Design
In order to define a TPB-based structural latent variable model explaining variance in breastfeeding intentions and behaviors, this research team employed a longitudinal descriptive design. Questionnaire data were obtained using a modified version of Manstead’s reconstructed Predictive [Breastfeeding] Questionnaire (Manstead et al., 1983).
Setting
Data were collected in the Midwest region of the United States as part of the Mother’s Milk for Michigan Infants project from April 2018 - February 2019. The study was approved by Hope College’s Human Subjects Review Board. All enrolled participants were provided written research materials and consented prior to participation in the study. All the human subjects data was obtained in accordance with the guidelines of Hope College Human Subjects Review Board.
Sample
One hundred women with low-risk pregnancies and the intention to breastfeed were enrolled via convenience sampling after 30 weeks gestation and completed three questionnaires (antepartum, 10 and 60 days postpartum). Women were eligible to participate if they were 21 years of age or older, English proficient, intended to breastfeed with a singleton gestation, and lived within a 75-mile radius of the study site (see Table 1). The sample was recruited via social media, recruitment materials posted at local hospitals and businesses, and snowballing. Participants were provided with a $20 USD store card as partial compensation for their time and participation.
Measurement
Antepartum Questionnaire. The Antepartum Questionnaire collected participant demographic data and included the Predictive Breastfeeding questions developed by Manstead et al. (1983) (see Table 2) using the TPB.
Questions were reviewed for consistency with previously published breastfeeding studies using similar questionnaire data (Cabieses et al., 2014, Swanson, 2005). Participant demographic information included the maternal date of birth, marital status, annual household income, insurance type, the highest level of education, current employment status, race/ethnicity, and employment plans after the baby’s birth. The Predictive Breastfeeding questions examined participants’ attitudes, beliefs, social norms regarding breastfeeding, and behavior using a series of Likert-style questions. Breastfeeding intentions were measured using the Infant Feeding Intentions Scale (Nommsen-Rivers et al., 2010, Nommsen-Rivers & Dewey, 2009). Additional questions were included to address perceived behavioral control. The median duration to complete the Antepartum Questionnaire was 13.4 minutes (first and third quartile: 9.63, 17.52).
Day 10 & Day 60 Questionnaires. The targeted behaviors were exclusive breastfeeding at Day 10 and Day 60 postpartum. The Day 10 and Day 60 Questionnaires measured participants’ feeding practices postpartum. A series of multiple-choice questions measured feeding method, mode of milk expression, and frequency of feeding to conceptualize exclusivity and duration of breastfeeding (see Table 2). The median durations to complete the Day 10 and Day 60 Questionnaires were 7.2 minutes (4.92, 11.72) and 5.1 minutes (3.03, 14.93), respectively.
Data Collection
The consent form and questionnaires were administered electronically via QualtricsXM(R). Eligible participants were invited to review and complete an online consent form. Once the consent form was signed, participants were directed to complete the Antepartum Questionnaire. Participants were instructed to notify the research team when they gave birth. Based on the provided birth date, Day 10 and Day 60 Questionnaires were scheduled to be distributed to participants. No more than two reminders to complete any of the three questionnaires were sent to participants.
Data Analysis
Descriptive data were analyzed using IBM SPSS (Version 24). Additional data analysis were completed using SAS University Edition and the lavaan package in R for latent variable modeling (Rosseel, 2012). Analyses included confirmatory factor analysis, exploratory factor analysis including minimum average partial (MAP), very simple structure (VSS), parallels, and varimax rotation for the factor analysis. Finally, structural equation models (SEM) were utilized. MAP, VSS, and parallels provide the researcher a sense of how many factors may be present within any given set of data if one does not know how many should be present. A varimax factor rotation assumes that the different factors are not correlated with one another, thereby decreasing the possibility of a manifest variable strongly loading on more than one factor domain. The strategy was to begin with confirmatory analysis, fit a structural model, and then trim the model for best fit (See Figure 1).
When it was clear confirmatory factor analysis was insufficient, the strategy became exploratory factor analysis, fit latent variables (LVs) and trim manifest measures if necessary, and fit a structural model. Authors primarily relied on the Bayesian information criterion (BIC) to compare model fit between the first-order LV, hierarchical LV, and bifactor LV (described in more detail below). Model fit for each latent construct was assessed using cutoffs suggested by Schreiber and colleagues (2006) and the joint criteria suggested by Hu and Bentler (1999). Schreiber and colleagues (2006) suggest a Comparative Fit Index (CFI) of ≥ .95, Tucker‐Lewis Index (TLI) of ≥ .95, a Root Mean Squared Error of Approximation (RMSEA) of ≤ .06, and Standardized Root Mean Squared Residual (SRMR) of ≤ .08, and Hu and Bentler (1999) suggest either CFI ≥ .96 and SRMR ≤ .09 or SRMR ≤ .09 and RMSEA ≤ .06. Scale reliability was assessed using McDonald’s omega scores instead of alpha scores (McDonald, 2013, Revelle & Zinbarg, 2009, Zinbarg et al., 2005), and any missing data were handled with full information maximum likelihood in the structural models (Beaujean, 2014, Graham, 2009).