Bibian van der Voorn

126 CHAPTER 9 In STEP-2, total serum cortisol was measured using a competitive immunoassay (Luminescence Advia Centaur, Siemens Medical Solutions Diagnostics, USA) with an intra-assay CV of 3% (at a level of 700 nmol/L), an inter-assay CV of 7%, 6% and 6% (at levels of 80, 300 and 1,000 nmol/L, respectively), and a LLOQ of 30 nmol/L. Salivary samples were stored at −80°C and thawed just before analyses. Free cortisol in saliva was measured using a competitive immunoassay (Luminescence Architect, Abbott Laboratories, Diagnostics Division Abbott Park, Illinois, USA) with an intra- assay CV of 9% (at a level of 5 nmol/L), an inter-assay CV of 11% (at a level of 7 nmol/L), and a LLOQ of 1 nmol/l. All laboratory tests were performed by the Endocrine Laboratory of VUMC, Amsterdam, The Netherlands. DATA ANALYSIS To assess the effect of attrition at follow-up, baseline characteristics of participants, non-participants and excluded subjects were compared using one-way ANOVA, Chi- square or Kruskal-Wallis tests, with STEP-2 participants as reference group ( Table 1 ). Since cortisol is associated with BMI, we compared BMI at age 8 y between the groups. No significant differences were found between SGA and AGA groups ( p =0.560), and we therefore decided not to adjust our cross-sectional analyses at age 8 y for BMI. SERUM CORTISOL To test whether groups differed at any of the time points, cross-sectional, univariable, linear regression analyses were performed with either cortisol at term age, 3 mo. or 6 mo. corrected age, or 8 y as dependent factor. Subsequently, generalized estimating equations (GEEs) were used for longitudinal analysis of cortisol, i.e., the assessment of differences between groups, adjusted for intra-individual variation over time. We assumed that attrition at age 8 y resulted in ‘missings completely at random’ and therefore used all available data of participants of the original RCT (n=152) without exclusion of dropouts, while accounting for ‘missings completely at random’ by use of GEE. GEE is designed for the handling of missing data, provided that missingness is completely at random 20,21 . In addition, GEE adjusts for grouped samples, collected from the same subject at different times, by using a correlation structure. For our data analyses, we chose an exchangeable correlation structure, in which one average within-subject correlation between samples over time is assumed. Stepwise GEEs were performed with cortisol over time (at term age, 3 mo. corrected age, 6 mo. corrected age, and 8 y) as dependent, continuous factor. First, the association between birth weight SDS and cortisol was