Abstract
Hypotheses about the link between cannabis use and psychosis apply to the within-person level but have been tested mostly at the between-person level. We used a within-person design, in which a person serves as his own control, thus removing the need to consider confounding by any fixed (genetic and nongenetic) characteristic to study the prospective association between cannabis use and the incidence of attenuated psychotic experiences, and vice versa, adjusted for time-varying confounders. We combined 2 general population cohorts (at baseline: Early Developmental Stages of Psychopathology Study, n = 1395; Netherlands Mental Health Survey and Incidence Study-2, n = 6603), which applied a similar methodology to study cannabis use and attenuated psychotic experiences with repeated interviews (T0, T1, T2, and T3) over a period of approximately 10 years. The Hausman test was significant for the adjusted models, indicating the validity of the fixed-effects model. In the adjusted fixed-effects model, prior cannabis use was associated with psychotic experiences (aOR = 7.03, 95% CI: 2.39, 20.69), whereas prior psychotic experiences were not associated with cannabis use (aOR = 0.59, 95% CI: 0.21, 1.71). Longitudinal studies applying random-effects models to study associations between risk factors and mental health outcomes, as well as reverse causality, may not yield precise estimates. Cannabis likely impacts causally on psychosis but not the other way round.
Keywords: cannabis, psychosis, within-person effects, epidemiology, drug use
Introduction
The pooled analysis of the association between cannabis and psychosis indicates approximately a 4-fold increase in risk for the heaviest users and a 2-fold increase for the average cannabis user in comparison to nonusers.1,2 Observational studies on the association between cannabis use and psychosis outcomes commonly suffer from so-called omitted variable bias. This refers to the possibility that the observed link is due to any of a range of plausible confounders, beyond simple demographics and direct confounders, that can never be measured precisely in combination, such as genetic vulnerabilities, cognitive abilities, fetal exposures, early development, and parental rearing practices.
One way of comprehensively addressing omitted variable bias is to study within-person associations of cannabis and psychosis using panel data. Panel data (the same subjects measured at 2 or more points in time) consist of repeated measures that are nested within individuals. The repeated measures are referred to as the within-person level or level 1 units, whereas the persons are referred to as the between-person level or level 2 units. Studies on the association between cannabis use and psychosis over time are usually analyzed at the between-person level or the mixed within-person and between-person level. However, the relationship between cannabis use and psychosis may well differ across levels (ie, between-person versus within-person levels). A classic example involves the relationship between physical exercise and heart rate: the within-person analysis will reveal a positive association in that more strenuous exercise leads to a higher heart rate. However, in contrast, this association is likely to be reversed at the between-person level, as people who exercise a lot, on average, tend to have a lower heart rate due to their overall better fitness.3
Panel data of individuals measured repeatedly over time allow control for fixed characteristics, including full nongenetic and genetic risk, whether they are measured/measurable or not, by analyzing, over time, associations with an exposure in the fixed-effect within-person model (FE) and contrasting these with the traditionally used random-effects model (RE) that combines within-person and between-person effects. The within-person effect compares psychosis outcomes in the same person across time when switching from using to not using cannabis and, therefore, cannot be confounded by stable characteristics, including nongenetic and genetic variation. To date, this has been examined in only very few studies. Fergusson et al studied 1055 young adults over time and found, using a within-person analysis, that cannabis use predicted psychotic symptoms and that the direction of causality was from cannabis use to psychotic symptoms rather than the other way round.4 Bechtold et al, studying a sample of 1009 adolescent boys, repeatedly assessed over time and using a within-person analysis, found that cannabis use predicted subthreshold psychotic experiences but not the other way round.5
Although these 2 studies used relatively small samples, the results are consistent. The question arises, therefore, to what degree the results of within-person analyses may differ from traditional random-effects models that have shown more variable results, particularly with regard to the self-medication hypothesis.6 Previous within-person analyses did not conduct such a comparison.4,5 Also, a problem of fixed-effect exposure-outcome analysis is that it requires very large samples, as only individuals exhibiting a within-person change of exposure/outcome are informative. It is, therefore, important to further study longitudinal associations between cannabis use and psychosis, in both directions and in larger samples than available to date, at the within-person level, using panel data that allow for the use of participants as their own controls, and to compare the results with commonly used random-effects models.
With binary dependent variables, within-person analysis can be done with the use of conditional logit/fixed-effects logit models.7 There are several prerequisites for using fixed-effects methods to study prospectively the association between cannabis use and psychosis: (1) the psychosis outcome must be measured on at least 2 occasions for each individual, one with and one without preceding cannabis use; (2) the cannabis exposure must change across time for a substantial portion of the individuals; (3) in the causal model (cannabis→psychosis), cannabis use must precede psychosis and, in the self-medication model (psychosis→cannabis), psychosis must precede cannabis use. In order to meet these 3 requirements together, panel data are required that consist of a baseline assessment and at least 3 follow-up assessments.
Attenuated psychotic states, the main component of “clinical high-risk” states, are the focus of much research focused on early intervention and prevention.8–11 Here, we report (1) the unconfounded, longitudinal, within-person association between cannabis use and attenuated psychotic states, (2) the unconfounded, longitudinal within-person association between attenuated psychotic states and cannabis use (reverse causation), and (3) comparing the results with the traditional random-effects model for panel data, using 2 population cohorts with 4 interview waves and similar methodology, and together covering the entire relevant age range to assess attenuated psychosis outcomes, cannabis use, and possible time-varying confounders, thus evaluating the potential roles of the 3 explanations that can be brought to bear on the apparent association between cannabis use and psychosis.
Methods
Samples
The Early Developmental Stages of Psychopathology Study (EDSP) and Netherlands Mental Health Survey and Incidence Study-2 (NEMESIS-2) are longitudinal cohort studies of the prevalence, incidence, course, and consequences of mental disorders in the Dutch (NEMESIS-2) and German (EDSP) general populations. Both had a baseline interview (T0) and 3 follow-up interviews (T1–T3), using a version of the Composite International Diagnostic Interview (CIDI),12 and both are based on multistage, random sampling procedures of municipalities and households. Detailed information about the characteristics of these studies was published elsewhere.13–16
NEMESIS-2 was conducted to study the prevalence, incidence, course, and consequences of mental disorders in a random representative sample of the Dutch adult general population. The baseline data of NEMESIS-2 were collected from 2007 to 2009, follow-up was until 2018. Details of NEMESIS-2 are provided elsewhere.13,17 The first wave (T0) enrolled 6,646 participants (response rate 65.1%; average interview duration: 95 min), who were followed up in 3 visits within 9 years: successive response rates at year 3 (T1), year 6 (T2), and year 9 (T3) were 80.4% (n = 5303; excluding those who deceased; interview duration: 84 min), 87.8% (n = 4618; interview duration: 83 min), and 86.8% (n = 4007; interview duration: 102 min), respectively. Prevalence rates of mental disorders at baseline reflect lifetime occurrence; rates at T1–T3 reflect interval (baseline–T1, T1–T2, and T2–T3) occurrence of approximately 3 years. Attrition between T0 and T3 was not significantly associated with any of the mental disorders at T0, after controlling for sociodemographic characteristics.18,19
The EDSP is a prospective-longitudinal study in a German general population sample aged 14–24 years at baseline. The EDSP study collected information on the prevalence, incidence, risk factors, comorbidity, and course of mental symptoms and syndromes in a random representative community sample of 3021 adolescents and young adults living in the Munich area (aged 14–24 years at baseline) at 4 waves: at baseline (T0) and at 3 follow-ups after on average 1.6 (T1), 3.5 (T2), and 8.4 years (T3) after baseline, respectively. The first follow-up wave was conducted only in the subsample of respondents aged 14–17 years at T0, whereas the second and third follow-ups were again conducted for all respondents. Details are provided elsewhere.14,15 As the requirement of 4 interview waves only applied to the respondents aged 14–17 years at baseline, only this subgroup of EDSP respondents (n = 1395; 46% of the sample at baseline) was included in the analyses.
In both studies, participants were interviewed using a version of the CIDI.12 This is a comprehensive and standardized diagnostic interview assessing symptoms, syndromes, and diagnoses of mental disorders according to the diagnostic criteria of a version of the Diagnostic and Statistical Manual of Mental Disorders (DSM). The instrument is designed for trained interviewers who are not clinicians. Questions are read in a standardized way and participants’ answers are recorded by the interviewer. Therefore, the CIDI is essentially a self-report instrument.20 Both the validity21,22 and the test–retest reliability23 have been established, showing that the CIDI provides valid diagnoses for almost all nonpsychotic disorders with good to excellent kappa coefficients for most diagnostic sections.
The EDSP project was approved by the Ethics Committee of the Technische Universität Dresden. NEMESIS-2 was conducted with the approval of the ethics committee of the Netherlands Institute of Mental Health and Addiction, Utrecht, the Netherlands. All respondents provided informed consent at each wave.
Psychotic Experiences
In NEMESIS-2, a psychosis add-on instrument based on the G section of the previous versions of the CIDI version 1.1 was included. This add-on instrument consists of 20 psychotic symptoms corresponding to the symptoms assessed in EDSP. In NEMESIS-2, an experienced clinician did a follow-up telephone interview with participants reporting a psychotic symptom to assess whether the self-report information about the psychotic symptom also qualified as psychotic according to the clinicians rating. In NEMESIS-2, a mean of 79% of all participants eligible for follow-up clinical interview at each of the 4 assessment waves was reassessed, and the CIDI rating was adjusted according to a clinical follow-up interview. Given similarities between CIDI self-reported and clinically validated psychotic experiences (PE), in terms of associations, predictive value, and outcome,24–26 CIDI self-reported PE were preserved in the case clinical follow-up interview had not taken place, thus increasing statistical power. PE were dichotomized into “present” versus “absent,” consistent with previous work in NEMESIS-2.27,28
In the EDSP, different assessments for PE were carried out at assessment waves T0 and T1 on the one hand, and at T2 and T3 on the other. At T2 and T3, PE were assessed using the G section of the M-CIDI, as described previously.29,30 This section included 14 items about delusions and 5 items about hallucinations corresponding to classic psychotic symptoms like persecution, various hallucinations, and thought interference. The interviews were conducted by trained psychologists who were allowed to probe with follow-up clinical questions. At T0 and T1, as described previously,31–33 a binary measure of PE was constructed using the psychoticism and paranoid ideation dimensions of the SCL-90-R, combining these 2 dimensions into a single “psychosis” dimension and using the 80th percentile cutoff to create a binary measure of PE. The 80th percentile cutoff was chosen to approximately match the prevalence of PE assessed with the CIDI at T2 and T3.
Cannabis Use
In NEMESIS-2, cannabis use was assessed in the section Illegal Substance Use of the CIDI 3.0. Conform previous work,34 the cutoff of use of once per week or more in the period of most frequent use was used to define a binary variable for regular lifetime or interval cannabis use. In EDSP, cannabis use was assessed with the L-section of the DIA-X/M-CIDI using the question “Have you ever used cannabis five times or more?” to define cannabis exposure (lifetime use at baseline and interval use at follow-up interviews). Conforming to previous work,32,35 the DIA-X/M-CIDI cutoff of use of 5 times or more was used to define the binary variable for cannabis exposure.
Other Drug Use
In both EDSP and NEMESIS-2, a binary variable was constructed at each time point to denote any lifetime or interval use (given very low rates of use) of drugs associated with psychotic disorder: cocaine, stimulants, psychedelics/hallucinogens (mescaline, PCP, XTC, ketamine, LSD, and others).
Adulthood Stressful Life Events
In NEMESIS-2, based on the “Brugha Life events section”,36 participants were asked whether they experienced at least one of 9 life events within the last 12 months (T0) or since the last interview (T1 to T3). Examples of items are serious sickness, death of a family member or close friend, and serious financial problems. Conforming to our previous analyses in this cohort, life events were dichotomized into the present (≥1 life events) and absent (0 life events).37 In EDSP, the Munich Interview for the Assessment of Life Events and Conditions (Münchner Ereignis Liste—MEL)38 was used to assess baseline and interval occurrence of recent adversity. The MEL is a 3-step interview procedure for assessing recent adversity through recognition, rather than free recall, using a list of 84 very detailed and specific descriptions of positive and negative life events encompassing 11 dimensions.38
Any 12-Month DSM Diagnosis
In NEMESIS-2, the following CIDI, version 3.0, nonpsychotic DSM-IV diagnoses were assessed for the period of the last 12 months at each wave: major depression, dysthymia, bipolar disorder, any anxiety disorder, and any substance use disorder. In EDSP, any DSM-IV diagnosis was similarly defined as a 12-month diagnosis at each wave of any major depression, dysthymia, bipolar disorder, any anxiety disorder, and any substance use disorder.
Analyses
Risk Set
Individuals with a diagnosis of psychotic disorder at baseline (Nemesis-2: n = 43, EDSP: n = 0) were excluded from the analysis. Individuals with a lifetime history of psychotic experiences at baseline were per definition excluded from the models testing the association between preceding cannabis use at timepoint t-1 and subsequent psychotic experiences at timepoint t, as the requirement was that individuals did not display evidence of psychosis at timepoint t.
Similarly, individuals at baseline with a lifetime history of cannabis use (and thus including the subgroup with cannabis use disorder) were per definition excluded from the models testing the association between preceding psychosis at timepoint t-1 and subsequent cannabis use at timepoint t, as the requirement was that individuals did not display evidence of cannabis use at timepoint t-1. The risk set thus included 1395 individuals at baseline in the EDSP study, with 5580 observations over the follow-up assessments, and 6603 individuals at baseline in NEMESIS-2, yielding 20455 observations.
Models
Analyzing panel data using random intercept models may not be consistent; therefore, if unbiased estimation is essential, these models may not be suitable. However, random-effects models have the advantage of efficiency and the ability to model heterogeneous effects, and sometimes an answer with a small bias and higher precision may be more useful than an unbiased model with poor precision. Wooldridge indicates that, in practice, researchers often estimate both the random-effects (RE) model and the fixed-effects (FE) model, and then use the Hausman test to determine which model is more appropriate.39,40 In the analysis of panel data (the analysis of data over time), the Hausman test can be used as a test for model misspecification, helping to choose between a fixed-effects model or a random-effects model. The null hypothesis is that the preferred model is random effects; the alternate hypothesis is that the model is fixed effects. The tests look to see if there is a correlation between the unique errors and the regressors in the model. The null hypothesis is that there is no correlation between the two. Interpreting the result from a Hausman test is as follows: if the P-value is small (less than 0.05), the null hypothesis can be rejected. Therefore, when the Hausman test is significant, the FE model should be used; otherwise, the RE model should be preferred.39 Analyses were carried out in Stata 1641 using the xtlogit module, yielding ORs and their 95% CIs, with the FE option (within-person fixed effect) and RE option (random-effects model), followed by the “hausman” module to compare FE and RE models. FE models were adjusted for the following time-varying variables: age (in years), life events (binary), other drug use (binary), and any DSM diagnosis (binary). RE models were adjusted for the same variables.
Planned Sensitivity Analyses
The sample consisted of a merger of 2 separate cohorts with slightly different ways of data collection, interviewing, and variable definition. For the FE model, this cannot act as a confounder as a study sample does not vary within persons. Nevertheless, a planned sensitivity analysis was carried out, calculating the association between prior cannabis use at t-1 and PE at t, separately for NEMESIS-2 and EDSP. For the RE model, a planned sensitivity analysis was carried out, calculating the association between prior cannabis use at t-1 and PE at t whilst additionally adjusting for sample.
Results
Sample characteristics are displayed in Table 1. Rates of cannabis use and psychotic experiences were, as predicted, given their adolescent age, higher in the EDSP sample. Thus, in the NEMESIS-2 sample, the prevalence of psychotic experiences and cannabis use at baseline, in those aged under 20 years, was 16.2% and 8.9% (not in table), respectively, not dissimilar to the corresponding rates at baseline in the EDSP sample aged 14–17 years at baseline at, respectively, 25.8% and 6.1% (not in table). Tables 2 and 3 show that both cannabis use status and psychotic experiences status over time were sufficiently dynamic—ie, moving in and out of status over time—to allow within-person analyses.
Table 1.
Sample demographics of NEMESIS-2 study (Netherlands) and EDSP study (Germany)
| NEMESIS-2 | EDSP | Total | ||||
|---|---|---|---|---|---|---|
| Mean | SD | Mean | SD | Mean | SD | |
| Age at baseline | 44.3 | 12.5 | 15.1 | 1.1 | 42.8 | 16.7 |
| Sex | N a | % | N a | % | N a | % |
| Male | 9143 | 44.7 | 2856 | 51.2 | 11 999 | 46.1 |
| Female | 11 312 | 55.3 | 2724 | 48.8 | 14 036 | 53.9 |
| Total | 20 455 | 100 | 5580 | 100 | 26 035 | 100 |
| Baseline educational levela | ||||||
| High | 7591 | 37.1 | 3076 | 55.1 | 10 667 | 41 |
| Medium | 6588 | 32.2 | 1548 | 27.7 | 8136 | 31.3 |
| Low | 6276 | 30.7 | 956 | 17.1 | 7232 | 27.8 |
| Total | 20 455 | 100 | 5580 | 100 | 26 035 | 100 |
| T0–T3 Psychotic experiences | ||||||
| No | 19 280 | 94.3 | 3865 | 80.6 | 23 145 | 91.7 |
| Yes | 1175 | 5.7 | 930 | 19.4 | 2105 | 8.3 |
| Total | 20 455 | 100 | 4795 | 100 | 25 250 | 100 |
| T0–T3 Cannabis use | ||||||
| No | 19 454 | 97.9 | 3802 | 81.3 | 23 256 | 94.8 |
| Yes | 412 | 2.1 | 873 | 18.7 | 1285 | 5.2 |
| Total | 19 866 | 100 | 4675 | 100 | 24 541 | 100 |
| T0–T3 Other drug use | ||||||
| 0 | 19 648 | 96.1 | 5188 | 93 | 24 836 | 95.4 |
| 1 | 807 | 3.9 | 392 | 7.0 | 1,199 | 4.6 |
| Total | 20 455 | 100 | 5580 | 100 | 26 035 | 100 |
| T0–T3 Life event | ||||||
| No | 10 711 | 52.4 | 2426 | 43.5 | 13 137 | 50.5 |
| Yes | 9744 | 47.6 | 3154 | 56.5 | 12 898 | 49.5 |
| Total | 20 455 | 100 | 5580 | 100 | 26 035 | 100 |
| T0–T3 any 12-month diagnosis | ||||||
| No | 17 784 | 86.9 | 4368 | 78.3 | 22 152 | 85.1 |
| Yes | 2671 | 13.1 | 1212 | 21.7 | 3883 | 14.9 |
| Total | 20 455 | 100 | 5580 | 100 | 26 035 | 100 |
Highest academic achievement: high = at least higher secondary; medium = lower secondary; low = no/primary.
EDSP, Early Developmental Stages of Psychopathology Study; NEMESIS-2, Netherlands Mental Health Survey and Incidence Study-2.
aNumber of observations clustered within individuals (participants were interviewed 4 times over time).
Table 2.
Transitions in cannabis use over time from any t-1 to t (ie, over any 1 consecutive wave), any t-2 to t (ie, over any 2 consecutive waves), and any t-3 to t (ie, over any 3 consecutive waves) in a combined sample of NEMESIS-2 study and EDSP study
| Transition from t-1 to t | Transition from t-2 to t | Transition from t-3 to t | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Cannabis use t | Cannabis use t | Cannabis use t | ||||||||||||
| Cannabis use t-1 | 0 | 1 | Total | Cannabis use t-2 | 0 | 1 | Total | Cannabis use t-3 | 0 | 1 | Total | |||
| No | N a | 15 305 | 378 | 15 683 | No | N a | 9,404 | 426 | 9830 | No | N a | 4275 | 306 | 4581 |
| % | 97.6 | 2.4 | 100.0 | % | 95.7 | 4.3 | 100.0 | % | 93.3 | 6.7 | 100.0 | |||
| Yes | N a | 278 | 362 | 640 | Yes | N a | 207 | 160 | 367 | Yes | N a | 157 | 48 | 205 |
| % | 43.4 | 56.6 | 100.0 | % | 56.4 | 43.6 | 100.0 | % | 76.6 | 23.4 | 100.0 | |||
| Total | N a | 15 583 | 740 | 16 323 | Total | N a | 9,611 | 586 | 10 197 | Total | N a | 4432 | 354 | 4786 |
| % | 95.5 | 4.5 | 100.0 | % | 94.3 | 5.8 | 100.0 | % | 92.6 | 7.4 | 100.0 | |||
aNumber of observations clustered within individuals (participants were interviewed 4 times over time).
Table 3.
Transitions in psychotic experiences over time from any t-1 to t (ie, over any 1 consecutive wave), any t-2 to t (ie, over any 2 consecutive waves), and any t-3 to t (ie, over any 3 consecutive waves) in a combined sample of NEMESIS-2 study and EDSP study
| Transition from t-1 to t | Transition from t-2 to t | Transition from t-3 to t | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Psychotic experiences t | Psychotic experiences t | Psychotic experiences t | ||||||||||||
| Psychotic experiences t-1 | No | Yes | Total | Psychotic experiences t-2 | No | Yes | Total | Psychotic experiences t-3 | No | Yes | Total | |||
| No | N a | 14 967 | 608 | 15 575 | No | N a | 9250 | 423 | 9673 | No | N a | 4241 | 163 | 4404 |
| % | 96.1 | 3.9 | 100.0 | % | 95.6 | 4.4 | 100.0 | % | 96.3 | 3.7 | 100.0 | |||
| Yes | N a | 1026 | 465 | 1491 | Yes | N a | 743 | 261 | 1004 | Yes | N a | 503 | 104 | 607 |
| % | 68.8 | 31.2 | 100.0 | % | 74.0 | 26.0 | 100.0 | % | 82.9 | 17.1 | 100.0 | |||
| Total | N a | 15 993 | 1073 | 17 066 | Total | N a | 9993 | 684 | 10 677 | Total | N a | 4744 | 267 | 5011 |
| % | 93.7 | 6.3 | 100.0 | % | 93.6 | 6.4 | 100.0 | % | 94.7 | 5.3 | 100.0 | |||
aNumber of observations clustered within individuals (participants were interviewed 4 times over time).
Table 4 shows that the rate of incident psychotic experiences at timepoint t as a function of cannabis at timepoint t-1 was consistently higher for the group using cannabis at t-1. Table 5 shows the same with regard to psychotic experiences at t-1 predicting incident cannabis use at timepoint t.
Table 4.
Incident psychotic experiences as a function of cannabis use at the preceding interview over the periods T0–T1, T1–T2, and T2–T3 in a combined sample of NEMESIS-2 study and EDSP study
| Prediction from T0 to T1 | Prediction from T1 to T2 | Prediction from T2 to T3 | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| T1 psychotic experiences | T2 psychotic experiences | T3 psychotic experiences | ||||||||||||
| T0 cannabis use | No | Yes | Total | T1 cannabis use | No | Yes | Total | T2 cannabis use | No | Yes | Total | |||
| No | N a | 5141 | 166 | 5307 | No | N a | 4797 | 236 | 5033 | No | N a | 4163 | 118 | 4281 |
| % | 96.9 | 3.1 | 100.0 | % | 95.3 | 4.7 | 100.0 | % | 97.2 | 2.8 | 100.0 | |||
| Yes | N a | 238 | 13 | 251 | Yes | N a | 101 | 36 | 137 | Yes | N a | 139 | 21 | 160 |
| % | 94.8 | 5.2 | 100.0 | % | 73.7 | 26.3 | 100.0 | % | 86.9 | 13.1 | 100.0 | |||
| Total | N a | 5379 | 179 | 5558 | Total | N a | 4898 | 272 | 5170 | Total | N a | 4302 | 139 | 4441 |
| % | 96.8 | 3.2 | 100.0 | % | 94.7 | 5.3 | 100.0 | % | 96.9 | 3.1 | 100.0 | |||
aNumber of participants.
Table 5.
Incident cannabis use as a function of psychotic experiences at the preceding interview over the periods T0–T1, T1–T2, and T2–T3 in a combined sample of NEMESIS-2 study and EDSP study
| Prediction from T0 to T1 | Prediction from T1 to T2 | Prediction from T2 to T3 | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| T1 cannabis use | T2 cannabis use | T3 cannabis use | ||||||||||||
| T0 psychotic experiences | 0 | 1 | Total | T1 psychotic experiences | 0 | 1 | Total | T2 psychotic experiences | 0 | 1 | Total | |||
| No | N a | 5190 | 77 | 5267 | No | N a | 4856 | 106 | 4962 | No | N a | 4126 | 101 | 4227 |
| % | 98.5 | 1.5 | 100.0 | % | 97.9 | 2.1 | 100.0 | % | 97.6 | 2.4 | 100.0 | |||
| Yes | N a | 608 | 39 | 647 | Yes | N a | 282 | 18 | 300 | Yes | N a | 232 | 36 | 268 |
| % | 94.0 | 6.0 | 100.0 | % | 94.0 | 6.0 | 100.0 | % | 86.6 | 13.4 | 100.0 | |||
| Total | N a | 5798 | 116 | 5914 | Total | N a | 5138 | 124 | 5262 | Total | N a | 4358 | 137 | 4495 |
| % | 98.0 | 2.0 | 100.0 | % | 97.6 | 2.4 | 100.0 | % | 97.0 | 3.1 | 100.0 | |||
aNumber of participants.
Fixed- and Random-Effects Models of Cannabis-Psychosis Associations
Table 6 displays the unadjusted and adjusted longitudinal associations of cannabis at timepoint t-1 with psychotic experiences at timepoint t in the risk set of those without psychotic experiences at timepoint t-1 (top half; Table 6), and, vice versa, the unadjusted and adjusted longitudinal associations of psychotic experiences at timepoint t-1 with cannabis use at timepoint t in the risk set of those without cannabis use at timepoint t-1 (bottom half; Table 6).
Table 6.
Cannabis-psychosis causal hypothesis and cannabis-psychosis self-medication hypothesis: within-person fixed-effect model compared with random-effects model in a combined sample of NEMESIS-2 study and EDSP study
| Predicted outcome | Model | Adjustmentb | OR | 95% LL | 95% UL | P | Observationsc | Individualsc | Hausman chi2 | Hausman P |
|---|---|---|---|---|---|---|---|---|---|---|
| Incidenta PE at t predicted by preceding cannabis use at t-1 | FE | Not adjusted | 9.16 | 3.23 | 25.97 | 0.000 | 827 | 375 | 1.01 | .32 |
| RE | 5.58 | 3.79 | 8.23 | 0.000 | 15 169 | 6240 | ||||
| FE | Adjusted | 7.03 | 2.39 | 20.69 | 0.000 | 827 | 375 | 84.18 | .000 | |
| RE | 2.29 | 1.60 | 3.28 | 0.000 | 15 169 | 6240 | ||||
| Incidenta cannabis use at t predicted by preceding PE at t-1 | FE | Not adjusted | 0.59 | 0.37 | 0.93 | 0.022 | 557 | 225 | 115.37 | .000 |
| RE | 4.27 | 3.23 | 5.64 | 0.000 | 15 671 | 6241 | ||||
| FE | Adjusted | 0.59 | 0.21 | 1.71 | 0.33 | 557 | 225 | 95.23 | .000 | |
| RE | 1.48 | 1.13 | 1.93 | 0.004 | 15 671 | 6241 |
PE, psychotic experiences; FE, fixed-effects model; RE, random-effects model; 95% LL, 95% CI lower bound; 95% UL, 95% CI upper bound; Hausman, Hausman test comparing FE with RE model.
aExcluding individuals with psychotic experiences, respectively, cannabis use at preceding interview wave.
bAdjusted for age, life events, other drugs use, and any DSM diagnosis.
cObservations are clustered within individuals, given repeated measurements over time.
For 3 of the 4 models, and for all adjusted models, the Hausman test comparing FE and RE models was significant, indicating that the FE model should be used. RE models showed bidirectionality—cannabis predicting psychotic experiences and vice versa psychotic experiences predicting cannabis use (Table 6). However, FE models showed that cannabis use predicted psychotic experiences—at greater effect sizes than the RE models, whereas the reverse did not hold. If anything, prior psychotic experiences appeared to protect against cannabis use at follow-up (Table 6).
Planned Sensitivity Analyses
The results of the adjusted FE model of prior cannabis use predicting psychotic experiences (comparable to adjusted FE model; Table 6; top half) separately by sample showed similar effect size for NEMESIS-2 (OR = 7.38, 95% CI: 0.60, 91.09) and EDSP (OR = 6.10, 95% CI: 1.80, 20.65). The results of the adjusted RE model of prior cannabis use predicting psychotic experiences (comparable to adjusted RE model; Table 6; top half), with additional adjustment for sample, revealed a small further reduction of the effect size of the adjusted RE model (OR = 2.04, 95% CI: 1.40, 2.96).
Discussion
Interpretation of Findings
In a within-person analysis of the association between cannabis use and psychotic experiences, we showed a significant divergence between the results of traditional random-effects models and within-person fixed-effects models. Random-effects models showed bidirectionality in that cannabis use predicted psychosis and, vice versa, psychosis predicted cannabis use, similar to previous work in this area.6 However, the preferred within-person model showed a strong association between preceding cannabis use and later psychosis, but not the other way round. If anything, preceding psychosis appeared to be protective for later cannabis use. Although effect sizes were higher in the FE model, CIs were largely overlapping with those in the RE model.
The divergence between the FE and RE models is important, given the fact that hypotheses in mental health research, such as the link between cannabis use and psychosis, are posited almost exclusively at the within-person level, whereas conventional testing of the hypotheses relies almost exclusively on between-person or mixed approaches.42 The reason for the divergence between the 2 models likely has to do with conceptual divergence of the association in the between-person part of the model, just as the association between heart rate and physical exercise. Thus, at the within-person level, prior cannabis use increases the risk of psychosis, whereas prior psychosis tends to reduce the probability of cannabis use. Thus, the fact that there was a weaker association between prior cannabis use and psychosis in the random-effects model (the result of both within-person and between-person effects) may be due to the fact that individuals with psychosis proneness tend to avoid cannabis use (as suggested by the FE model) so that, at the population level, cannabis use becomes less positively associated with psychosis.
Comparison With Previous Work
The sample in the current study was larger than previous work4,5 and statistically compared within-effects and random-effects approaches, finding the former more appropriate, which is relevant in view of divergent results in relation to the self-medication hypothesis. It concurred in finding evidence for a within-person person association from cannabis to psychosis but not the other way round. One of the studies, similar to ours, reported a negative association between psychotic symptoms and subsequent cannabis use.4 Relative risk effect sizes in previous work ranged from 2 to 5, not much different from reported effect sizes here. The fact that our effect sizes were somewhat higher may be related to the fact that we, having a larger sample, were able to model true incident outcomes, whereas previous work adjusted for baseline status of the outcome.4,5
Implications for Psychiatric Epidemiology
All time-varying environmental risks associated with mental health outcomes are amenable to within-person designs. For example, obstetric complications can be studied following multiple pregnancies over time and comparing pregnancies with and without complications within the same woman. Life events, drug use, social circumstances, urban environment, and other dynamically changing environmental exposures have only rarely been examined in a within-person design. Our findings suggest that increased reliance on within-person designs may be in order for investigating time-varying environmental factors.
Limitations
While fixed-effect models do not suffer from omitted variable bias, they have low power as only individuals with within-person changes in exposure/outcome status over time are informative. Although we showed that the sample was sufficiently enriched with these dynamic changes in exposure and outcome, the within-person analysis of the cannabis-psychosis association was carried out in 827 observations pertaining to 375 individuals, whereas the random-effects analysis was carried out in 15 169 observations pertaining to 6240 individuals (Table 6). As a result, CIs were wider for the fixed-effect analysis results. In addition, fixed-effect analyses may still be confounded by time-varying factors, not all of which may have been captured in the current analyses. For example, high collinearity between cannabis and tobacco smoking did not allow for the control of tobacco smoking, which has been suggested as a potential risk factor for psychosis even though its psychoactive effects do not include psychosis.
Attenuated psychotic states, the main component of “clinical high risk” states, are the focus of much research focused on early intervention and prevention.8–11 The current results, therefore, contribute to this rapidly developing area of research. As hypotheses were examined at the level of attenuated psychotic experiences and not clinical psychosis, results may not be generalizable to the full clinical syndrome. However, attenuated psychotic experiences show temporal, genetic, and broader etiological continuity with clinical psychosis43–45—which is why they are commonly used to investigate genetic and nongenetic hypotheses regarding the psychosis phenotype, given evidence that they pertain to the same spectrum as the psychotic syndrome, just as anxiety and depression phenotypes are thought to represent spectrum phenotypes with a half-normal distribution in the general population.44
Acknowledgments
The authors have no conflicts of interest to declare
Funding
NEMESIS-2 was supported by the Ministry of Health, Welfare and Sport, with supplementary support from the Netherlands Organization for Health Research and Development (ZonMw). The EDSP is funded by the German Ministry of Research and Education (BMBF) as part of the Addiction Research Initiative (01EB94056). This work was supported by the European Community’s Seventh Framework Program under grant agreement no. HEALTH-F2-2009-241909 (Project EU-GEI).
References
- 1. Marconi A, Di Forti M, Lewis CM, Murray RM, Vassos E. Meta-analysis of the association between the level of cannabis use and risk of psychosis. Schizophr Bull. 2016;42:1262–1269. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Hasan A, von Keller R, Friemel CM, et al. . Cannabis use and psychosis: a review of reviews. Eur Arch Psychiatry Clin Neurosci. 2020;270(4):403–412. [DOI] [PubMed] [Google Scholar]
- 3. Hoffman L. Longitudinal Analysis: Modeling Within-Person Fluctuations and Change. New York, NY: Taylor & Francis; 2014. [Google Scholar]
- 4. Fergusson DM, Horwood LJ, Ridder EM. Tests of causal linkages between cannabis use and psychotic symptoms. Addiction. 2005;100(3):354–366. [DOI] [PubMed] [Google Scholar]
- 5. Bechtold J, Hipwell A, Lewis DA, Loeber R, Pardini D. Concurrent and sustained cumulative effects of adolescent marijuana use on subclinical psychotic symptoms. Am J Psychiatry. 2016;173(8):781–789. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Ferdinand RF, Sondeijker F, van der Ende J, Selten JP, Huizink A, Verhulst FC. Cannabis use predicts future psychotic symptoms, and vice versa. Addiction. 2005;100(5):612–618. [DOI] [PubMed] [Google Scholar]
- 7. Allisson P. Fixed Effects Regression Models. Thousand Oaks, CA: SAGE; 2009. [Google Scholar]
- 8. Allswede DM, Addington J, Bearden CE, et al. . Characterizing covariant trajectories of individuals at clinical high risk for psychosis across symptomatic and functional domains. Am J Psychiatry. 2020;177(2):164–171. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Gordon JA, Heinssen RK. Collaborative approaches to the clinical high-risk state: from data to mechanism to intervention. Biol Psychiatry. 2020;88(4):287–288. [DOI] [PubMed] [Google Scholar]
- 10. McGorry PD, Nelson B. Clinical high risk for psychosis-not seeing the trees for the wood. JAMA Psychiatry. 2020;77(7):559–560. [DOI] [PubMed] [Google Scholar]
- 11. Mongan D, Focking M, Healy C, et al. . Development of proteomic prediction models for transition to psychotic disorder in the clinical high-risk state and psychotic experiences in adolescence. JAMA Psychiatry. 2021;78:77–90. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. World Health Organization. Composite International Diagnostic Interview (CIDI). Geneva, Switzerland: World Health Organization, Division of Mental Health; 1990. [Google Scholar]
- 13. de Graaf R, Ten Have M, van Dorsselaer S. The Netherlands mental health survey and incidence study-2 (NEMESIS-2): design and methods. Int J Methods Psychiatr Res. 2010;19(3):125–141. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Lieb R, Isensee B, von Sydow K, Wittchen HU. The early developmental stages of psychopathology study (EDSP): a methodological update. Eur Addict Res. 2000;6(4):170–182. [DOI] [PubMed] [Google Scholar]
- 15. Wittchen HU, Perkonigg A, Lachner G, Nelson CB. Early developmental stages of psychopathology study (EDSP): objectives and design. Eur Addict Res. 1998;4(1-2):18–27. [DOI] [PubMed] [Google Scholar]
- 16. Zimmermann P, Brückl T, Lieb R, et al. . The interplay of familial depression liability and adverse events in predicting the first onset of depression during a 10-year follow-up. Biol Psychiatry. 2008;63(4):406–414. [DOI] [PubMed] [Google Scholar]
- 17. de Graaf R, Ten Have M, van Gool C, van Dorsselaer S. [Prevalence of mental disorders, and trends from 1996 to 2009. Results from NEMESIS-2]. Tijdschr Psychiatr. 2012;54:27–38. [PubMed] [Google Scholar]
- 18. De Graaf R, van Dorsselaer S, Tuithof M, Ten Have M.. Sociodemographic and Psychiatric Predictors of Attrition in the Third Follow-up of the Netherlands Mental Health Survey and Incidence Study-2 (NEMESIS-2). Utrecht, the Netherlands: Trimbos Instituut; 2018. [Google Scholar]
- 19. de Graaf R, van Dorsselaer S, Tuithof M, ten Have M. Sociodemographic and psychiatric predictors of attrition in a prospective psychiatric epidemiological study among the general population. Result of the Netherlands Mental Health Survey and Incidence Study-2. Compr Psychiatry. 2013;54(8):1131–1139. [DOI] [PubMed] [Google Scholar]
- 20. Eaton WW, Neufeld K, Chen LS, Cai G. A comparison of self-report and clinical diagnostic interviews for depression: diagnostic interview schedule and schedules for clinical assessment in neuropsychiatry in the Baltimore epidemiologic catchment area follow-up. Arch Gen Psychiatry. 2000;57(3):217–222. [DOI] [PubMed] [Google Scholar]
- 21. Reed V, Gander F, Pfister H, et al. . To what degree does the Composite International Diagnostic Interview (CIDI) correctly identify DSM-IV disorders? Testing validity issues in a clinical sample. Int J Methods Psychiatr Res. 1998;7:142–155. [Google Scholar]
- 22. Haro JM, Arbabzadeh-Bouchez S, Brugha TS, et al. . Concordance of the Composite International Diagnostic Interview Version 3.0 (CIDI 3.0) with standardized clinical assessments in the WHO World Mental Health surveys. Int J Methods Psychiatr Res. 2006;15(4):167–180. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Wittchen HU. Reliability and validity studies of the WHO–Composite International Diagnostic Interview (CIDI): a critical review. J Psychiatr Res. 1994;28(1):57–84. [DOI] [PubMed] [Google Scholar]
- 24. van der Steen Y, Myin-Germeys I, van Nierop M, et al. . ‘False-positive’ self-reported psychotic experiences in the general population: an investigation of outcome, predictive factors and clinical relevance. Epidemiol Psychiatr Sci. 2019;28(5):532–543. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. van Nierop M, van Os J, Gunther N, et al. . Phenotypically continuous with clinical psychosis, discontinuous in need for care: evidence for an extended psychosis phenotype. Schizophr Bull. 2012;38(2):231–238. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Bak M, Delespaul P, Hanssen M, de Graaf R, Vollebergh W, van Os J. How false are “false” positive psychotic symptoms? Schizophr Res. 2003;62(1-2):187–189. [DOI] [PubMed] [Google Scholar]
- 27. Radhakrishnan R, Guloksuz S, Ten Have M, et al. . Interaction between environmental and familial affective risk impacts psychosis admixture in states of affective dysregulation. Psychol Med. 2019;49(11):1879–1889. [DOI] [PubMed] [Google Scholar]
- 28. Pries LK, Guloksuz S, Ten Have M, et al. . Evidence that environmental and familial risks for psychosis additively impact a multidimensional subthreshold psychosis syndrome. Schizophr Bull. 2018;44(4):710–719. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Honings S, Drukker M, van Nierop M, et al. . Psychotic experiences and incident suicidal ideation and behaviour: disentangling the longitudinal associations from connected psychopathology. Psychiatry Res. 2016;245:267–275. [DOI] [PubMed] [Google Scholar]
- 30. Dominguez MD, Saka MC, can Saka M, Lieb R, Wittchen HU, van Os J. Early expression of negative/disorganized symptoms predicting psychotic experiences and subsequent clinical psychosis: a 10-year study. Am J Psychiatry. 2010;167(9):1075–1082. [DOI] [PubMed] [Google Scholar]
- 31. Cougnard A, Marcelis M, Myin-Germeys I, et al. . Does normal developmental expression of psychosis combine with environmental risk to cause persistence of psychosis? A psychosis proneness-persistence model. Psychol Med. 2007;37(4):513–527. [DOI] [PubMed] [Google Scholar]
- 32. Henquet C, Krabbendam L, Spauwen J, et al. . Prospective cohort study of cannabis use, predisposition for psychosis, and psychotic symptoms in young people. BMJ. 2005;330(7481):11. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Spauwen J, Krabbendam L, Lieb R, Wittchen HU, van Os J. Impact of psychological trauma on the development of psychotic symptoms: relationship with psychosis proneness. Br J Psychiatry. 2006;188:527–533. [DOI] [PubMed] [Google Scholar]
- 34. van Winkel R; Genetic Risk and Outcome of Psychosis (GROUP) Investigators . Family-based analysis of genetic variation underlying psychosis-inducing effects of cannabis: sibling analysis and proband follow-up. Arch Gen Psychiatry. 2011;68(2):148–157. [DOI] [PubMed] [Google Scholar]
- 35. Kuepper R, van Os J, Lieb R, Wittchen HU, Höfler M, Henquet C. Continued cannabis use and risk of incidence and persistence of psychotic symptoms: 10 year follow-up cohort study. BMJ. 2011;342:d738. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Brugha T, Bebbington P, Tennant C, Hurry J. The list of threatening experiences: a subset of 12 life event categories with considerable long-term contextual threat. Psychol Med. 1985;15(1):189–194. [DOI] [PubMed] [Google Scholar]
- 37. Pries LK, van Os J, Ten Have M, et al. . Association of recent stressful life events with mental and physical health in the context of genomic and exposomic liability for Schizophrenia. JAMA Psychiatry. 2020;77(12):1296–1304. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Wittchen HU, Essau CA, Hecht H, Teder W, Pfister H. Reliability of life event assessments: test-retest reliability and fall-off effects of the munich interview for the assessment of life events and conditions. J Affect Disord. 1989;16(1):77–91. [DOI] [PubMed] [Google Scholar]
- 39. Wooldridge JM. Introductory Econometrics: A Modern Approach. 5th ed. Mason, OH: South-Western, Cengage Learning; 2013. [Google Scholar]
- 40. Hamaker EL, Muthén B. The fixed versus random effects debate and how it relates to centering in multilevel modeling. Psychol Methods. 2020;25(3):365–379. [DOI] [PubMed] [Google Scholar]
- 41. StataCorp. Stata Longitudinal-Data/Panel-Data Reference Manual Release 16. STATA Statistical Software: Release 16. College Station, TX: StataCorp LLC; 2019. [Google Scholar]
- 42. Curran PJ, Bauer DJ. The disaggregation of within-person and between-person effects in longitudinal models of change. Annu Rev Psychol. 2011;62:583–619. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43. Guloksuz S, Pries LK, Ten Have M, et al. . Association of preceding psychosis risk states and non-psychotic mental disorders with incidence of clinical psychosis in the general population: a prospective study in the NEMESIS-2 cohort. World Psychiatry. 2020;19(2):199–205. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44. Linscott RJ, van Os J. An updated and conservative systematic review and meta-analysis of epidemiological evidence on psychotic experiences in children and adults: on the pathway from proneness to persistence to dimensional expression across mental disorders. Psychol Med. 2013;43(6):1133–1149. [DOI] [PubMed] [Google Scholar]
- 45. Zavos HM, Freeman D, Haworth CM, et al. . Consistent etiology of severe, frequent psychotic experiences and milder, less frequent manifestations: a twin study of specific psychotic experiences in adolescence. JAMA Psychiatry. 2014;71(9):1049–1057. [DOI] [PMC free article] [PubMed] [Google Scholar]