The cohort comprised offspring of women participating in the Kaiser Foundation Health Plan (KFHP) who received obstetric care and were delivered in Alameda County, California, between 1959 and 1967, and who agreed to participate in the Child Health and Development Study (CHDS) (n
=19 044). The KFHP was one of the first and largest health insurance plans in the USA, and nearly all pregnant women who were members participated in the CHDS. A comparison of these participants' demographic data with US census data demonstrated that although both income extremes were underrepresented, members were demographically similar with respect to ethnicity, occupation and educational attainment (Krieger, 1992
The study reported here is restricted to a cohort referred to as the Prenatal Determinants of Schizophrenia (PDS) study cohort, which is a subsample of the offspring of those participating in the CHDS. The design for the PDS study cohort has been previously described in full (Susser et al, 2000
); thus, only a summary is provided here. The PDS study cohort was designed to follow up and assess participants for the presence of schizophrenia-spectrum disorders in order to evaluate developmental determinants of schizophrenia. It included offspring of mothers from the CHDS cohort who were members of KFHP from 1 January 1981 to 31 December 1997 (n
=12 094). These dates correspond to the period of case ascertainment, which commenced when KFHP began using computerised records, making it feasible to identify all members who had accessed mental health services. Participants in the PDS cohort are comparable with the CHDS sample, with two exceptions: mothers of African American ethnicity were more likely to be included in the PDS cohort (CHDS 16%, PDS 28%) and offspring of low-income unmarried mothers were somewhat less likely to be included (CHDS 26, PDS 20%).
Height in inches to the nearest sixteenth of an inch (1.6 mm) was measured and recorded during regular paediatric visits from birth to age 13 years. These measurements were systematically abstracted from medical records by trained CHDS abstractors. The present analyses were restricted to data from birth to age 9 years, since reliable data on the pubertal stage of development were not available. The average number of height measurements between birth and age 2½ years was 9 for both the schizophrenia-spectrum disorder (SSD) group and the non-SSD group ranging from 3 to 15 measurements for the former and 1 to 31 for the latter. Between ages 2½ years and 9 years the average number of height measurements was 6 for both groups, ranging from 1 to 19 measurements for the SSD group and 1 to 41 for the non-SSD group. To identify outliers, height was standardised by computing gender- and age-adjusted means using the least mean squares (LMS) method (Cole, 1990
). This method is advantageous for standardising height because it uses a series of calculations to reduce asymmetry in skewed data. Height measurements that were greater than 4 standard deviations above or below the mean were considered to be outliers and were excluded from the analyses; 1.0% of all measurements were excluded.
Assessment of potential confounders
Potential confounding factors were determined a priori based on characteristics that have been shown to be associated with both height and schizophrenia. These included gender, maternal race and education, standardised maternal height, pre-pregnancy body mass index (BMI), gestational age at birth and birth weight. (It is unnecessary to control for birth length in the adjusted analysis because it is used to estimate growth patterns.) Demographic measures were assessed through maternal interview, which was completed during the first prenatal visit.
Maternal race was categorised as White, Black or other. Maternal education was rated on a seven-point scale reflecting the highest level of education achieved. Because not all levels of education were represented among the SSD sample, the categories were collapsed as follows: less than high-school diploma, high-school graduate with or without trade school, high-school graduate plus 1–3 years of college, and college graduate.
Maternal height was measured during the maternal interview. Because paternal height was based on maternal report and we had insufficient data on paternal height, standardised maternal height was used as a proxy for the child's genetic growth potential. Standardisation of maternal height was accomplished using the LMS method. Because adult height is not typically achieved until 20 years of age, z
-score transformations of height for mothers aged 15–19 years were calculated separately at 1-year intervals. Mothers aged 20 years or over were assumed to have attained their adult height and were standardised as one group irrespective of age. Maternal pre-pregnancy BMI was calculated using self-reported weight prior to pregnancy and was classified based on categories used in a previous study assessing the association between maternal BMI and schizophrenia-spectrum disorders in this cohort: low, ≤ 19.9 kg/m2
; average, 20.0–26.9 kg/m2
; greater than average, 27.0–29.9 kg/m2
; high, ≥ 30.0 kg/m2
) (Schaefer et al, 2000
). Maternal prepregnancy BMI data were missing in a proportion of the sample (8 SSD and 1166 non-SSD). To preserve sample size, the average maternal BMI value was substituted for missing data in the adjusted analyses. Gestational age at birth was calculated as the number of days between the last reported menstrual period and birth.
Ascertainment and diagnosis
We identified potential cases of schizophrenia-spectrum disorder through the KFHP computerised records of in-patient, out-patient and pharmacy registries. With regard to the hospitalisation registry, potential cases were first identified if the individual had received ICD–9 diagnosis codes of 295, 296, 297, 298 or 299 (World Health Organization, 1978
) or were not given a specific diagnosis. A review of psychiatric and medical records by a psychiatrist was conducted to determine whether individuals screened positive for evidence of a psychotic disorder. Individuals from the out-patient registry were considered to be screen-positive for SSD if they had diagnosis codes of 295, 297, 298 or 299. For the pharmacy registry, cases screened positive if the individual had received treatment with antipsychotic medication.
We identified 183 participants for further diagnostic assessment. Of those identified, 13 had died. Of the 170 remaining, 146 (86%) were successfully contacted and 107 (58% of those originally identified) completed the Diagnostic Interview for Genetic Studies (DIGS; Nurnberger et al, 1994
) administered by a trained research clinician. For the remaining 76 (42%) individuals who were not interviewed, a diagnosis was made based on review of the medical records by trained clinicians. All individuals provided written informed consent for participation prior to the diagnostic interview. Informed consent was approved by the institutional review boards of the New York State Psychiatric Institute and the Kaiser Permanente Division of Research. Diagnoses were made by consensus of three diagnosticians who independently reviewed all relevant material for each case. In total, 71 cases were identified (43 schizophrenia, 17 schizoaffective disorder, 5 schizotypal disorder, 1 delusional disorder and 5 other schizophrenia-spectrum psychosis). Forty-four individuals completed the DIGS and 27 were diagnosed by chart review.
The PDS cohort comprised 12 094 individuals – 71 SSD and 12 023 non-SSD. Because our analysis required information obtained through maternal interview during enrolment, people who had not completed the interview were excluded (n
= 2412), reducing the sample size to 9682. Since siblings represent non-independent observations, only one sibling per family was included in the analysis and siblings of offspring with schizophrenia-spectrum disorder were excluded (n
=1886). Two people in the SSD group were siblings, so one of them was randomly selected and the data excluded from the analysis. The final cohort consisted of 7795 persons. Please see Susser et al (2000)
for a further description of the method for deriving the analytic sample. From this cohort only 15 individuals (all from the non-SSD group) were excluded because they did not have at least one valid height measurement; thus, the analytic sample comprised 7780 persons (SSD n
=70; non-SSD n
=7710). A summary of the exclusion criteria for the analytical data-set is provided in .
Study profile (PDS, Prenatal Determinants of Schizophrenia).
The relationship between growth and adult schizophrenia was examined using multilevel growth models (McArdle & Aber, 2000
; Cohen et al, 2003
; Chen & Cohen, 2006
). The PROC MIXED function in the SAS statistical package (Littell et al, 1996
) was used to estimate the growth pattern of height. In these multilevel growth models, longitudinal data on individuals are considered the basic ‘random’ data, similar to a cross-sectional study where single individual variables are the basic units of analyses (Cohen et al, 2003
The first step in the analyses was to determine the best model fit for the relationship between growth and age from birth to 2½ years, which is referred to as the basic growth model. We first examined the mean level model, which is the between-participant differences in mean level of height (i.e. height=β0+error). The results suggested that between participants differences were found for height; the estimated variance of the mean height was 2.06 (s.e.=0.28, P<0.001). To determine the best model for predicting growth, we compared three additional models: linear change in height (height=β1+age+error), quadratic change in height (height=β2+ age+age2+error) and cubic change in height (height=β3+age+age2+age3+error), where random variation was permitted for each age term. To estimate model fit, we calculated the chi-squared value by subtracting the −2 log likelihood estimates from the subsequent models (i.e. mean level v. linear, linear v. quadratic, and quadratic v. cubic). Compared with the mean level model, the addition of the linear term significantly improved model fit (P<0.001). The addition of the quadratic term also significantly improved model fit compared with the linear model (P<0.001). Also, the cubic model appeared to significantly improve model fit compared with the quadratic model (P<0.001). However, the fixed effect of the cubic term was small and the significant improvement mostly due to the random effects. Also, the addition of the cubic term reduced the variance associated with the linear term to zero. Therefore, to facilitate interpretation, and in the interest of comparability and parsimony, we chose to use the quadratic model. Although fractional polynomial models are sometimes used in this type of analysis, a quadratic equation appears to be an accurate estimate of growth and is easier to interpret. In summary, the model comparisons suggested that the quadratic model provided the best model fit for estimating patterns of growth.
Once the basic model for growth was established, we fitted a conditional model to examine the effect of schizophrenia-spectrum disorders on growth. We added SSD (1, case, 0, non-case) as a covariate to examine mean level differences in height. To assess growth differences between the SSD and non-SSD groups, we included an interaction term for age and SSD and an interaction term for (age)2 and SSD in the quadratic model. All adjusted analyses controlled confounding effects of gender, maternal education, race, BMI and height, gestational age at birth and birth weight. Each model included a χ2-test of improvement of fit to the data.
Using the methods described above we also estimated the basic model for growth from age 2½ to 9 years. Separate analyses were performed for early life (birth to age 2½ years) and later childhood (2½ to 9 years) because growth during these periods is regulated by different mechanisms (Karlberg, 1987
; Reiter & Rosenfeld, 2003
). Additionally, previous research suggests that growth during the first 2 years of life is more variable than growth in later childhood and abnormal growth patterns may be indicative of problems during prenatal development (Tanner, 1994
; Reiter & Rosenfeld, 2003
). Hence, we predicted that early life (birth to 2½ years) would be associated with the most pronounced growth deficit in individuals who later developed schizophrenia. Previous studies also demonstrated gender differences in growth patterns in healthy samples (Karlberg, 1987
) and in the risk and correlates of schizophrenia (Goldstein et al 2002
; Aleman et al, 2003
). Thus, all analyses were performed on the entire cohort as well as being further stratified by gender in order to determine whether differences in growth patterns between individuals in the SSD and non-SSD groups varied by gender.
In addition to assessments using the multilevel growth model, we calculated the linear slope (linear change in height by age) for each participant by using maximum likelihood estimates from the basic growth models for both early life (birth to age 2½ years) and later childhood (2½ to 9 years). A feature of this program is the ability to output the empirical Bayesian estimates of the linear slope of height for each participant over the defined periods (McArdle et al, 2005
). The linear slopes provide additional information about differences between the SSD and non-SSD groups observed from the multilevel models.