The Maternity Experiences Survey is a population-based survey conducted by Statistics Canada between October 23, 2006, and January 31, 2007 on behalf of the Canadian Perinatal Surveillance System, Public Health Agency of Canada. The overarching goal of the MES was to fill knowledge gaps by collecting comprehensive information on practices, factors, and maternal experiences before, during and after pregnancy [28–30].
The MES target population consisted of biological mothers who were 15 years of age and older at the time of their babies' singleton live birth in Canada, and who also lived with their infants at the time of the survey . A stratified simple random sample was selected without replacement, using recent births drawn from the Census 2006 sampling frame. The sample was stratified on province or territory in which the mother resided at the time of the census and on maternal age (<20 years, > 20 years). Among 8,542 women selected from the census frame, the screening process for determining eligibility (i.e., did not have the baby in Canada or were not living with the baby by the time of the interview) was completed during data collection for 6,937 women, of which 234 were determined to be out-of-scope, leaving 6,703 women eligible to participate in the survey. Among these, 6,421 women who completed the questionnaire between five and fourteen months postpartum and gave Statistics Canada permission to provide their information to the Public Health Agency of Canada were classified as respondents . The number of eligible cases was estimated by applying the proportion of eligible cases among those screened (96%) to the unscreened, resulting in an estimated 8,244 eligible cases. It was estimated that 1,541 (0.96 * 1,605) of the unscreened cases would have been eligible to participate in the survey. The number of estimated in-scope cases was therefore 8,244 (6,421 + 1,541) [29, 30]. After applying the survey weights, these women represented approximately 76,500 Canadian women. The response rate was 77.9%. Compared to women who completed the questionnaires, non-respondents were more likely to be teenagers or older than 39 years, have a first language other than English or French, live in Toronto and be single. To improve the representation of the sample and reduce selection bias, total non-response was handled by adjusting the weight of individuals who responded to the survey to compensate for those who did not respond. For example, the proportion of mothers who did not speak English or French was 15.0% for the respondents and 27.0% for the non-respondents. After the weighting adjustments, the proportion was 16.9% for the respondents compared to 17.2% for the MES frame . Most questionnaires were completed in a 45 minute computer-assisted telephone interview by professional female interviewers in English, French and 13 non-official languages. In about 30 cases a telephone interview was not feasible and therefore a personal interview with a paper version of the questionnaire was used. Further details of the survey design and methods have been reported elsewhere [28–30].
Singleton preterm birth was defined as a live birth before 37 completed weeks of gestation and constructed based on self-reported gestational age.
Singleton small-for-gestational age (SGA) was defined as the sex- and gestational age-specific birthweight below the 10th percentile of a Canadian population-based reference  and constructed based on self-reported gestational age, infant sex and birthweight. The resulting rates of singleton PTB and SGA in the survey were consistent with national surveillance data based on birth certificates .
Postpartum Depression (PPD) was assessed using the Edinburgh Post-Natal Depression Scale, which was also administered by the interviewers over the telephone. The scale is a ten item screening tool to identify postpartum depression at the time of its administration . A score of 13 or more out of a maximum possible of 30 was used to indicate the presence of postpartum depression. Validation studies showed that the scale can detect depression in postpartum women with 86% sensitivity and 78% specificity [34, 35].
The MES contained a section to assess abuse and violence . The questions were adapted from the Violence Against Women Survey  to capture abuse and violence during the childbearing year. Women were asked whether in the two years prior to the interview "a spouse or partner or anyone else" had committed "acts of physical or sexual violence" and if they (i) were threatened with being hit; (ii) had something thrown at them that could have hurt them; (iii) were pushed, grabbed or shoved in a way that could have hurt them; (iv) were slapped; (v) were kicked or bit; (vi) were hit with a weapon; (vii) were beaten; (viii) were choked; (ix) were threatened with gun or knife or had one used on her; and (x) were forced into unwanted sexual activity.
Any abuse was considered to be present as any affirmative answer to any question i to x. These questions were categorised into two groups. Threats or potential hurting acts were defined as at least one affirmative answer to questions i to iii. Physical or sexual abuse was defined as at least one affirmative answer to questions iv to x. The combination of threats and physical or sexual abuse was used as a measure of severity. Women were also asked about frequency of events, categorised here as once, 2 to 5 times, and 6 times or more, and whether the perpetrator was their partner/husband/boyfriend or not (i.e., family member, friend/acquaintance or stranger). Finally, women were asked about the timing of these abusive episodes (before, during and/or after pregnancy).
There is a vast array of factors associated with both abuse and the outcomes [4, 24, 38–40]. We based our choice of covariates for confounder control based on a theoretical model assisted with the use of directed acyclic graphs (DAGs) . DAGs entail a set of rules aimed to identify a minimally sufficient set of covariates for confounding control when the objective is to obtain an unbiased causal effect estimate of an exposure on an outcome. DAGs are particularly appropriate to prevent overcontrol (e.g., adjusting for variables affected by the exposure) and when conventional criteria for confounding control are not met. The use of DAGs also makes explicit the investigators' assumptions about the relation between the variables.
The rationale for confounding control is as follows:
Young maternal age is associated with abuse and also with single marital status and these two are predictors of poor pregnancy outcomes [4, 42]. Prevalence of abuse varies according to ethnicity  and so does pregnancy outcomes. Controlling for maternal age, marital status and ethnicity removes the part of the effect of abuse on pregnancy outcomes that is due to differences in the distribution of these covariates. Low socioeconomic position (SEP) is a consistent predictor of both abuse and pregnancy outcomes [4, 43]. Abuse can increase stress and lead to substance use (e.g., tobacco smoking, alcohol drinking, legal and illegal drug use), either directly (through coercion by abusive partner or peer-pressure) and/or indirectly (as a mechanism to cope with the stress and anxiety produced by abuse) [5, 15]. Low SEP may also lead to stress and substance use. Controlling for SEP removes from the association between abuse and pregnancy outcomes the direct and indirect effects of SEP that are not related to abuse. Thus, after further adjustment for SEP the effect estimates of abuse reflect the total independent influence of abuse on pregnancy outcomes. This influence may operate through two non-exclusive pathways. A direct path may involve trauma leading to premature labour or rupture of membranes, placental abruption or a ruptured uterus, or to exacerbation of pre-existing conditions that could trigger adverse outcomes, and an indirect path through stress and/or substance use, that may involve unwanted pregnancy, poor health care utilization, poor maternal weight gain, anemia, an unhealthy diet, sexually transmitted diseases and other infections, lower social support and psychological morbidity [1, 11, 44]. This general conceptual model was applied to all three outcomes. However, the analyses on postpartum depression were restricted to women who had not been diagnosed with depression or took antidepressants before pregnancy to ensure that they were not depressed at the beginning of pregnancy.
Thus, variables for confounder control were maternal age (< 20 years, 20 to 34 years and 35 years and more), marital status (single, divorced, separated versus married or common-law), and immigrations status (Canadian-born and foreign-born). Socioeconomic position (SEP) was approximated by the Low income after-tax cut-off (LICO-AT) (Above LICO-AT, at or below LICO-AT and missing) . The LICO-AT reflects whether the respondent lived in a household spending 20 percentage points more of their after-tax income than the average family on food, shelter and clothing, thus leaving less income available for other expenses such as health, education, transportation and recreation. The low income after-tax cut-offs are set at after-tax income levels, differentiated by size of family and area of residence.
The MES is based upon a sample design involving stratification and unequal probabilities of selection of respondents. Survey weights were used to account for the unequal probabilities of selection of respondents and thus obtain unbiased point estimates representative of the Canadian population. However, because the stratification of the sample's design affects the variance estimates calculated using the sampling weights alone, replicate methods were used to calculate variance estimates. Proportions and Odds Ratios (with 95% confidence intervals) were weighted and calculated with the jackknife method of variance estimation .
Frequency of violence was entered in the models as a continuous variable and therefore its Odds Ratios are interpreted as the change in the odds of the event associated with one unit increase in the number of reported episodes of violence.
Missing data were very low for most variables and therefore were not considered in the analyses, with the exception of low income, for which we created a category labelled "Missing" to prevent a significant drop in the sample size.
The data were analysed at the Toronto regional Statistics Canada Research Data Centre. All analyses were conducted with SAS 9.2® (SAS Institute Inc., Cary, NC). The MES reporting guidelines prohibit the reporting of estimates based on counts less than 5 and recommend the reporting of weighted counts rounded to the nearest 100. The degree of sampling error affecting estimates was based on the coefficient of variation (CV). Estimates with a CV in excess of 33.3% are considered unreliable and therefore not reported .