Exposure to methamphetamine

Amphetamine-type stimulant exposure since the previous wave of data collection was ascertained at age 18, 21, 25, 30, and 35 for the age periods 16–18; 18–21; 21–25; 25–30; and 30–35 respectively. Participants were asked whether they had ever used ‘methamphetamine, speed, P, ice, etc’ in that period, and if so, how often they had used. Most of this stimulant use is likely to have been methamphetamine, since most participants lived in New Zealand during the follow-up period and methamphetamine was by far the most commonly used amphetamine-type stimulant available in New Zealand over that time (Wilkins, Sweetsur, & Griffiths, Reference Wilkins, Sweetsur and Griffiths2017). An ordinal methamphetamine use variable (no use; once or twice only; no more than monthly; weekly or more frequent) was used in a subsequent analysis on the dose−response profile summed across follow up periods (see Statistical Analysis). Repeated measures analyses used a dichotomous methamphetamine use variable (use/no use) due to the low number of participants who reported using methamphetamine on a regular basis.

Psychosis outcomes

At each assessment from age 18 to age 35, cohort members were administered a comprehensive mental health interview designed to assess a number of aspects of the individual's mental health and psychosocial adjustment. As part of this interview, participants were questioned regarding psychotic symptomatology. For ages 18, 21, and 25, cohort members were asked to report on their symptoms over the past month. At ages 30 and 35, participants were asked to report on their symptoms since the previous assessment (the difference in time frames was corrected in the statistical analyses; see below). At ages 18, 21, and 25 these questions were derived from the by the Symptom Checklist 90 (SCL-90; Derogatis, Lipman, & Covi, Reference Derogatis, Lipman and Covi1973), while at ages 30 and 35 questions were derived from the Diagnostic Interview Schedule (DIS) (Robins, Cottler, Bucholz, & Compton, Reference Robins, Cottler, Bucholz and Compton1995). These items spanned a range of symptom areas, including: hallucinations and delusions; paranoid ideation; and related symptoms. Confirmatory factor analysis of the SCL-90 item set has shown previously that the items formed a unidimensional scale reflecting the extent of psychotic symptomatology (Fergusson, Horwood, & Swain-Campbell, Reference Fergusson, Horwood and Swain-Campbell2003). These measures were used to generate two outcome measures: a dichotomous measure indicating whether the participant reported at least one symptom at each assessment; and a count of the number of symptoms of psychosis experienced by each participant during each assessment period. The reliability of the scales at each age was moderate (coefficient α ranged from 0.71 to 0.75). The items used for each measure are presented in the online Supplementary Tables S1 and S2.

Individual factors and early environmental confounders

A series of covariate factors was selected from the CHDS database on the basis of their known or posited association with the exposure variable or outcomes. These factors were:

1. 1. Family socio-demographic background: family living standards (0–10 years).

2. 2. Family functioning: family instability (changes of parents 0–16 years); parental intimate partner violence (0–16 years); parental history of alcohol problems, criminality and illicit drug use; quality of parenting/relationship with parents (parental bonding 0–16 years).

3. 3. Child abuse: severity of exposure to childhood sexual abuse (0–16 years) and physical abuse (parental use of physical punishment 0–16 years).

4. 4. Childhood characteristics: sex; severity of childhood conduct and attention problems (7–13 years).

5. 5. Adolescent adjustment: self-esteem (15 years); novelty seeking (16 years); history of conduct disorder (14–16 years), anxiety disorder (14–16 years); and affiliations with delinquent or substance using peers (15 years).

Detailed description of these measures is provided in the Online Supplementary Material.

Time-dynamic covariate factors (ages 18–21, 21–25, 25–30, and 30–35 years)

To control for time-dynamic factors contributing to the associations between methamphetamine and symptoms of psychosis during each assessment period (ages 16–18, 18–21, 21–25, 25–30, and 30–35 years), covariate factors were chosen from the study database. These included: major depression; anxiety disorder; alcohol use disorder; frequency of cannabis use; and life stress; (see Online Supplementary Material for details).

Statistical analyses

First, unadjusted analyses were performed, tabulating the raw percentage of participants reporting symptoms of psychosis in each of the five assessment periods, according to whether the participant reported any methamphetamine use in that time period. The population-averaged odds ratios (ORs) for the dichotomous outcome, and incidence rate ratio (IRR) for the count measure across the five assessment periods were then estimated using logistic (for the dichotomous measure) and negative binomial (for the count measure) generalized estimating equations (GEE; Liang & Zeger, Reference Liang and Zeger1986), fitted in Stata version 15.0. Models treated the five repeated measures of methamphetamine exposure and the five repeated measures of symptoms of psychosis as time-dynamic variables. GEE models deal with the issue of missing data by including all participants with at least one point of repeated measures data.

The models fitted were of the form:

(1)$$f\,{\rm Yit\ } = {\rm B}0 + {\rm B}1{\rm Xit}$$

where Yit was log odds (for the dichotomous measure) or the log rate (for the count measure) of the outcome for person i at time t, B0 was the intercept term, and B1Xit was the estimate of the effect of methamphetamine on psychotic symptomatology for person i at time t (error terms not shown). Estimates of the OR and IRR were obtained by exponentiation of the estimate B1 (e^b).

Second, two series of adjusted GEE analyses were then performed, including covariates representing (a) individual factors and early environmental exposures likely to confound the association between methamphetamine use and psychotic symptoms (Model 1; see above for list of these factors), and (b) those covariates listed in (a), plus covariates representing time-dynamic exposures from age 16–35 (Model 2; see above for list of these factors). These models were of the form:

(2)$$f\,{\rm Yit\ } = {\rm B}0 + {\rm B}1{\rm Xit\ } + \Sigma {\rm BjXj\ } + \Sigma {\rm BkXit}$$

where ΣBjXj represents the effects of confounding factors on the association (Model 1), and ΣBkXit represents time-dynamic covariation for person i at time t (Model 2).

These models also included a scale (SCL-90/DIS)*methamphetamine exposure interaction term, to ensure that the strength of the association between methamphetamine exposure and psychosis outcomes did not differ according to which scale was used to assess psychotic symptomatology (SCL-90 at ages 18, 21, 25; DIS at ages 30, 35). An estimate of the E-value, which estimates the amount of unmeasured confounding that would render the methamphetamine-psychosis link to be rendered null, was obtained via the “evalue’ procedure in Stata using the OR model.

Third, a model assessing the dose–response profile of methamphetamine and symptoms of psychosis was fitted. Data sparseness for the more frequent methamphetamine use categories (particularly in the two later age periods) prevented repeated measures analyses from being performed on this series. Instead, a lifetime (age 16–35) maximum methamphetamine use frequency was computed according to the most frequent use reported in any of the five age periods. A binary variable representing the lifetime (age 16–35) reporting of at least one symptom of psychosis was also created. The associations between greatest methamphetamine use frequency and psychotic symptoms were then estimated using logistic regression, after adjustment for fixed confounding factors selected on the basis that these were statistically significant (p < 0.05) multivariate predictors of psychosis outcomes in Model 1 described above. Dummy variables were used to represent levels of methamphetamine use, with pairwise tests of statistical significance derived from chi-squared likelihood ratio tests.

For the purposes of conducting a sensitivity analysis, the two most commonly reported SCL-90 items that may lack face validity (‘having ideas or beliefs that others do not share’ and ‘never feeling close to another person’) were removed from the scale scores, and the analyses were repeated using the revised scale scores.

Estimates of statistical power

With an effective sample size of n = 1056, it was calculated that the analyses had at least 90% power to detect odds ratios of 1.2 or higher.

Population attributable fraction (PAF) estimate

The PAF is a measure of the extent to which (assuming causality), the amount of an exposure that would be prevented if the exposure had not existed. The PAF for methamphetamine in relation to the outcomes considered in this study was estimated using the following formula (Khosravi & Mansournia, Reference Khosravi and Mansournia2019):

(3)$${\rm PAF} = pc( {{\rm AO}{\rm R}_{{\rm dj}}- 1} ) /{\rm AOR}$$

where pc is the lifetime prevalence of methamphetamine use among those with the outcome, and AOR is the adjusted odds ratio of the outcome for those with a history of methamphetamine use compared to those without. The population-averaged estimate of the association between methamphetamine use and symptoms of psychosis at age 16–35 from Model 2 (see above) was used to calculate the AOR.

Proximal/distal use of methamphetamine

In order to examine whether earlier methamphetamine use has an effect on psychotic symptomatology later in life, the cohort members were classified into three groups: those never using methamphetamine; those who used methamphetamine prior to age 30, but not during the period 30–35 years; and those who used methamphetamine both prior to age 30 and during the period 30–35 years. The percentage reporting at least one symptom of psychosis during the period was compared across these groups using the Mantel−Haenszel test of significance. The comparison was then repeated using only those cohort members who had reported at least one symptom of psychosis prior to age 30.

Sample size and sample bias

The sample sizes for the present analyses were based on cohort members with at least one completed wave of data at ages 18, 21, 25, 30, and 35 years. These sample sizes were 1025 (age 18), 1011 (age 21), 1003 (age 25), 987 (age 30) and 962 (age 35), representing 79–82% of the surviving cohort at each age. To examine the effects of sample losses on the representativeness of the sample, the obtained samples with complete data at each age, were compared with the remaining sample members on a series of socio-demographic measures collected at birth. This analysis suggested that there were statistically significant (p < 0.01) tendencies for the obtained samples to under-represent individuals from socially disadvantaged backgrounds characterized by low parental education, low socio-economic status and single parenthood. To address this issue, the data weighting methods described by Carlin, Wolfe, Coffey, and Patton (Reference Carlin, Wolfe, Coffey and Patton1999) were used to examine the possible implications of selection effects arising from the pattern of missing data. These analyses produced essentially the same pattern of results to those reported here, suggesting that the conclusions of this study were unlikely to have been influenced by selection bias.