Cognitive Performance Profiles by Latent Classes of Drug Use

Abstract

Background and Objectives

The relationship between substance use and cognitive deficits is complex and requires innovative methods to enhance understanding. The present study is the first to use LCA to examine associations of drug use patterns with cognitive performance.

Methods

Cocaine/heroin users (N=552) completed questionnaires, and cognitive measures. LCA identified classes based on past-month drug use and adjusted for probabilities of group membership when examining cognitive performance. Latent indicators were: alcohol (ALC), cigarettes (CIG), marijuana (MJ), crack smoking (CS), nasal heroin (NH), injection cocaine (IC), injection heroin (IH), and injection speedball (IS). Age and education were included as covariates in model creation.

Results

Bootstrap Likelihood Ratio Test (BLRT) supported a 5-class model. Prevalent indicators (estimated probability of over 50%) for each class are as follows: “Older Nasal Heroin/Crack Smokers” (ONH/CS, n=166.9): ALC, CIG, NH, CS; “Older, Less Educated Polysubstance” (OLEP, n=54.8): ALC, CIG, CS, IH, IC, and IS; “Younger Multi-Injectors” (MI, n=128.7): ALC, CIG, MJ, IH, IC, and IS; “Less Educated Heroin Injectors” (LEHI, n=87.4): CIG, IH; and “More Educated Nasal Heroin” users (MENH, n = ALC, CIG, NH). In general, all classes performed worse than established norms and older, less educated classes performed worse, with the exception that MENH demonstrated worse cognitive flexibility than YMI.

Discussion and Conclusions

This study demonstrated novel applications of a methodology for examining complicated relationships between polysubstance use and cognitive performance.

Scientific Significance

Education and/or nasal heroin use are associated with reduced cognitive flexibility in this sample of inner city drug users.

INTRODUCTION

The relationship between substance use and cognitive deficits is complex and requires innovative methods to enhance understanding. Cognitive deficits, particularly disinhibition, may lead to the development of substance dependence [1, 2]. Further, drugs themselves can cause cognitive deficits. This contributes to the chronic nature of substance dependence [3-6]. Most evidence suggests associations with neuropsychological deficits for both cocaine dependence [7, 8] and opioid dependence [9-13]. However, due to the high prevalence of polydrug use among cocaine and heroin users [14], it is difficult to discriminate which drugs may be most associated with the deficits.

Previous research has attempted to examine deficits associated with polydrug use by various methodological means. Ornstein and colleagues examined set-shifting, i.e., shifting between relevant/irrelevant stimuli, among non-drug users compared to amphetamine users and heroin users after withdrawal symptoms had been resolved [13]. This intra/extra-dimensional set-shift is similar to the set-shifting in the Wisconsin Card Sorting Task. Though both groups demonstrated memory impaiments, heroin users performed worse in identifying the initial relevant stimuli, while amphetamine users demonstrated deficits with relevant stimuli [13].

Verdejo-Garcia and colleagues examined groups of polysubstance cocaine and opioid users on tasks measuring cognitive impulsivity. Though polysubstance use among opioid users was greater than cocaine users, cocaine users performed significantly worse on selective attention, impulsivity, response inhibition, and cognitive flexibility [15, 16]. Both groups of polysubstance users performed worse than controls on decision-making, impulsivity, working memory and reasoning. This is of particular concern as impulsivity and disinhibition may lead to more severe substance use and possible substance dependence [1, 2]. However, it is difficult to discern which patterns of substance use are associated with which deficits.

In addition, most research ignores drug routes of administration (ROAs), e.g., smoking, nasal, injection, despite associations with unique health consequences for different ROAs [17-19]. For example, smoking is linked with greater medical complaints, higher mortality rates, and greater risk of injection drug use (IDU) [19-23]. Smoking crack cocaine, as opposed to nasal use, is associated with increased risk of dependence [24]. Injection is associated with increased risk for drug dependence [25, 26], overdose [27], and HIV and hepatitis C infection [28, 29].

Although research in this area is limited, there are reasons to believe ROAs may be differentially associated with cognitive functioning for both cocaine and heroin. Injection or smoking produces larger highs and greater pain reduction than nasal or oral use [30-32], suggesting that cognitive performance may be differentially impacted as well. Further, given that injection is associated with well-known negative consequences, injection drug users may be particularly impulsive and severe [33, 34]. Indeed, academic failure is associated with heroin injection severity [35] and impaired planning ability is associated with increased cocaine and heroin injection practices [36]. Thus, there is some limited evidence to suggest cognitive functioning and ROAs are differentially associated, but further research is needed.

Very little research has used latent variables to examine complicated questions regarding drug use and cognitive performance. Latent modeling – a technique that infers unobserved, “latent” variables from observed variables –is a promising tool to assess cognitive performance and enhance understanding of effects of multiple drug use. For example, item-centered analyses, such as Principal Component Analysis (PCA) provided useful information on identifying cognitive domains that are damaged in substance abuse [16] and Exploratory Factor Analysis (EFA) provided information on beliefs about substance use and relationship to executive deficits [37]. However, only one paper used person-centered analyses, such as Latent Profile Analysis (LPA). This recent paper used LPA to examine cognitive performance profiles, finding three levels of intellectual and executive functioning, viz., impaired, intact, and high [38]. However, it appears no one has used person-centered analyses, such as Latent Class Analysis (LCA), to address associations between polydrug use and cognitive performance.. Effective use of LCA has allowed for furthering our understanding of the development of drug dependence [20, 39-46] . Although LCA has been used in increasing frequency, the present study is the first to use LCA to examine associations of drug use patterns with cognitive performance. More extreme drug use patterns (e.g., injection drug use, polysubstance use) are expected to be associated with poorer executive functioning.

MATERIALS AND METHODS

Participants

The present study sample consisted of 552 Black and White drug users in Baltimore, Maryland, from the NEURO-HIV Epidemiologic study [21, 47, 48]. Upon arrival, participants provided informed consent, blood and urine samples, completed the HIV-Risk Behavior Interview – a semi-structured interview about drug use and sexual practices – and performed cognitive assessments described below. The study was approved by the Institutional Review Board at the University of Florida.

Assessments

Latent class indicators included all routes of administration (ROAs) and substances used in the past month by 20% or more of the sample. Self-report results of prior 30-day drug use were consistent with urinalysis drug results [20]. The eight binary substance use indicators included in the analyses were: alcohol, cigarettes, smoking marijuana, smoking crack, snorting heroin, injecting heroin, injecting cocaine, and injecting “speedball” (cocaine and heroin mixed). In addition, since age and education appeared to be important variables (see Table 1), dichotomous median-split variables for age (33) and education (highest grade completed, 11) were included in model creation using the “c on” command in M-plus. Other variables included in separate analyses were race (Caucasian or African-American), gender, current participation in drug treatment, and lifetime history of regular drug use for all indicators except cigarette smoking.

Table 1.

Fit statistics and entropy for a latent class analysis of 8 substances used in the past month, controlling for age and education, among 549 users of cocaine and heroin in the past 6 months.

Classes	LLa	parameters	LMRb	BLRTc	s.c. r.f. (f)d	Entropy
1	−3482.92	12	NAe	NAe	NAe	NAe
2	−2447.87	19	p < .001	p < 0.001	.48 (266.0)	0.779
3	−2396.44	30	p=.013	p < 0.001	.22 (120.3)	0.742
4	−2369.49	41	p = .021	p < 0.001	.10 (55.5)	0.759
5	−2347.95	52	p<.0500	p < 0.001	.10 (54.8)	0.769
6	−2328.51	63	p=0.204	p<.0001	.09 (51.4)	0.779

^aLog Likelihood;

^bLo-Mendell-Rubin adjusted likelihood ratio test;

^cparametric Bootstrap Likelihood Ratio Test;

^dsmallest class relative frequency (frequency);

^eNot Applicable

Ideal number of classes based on fit statistic shown in bold

Cognitive measures were selected to assess major domains implicated in substance abuse. These tests included measures of intelligence, memory, and three measures of executive functioning. Executive functioning refers to cognitive processing of the initiation and maintenance of actions. It is believed to be critical for the initiation, maintenance, and cessation of substance abuse behavior [1-6]. The cognitive measures used included measures of general intelligence, viz., the Shipley Institute of Living Scale (SILS) [49], working memory, viz., the Digit-span subtest of the WAIS-R [50], attention, viz., the Trail Making Test [51], behavioral inhibition, viz., the Stroop test [52], and cognitive flexibility, viz., the Wisconsin Card Sorting Task [53]. .

The cessation of substance abuse behavior theoretically requires adequate intelligence to understand the problematic nature of the behavior and to gain value from instructions from others [54]. To measure overall intelligence, we used the Shipley Institute of Living Scale (SILS). The SILS consists of a vocabulary subtest and an abstraction subtest. For the verbal subtest of the SILS, participants were given 20 minutes to complete 40 multiple choice vocabulary questions. For the abstraction section of the SILS, participants were given 20 minutes to complete 20 abstract reasoning questions. Raw scores were calculated according to the number of questions answered correctly in the allotted time. Adjusted total T-scores controlling for age/education were used in the present analysis [49]. As SILS scores are sometimes used as a proxy for pre-morbid intellectual functioning among substance using samples [55], we further report predicted IQ based on the Wechsler Adult Intelligence Scale – Revised [50]. This provides a full scale IQ using the SILS age-stratified raw score [49].

To cease substance use behavior, adequate working memory is needed to continually work towards goals [56]. Working memory was assessed using the Digit Span subtest of the WAIS-R [50]. Participants were asked to repeat a list of numbers read aloud by the administrator either in the same order (forward) or in reverse (backward) order. Individuals’ ability to manipulate increasingly long strings of numbers, reversing the order of two to seven digits, reflects their ability to retain information and manipulate it [57]. The Digit Span subtest is commonly used in studies of cognitive functioning and substance use [58-61].

Attention is another domain theoretically important in the reduction of substance use behavior [62]. Two sections of the Trail Making Test were included in this study. In the first trial (Part A), the time to connect the dots with incremental numbers was recorded. In the second trial (Part B), the time to connect the dots from numbers to letters, both increasing by one digit or letter was recorded. While the first test largely examines psychomotor speed, the second test requires careful attention due to its relatively higher level of complexity and novelty. Participants were asked to connect the dots as fast as they could in each of the trials [51]. The Trailmaking Test is commonly used to assess cognitive impairment amongst substance users [63]. The scores in the table reported here reflect the time spent on the task and thus, higher scores are indicative of worse performance. However, for the sake of clarity, the scores in the figure are reversed z-scores and thus, similar to the other measures reported, higher scores are indicative of better performance.

Behavioral inhibition seems to be critically important in the reduction of substance use. Individuals in recovery must inhibit drug abuse, a highly rewarding behavior [64]. The Stroop test is a widely used measure of behavioral inhibition [65]. The Stroop test was administered in three parts in this study [52]. Participants were first asked to read words printed in black, and second name colors printed without words as fast as they could. Third, participants were asked to name the color of the word printed, while ignoring the word that was printed. “Interference” scores report the number of correct answers in the third trial adjusted for the number of correct answers on the first two trials [66]. Thus, higher interference scores represent better performance. The Stroop appears to be associated with an inability to inhibit behavior amongst substance abusers [11, 67].

Finally, cognitive flexibility is needed to reduce substance use [68]. When substance use is no longer available as a coping strategy, recovering individuals need to learn a new set of rules for their behavior. We used a computerized version of the WCST to measure cognitive flexibility. In the WCST, participants must match 128 cards; after 10 correct matches, the rules for matching change without alerting the participant [53]. The number of times the participant continues using the previous matching rules are called “perseverative errors”; this has been described as the most sensitive executive function score in the WCST [69]. Perseverative error scores were standardized for age and education [53]. These standard scores are also reversed so that higher scores indicate better performance.

Statistical Analyses

Mplus version six was used for the Latent Class Analysis (LCA) modeling, which assumes that “latent classes”, i.e., an underlying grouping of individuals, exist and can explain the patterns of reporting by individuals [70]. Eight drug use indicator variables were entered into the LCA model. Starting with a one class model and incrementally increasing the number of classes, a series of LCA models were fit to the data. As part of LCA, Mplus begins with a random value, i.e, a “random start”, and then performs several iterations until it reaches ideal values, i.e., “maxima”. However, any set of maxima determined after several iterations of a random start could potentially be missing a superior set of maxima on a more global level. As suggested in the M-plus manual, we used a minimum of 500 random starts to ensure that global, rather than local, maxima were reached [70]. If the log likelihood was not replicated at least five times, the number of starts was increased until the log likelihood was replicated a minimum of five times.

Multiple fit statistics were used to determine the best-fitting, most parsimonious model, including the Bayesian Information Criteria (BIC) [71], the parametric bootstrap likelihood ratio test (BLRT) [72], and the Lo-Mendell-Rubin adjusted likelihood ratio test (LMR) [73]. The value and utility of the resultant classes was assessed using entropy [74, 75]. Entropy uses individual estimated posterior probabilities to summarize the degree to which the latent classes are distinguishable and the precision of assignment of individuals into classes. Entropy ranges from 0 to 1; with higher values indicating better class separation. Finally, the choice of latent class solution presented was also informed by a priori substantive criteria, such as meaningfulness in terms of the current epidemiology of drug use. Higher class models were preferred to increase specificity and understanding of the relationships with specific drugs and cognitive dysfunction. Mplus uses a full information maximum likelihood estimation with the assumption the data is missing at random [76], a widely accepted approach [77], so participants were not removed due to missing data on latent class indicators. There were no more than 2 participants with missing data on any of the latent class indicators. Covariance coverage ranged from 0.993-1.0, well over minimum thresholds for adequate coverage, e.g., 0.10 [78].

The n reported in each class is a total based on individual probabilities of class membership, i.e., a number ranging from 0.0, or no chance of membership in that specific class for that specific individual, and 1.0, or 100% certainty of class membership. Thus, these numbers are decimals, rather than whole numbers. However, in order to compare groups on categorical demographic variables, such as race, we used Most Likely Class Membership to add group assignment to IBM SPSS 19.0 and then conducted chi-square analyses. Similar methodology has been used elsewhere [46].

After deciding on the appropriate number of classes that best fit the data, we examined the association between class membership and several demographic, predictor and outcome variables utilizing the auxiliary option [79-81]. This option was used to study the association of classes with cognitive performance measures without changing the unconditional latent class model [70]. The AUXILIARY (e) option was used to examine the extent to which cognitive variables and continuous demographic variables varied as a function of latent class membership by testing the equality of means across latent classes using posterior probability-based multiple imputations. LCA was used to identify classes based on past-month drug use and adjust for probabilities of group membership when examining cognitive performance. Results obtained are reported with standard errors based on probabilities of class membership. The inclusion of auxiliary information in mixture analysis helps us to understand and evaluate the fidelity and utility of the resultant profiles [82], as well as providing useful information for understanding relationships between drug use and cognitive performance.

RESULTS

Lo-Mendel Rubin (LMR) test supported a 5-class model. Monte carlo simulations support the LMR and Bootstrap Likelihood Ratio Test (BLRT) over other fit statistics [83]. However, according to the BLRT, all 6 models were superior to the models with less classes. Since a large number of classes generally decreases the usefulness of the model, we used the LMR-supported 5-class model. Although in prior research we chose a three class model [20], we chose to use a five class model in the present study so as to provide further information above and beyond prior research and increase specificity when looking at associations between drugs and cognitive factors. LCA assumes local independence. This assumption holds if bivariate residuals are all less than 3.84 [84, 85]. In the 5-class model presented here, bivariate residuals ranged from 0.000-0.039. Entropy was relatively high (0.759) indicating good class separation.

As can be seen in figure 1a, prevalent indicators (having an estimated probability of over 50%) for each class are as follows: “Nasal Heroin/Crack Smokers” (n=166.9 ): cigarettes, alcohol, nasal heroin, and crack smoking; “Polysubstance” ( n=54.8 ): cigarettes, alcohol, crack smoking, injection heroin, injection cocaine, and injection speedball; “Multi-Injectors” (n=128.7 ): cigarettes, alcohol, marijuana, injection heroin, injection cocaine, and injection speedball; “Heroin Injectors” (n=135.7 ): cigarettes and injection heroin; and “Nasal Heroin” (n=62.9): cigarettes, alcohol, and nasal heroin.

Figure 1.

a) Five class solution of a latent class analysis of 552 users of cocaine and/or heroin in Baltimore, MD. Estimated probabilities for past-month drug use are graphed based on latent class membership. Error bars indicate standard errors of estimated probabilities. b) Cognitive performance by class. Intelligence is based on total t-score from Shipley Institute of Living Scales. Memory is based on total score from Digit Span task. Attention is based on number of seconds to complete Trails B task. Behavioral Inhibition is based on Interference score from the Stroop task. Cognitive Flexibility is based on perseverative errors on Wisconsin Card Sorting Task. For illustrative purposes, z-scores were computed and scores from Trails B and Wisconsin Card Sorting Task were reverse-scored. Thus, higher scores always indicate better performance.

The groups differed on demographic characteristics (Table 2). “Crack Smoke/Nasal Heroin” and “Polysubstance” were both significantly older than “Nasal Heroin” users and “Heroin Injectors”. “Multi-Injectors” were significantly younger than the other four groups. “Polysubstance” users and “Heroin Injectors” were significantly less educated than “Crack Smoke/Nasal Heroin” and “Multi-Injectors”, whereas “Nasal Heroin” users were significantly more educated. Since age and education were included in model creation, these variables were used in class naming. There were also significant differences in gender, race, and current drug treatment.

Table 2.

Demographics and drug use characteristics by latent class.

	Older, Less Educated Polysubstance	Older Crack Smoke / Nasal Heroin	Less Educated Heroin Injectors	More Educated Nasal Heroin	Younger Multi- injectors
	M (SE) / N (%)	M (SE) / N (%)	M (SE) / N (%)	M (SE) / N (%)	M (SE) / N (%)
Age***	36.42 (0.98) b	36.01 (0.48) a,b	32.65 (0.75) a	32.43 (1.04) a	28.90 (0.68) c
Highest Grade Completed***	10.17 (0.28) b	11.16 (0.15) c	10.29 (0.18) b	12.44 (0.16) a	10.87 (0.16) c
Female**	22 (44.0%)	82 (51.2%)	61 (43.3%)	25 (35.7%)	39 (30.5%)
African-American***	31 (62.0%)	125 (78.1%)	47 (33.3%)	49 (70.0%)	22 (17.2%)
Currently in Drug Treatment**	17 (34.0%)	42 (26.6%)	48 (35.3%)	16 (23.9%)	16 (12.6%)
Lifetime History of Regular Use
Drink Alcohol**	34 (68.0%)	101 (63.5%)	65 (46.4%)	37 (52.9%)	81 (64.3%)
Smoke Marijuana***	29 (58.0%)	92 (57.5%)	66 (48.2%)	34 (49.3%)	94 (73.4%)
Smoke Crack***	22 (44.0%)	106 (66.2%)	20 (14.2%)	11 (15.7%)	36 (28.1%)
Nasal Heroin**	28 (56.0%)	99 (61.9%)	63 (44.7%)	46 (65.7%)	79 (61.7%)
Inject Heroin***	40 (80.0%)	39 (24.4%)	104 (73.8%)	14 (20.0%)	119 (93.0%)
Inject Cocaine***	20 (40.0%)	9 (8.3%)	29 (20.6%)	1 (1.4%)	49 (38.3%)
Inject Speedball***	24 (48.0%)	13 (8.1%)	27 (19.1%)	4 (5.7%)	50 (39.1%)

^*Omnibus test (degrees of freedom = 4), p < .05

^**Omnibus test (df = 4), p < .01

^***Omnibus test (df = 4), p < .001

Subscripts with same letter are not significantly different at p < .05, as revealed by post-hoc pairwise tests (df = 1).

If at least one class is significantly lower than another, class(es) with lowest number on variable shown in italics.

If at least one additional class is significantly higher than another, class(es) with highest number on variable shown in bold.

In addition, there were significant differences in lifetime history of drug use (Table 2). These differences were in the anticipated directions. For example, more individuals in the “More Educated Nasal Heroin” group had a lifetime history of regularly snorting heroin (65.7%) than members of the “Less Educated Heroin Injectors” group (44.7%). This demonstrates that for most individuals, the past-month drug use was similar to their lifetime history of drug use, although there is some important variability. For example, a majority of “Older, Less Educated Polysubstance” users, “Older Crack Smoke / Nasal Heroin” users, “Younger Multi-injectors”, and “More Educated Nasal Heroin” users reported drinking alcohol regularly, compared to a minority of “Less Educated Heroin Injectors”

Cognitive performance results are presented on Table 3 and graphically, using z-scores, in Figure 1b. The results indicated profound impairment among this sample of cocaine and heroin users in comparison to published norms [61, 86-89]. In general, younger and more educated groups out-performed older and less educated groups, but with some notable variations discussed further below.

Table 3.

Cognitive performance by class.

	Older, Less Educated Polysubstance	Older Crack Smoke / Nasal Heroin	Less Educated Heroin Injectors	More Educated Nasal Heroin	Younger Multi- injectors	Published Norms
	M (SE)	M (SE)	M (SE)	M (SE)	M (SE)	M (SE)
Intelligence (Shipley)
Verbal ***	36.42 (1.62) a	36.42 (0.91) a	37.40 (1.05) a,b	41.50 (1.81) c	41.84 (0.91) b,c	---
Abstract ***	44.57 (1.35) a	46.55 (0.73) a	45.77 (0.83) a	49.89 (1.35) b	49.18 (0.84)b	---
Predicted IQ ***	83.57 (1.94) a	85.22 (1.06) a	84.54 (1.27) a	91.53 (2.13) b	90.07 (1.24) b	102.5 (0.92)
Memory (Digit Span Task)
Forward	9.63 (0.30)	9.71 (0.19)	9.68 (0.22)	10.10 (0.37)	10.01 (0.20)	10.2 (0.52)
Backward *	5.20 (0.29) a	5.55 (0.19) a,b	5.41 (0.20) a	6.38 (0.38) b	5.78 (0.20) a,b	7.20 (0.45)
Total †	14.73 (0.51) a	15.18 (0.33) a,b	15.08 (0.35) a,b	16.48 (0.66) b	15.80 (0.35) a,b	---
Attention (Trail Making)
Trails A (seconds)	27.29 (1.53)	27.55 (0.92)	28.72 (1.15)	27.15 (1.66)	25.95 (0.92)	24.40 (1.52)
Trails B (seconds)*	92.92 (6.95) a,b	87.99 (3.89) a,b	92.91 (4.73) b	80.64 (6.00) a,b	78.24 (4.00) a	50.68 (2.15)
Beh Inhibition (Stroop)
Word Score	41.26 (1.21)	41.50 (0.69)	40.89 (0.76)	42.21 (1.39)	42.74 (0.77)	---
Color Score	43.72 (1.24)	42.36 (0.70)	42.27 (0.77)	43.13 (1.24)	43.56 (0.75)	---
Interference Score †	45.95 (1.33) a	46.55 (0.66) a	49.30 (0.73) b	48.83 (1.31) a,b	49.73 (0.76) b	50.7 (1.0)
Cog Flexibility (WCST)
Categories Completed *	3.75 (0.33) a	4.29 (0.17) a	4.38 (0.20) a,b	4.47 (0.29) a,b	4.87 (0.20) b	5.0 (0.2)
Non-perseverative Errors***	82.35 (2.82) a	87.74 (1.33) a,b	89.92 (1.69) b,d	88.05 (2.04) a,d	95.60 (1.80) c	---
Perseverative Errors***	90.98 (3.78) a,b	91.55 (1.80) a	98.37 (2.44) b	90.41 (3.10) a,b	106.58 (2.79) c	---

^†Omnibus test (degrees of freedom = 4), p < .1

^*Omnibus test (df = 4), p < .05

^**Omnibus test (df = 4), p < .01

^***Omnibus test (df = 4), p < .001

Subscripts with same letter are not significantly different at p < .05, as revealed by post-hoc pairwise tests (df = 1).

If at least one class is significantly worse than another, class(es) with worst performance on variable shown in italics.

If at least one additional class is significantly better than another, class(es) with best number on variable shown in bold.

Note: Higher scores on all tasks, except for Trail Making Task, indicate better performance

DISCUSSION

This is the first study of which we are aware to use latent classes to elucidate the complex relationship between polydrug use and cognitive performance. Five classes were found: Older-Less-Educated-Polysubstance, Older-Crack-Smoke/Nasal-Heroin, Less-Educated-Heroin-Injectors, More-Educated-Nasal-Heroin, and Younger-Multi-Injectors (YMI). These classes are largely consistent with similar latent class analyses [40, 44, 46] and clinician reports [14]. The associations with age and education are novel and may be due to cohort/demographic effects, as suggested by significant differences in race and gender.

Although a non-drug user control group was not available for comparison in this data set, there was evidence to suggest that substance use in itself was associated with worse cognition. Cognitive performance on nearly all tests was substantially worse than published normative values. Though age- and education-matched mean time on the Trail-Making Test Part A was comparable, Part B time was considerably higher in our study population, with a range of 80.7 – 92.9 seconds, compared to the normative time of 50.7 seconds [90]. This finding is consistent with the different degree of difficulty on these two tasks, as Part B is a more difficult cognitive task due to its increase in visually interfering stimuli and consequent increased demands in psychomotor speed and visual search [91]. Similarly, the Stroop Color and Word Test Manual reports predicted word and color scores for age 30-35, with 11 years of education to be approximately 99 and 75 points, respectively, which indicates a far greater number of items correctly completed than the scores of our study population [52]. Similar differences are seen in the Wisconsin Card Sorting Task (WCST) [53].

This study demonstrated novel applications of a methodology for examining the complicated relationships between polysubstance use and cognitive performance. The analysis found associations between age, education, and polydrug patterns. In general, younger classes performed better on cognitive measures. Age-related decrements in cognitive performance, including executive functioning, are documented for those 55 years and older [69] and many argue that cognitive decline begins when individuals are in their 20’s or 30’s [92].

More educated classes performed better on intelligence and memory, but not significantly so on attention. Further, the More-Educated-Nasal-Heroin class performed significantly worse than Young-Multi-Injectors on both behavioral inhibition and cognitive flexibility and did not differ significantly from Less-Educated-Heroin-Injectors. This is difficult to interpret, but some interesting and important possibilities exist. It is possible that education and/or nasal heroin use are particularly associated with lower cognitive flexibility. Either of these possibilities deserve further research. If education received by this sample of inner-city drug users did not enhance, and perhaps even impaired, cognitive flexibility, other interventions may be needed [93-95]. If nasal heroin use is associated with more extreme decrements in cognitive performance than injection use, this is important for clinicians to understand for purposes of addressing limitations in particular subpopulations and for harm reduction efforts. Alternatively, as the multi-injectors were the youngest class, it is possible that age overcame effects of education or other drug use, which would have important implications for understanding changes in drug dependence as users age.

There are some limitations to this study. Although these classes were different with respect to specific cognitive measures, this may only be a result of in-group variabilities. The sample consisted of a population that had varying times of last drug use. This introduces a type of random error into the analysis, which may make finding significant differences difficult. Further, as with all cross-sectional studies, no causal relationships can be established from the data presented. Experimental or longitudinal research is needed to address these issues.

LCA provides a novel method for exploring the relationships between polydrug use and cognitive function. This is necessary as polydrug use is the norm among illicit drug users [14]. As all human cognitive functioning data is contaminated by confounders, it is anticipated that latent modeling will become more common and more sophisticated so as to address these difficult issues. Future research should develop this potential to further refine and explore this methodology in different populations and gain deeper insights into associations between executive functioning and drug use.

Table 4.

Cognitive performance by class.

	Older, Less Educated Polysubstance	Older Crack Smoke / Nasal Heroin	Less Educated Heroin Injectors	More Educated Nasal Heroin	Younger Multi- injectors	Published Norms
	M (SE)	M (SE)	M (SE)	M (SE)	M (SE)	M (SE)
Intelligence (Shipley)
Verbal ***	36.42 (1.62) a	36.42 (0.91) a	37.40 (1.05) a,b	41.50 (1.81) c	41.84 (0.91) b,c	--
Abstract ***	44.57 (1.35) a	46.55 (0.73) a	45.77 (0.83) a	49.89 (1.35) b	49.18 (0.84)b	---
Predicted IQ ***	83.57 (1.94) a	85.22 (1.06) a	84.54 (1.27) a	91.53 (2.13) b	90.07 (1.24) b	102.5 (0.92)
Memory (Digit Span Task)
Forward	9.63 (0.30)	9.71 (0.19)	9.68 (0.22)	10.10 (0.37)	10.01 (0.20)	10.2 (0.52)
Backward *	5.20 (0.29) a	5.55 (0.19) a,b	5.41 (0.20) a	6.38 (0.38) b	5.78 (0.20) a,b	7.20 (0.45)
Total †	14.73 (0.51) a	15.18 (0.33) a,b	15.08 (0.35) a,b	16.48 (0.66) b	15.80 (0.35) a,b	---
Attention (Trail Making)
Trails A (seconds)	27.29 (1.53)	27.55 (0.92)	28.72 (1.15)	27.15 (1.66)	25.95 (0.92)	24.40 (1.52)
Trails B (seconds)*	92.92 (6.95) a,b	87.99 (3.89) a,b	92.91 (4.73) b	80.64 (6.00) a,b	78.24 (4.00) a	50.68 (2.15)
Beh Inhibition (Stroop)
Word Score	41.26 (1.21)	41.50 (0.69)	40.89 (0.76)	42.21 (1.39)	42.74 (0.77)	---
Color Score	43.72 (1.24)	42.36 (0.70)	42.27 (0.77)	43.13 (1.24)	43.56 (0.75)	---
Interference Score †	45.95 (1.33) a	46.55 (0.66) a	49.30 (0.73) b	48.83 (1.31) a,b	49.73 (0.76) b	50.7 (1.0)
Cog Flexibility (WCST)
Categories Completed *	3.75 (0.33) a	4.29 (0.17) a	4.38 (0.20) a,b	4.47 (0.29) a,b	4.87 (0.20) b	5.0 (0.2)
Non-perseverative Errors***	82.35 (2.82) a	87.74 (1.33) a,b	89.92 (1.69) b,d	88.05 (2.04) a,d	95.60 (1.80) c	---
Perseverative Errors***	90.98 (3.78) a,b	91.55 (1.80) a	98.37 (2.44) b	90.41 (3.10) a,b	106.58 (2.79) c	---

^†Omnibus test (degrees of freedom = 4), p < .1

^*Omnibus test (df = 4), p < .05

^**Omnibus test (df = 4), p < .01

^***Omnibus test (df = 4), p < .001

Subscripts with same letter are not significantly different at p < .05, as revealed by post-hoc pairwise tests (df = 1).

If at least one class is significantly worse than another, class(es) with worst performance on variable shown in italics.

If at least one additional class is significantly better than another, class(es) with ^best number on variable shown in bold.

Note: Higher scores on all tasks, except for Trail Making Task, indicate better performance