms or criteria were used this cannot be determined to be the sole cause for differential results, as these studies employed different samples (some combining men and women together), as well as inclusion of covariates and other predictors. In fact, other analyses have shown concordant results between symptom and criteria data. Replication of a structural equation modeling technique with symptom items has been examined by Harford and Muthen (2000). However, a direct comparison of LCAs with symptom and criteria items has not been undertaken, to our knowledge. As was discussed above, LCA is a technique particularly susceptible to the number and quality of inputted indicators, so it is likely that class structure is significantly impacted by data aggregation. Separating the effect of data aggregation from the effect of gender differences is impossible by just examination of the literature, this paper will conduct parallel analyses, examining the effect of sample and data structures separately.

Latent Class Analysis

Latent class analysis (LCA) is special case of the finite mixture model and a common tool for characterizing a latent construct measured by categorical indicators. LCA models a number of unobserved subpopulations to identify categories of a construct (also referred to as groups or classes), with the goal of fully explaining all the variance in the data. LCA operates by iteratively finding maximum likelihood parameter estimates of the population proportion and the “mean profile” for each cluster (i.e., the probability of endorsing each indicator). Classes in an LCA are assumed to be internally homogeneous, meaning every individual within a class has an identical set of item endorsement probabilities. Classes are also assumed to be locally independent, meaning the indicators do not covary conditional on class membership (for more detail on the computation of LCAs, see Bartholomew, Knott, & Moustaki, 2011). Researchers have examined plots of the mean profiles to determine whether a construct exhibits severity-distinguished structure, combined with incrementally better-fitting solutions as the number of classes increases. Such behavior has been found in LCAs of AUD, and some have thus argued for a dimensional conceptualization of the disorder (Morey, Skinner & Blashfeld, 1984; Bucholz et al, 1996). The following analyses will help determine if these conclusions were partly based on the sex differences and data aggregation described in the following section.

Data Aggregation

LCA has been shown to provide more accurate results when more indicators are included in the analysis (Albert & Dodd, 2004; Brusco, 2004; Brusco & Cradit, 2001; Milligan, 1989; Yang, 2006). Increasing the number of indicators improves the results because the multivariate distribution of the cluster structure becomes clearer, making the clusters more separated. When symptom data are aggregated to form diagnostic criteria, this removes a large number of potentially informative indicators of the latent structure of the construct (e.g., AUD). Thus, a person centered technique like LCA may be optimally suited with a larger number of indicators without data aggregation.

However, additional indicators will diminish performance of the LCA when those indicators are non-informative with regard to the individuals’ group membership (Brusco, 2004; Raftery & Dean, 2006; Swanson, Lindenberg, Bauer, & Crosby, 2012). These “masking” variables (Fowlkes & Mallows, 1986), can obscure class structure that would otherwise be found by introducing noise in the multivariate distribution of the data.

Using LCA as an exploratory technique (i.e., when the number of clusters and their structure is unknown) means that it is impossible to definitively determine whether the inclusion of variables is helping or hurting the accuracy of the solution. We can, however, examine the results between analyses using more or fewer indicators (disaggregated versus aggregated data) for differences, to decide what substantive information is lost or gained between the two approaches. If the aggregation from symptoms to criteria is benign, the same structure should result in both.

Method

Analyses used the National Epidemiologic Survey of Alcohol and Related Conditions (NESARC; Grant, Dawson, Stinson, Chou, Dufour, & Pickering, 2004). NESARC was collected by the US Bureau of the Census and the National Institute on Alcohol Abuse and Alcoholism, and assessed a very large, nationally representative sample via in-person interviews, the first wave being collected in 2001-2002, the second wave in 2004-2005, with a cross-sectional sample collected in 2012-2013. NESARC is racially diverse, with an oversampling of Blacks and Hispanics (25% Hispanic, 19% Black), as well as geographically diverse (using all US census regions; for more detail on sampling and interview procedures see http://niaaa.census.gov/). Analyses used Wave 2 data and excluded those who reported abstaining from alcohol use in the past year (28.61% of men, 41.35% of women), resulting in a final sample size of N = 22177, 10395 men and 11782 women. Observations were weighted in analyses to obtain estimates representative of the U.S. population.

Measures

DSM-5 (American Psychiatric Association, 2013) criteria items are assessed with the Alcohol Use Disorders and Associated Disabilities Interview Schedule (AUDADIS-IV; Grant et al, 2001). What will be referred to as the “disaggregated” data is the set of 34 items which are used in the AUDADIS algorithm to determine whether the respondent exhibits a particular DSM-5 AUD criterion (this excludes any items unrelated to the “legal consequences” DSM-IV criterion or any other construct unrelated to DSM-5 AUD). A description of these items, the criteria they are used to measure, and endorsement rates for the analysis sample are given in Tables 2 and 3. What will be referred to as the “aggregated” data is the reduced set of 11 diagnostic criteria of AUD created by an individual’s endorsement of an adequate number of criteria-related symptoms. Per the AUDADIS-IV algorithm, all criteria except withdrawal are coded as positive if the individual endorses at least one of the criteria-related symptoms. The withdrawal criterion is coded positive if the individual exhibits all of the first eight withdrawal-related symptoms shown in Table 2, or at least one of the last two. If these are not satisfied, the individual does not meet the criterion. The diagnostic criteria, their descriptions, and the endorsement rates for the samples considered in the following analyses are given in Table 4.

Table 2.

DSM-5 Symptom Items

	Abbreviation	Description	% Men Endorsed	% Women Endorsed
Tolerance	LESSEFFECT	Find that your usual number of drinks had much less effect on you than it once did	0.069	0.049
	MUCHMORE	Find that you had to drink much more than you once did to get the effect you wanted	0.038	0.023
	FIFTH	Drink as much as a fifth of liquor in one day	0.045	0.007
	INCREASE	Increase your drinking because the amount you used to drink didnt give you the same effect anymore	0.023	0.014
Quit	STOP	More than once want to stop or cut down on your drinking	0.160	0.091
	TRY	More than once try to stop or cut down on your drinking but found you couldnt do it	0.038	0.017
Larger Longer	LARGER	Have a period when you ended up drinking more than you meant to	0.149	0.092
	LONGER	Have a period when you kept on drinking for longer than you had intended to	0.121	0.062
Withdrawal	ASLEEP	Have trouble falling asleep or staying asleep (when the effects of alcohol were wearing off)	0.056	0.041
	SHAKING	Find yourself shaking (when the effects of alcohol were wearing off)	0.020	0.010
	ANXIOUS	Feel anxious or nervous (when the effects of alcohol were wearing off)	0.033	0.015
	SICK	Feel sick to your stomach or vomit (when the effects of alcohol were wearing off)	0.085	0.072
	RESTLESS	Feel more restless than is usual for you (when the effects of alcohol were wearing off)	0.049	0.031
	SWEAT	Find yourself sweating or your heart beating fact (when the effects of alcohol were wearing off)	0.044	0.024
	HALLUCINATE	See, feel, or hear things that werent really there (when the effects of alcohol were wearing off)	0.007	0.003
	SEIZURES	Have fits or seizures (when the effects of alcohol were wearing off)	0.001	0.001
	DRUGOVER	Take a drink or use any drug or medicine, other than aspirin, Advil, or Tylenol to get over any of the bad effects of drinking	0.034	0.020
	DRUGKEEP	Take a drink or use any drug or medicine, other than aspirin, Advil, or Tylenol to keep from having any of these bad aftereffects of drinking	0.023	0.014
Time spent	TIMEDRINK	Have a period when you spent a lot of time drinking	0.036	0.015
	TIMESICK	Have a period when you spent a lot of time being sick or getting over the bad aftereffects of drinking	0.012	0.007
Give up	IMPORTANT	Give up or cut down on activities that were important to you in order to drink, like work, school, or associating with friends or relatives	0.011	0.004
	PLEASURE	Give up or cut down on activities that you were interested in or that gave you pleasure in order to drink	0.011	0.004
Continued	DEPRESSED	Continue to drink even though you knew it was making you feel depressed, uninterested in things, or suspicious or distrustful of other people	0.027	0.013
	HEALTH	Continue to drink even though you knew it was causing you a health problem or making a health problem worse	0.047	0.020
	BLACKOUT	Continue to drink even though you had experienced a prior blackout, that is, awakened the next day not being able to remember some of the things you did while drinking or after drinking	0.032	0.014
Social problems	FAMILY	Have a period when your drinking or being sick from drinking often interfered with taking care of your home or family	0.011	0.006
	SCHOOL	Have job or school troubles because of your drinking or being sick from drinking, like missing too much work, not doing your work well, being demoted or losing a job, or being suspended, expelled, or dropping out of school	0.007	0.002

Table 3.

DSM-5 Symptom Items, Continued

	Abbreviation	Description	% Men Endorsed	% Women Endorsed
Hazard	DRIVEWHILE	More than once drive a car or other vehicle while you were drinking	0.107	0.040
	DRIVEAFTER	More than once drive a car, motorcycle, truck, boat, or other vehicle after having too much to drink	0.036	0.011
	HURT	Get into situations while drinking or after drinking that increased your chances of getting hurt, like swimming, using machinery, or walking in a dangerous area or around heavy traffic	0.026	0.009
Social	FAMILYFRIENDS	Continue to drink even though you knew it was causing you trouble with your family or friends	0.017	0.005
	FIGHT	Get into physical fights while drinking or right after drinking	0.012	0.003
Craving	BADLY	Want to drink so badly couldn’t think of anything else	0.012	0.005
	DESIRE	Feel very strong desire to drink	0.054	0.028

Table 4.

DSM-5 AUD Criteria

Abbreviation	Description	% Men Endorsed	% Women Endorsed
TOL	Tolerance of alcohol	0.111	0.060
WITHD	Withdrawal after using alcohol	0.093	0.065
LARGERLONGER	Substance used more than intended	0.176	0.105
CUTD	Desire or efforts to cut down on use	0.165	0.093
TIME	Spent all day using substance or recovering from effects	0.040	0.018
GIVEUP	Important activities given up because of use	0.015	0.005
CONTINUE	Continued use despite health problems	0.072	0.033
ROLE	Failure to fulfill role obligations	0.015	0.007
HAZARD	Use in physically hazardous situations	0.134	0.051
SOCIAL	Use despite social/interpersonal problems	0.025	0.008
CRAVING	Craving or a strong desire or urge to use alcohol	0.056	0.029

Analysis

The following series of LCAs were conducted in Mplus Version 7 (Muthen & Muthen, 2012). Each LCA was initialized 50 times (with 50 iterations for the initial stage of estimation). LCAs were fit with 1-10 classes and the solutions were evaluated for (1) model fit, (2) substantive quality, and (3) classification agreement. The Bayesian Information Criterion (BIC; Schwarz, 1978) was used as a measure of model fit, which selects a model to maximize the probability of observing the data given the model, while controlling for the capitalization on chance by penalizing the likelihood function using the number of parameters in the LCA. The BIC has shown to be adequate in the selection of the correct number of clusters in an LCA (Lin & Dayton, 1997; Nylund, Asparouhov, & Muthen, 2007). In addition to the BIC, the plots of the mean profiles were created to consider the substantive quality of the solution. The structure of these profiles can indicate whether the classes are distinguished by increasing probabilities of item endorsement, meaning that the classes only differ in their severity of AUD (Ferguson, 1983). To determine whether differential fit led to a highly differentiated solution, we used the Adjusted Rand Index (ARI; Hubert & Arabie, 1985) to assess classification agreement between alternative models and data aggregation levels. The ARI is a measure of cluster solution agreement that can be used to compare the results of cluster solutions with different numbers of clusters by analyzing the number of pairs clustered together in both solutions, apart in both solutions, and the number of times pairs of individuals are discordantly clustered (different clusters in one, same in the other, or vice versa). There are established conventions for the ARI that denote the adequacy of the agreement of the solutions at different magnitudes of the ARI. They are: > 0.9 indicates “excellent” agreement, > 0.8 “good”, > .65 “moderate”, and < .65 is considered “poor” agreement between the two solutions (Steinley, 2004).

Results and Discussion

Table 4 shows the BICs by aggregation level and sex inclusion. Perhaps most striking about the results is that many more classes are found for the data containing disaggregated symptoms, and the number of classes found is not consistent across sex inclusion (9 classes for men, 6 classes for women, and 9 classes for sexes combined). When the data are aggregated into diagnostic criteria, the class structure simplifies into consistent, severity-graded classes (4 classes for men, women, and men and women combined). The 4-class structure in the aggregated criteria largely resembles the diagnostic structure in the DSM-5 (i.e., no diagnosis, mild, moderate, and severe).

An examination of the mean profile plots in Figures 1 and 2 highlights the nature of the differences in results in LCAs based on aggregated (i.e., criterion) versus disaggregated (i.e., specific symptoms). Most importantly, the classes in the disaggregated data are not distinguished by only increasing probabilities of item endorsement. The “crossover” of classes is not seen in the mean profile plots of the aggregated data, whereas there are several instances of this in the disaggregated symptoms (e.g., LESSEFFECT, MUCHMORE, among others)¹. When the data are aggregated, individuals in “high tolerance” classes (Class 7 for men and Class 6 for women, distinguished by having a high probability of endorsing one or more tolerance items but little other symptoms) are classified differentially. 71% of women in the high tolerance class are categorized into Class 4 when data are aggregated, which is the lowest class (if we are using the parlance of the DSM-5, this would be the class with no diagnosis), with the next largest proportion allocated into Class 2, the moderately high class (26%). With men, most in the “high tolerance” class are classified into the low group as well (61%), with the next highest proportion being categorized into Class 1, the “high” class (25%). Note, however, these alternatively classified individuals represented a small proportion of the entire sample and the classification agreement between the aggregated and disaggregated data is still in the adequate range (ARI = .78 for men, .80 for women, and .81 for men and women combined).

Figure 1.

Mean Profile Plots for Disaggregated Data

Figure 2.

Mean Profile Plots for Aggregated Data

In general, the larger number of classes derived from disaggregated data serve to resolve intermediate levels of severity not captured in the solutions based on the aggregated data. In addition, in the solutions based on the disaggregated data, there are instances where some symptoms of a criterion appear to diverge in endorsement likelihood from other symptoms of that criterion. Whether or not such differences in overall severity or in symptom configuration potentially provide a foundation for improving clinical diagnosis is an empirical issue that rests upon further construct validation. In the absence of such validation, the clinical utility of the more differentiated solutions must be viewed cautiously owing to both the difficulty of working with so many diagnostic subtypes in a clinical settings and the very low base rates of some classes. Nevertheless, it could be that these more refined classes do reveal important etiological or clinical heterogeneity and, consequently, warrant further investigation.

The results also show that combining men and women has a differential effect based on whether the data are aggregated. When the data are disaggregated, combining genders led to two classes (instead of one) which exhibit low probabilities on most symptoms with the exception of tolerance items (Classes 4 and 8). However, women largely maintain their classifications–only being classified into 6 of the 9 classes (the remaining 3 classes are comprised entirely of men). When the data are aggregated, the classes look almost identical between the data with men, women, and the sexes combined. Examining the classifications between these solutions, however, shows that even though the mean profiles look nearly identical, when men and women are combined, women are pushed into lower severity classes (10% of women are alternatively classified when men are included in the sample).

These results suggest that the impact of aggregating symptoms into criteria removes a group for whom their only AUD symptoms are tolerance-related. These “high tolerance” individuals have heterogeneous classifications when data are aggregated. Some are classified into a class receiving a DSM-5 AUD diagnosis (39% of high tolerance men are in a class which would receive a diagnosis, 29% of high tolerance women). When men and women are combined into a single dataset, men maintain their class assignment, but women are typically alternatively classified, some remaining in the highest severity class, some moved into a lower severity class.

Post Hoc Analysis: Split-Half Replication

The cluster solutions were examined with split-half replication. That is, the data sets were each split in half randomly, and the same procedures employed above were carried out with both random halves. The BICs for the disaggregated symptoms are presented in Table 5 and the aggregated criteria in Table 6.

Table 5.

BICs by Sex and Data Aggregation Level

Classes	DisagB	DisagM	DisagW	AgB	AgM	AgW
1	195078	113802	78606	104091	59898	42518
2	151201	89314	61094	83911	48687	34566
3	143504	84664	58489	81604	47340	33788
4	142016	83854	58048	81341	47198	33775
5	140921	83360	57758	81342	47239	33809
6	140243	83166	57713	81389	47314	33878
7	140149	83106	57771	81467	47385	33948
8	149980	83099	57902	81545	47461	34036
9	139907	83087	58057	81633	47535	34121
10	139922	83170	58250	81717	47624	34217

B - Both men and women, M - Men only, W - Women only

Table 6.

Split Half BICs - Disaggregated Symptoms

Classes	Men0	Men1	Men2	Women0	Women1	Women2	Both0	Both1	Both2
1	113802	63467	59454	78606	44599	41462	195078	96929	98415
2	89314	50264	47517	61094	34501	33246	151201	75690	76027
3	84664	47706	45083	58489	32306	32388	143504	72218	72087
4	83854	46240	44556	58048	31216	32418	142016	71568	71514
5	83360	46204	44559	57758	31409	32520	140921	71191	71052
6	83166	46338	44654	57713	31669	32827	140243	71035	70953
7	83106	46622	44944	57771	32061	33225	140149	71039	70913
8	83099	46973	45369	57902	32524	33647	149980	71059	70979
9	83087	47329	45690	58057	32984	34190	139907	71176	71032
10	83170	47765	46172	58250	33437	34601	139922	71302	71147

0 - Full sample, 1 - First split half, 2 - Second split half DNC - Did Not Converge

The most striking observation from this analysis is that the cluster structure is not consistent between the two split halves for men, women, or the two combined when the data are the disaggregated symptoms. With the aggregated criteria, the best-fitting number of clusters was consistent between halves for men, and men and women combined, but not for women alone. A potential explanation for the lack of split-half replication in the aggregated data containing women could be related to the lower prevalences of the criteria in this group. When the data are split in half, the most severe class for women (comprising only .7% of the full sample) is so small as to be nearly nonexistent in the split halves, resulting in less classes being necessary for the full description of the data. The split half solutions were then combined into a single dataset, and the classifications compared using the ARI between the solutions obtained via split half and the combined data (see Analysis section).

The results of this ARI comparison show another instance where the combining of men and women leads to different conclusions about the impact of aggregating data. Specifically, when the sexes are analyzed separately, the split halves have much more consistent structure than when the sexes are combined. When data were aggregated, even though these were all four-cluster solutions, the split halves of the data with sexes combined had poor classification consistency (ARI = .33). When the sexes were separated, however, agreements were good between the split halves (.84 for men, .74 for women). When data are disaggregated symptoms, no solution had highly consistent classifications (ARI = .37 for men, .27 for women, .37 for men and women combined).

Conclusion

The best fitting class structure when analyzing AUD criteria was found to be a four-category, severity-graded solution. However, very different solutions were found when analyzing disaggregated symptoms, and these structures varied between males and females. Although aggregating the data provided an arguably “cleaner” structure of AUD with a smaller number of easily interpretable classes, these findings of severity-graded classes are of limited theoretical interest in that they suggest subtypes that tend to differ more “quantitatively” (i.e., by severity) than “qualitatively” (i.e., by unique configurations of indicators). Moreover, there were patterns of AUD presentation of theoretical and clinical interest that were lost when the symptoms are aggregated to criteria. Specifically, a class of individuals in both men and women arises for which there is little symptomatology but elevated tolerance (i.e., a high probability to state that they have diminished effects from their usual number of drinks and that they have to drink much more to experience the same effect). In analyses of the aggregated data, these individuals are largely categorized into the “no diagnosis” group, for which the probability of endorsing the tolerance criterion for the class is less than .05. To the extent that such individuals have different courses or complications than those without other symptoms, obscuring this group represents a loss of important etiological information. While high tolerance by itself may not seem like a problematic phenotype, it could presage a greater likelihood of progression of AUD and/or increased likelihood of medical or psychiatric complications associated with heavy consumption.

The results from the aggregated data appear to map on to the diagnostic structure of the DSM-5 (that is, 4 clusters separated by an increasing probability of endorsing the criteria). The stability of this cluster solution, however, only exists when men and women are analyzed separately. This could be due to the fact that high-severity women are very rare (.7% of the full sample), leading to fewer classes which are needed to describe the lower variance. Further research should examine these symptom items for whether or not they are masking cluster structure (Steinley & Brusco, 2008). This could serve to identify subsets of symptoms that produce a more consistent cluster structure thereby aiding in the reproducibility of health-related research (Collinsa & Tabak, 2014).

Additionally, there is a lack of general consensus of how to “lump” or “split” criteria across different diagnostic systems. For example, the two DSM-IV criteria of TIMESPENT and GIVEUP are combined into a single criterion in the International Classification of Diseases - 10^th Edition (ICD-10; World Health Organization, 1992) and it appears that a similar “lumping” of the DSM LARGERLONGER and CUTD occurs as well. Regardless of how they are operationalized, it is not clear how “narrow” or “broad” each diagnostic criterion should be. Although systematically explored in the current studies, this basic issue could also materially affect the nature of solutions obtained in the LCA. Thus, resolution of the issues raised here go beyond whether to aggregate symptom level data into criteria but also whether to aggregate or disaggregate criteria into small or larger criteria sets. The type of diagnostic research we are proposing involves numerous types of permutations in order to more fully understand the conditional nature of findings and point the way to the most robust and meaningful typologies.

Recommendations

Results from the above analyses lend themselves to recommendations improving the state of the art of LCA in research on AUD and related topics. As is documented in basic research, the number and type (i.e., whether an indicator is “masking” or not) of input indicators in an LCA is pivotal in finding accurate results. However, applied researchers do not know the optimal set of indicators, so in this situation it seems sensible to explore a wide set of solutions, including several levels of data aggregation and examining their correspondence with external covariates to obtain construct validity. Although researchers often use aggregated sets of indicators, the aggregation strategy (as it relates to data reduction) is often arbitrary and does not necessarily “carve nature at her joints”. In fact, it has been shown that sum scores, as well as other approaches to linearly combining variables, are likely to degrade the ability to uncover cluster structure, as compared to a more thoughtful approach to assessing the importance of individual indicators (Steinley, Brusco, & Henson, 2012). Considering the sensitivity of LCA to the number of indicators, shown in a line of methodological research and supported by the results in this paper, it is recommended that data are not aggregated prior to analysis.

Limitations

The data used for this analyses, although providing a very large data set and a diverse sample, is an older dataset (being collected between 2004 and 2005). This renders the data susceptible to period and cohort effects (see Kerr et al, 2004). Results should be replicated with newer data, including the recently released NESARC III (http://www.niaaa.nih.gov/research/nesarc-iii) data set which provides a more contemporary survey of AUD symptoms in the general population.

Note that consistent with drinking patterns of men and women in the United States, NESARC Wave II contains many more female abstainers than male abstainers (41.35% of women had abstained from alcohol in the last 12 months before interview, 28.61% of men). Abstainers represent a type of missing in that were they to drink, we might observe drinking patterns not manifested in our current sample of drinkers. Future research could examine the same issues in cultures where the rates of abstinence are both more similar between men and women and also lower than in the United States (e.g., Australia, see Wilsnack et al, 2000; Wilsnack et al, 2009). Replication in such a population would suggest that the findings presented here are not conditioned upon censoring occasioned by differentially abstinence rates.

Additionally, we did not examine other commonly used measures of AUD such as the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I; First et al, 2002) or the Semi-Structured Assessment for the Genetics of Alcoholism (SSAGA-II; Bucholz et al, 1994). These alternative interviews use different items and aggregation methods to determine whether individuals exhibit one or more diagnosis criteria. Similar analyses should be conducted with these and other alternative measures to make a more general statement on the effect of aggregating AUD symptom data to criteria.

We did not attempt to determine the possible reasons for the differences between men and women. This phenomenon may exist across other demographic groups. For example, studies of AUD using older individuals have found different latent variable models than those using a younger sample within the same data set (Sacco et al, 2009; Saha et al, 2007), these differences could be an artifact of combining these population subgroups. Future research should inspect the effect of aggregation within these alternative groupings as well.

Finally, LCA is a widely used technique to model AUD, but techniques such as Factor Mixture Modeling and Mixtures of Factor Analyzers have become more popular in recent years, combining the ideas of a dimensional and categorical structure of AUD (Lubke & Muthen, 2005; McLachlan & Peel, 2000). A similar treatment of these techniques to determine whether the structure of the data is unduly influencing these results would be a useful future direction for this line of research.

Table 7.

Split Half BICs - Aggregated Criteria

Classes	Men0	Men1	Men2	Women0	Women1	Women2	Both0	Both1	Both2
1	59898	30423	29547	42518	21303	21298	104091	51838	52338
2	48687	24679	24155	34566	17162	17555	83911	42003	42080
3	47340	24077	23486	33788	16755	17270	81604	41014	40860
4	47198	24060	23460	33775	16784	17316	81341	40937	40769
5	47239	24104	23530	33809	16850	17370	81342	40978	40813
6	47314	24177	23608	33878	16925	17440	81389	41048	40888
7	47385	24251	23684	33948	17006	17523	81467	41123	40967
8	47461	24326	23761	34036	17086	17598	81545	41205	41057
9	47535	24406	23836	34121	17177	17678	81633	41288	41149
10	47624	24498	23916	34217	17264	17760	81717	41362	41233

0 - Full sample, 1 - First split half, 2 - Second split half

DNC - Did Not Converge