The Cannabis Effects Expectancy Questionnaire–Medical (CEEQ–M): Preliminary Psychometric Properties and Longitudinal Validation within a Clinical Trial

Abstract

The use of cannabis for medical symptoms is increasing despite limited evidence for its efficacy. Expectancies—prior beliefs about a substance or medicine—can modulate use patterns and effects of medicines on target symptoms. To our knowledge, cannabis expectancies have not been studied for their predictive value for symptom relief. The 21-item Cannabis Effects Expectancy Questionnaire–Medical (CEEQ–M) is the first longitudinally validated measure of expectancies for cannabis used for medical symptoms. The questionnaire was developed for a randomized clinical trial of the effect of state cannabis registration (SCR) card ownership on symptoms of pain, insomnia, anxiety, and depression in adults (N = 269 across six questionnaire administrations). Item-level analyses (n = 188) demonstrated between-person stability of expectancies and no aggregate, within-person expectancy changes three months after individuals gained access to SCR cards. Exploratory factor analysis (n = 269) indicated a two-factor structure. Confirmatory factor analysis at a later timepoint (n = 193) demonstrated good fit and scalar invariance of the measurement model. Cross-lagged panel models across three and twelve months (n = 187 and 161, respectively) indicated that CEEQ–M-measured expectancies did not predict changes in self-reported cannabis use; symptoms of pain, insomnia, anxiety, and depression; and well-being. However, greater baseline cannabis use predicted more positive expectancy changes. The findings suggest that the CEEQ–M is psychometrically sound. Future work should clarify at what timescales cannabis expectancies have predictive value and how cannabis expectancies for medical symptoms are maintained and diverge from other substance use expectancies.

In the United States, cannabis use for medical symptoms has been increasing for the past two decades as legal commercial markets for smoked, vaped, ingested, and other cannabis products have grown (Compton et al., 2017; Hasin, 2018). Cannabis products used for medical purposes are most commonly used for common, chronic conditions such as pain, anxiety, depression, and insomnia (Kosiba et al., 2019; Lintzeris et al., 2020). However, there is no high-quality evidence for efficacy of commercial cannabis products for these conditions, especially for psychiatric disorders (Allan et al., 2018; Sarris et al., 2020; Whiting et al., 2015). As the use of cannabis for medical symptoms increases, individuals either currently using or interested in using cannabis find themselves in a highly variable landscape of external factors that may influence personal beliefs about the risks and benefits of cannabis use and its effects on medical symptoms, including legal and regulatory constraints, the beliefs of clinicians, and media coverage (Azcarate et al., 2020; Gedin et al., 2022; Karanges et al., 2018; Takakuwa et al., 2020; Zolotov et al., 2018). To understand these beliefs and their influence on the effects of cannabis use on medical symptoms, validated measures are needed.

Expectancies and their Clinical Relevance

The conceptual framework of response expectancies (Kirsch, 1985), as a subset of social learning theory (SLT; Bandura & Walters, 1977), provides a psychological lens through which to examine how personal perceptions of the risks and benefits of cannabis use affect whether and how individuals start and maintain cannabis use and potentially obtain symptom relief from cannabis. SLT has previously been applied to psychoactive drugs such as alcohol (Maisto et al., 1999); these applications have emphasized response expectancies, i.e., beliefs about the effects of a drug, as determinants of initiation, maintenance, and cessation of drug use. Further, the experienced effects of a medicine or placebo can be directly shaped by expectancies (see, e.g., Kaptchuk & Miller, 2015). Prior studies have reported an effect of response expectancies on the subjective and physiological effects of medicines and placebos in observational (J. A. Chen et al., 2011) and experimental (Bingel et al., 2011; Faria et al., 2017) settings. This evidence conveys the importance of assessing expectancies when assessing efficacy of a treatment.

Although there are few rigorous experimental studies of cannabis expectancies for medical symptoms, studies suggest that expectancies modulate cannabis’ effects. Specifically, it has been established that medically related cannabis expectancies modulate pain-related outcomes (De Vita et al., 2021) and that these expectancies might play a role in pain reduction in response to placebo in randomized clinical trials of cannabinoids (Gedin et al., 2022).

Therefore, cannabis expectancies merit investigation as potential predictors of cannabis use and changes in symptoms for which cannabis is sought as relief. In clinical trial settings, a validated measure of cannabis expectancies could be used to measure relationships among expectancies and cannabis use and/or clinical outcomes. Additionally, knowledge about how cannabis expectancies affect clinical outcomes would be useful, alongside safety and efficacy data, for the development of clinical guidelines for prescribing cannabis and discussing cannabis with patients. Further, a reciprocal effect might also exist by which cannabis use impacts cannabis expectancies and thus subsequent trajectories of cannabis use; such an effect has been found for alcohol (Sher et al., 1996) and, for cannabis, might help to explain changes over time in the cannabis use behaviors within an individual, as well as potentially stable differences between individuals.

Existing Measures of Cannabis Expectancies

Several validated constructs of recreational cannabis expectancies have been described in the general substance use literature (e.g., Connor et al., 2011; Galen & Henderson, 1999; Heishman et al., 2001, 2009; Schafer & Brown, 1991; Torrealday et al., 2008). However, we posit that expectancies for cannabis for medical symptoms may be distinct from existing constructs because the motives for using cannabis medically are distinct—though likely not orthogonal—and the population using cannabis for medical purposes is also different from the population using for recreational purposes. Those using for medical purposes tend to be older, sicker, have lower incomes, and generally do not have a history of substance use (e.g., Furler et al., 2004; Lin et al., 2016; Roy-Byrne et al., 2015; Woodruff & Shillington, 2016). A recent study reported in a large, geographically diverse sample that sleep, social anxiety, and coping motives were associated with greater odds of cannabis for medical symptoms, while motives related to boredom, enjoyment, simultaneous alcohol use, and celebration were associated with greater odds of recreational use (Vedelago et al., 2020). Therefore, expectancy measures developed specifically for the population using cannabis for medical purposes are needed.

One instrument has been developed to address this need: the Medical Cannabis Expectancy Questionnaire (MCEQ; Morean & Butler, 2019). The MCEQ comprises 27 binary yes/no items that represent two factors corresponding to positive (e.g., symptom relief) and negative (e.g., side effects) cannabis expectancies. Although the latent variables represented by the MCEQ are invariant across product type, sex, and whether respondents also consume cannabis recreationally, the respondents considered for its development and validation were required to have a state cannabis registration card and have used cannabis within the past 30 days. The expectancies of individuals with little or no cannabis use history are thus not reflected in the MCEQ; this fact limits the generalizability of the MCEQ to active cannabis users and therefore limits the MCEQ from predicting cannabis use initiation and maintenance for current non-users, which is a desirable feature for expectancy measures (see, e.g., Montes et al., 2019).

In addition, the dichotomous item structure of the MCEQ does not lend itself to quantifying cannabis expectancies and thus precludes measurements of expectancy intensities. An ability to more dimensionally quantify expectancies, e.g., with Likert-type items, is useful for assessing expectancy change over time, measuring associations between expectancies and quantitative outcomes of interest, and comparing expectancy intensities across groups; see, e.g., the comparison of the dichotomously scored Cocaine Effects Expectancy Questionnaire (Schafer & Brown, 1991) and the Likert-scored Cocaine Expectancy Questionnaire (Jaffe & Kilbey, 1994) by the authors of the latter for an illustration of the advantages of Likert-type items over dichotomously scored items. While validation of the MCEQ in non-cannabis-using populations and with Likert-type items would be desirable, we were unable to perform such a validation study because the MCEQ was published after data collection began for this project.

Further, the MCEQ has not been validated longitudinally, and thus it is unclear if expectancies measured by the MCEQ are stable over time and to what extent changes in cannabis use and symptoms might induce changes in cannabis expectancies and vice versa. Finally, although the MCEQ captures expectancies corresponding to a broad range of medical conditions from which individuals might seek relief, several MCEQ items correspond to recreational cannabis use, e.g., “Keyed-Up,” “Guilty,” “Sociable,” and “Increased Sex Drive.” Therefore, because of the inclusion of recreational items, biases in the predictive value of the MCEQ for symptom relief arising from differences in recreational expectancies cannot be ruled out. A cannabis expectancy measure consisting only of items that pertain to expected medical effects would carry less risk for such biases.

Project Overview

We present the 21-item Cannabis Effects Expectancy Questionnaire–Medical (CEEQ–M), a new measure of cannabis expectancies that (1) reflects the expectancies of individuals with and without previous cannabis use, (2) permits quantification of expectancy intensities, (3) has been validated with longitudinal samples to assess stability of expectancies and potential associations of expectancies with changes in cannabis use and clinical outcomes, and (4) considers only dimensions of expectancy that correspond to use of cannabis for medical symptom relief. First, we describe the item-level properties of the CEEQ–M, including the change of these properties over time among participants of a randomized controlled trial (RCT) that randomized individuals interested in cannabis to receive a state cannabis registration (SCR) card or to a waitlist control (Gilman et al., 2022). We then ascertain the factor structure of the CEEQ–M and test it for invariance and explore longitudinal effects among expectancies measured by the CEEQ–M and self-reported cannabis use; symptoms in the domains of pain, sleep, anxiety, and depression; and measures of physical and mental well-being.

Overview of Methods

Studies

This project proceeded through four sequential studies, each of which had largely overlapping (although not identical) analytic samples of participants (total N = 269). Study 1 reported the item-level characteristics of the CEEQ–M, i.e., response distributions and descriptive statistics for individual items, as well as item-specific within-individual changes across time in a clinical trial of SCR cards. Study 2 ascertained the latent variable structure of the CEEQ–M through exploratory factor analysis. Study 3 used confirmatory factor analysis to (1) verify the latent variable structure suggested by exploratory factor analysis and (2) perform measurement invariance testing for the hypothesized latent variable structure across time and demographic groups. Finally, study 4 explored longitudinal associations among cannabis expectancies measured by the CEEQ–M and clinical outcomes of interest, including self-reported cannabis use; self-reported symptoms in the domains of sleep, pain, anxiety, and depression; and measures of general well-being. Figure 1 outlines the questions and analyses comprising the four studies of this project.

Figure 1. Summary of the Sequence of Studies and their Questions and Main Analyses.

Note. SCR = state cannabis registration.

Participants

All participants were subjects in a single-site, longitudinal RCT (NCT03224468, conducted from July 1, 2017, to July 31, 2020; Gilman et al., 2022) that recruited adults, aged 18 to 65, who were considering using cannabis for symptoms of pain, sleep disorders, depression, or anxiety. Participants were recruited by study staff at the Massachusetts General Hospital Center for Addiction Medicine through advertising by email, web, and bulletin board announcements in the community, as well as through advertising on public transportation and on radio advertisements. Trial participants were randomized to either a group which could receive a SCR card and start using cannabis immediately (SCR card group) or a waitlist control (WLC group) who agreed to wait for three months after randomization before obtaining a SCR card and initiation of cannabis product use. Randomization assignments were stratified by sex, age, and chief concern (i.e., sleep, pain, or affective symptoms [depression and/or anxiety]) in a 2:1 ratio (SCR card group to WLC group) within each stratum. Participants were free to select and change their cannabis product, dose, use frequency, and manner of consumption throughout the trial. The trial did not provide or pay for SCR cards or cannabis products. Criteria for study exclusion included daily pre-enrollment cannabis use, current substance use disorder (including cannabis use disorder but not including nicotine dependence or light to moderate alcohol use), pregnancy, and major medical and psychiatric conditions. Table 1 shows participant characteristics for the analytic samples used in each study. See Gilman et al. (2021, 2022) and Tervo-Clemmens et al. (2023) for further details about participant characteristics and recruitment and study design and procedures. The analytic samples were relatively homogeneous with respect to race and ethnicity, and the average educational attainment of participants was high.

Table 1.

Characteristics of Participants Comprising Analytic Samples

Variable	Studies 1 + 2 Screening sample	Study 3 1-month sample	Studies 1 + 4 3-month sample	Study 4 12-month sample
Number of participants	269	193	188/187a	161
SCR card group	N/Ab	111	109	N/Ac
Waitlist control group	N/Ab	82	78	N/Ac
Primary concern
Pain	91	60	60	55
Insomnia	55	44	44	37
Anxiety or depression	123	89	83	69
Biological sex
Female	180	124	124	108
Male	89	69	63	53
Race (% Caucasian)	77.7%	81.3%	81.8%	83.9%
Ethnicity (% Hispanic or Latino)	7.8%	7.3%	7.5%	6.2%
Age in years	37.0 (14.3)	37.4 (14.3)	37.2 (14.3)	37.4 (14.5)
Education in years	16.4 (2.66)	16.6 (2.67)	16.6 (2.62)	16.5 (2.51)

Note. Primary concerns were self-reported. For age and education, the mean is given with the standard deviation in parentheses. SCR = state cannabis registration.

^aOne participant completed the CEEQ–M at the three-month visit but did not provide data required for analyses in study 4.

^bThe screening visit occurred before randomization in the parent trial.

^cThe twelve-month visit occurred after the randomized period of the parent trial.

Measures

The Cannabis Effects Expectancy Questionnaire–Medical (CEEQ–M; for instructions and items, see Appendix in Supplemental Materials) was developed by investigators for use in the parent clinical trial (NCT03224468), where it was reported as the Medical Marijuana Effects Expectancy Questionnaire (MMEEQ) and contained items with the word “marijuana” instead of “cannabis,” which latter we recommend as more appropriate for future questionnaire administrations (cf. Solomon, 2020). The CEEQ–M contains 21 Likert-type items scored from 1 to 10, where numerical scores are indicated on the questionnaire to correspond to the following ordered categories: 1 and 2, “Strongly disagree;” 3 and 4, “Disagree;” 5 and 6, “Neutral;” 7 and 8, “Agree;” 9 and 10, “Strongly agree.” The items of the CEEQ–M were largely selected from Reinarman et al., 2011, which queried 1,746 patients in nine medical cannabis clinics about their reasons for use. The complete, 21-item version of the CEEQ–M was administered to participants during the screening and post-randomization baseline visits, as well as at the following visit timepoints following randomization: one month, three months, six months, and one year (intervals correspond to time after randomization). For participants randomized to the SCR card group, the baseline visit was scheduled to cooccur as closely as possible with SCR card receipt by the participant. At each of these visits, substance use history over the past month, including cannabis, was captured using a timeline follow-back method that asked participants to classify their use according to the following options: “Less than once a month,” “Less than once every two weeks,” “Less than once a week,” “1-2 days a week,” “3-4 days a week,” “5-6 days a week,” and “Once or more per day.” For each post-randomization visit at which the CEEQ–M was administered, symptom self-reports with respect to pain, sleep, anxiety, and depression were obtained with the following instruments: pain, Brief Pain Inventory severity subscale (BPI; Cleeland & Ryan, 1994; Tan et al., 2004); sleep, Athens Insomnia Scale summed score (AIS; Soldatos et al., 2000); anxiety and depression, respective subscales of the Hospital Anxiety and Depression Scale (HADS; Snaith, 2003; Zigmond & Snaith, 1983). General physical and mental well-being were also measured at these visits with the 12-item Short-Form Health Survey (SF-12; Ware et al., 1996). To establish internal consistency for these instruments in our samples, we calculated Cronbach’s α as follows at the baseline visit: BPI-Severity, α = 0.911 (n = 102); AIS, α = 0.847 (n = 207); HADS-Anxiety, α = 0.853 (n = 207); HADS-Depression, α = 0.827 (n = 207); SF-12, standardized α for all items = 0.835 (n = 207). All instruments thus had “very good” internal consistency in our sample (DeVellis & Thorpe, 2021).

Procedure

All participants provided written informed consent and were financially compensated for participation. Visits for the parent study generally lasted between one and two hours. The Mass General Brigham Human Research Committee approved study procedures.

Analysis

Although it is ideal to utilize independent samples for each of the typical stages of questionnaire development and validation (i.e., item development, exploratory factor analysis, confirmatory factor analysis, and validity analyses), the current project was constrained to the longitudinal sample of the parent clinical trial. Therefore, to avoid bias arising from performing sequential analyses on the same set of questionnaire responses (i.e., overfitting), each study within this project considered questionnaire responses at different timepoints. This approach optimized the generalizability of the obtained results given the sampling constraints. All analyses were performed using R, version 4.0.1 (R Core Team, 2020). The analyses presented here were not preregistered; however, the parent clinical trial was registered on clinicaltrials.gov (NCT03224468). Data and analysis code for this study are available upon request pending scientific review and a completed data use agreement/material transfer agreement.

Study 1: Item-Level Analyses

Study 1: Methods

Participants.

The screening sample, representing randomized participants with complete CEEQ–M responses at the screening visit (n = 269), was used for analyses of item response distributions and descriptive statistics. The three-month sample, consisting of randomized participants who completed the CEEQ–M at the baseline and three-month visits (n = 188), was used for analyses of within-individual changes in item responses between these two visits. See Table 1 for participant characteristics.

Analysis.

Given the novelty of the CEEQ–M and the desirability of information about what specific effects are expected, and to what degree, by prospective users of cannabis products for medical complaints, the first phase of the current project sought to establish a comprehensive overview of the item-level characteristics of the CEEQ–M, including response distributions and descriptive statistics for each of its 21 items. For each item, the mean, standard deviation, median, skew, and kurtosis were computed from the screening sample, and relative frequency bar charts were generated to illustrate response distributions. Additionally, to assess whether gaining access to a SCR card was associated with changes in responses to particular CEEQ–M items between the baseline and three-month timepoints, i.e., the randomized period of the parent study, item-level within-individual response changes between these visits were tested for statistical significance between the two randomization groups using a two-way repeated measures ANOVA framework; in particular, the group-by-time interaction was assessed for significance for each item as an indicator of whether response changes across time differed discernibly between randomization groups.

Study 1: Results

Response distributions for the 21 items of the CEEQ–M at the screening timepoint varied across items (Figure S1, Table 2). Average responses, i.e., levels of agreement, were highest for the items asking whether “marijuana relieves” pain (item 1, mean 7.43) and anxiety (item 4, mean 7.49) and whether “marijuana improves” sleep (item 10, mean 7.58), relaxation (item 11, mean 8.13), and appetite (12, mean 7.91), while the lowest average response occurred for item 21: “marijuana cures cancer” (mean 2.97). Shapes of response distributions were qualitatively different: many items (e.g., 9, 15, 17, and 18) showed clear modes at 5, corresponding to “neutral” and indicative of central tendency bias. These qualitative differences are reflected in quantitative variation across items in skew and kurtosis (Table 2).

Table 2.

Descriptive Statistics for CEEQ–M Responses at the Screening Timepoint (n = 269)

Item	Mean	Standard deviation	Median	Skew	Kurtosis
1. Relieves pain	7.43	1.68	7	0.05	−0.99
2. Relieves spasms	6.47	1.85	6	0.41	−0.43
3. Relieves headaches	6.67	1.98	7	0.11	−0.66
4. Relieves anxiety	7.49	1.78	8	−0.35	−0.52
5. Relieves nausea	7.10	2.02	7	−0.14	−0.95
6. Relieves depression	6.51	1.91	6	0.10	−0.38
7. Relieves cramps	6.24	1.82	6	0.43	−0.09
8. Relieves panic	6.36	1.94	6	0.24	−0.66
9. Relieves diarrhea	4.70	1.61	5	0.37	2.85
10. Improves sleep	7.58	1.73	8	−0.35	−0.58
11. Improves relaxation	8.13	1.45	8	−0.57	−0.17
12. Improves appetite	7.91	1.71	8	−0.66	0.08
13. Improves concentration	4.97	2.12	5	0.44	−0.04
14. Improves energy	4.99	1.96	5	0.36	−0.04
15. Reduces medication side effects	5.77	1.97	5	0.16	0.48
16. Reduces anger	6.52	1.96	6	0.09	−0.64
17. Reduces involuntary movements	5.70	1.80	5	0.43	0.71
18. Reduces seizures	6.08	1.91	5	0.39	−0.06
19. Is a substitute for prescription medication	5.55	2.62	5	−0.12	−0.82
20. Is a substitute for alcohol	5.17	2.56	5	−0.03	−0.72
21. Cures cancer	2.97	2.30	2	0.97	0.07

The item-level within-person trajectories of CEEQ–M responses between the baseline and three-month visits of the parent study were not significantly affected by whether an individual had been randomized to gain access to a SCR card: two-way group-by-time repeated measures ANOVA tests yielded p > 0.05 with 1 and 186 degrees of freedom for the group-by-time interaction for all items. Figure 2 illustrates this result with visualizations of within-person change score distributions between the baseline and three-month visit timepoints by randomization group.

Figure 2. Within-Person Changes in Responses to CEEQ–M Items During the Randomized Study Period (n = 188).

Note. Change scores were computed for each participant by subtracting responses at the three-month timepoint from responses at the baseline timepoint. Circles indicate mean change scores. Error bars show 95% confidence bounds for change scores based on a t-distribution. SCR = state cannabis registration; WLC = waitlist control.

Study 2: Exploratory Factor Analysis (EFA)

Study 2: Methods

Participants.

Responses to the CEEQ–M from randomized participants at the screening visit (n = 269; see Table 1 for characteristics) served as the analytic sample.

Analysis

Assumptions.

Univariate, per-item normality of the responses was assessed visually with bar charts and statistically with a Shapiro-Wilk test for each item. Univariate normality was considered to be achieved if the Shapiro-Wilk p ≥ 0.05. Multivariate normality of the responses was assessed with Mardia’s skewness and kurtosis tests using the following criterion for multivariate normality: z-kurtosis < 5 and p ≥ 0.05 for skewness and kurtosis tests. The Kaiser-Meyer-Olkin statistic was computed for each item to assess whether factor analysis would account for an adequate amount of shared variance among the items, and the canonical statistic cutoffs (“marvelous” if ≥ 0.90, “meritorious” if ≥ 0.80, etc.; Kaiser & Rice, 1974) were used for interpretation. Bartlett’s test of sphericity was used to confirm whether correlations among the items were present; nonzero correlations were deemed present if p < 0.05.

Number of factors to be retained.

Per guidelines for obtaining a reliable factor solution (Nunnally, 1994), multiple criteria were used to determine a consensus for the number of factors to retain. Scree plots of factor eigenvalue versus factor number were visualized for presence of an “elbow” cutoff, Velicer’s minimum average partial (MAP) test was applied to identify a factor quantity that minimizes the MAP statistic, parallel analysis was conducted, and the Kaiser criterion (keeping all factors with an eigenvalue greater than one) was considered (see Courtney & Gordon, 2013 for an overview of these criteria).

Factor extraction and iterative solution construction.

After concluding that the assumptions of univariate and multivariate normality were violated by the responses, the factor extraction method was chosen as principal axis factoring, which is recommended for non-normal data (Brown, 2015). Factor extraction was performed using the fa function in the psych package of R (Revelle, 2022) with an oblique rotation (“oblimin”), which allowed the extracted factors to correlate. After initial factor extraction, the following criteria were used to drop individual items, and the factor analysis was repeated after dropping an item to determine whether any remaining items met the criteria for being dropped: low loadings onto all factors (< 0.35), high cross-loading(s) (item complexity > 1.8), or poor interpretability of an item with satisfactory loading onto a factor (≥ 0.35) in the context of other items with satisfactory loadings onto the same factor. When no items meeting the criteria for being dropped remained, the factors were named according to the shared meaning of the items with major loadings onto each factor.

Internal consistency analysis.

The internal consistency (reliability) of each factor was measured using Cronbach’s α and corrected item-total correlations using the alpha function in the psych R package; published cutoffs were used for interpretation of α (DeVellis & Thorpe, 2021), and items with an item-total correlation greater than 0.5 were kept. The values of Cronbach’s α if an item were to be dropped were also considered; items were candidates for being dropped if this latter value of α was greater than the original α for the factor with the item not dropped. Finally, from the resulting factor structure, the between-factor correlation(s) and percentage(s) of total variance explained by each factor were recorded.

Study 2: Results

Assumptions.

Shapiro-Wilk tests of univariate normality yielded p < 0.05 for each CEEQ–M item for the screening sample, indicating violation of univariate normality for each item, which agreed with visualizations of response distributions in this sample (Figure S1). For the same sample, Mardia’s skewness and kurtosis tests yielded p < 0.05, and z-kurtosis was 24.8, greater than the threshold of 5; together, these results indicated violation of the multivariate normality assumption by the joint 21-item response distribution. The overall Kaiser-Mayer-Olkin statistic for the 21 items was 0.90, and per-item Kaiser-Mayer-Olkin statistics ranged between 0.81 and 0.95, suggesting adequate sampling for factor analysis for each item. Bartlett’s test of sphericity yielded p < 0.05, thus indicating the presence of nonzero correlations among items.

Number of factors to be retained.

The scree plot showed an elbow at two factors. Velicer’s MAP test reached a minimum value with two factors. Parallel analysis with 1,000 iterations consistently suggested a two-factor latent variable structure. Two factors met the Kaiser criterion of having an eigenvalue greater than one (eigenvalues of the first two factors were 7.70 and 1.03, respectively). Given the agreement among the set of criteria used, two factors were selected as the quantity to be retained during factor extraction.

Factor extraction and iterative solution construction.

Following factor extraction with principal axis factoring, items 15 (“Marijuana reduces medication side effects”), 16 (“Marijuana reduces anger”), and 18 (“Marijuana reduces seizures”) were dropped because their complexity was greater than 1.8 (i.e., cross-loadings were too high). Additionally, item 9 (“Marijuana relieves diarrhea”) was dropped because of poor interpretability of the factor loading of this item in the context of other items with satisfactory loadings (≥ 0.35) onto the same factor. After these four items were successively dropped, no additional items meeting criteria for being dropped remained, leaving a two-factor, 17-item factor structure (Table 3). Factors were named “Symptom relief” (11 items) and “Atypical beliefs” (6 items) based on the items with major loadings onto each factor (Table 3).

Table 3.

Factor Structure of CEEQ–M Items from Iterative Exploratory Factor Analysis of Responses at the Screening Timepoint (n = 269)

Item	Symptom relief loading	Atypical beliefs loading	Complexity	Communality
1. Relieves pain	0.772	−0.053	1.01	0.554
2. Relieves spasms	0.526	0.254	1.44	0.487
3. Relieves headaches	0.703	0.085	1.03	0.567
4. Relieves anxiety	0.810	−0.114	1.04	0.568
5. Relieves nausea	0.647	−0.051	1.01	0.385
6. Relieves depression	0.520	0.220	1.35	0.443
7. Relieves cramps	0.657	0.082	1.03	0.497
8. Relieves panic	0.598	0.106	1.06	0.439
10. Improves sleep	0.651	0.010	1.00	0.431
11. Improves relaxation	0.764	−0.025	1.00	0.564
12. Improves appetite	0.602	−0.027	1.00	0.346
13. Improves concentration	−0.047	0.798	1.01	0.598
14. Improves energy	0.004	0.672	1.00	0.455
17. Reduces involuntary movements	0.247	0.422	1.61	0.353
19. Is a substitute for prescription medication	0.201	0.452	1.38	0.343
20. Is a substitute for alcohol	0.197	0.385	1.49	0.270
21. Cures cancer	0.018	0.570	1.00	0.336
Cronbach’s α	0.904	0.778
Percentage of total variance explained	31.0%	13.9%
Factor correlation	0.546

Note. Factor names (“Symptom relief” and “Atypical beliefs”) were derived from major loadings (bolded) onto each factor.

Internal consistency analysis.

Cronbach’s α was 0.904 (“very good” consistency) for the “Symptom relief” factor and 0.778 (“respectable” consistency) for the “Atypical beliefs” factor. Corrected item-total correlations exceeded 0.5 for all 17 items, which further suggested good internal consistency of each factor. The values of Cronbach’s α if an item were to be dropped were all lower than the initial values of Cronbach’s α for the respective factors, so no additional items were dropped, and the two-factor, 17-item structure remained as the final model. The between-factor correlation was 0.546, and the “Symptom relief” and “Atypical beliefs” factors accounted respectively for 31.0% and 13.9% of the total variance among the 17 items comprising the final factor structure.

Study 3: Confirmatory Factor Analysis (CFA) and Invariance Testing

Study 3: Methods

Participants.

Responses to the CEEQ–M from participants with responses at the baseline and one-month visits (n = 193; see Table 1 for characteristics) served as the analytic sample. This analytic sample was selected for CFA and invariance testing since the one-month visit occurred within the randomized period of the parent trial; therefore, CEEQ–M responses from trial participants at this timepoint could be tested for invariance across randomization group to permit robust conclusions about the generalizability of the factor structure of the CEEQ–M with respect to differences in cannabis use (ability to use in the SCR card group versus agreement not to use in the WLC group). Although this study could have used the three-month sample for its analyses, the three-month sample was reserved for subsequent testing of longitudinal effects of expectancies in study 4 with an intent to prevent bias arising from performing successive validation steps on the same analytic sample.

Analysis

All models in this study and study 4 were specified, fit, and assessed with the lavaan package in R, version 0.6-9 (Rosseel, 2012).

CFA of measurement model.

A single CFA with the factor structure obtained in study 2 was first fit to the entire set of responses at the one-month visit to assess the overall goodness of fit of the measurement model. A robust, diagonally weighted least squares estimator with a mean-adjusted chi-square test statistic (“WLSM” in lavaan) was used for all models in this study. The chosen estimator needed to be robust because of observed non-normality of the response distributions of individual items, and a weighted least squares estimator was selected because least squares methods have desirable properties, such as more accurate factor loading and standard error estimates, for models involving ordinal data (Li, 2016; Rhemtulla et al., 2012). Although the CFA models estimated in this study involved no ordinal variables, some structural models estimated in study 4 included an ordinal variable; therefore, given that model fit indices are meaningfully comparable only when the same estimator is used for all models to be compared (Xia & Yang, 2019), the WLSM estimator was used for all models to ensure comparability. The metrics and thresholds used to define CFA fit quality were as suggested in Bentler, 1990; Browne & Cudeck, 1992; and Hu & Bentler, 1999. Robust versions of CFI, TLI, and RMSEA were used for fit assessment because of use of a robust estimator. Cronbach’s α and McDonald’s ω were computed for each latent variable as measures of internal consistency. Although Cronbach’s α is a weaker measure of consistency than McDonald’s ω (McNeish, 2018), α was computed to allow comparisons to the values of α from study 2, as calculation of McDonald’s ω requires a fitted CFA model.

Cross-group and longitudinal Invariance testing.

For the latent variables measured by the CEEQ–M to be comparable across time, demographic groups, and clinical cohorts (in particular, cohorts of cannabis users and non-users), invariance of the factor structure obtained in study 2 needed to be tested separately across time and for each group of interest. Therefore, invariance testing was performed across randomization group (SCR card group vs. WLC group), sex, age (binned into groups of 18–25 years and 26–65 years to reflect randomization procedures), and time (from the baseline visit to the one-month visit). Additional timepoints were not included in longitudinal invariance testing because the sample size did not permit models with substantially more parameters to be estimated reliably. Invariance testing entailed fitting successively more restrictive CFA models to either the set of CEEQ–M responses from the one-month visit (for group invariance testing) or the joint set of responses from the baseline and one-month visits (for longitudinal invariance testing). Specifically, for each comparison, the following three nested models were fitted: a model with configural invariance, in which the same factor structure was imposed onto each group or timepoint; a model with metric (weak) invariance, in which equality of factor loadings was additionally imposed; and a model with scalar (strong) invariance, in which equality of item intercepts was additionally imposed (see, e.g., Putnick & Bornstein, 2016 for a summary of current conventions for invariance testing).

Criteria for invariance.

The changes in goodness of fit incurred with these successive equality constraints were assessed according to standard criteria (F. F. Chen, 2007) to determine whether a given level of invariance was present. As above, robust versions of CFI and RMSEA were used for fit assessment because of use of a robust estimator. Changes in the chi-square model test statistics and degrees of freedom between successive levels of invariance were also reported for the sake of consistency with other works using these statistics and the corresponding likelihood ratio test for assessing measurement invariance. However, given that the alternative goodness-of-fit indices outlined by F. F. Chen, (2007) have been shown to be more accurate and less dependent on sample size for invariance testing than the chi-square likelihood ratio test (Cheung & Rensvold, 2002; Meade et al., 2008), the changes in chi-square statistics and degrees of freedom were not considered when determining whether a given level of invariance was present.

Study 3: Results

CFA of measurement model.

CFA of the two-factor, 17-item measurement model for all respondents at the one-month timepoint showed good fit in terms of CFI, TLI, and SRMR, while RMSEA suggested adequate fit (Table S1). Cronbach’s α and McDonald’s ω were, respectively, 0.906 and 0.907 for the “Symptom relief” factor and 0.779 and 0.772 for the “Atypical beliefs” factor, indicating that the internal consistency of each factor in the one-month sample was similar to the corresponding value observed in the screening sample in study 2.

Cross-group and longitudinal invariance testing.

Invariance testing of the measurement model indicated scalar invariance, i.e., equality of factor structure, factor loadings, and item intercepts on each factor, across randomization group, age (18–25 years / 26–65 years), sex, and two successive questionnaire administrations (baseline and one-month visit timepoints) (Table S1).

Study 4: Exploratory Analyses of Longitudinal Effects of Expectancies

Study 4: Methods

Participants.

Randomized study subjects who completed the CEEQ–M and provided other outcome information at the baseline and three-month visits (n = 187) and at the baseline and twelve-month visits (n = 161) comprised the analytic samples. See Table 1 for participant characteristics.

Analysis

To test whether longitudinal associations exist between the latent variables measured by the CEEQ–M and outcomes of self-reported cannabis use; self-reported pain, insomnia, and affective symptoms; and self-reported overall physical and mental well-being, cross-lagged panel models (CLPMs) were fitted to the CEEQ–M measurement model and variables measuring these outcomes. Separate analyses were performed for the randomized and unrandomized analytic intervals of the parent trial (baseline to three months and baseline to twelve months, respectively) to maintain an ability to account for the effect of randomization group membership on cannabis use during the randomized study period. In particular, for each set of responses (baseline and three months and baseline and twelve months), a separate CLPM with a full set of cross-lagged paths between timepoints was fitted between the longitudinally-scalarly invariant CEEQ–M-measured latent variables and each outcome of interest: past-month cannabis use, measured by a timeline follow-back method, binned into “once or more per week” and “less than once a week;” self-reported symptoms of anxiety (HADS), depression (HADS), pain severity (BPI), and insomnia (AIS); and general physical and mental well-being (SF-12).

Separate models were fitted for each outcome in the interest of model simplicity and interpretability. For all latent and outcome variables that were endogenous in the models, i.e., variables for the three-month and twelve-month timepoints, age and sex were included as predictors to avoid detection of potential spurious relationships arising from uncontrolled effects of age and sex, which impact perceptions of cannabis (Pacek et al., 2015). To account for the effect of randomization group membership on cannabis access and use, for the CLPM fitted to cannabis use at baseline and three months, randomization group membership (encoded as 0 for the WLC group and 1 for the SCR card group) was added as a predictor of cannabis use at three months. CLPM fit was assessed according to Bentler & Bonett, 1980 and Hu & Bentler, 1999. As above, robust versions of CFI and RMSEA were used for fit assessment because of use of a robust estimator. Autoregressive and cross-lagged coefficients in CLPMs were considered significant if p < 0.05.

Study 4: Results

The fourteen CLPMs fit for seven outcomes and two timepoint pairs had adequate or better fit as measured by CFI and RMSEA, though SRMR values for all CLPMs marginally exceeded the threshold for adequate fit (Table S2). Autoregressive paths between latent and outcome variables were significant in all models. For the CLPM comparing CEEQ–M-measured latent variables and cannabis use between baseline and three months (Figure 3A), significant cross-lagged paths were present from baseline cannabis use to both latent expectancy variables at three months. For the CLPM of the same variables between baseline and twelve months (Figure 3B), a significant cross-lagged path was present from baseline use to the “Atypical beliefs” variable at twelve months. For all other CLPMs, all cross-lagged paths were nonsignificant. Figures S2–S7 show CLPMs with standardized solutions for models with fully nonsignificant sets of cross-lagged paths (CLPMs between CEEQ–M-measured expectancies and symptoms of pain, insomnia, anxiety, and depression, as well as physical and mental well-being).

Figure 3. Cross-Lagged Panel Models of CEEQ–M-Measured Latent Variables and Self-Reported Past-Month Use of Cannabis Between (A) Baseline and Three-Month Timepoints (n = 187) and (B) Baseline and Twelve-Month Timepoints (n = 161).

Note. Randomization group was encoded as 1 for the active state cannabis registration card group and 0 for the waitlist control. Cannabis use was binned into two categories: “once or more per week” and “less than once a week.” Numeric values correspond to the standardized model solution. Age and sex were included as predictors of three-month and twelve-month variables (paths not shown). THC = tetrahydrocannabinol.

Solid paths are significant (p < 0.05); dashed paths are nonsignificant (p ≥ 0.05).

Discussion

The current project is the first investigation into the psychometric properties and associations with clinical outcomes of the CEEQ–M. The CEEQ–M is the first questionnaire for expectancies for cannabis use for medical symptoms use that reflects the expectancies of individuals with and without previous cannabis use and permits quantification of expectancy strengths. Through a sequence of four studies, the psychometric properties and longitudinal predictive capabilities of the CEEQ–M were ascertained.

Study 1 found that gaining access to a state cannabis registration card (a proxy for initiation of cannabis use) is not associated with within-person response changes to the items of the CEEQ–M. This result suggests that medically related cannabis expectancies might be formed before initiation of cannabis use (although they might change in response to initial patterns of cannabis use for medical symptom relief, as the findings of study 4, discussed below, suggest). Study 2 revealed that the CEEQ–M has a two-factor latent variable structure and that the two factors in the 17-item measurement model correspond to two related but distinct dimensions: symptom relief (shorter-term, specific outcomes of cannabis use) and atypical beliefs (optimistic, possibly fantastical views of the medical effects of cannabis use). Study 3 determined that the 17-item, two-factor measurement model has good fit to CEEQ–M responses at another timepoint and suggested scalar invariance of the measurement model across SCR card access, age, sex, and a one-month time interval. Results of study 4 suggested that the expectancies measured by the CEEQ–M do not predict changes in clinical outcomes—symptoms of pain, insomnia, anxiety, and depression; and general physical and mental well-being—across the time intervals tested (three and twelve months).

Study 4 demonstrated that greater baseline cannabis use, defined as use weekly or greater, was associated with positive changes in “Atypical beliefs” scores at three and twelve months and with positive changes in “Symptom relief” scores at three months. No other outcome had significant predictive paths to the CEEQ–M-measured expectancy variables at either time interval. The significant paths from baseline use to expectancies at three and twelve months suggest that prior experience with cannabis, e.g., the amount of cannabis used before an individual enrolled in the trial, might shape the ways in which initial use of cannabis for medical symptoms is experienced. Initial positive cannabis experience might serve to make expectancies more positive among individuals who start using cannabis for medical symptoms with higher levels of baseline cannabis use, while individuals with lower levels of baseline use might have cannabis expectancies that are less mutable or slower to change in the context of initiation of cannabis use. Together with the finding in study 1 that gaining access to a SCR card is not associated with within-person response changes to any individual CEEQ–M item, the association of baseline cannabis use with later changes in medically related cannabis expectancies suggests that prior cannabis use patterns of an individual might be a strong predictor of the experiences of an individual in initial months with use of cannabis for medical symptoms. A predictive link between higher baseline use and more positive subsequent expectancy changes has also been observed in a natural history study of adolescent smokers (Wahl et al., 2005).

Given the lack of questionnaires that measure strictly medically related cannabis expectancies, comparison of the psychometric properties of the CEEQ–M to other questionnaires proves difficult. The only similar questionnaire, the Medical Cannabis Expectancy Questionnaire (MCEQ), has a differently patterned two-factor structure reflecting positive and negative expectancies that arises from its use of both positively and negatively worded items; therefore, since the CEEQ–M does not contain negatively worded items, the MCEQ does not provide a useful comparison in terms of factor structure. However, both measure expectancies that are not associated with measures of general well-being/quality of life when demographic variables are controlled for (for corresponding models in this work, see Figures S6 and S7; see also Morean & Butler, 2019).

In the general context of previous findings that expectancies can predict substance use and its subjective effects for a range of substances, including cannabis (e.g., Kirk et al., 1998; Metrik et al., 2009), the findings of study 4—that the baseline cannabis expectancies captured by the CEEQ–M have no discernible predictive capability for symptom changes—suggest two possible explanations that merit further study. First, although previous investigations of cannabis expectancies have largely treated cannabis as a recreational substance—except for the MCEQ, which includes medical and recreational items—upwards of half and up to 80% of those who use cannabis for medical symptom relief concurrently use cannabis for recreation (Morean & Lederman, 2019; Turna et al., 2020). Thus, while the clinical distinction between motivations for cannabis use has logically resulted in psychometric distinctions between expectancies for cannabis intended for recreational effects and for medical symptom relief, it is likely that the expectancies that drive cannabis use behaviors and shape experienced symptom relief and well-being are not exclusively medical but instead a combination of recreational and medical dimensions.

Second, the baseline expectancies for medical effects of cannabis that are captured by the CEEQ–M might be stable, trait-like characteristics that are formed before an individual begins using cannabis and that do not influence cannabis use or experienced effects of cannabis until updated after initial use experiences. That expectancies can be formed before initial use of a substance is clear from studies of alcohol expectancies in children (e.g., Miller et al., 1990); cannabis expectancies may operate similarly. Further, the potential for discrepancies to persist between expectancies and actual experiences of substance use has been established for alcohol (Morean et al., 2015), but nothing is known to date about whether cannabis expectancies, recreational or medical, exhibit the same ability to remain stable when experienced effects, e.g., symptom relief or lack thereof, contradict expectancies.

This possibility—that expectancies for medical effects of cannabis are stable in response to cannabis use experiences—is particularly deserving of further attention considering the finding in study 1 that gaining access to a SCR card does not significantly change responses to individual CEEQ–M items at three months after gaining card access. Indeed, the potential stability of positive cannabis expectancies might serve as a pathway through which cannabis use can be reinforced in individuals who do not experience symptom relief from cannabis; such individuals might represent a large portion of cannabis users, especially among individuals using cannabis for pain and affective disorders (Gilman et al., 2022). Further, given that individuals who use cannabis for affective disorders represent a disproportionately large proportion of those who use cannabis both recreationally and for medical symptom relief (Turna et al., 2020), the possibility that cannabis expectancies are unresponsive to use experiences is of particular concern for this subpopulation, as this possibility represents a mechanism through which problematic levels of cannabis use could be promoted and reinforced.

Limitations and Constraints on Generality

This project has limitations. First, given the design of the parent clinical trial, all studies were performed on non-independent, largely overlapping sets of study participants at different timepoints. Further, these analytic samples were relatively homogeneous with respect to demographic and socioeconomic factors such as race, ethnicity, and educational attainment. The non-independence and homogeneity of the analytic samples in this project limit the generalizability of the CEEQ–M measurement model. These biases were mitigated to the extent possible through the performance of analyses on CEEQ–M responses at different timepoints—e.g., exploratory factor analysis was conducted on the screening sample, while confirmatory factor analysis involved the baseline and one-month samples. Still, the lack of generality arising from sampling constraints cannot be eliminated through this procedure.

Separately, we are unable to report on multiple aspects of the content validity of the CEEQ–M. In particular, the expectancies in this questionnaire are not inclusive of all possible expectancies, especially negative expectancies such as those related to side effects of cannabis use for health concerns. We undertook a pragmatic, approach to item development, which drew on previously described motives for use (Reinarman et al., 2011), to meet the needs of the parent trial. Thus, the items of the CEEQ–M did not undergo a systematic development process. Relatedly, since only the original, 21-item version of the CEEQ–M was administered in this project, we were unable to characterize the 17-item version that emerged from studies 2 and 3. We were also unable to test the effect of changing “marijuana” to “cannabis” in each item of the questionnaire, as all administered questionnaires in this work used “marijuana.”

The incremental and discriminant validity of the CEEQ–M in comparison to existing cannabis expectancy questionnaires also remain unclear, as we did not ask participants to complete other such questionnaires. As motives for cannabis use for recreational and medical purposes are likely non-orthogonal, questions persist about whether the CEEQ–M predicts cannabis use for medical concerns above and beyond existing cannabis expectancy questionnaires and whether the CEEQ–M measures distinct expectancies from existing cannabis expectancy questionnaires developed for recreational use.

Future Work

The CEEQ–M would benefit from additional validation and refinement for its psychometric properties and predictive capabilities to be more precisely and accurately known. Large, diverse samples could be used to verify the factor structure and degrees of measurement invariance. Such samples should include individuals with and without the health conditions common among those seeking cannabis for health concerns. To establish the incremental, convergent, and discriminant validity of the CEEQ–M, future work should also ask respondents to complete related instruments such as the MCEQ (Morean & Butler, 2019) and recreational cannabis expectancy questionnaires such as the Marijuana Effect Expectancy Questionnaire (Schafer & Brown, 1991). The content validity of CEEQ–M items should also be established more robustly through analyses of incremental validity that draw on an expanded item pool, including items measuring negative expectations, and testing of word choice effects, especially between “marijuana” and “cannabis.” Additionally, future work should gather information about the type(s) of cannabis product(s) used by respondents to test whether expectancies for medical effects of cannabis differ by cannabis product, as Morean and Butler (2019) found.

From a methodological perspective, future work might gain additional insights from using alternative longitudinal modeling methods that explicitly separate within-person and between-person effects. One promising alternative is the random intercepts cross-lagged panel model (RI-CLPM; Hamaker et al., 2015), which has recently received attention as a model more useful than the traditional CLPM for representing trait-like, stable characteristics, of which cannabis expectancies might be an example. Use of RI-CLPMs, for instance, might clarify to what extent expectancies for medical effects of cannabis demonstrate within-person and between-person stability over time.

Finally, an improved understanding of the temporal characteristics of cannabis expectancies remains essential for the potential use of cannabis expectancy theory in clinical settings. Experimental modulation of expectancies for medical effects of cannabis, e.g., with balanced-placebo designs, could help to define time intervals in which cannabis expectancies have predictive value.

Conclusion

The CEEQ–M measures latent variables reflecting two dimensions of expectancies of medical effects of cannabis products. The measurement model represented by this questionnaire is psychometrically sound, with good internal consistencies and scalar invariance across demographic groups and time. However, the CEEQ–M lacks longitudinal predictive capabilities for hypothesized expectancy effects on cannabis use frequency and attenuation of medical symptoms (i.e., symptom relief from pain, insomnia, anxiety, and depression) or improvements to general physical and mental well-being. While the expectancies measured by the CEEQ–M appear to be stable in response to changes in symptoms and well-being, some evidence exists for strengthening of cannabis expectancies during initial months of cannabis use for medical symptom relief among individuals with higher baseline levels of cannabis use. These findings suggest avenues of further study of the properties of cannabis expectancies for medical symptoms: their connections to other substance use expectancies (e.g., those for recreational cannabis), the intervals of time at which they act, and the potential for incongruities to persist between expectancies and experienced effects on symptom severity and burden. In sum, the CEEQ–M shows promise as a questionnaire for expectancies for cannabis used for medical symptoms, particularly for between-person, population-level differences. Administering the CEEQ–M in additional samples will strengthen evidence for its psychometric structure and longitudinal predictive value.

Public Significance Statement.

Commercially available cannabis is now widely used to self-treat medical symptoms, despite lack of controlled trial data demonstrating efficacy. Beliefs about the effects of cannabis have not been assessed systematically in trials of cannabis for medical symptoms. Our questionnaire, which measures beliefs about the effects of cannabis on medical symptoms, will help researchers and clinicians understand how these beliefs impact the perceived effects of cannabis on various symptoms.

Data Sharing Statement:

All data, code, and materials used in the analyses can be provided by Jodi Gilman and Massachusetts General Hospital pending scientific review and a completed data use agreement/material transfer agreement. Requests for all materials should be submitted to Jodi Gilman.