2.5. Statistical Analyses

Frequencies and descriptive statistics of all women were expressed as n (%) or as means (standard deviation, SD). Median (inter-quartile range, IQR) was reported when continuous variables were not normally distributed. The impact of each trace element (in categories) on TTP was investigated using accelerated failure time models, as previously published [19,20] with log normal distribution to estimate time ratios (TR) and 95% CIs. Adjusted marginal estimates for median TTP according to trace element concentrations were calculated using the Stata ‘margins’ command, together with corresponding 95% CIs. Alternative distributions were investigated but the log normal distribution was selected based on providing the lowest −2 log likelihood and Akaike information criterion value. These TRs can be interpreted as the ratios of the median values of the duration (in months) to achieve pregnancy between the compared groups. A TR above 1 implies that a given exposure is associated with longer TTP, whereas a TR below 1 indicates a shorter TTP.

We used causal diagrams (directed acyclic graphs) to guide selection of potential confounders for which to control, based on a priori selection of variables associated with trace element concentrations and fertility. Three separate adjusted models were used. Model 1: maternal age, maternal body mass index (BMI), ethnicity, socioeconomic status, plasma CRP, pre-pregnancy alcohol consumption, pre-pregnancy smoking status, frequency of sexual intercourse prior to conception, multivitamin use in first trimester, and trace element of interest in multivitamin; Model 2: Model 1 plus pre-pregnancy intake of fast food, green leafy vegetables, fruit, and fish; Model 3: Model 2 plus paternal age and paternal BMI. This was a complete case analysis.

Statistical significance was defined as a two-sided p-value of <0.05. All statistical analyses were undertaken using Stata IC 14 (Stata, College Station, TX, USA).