Results
Missing Data
Because of attrition over time and a small number of data-collection errors, several pieces of data were missing at both periods. At Time 1, social support data were missing for 1 participant and WMCI data were missing for 5 participants. At Time 2, 17 participants did not complete any measures. In addition, 12 par- ticipants did not complete the Strange Situation be- cause of an inability to participate in the taped portions of the interview (e.g., living out of state). To maximize statistical power, missing data were esti- mated using the hot-deck imputation method from Lisrel 8.5/Prelis 2.0 (Jo¨ reskog & So¨ rbom, 2001) after all recodings and data transformations. This method of estimation substitutes real values for missing values; real values are obtained from another participant’s responses that are most closely matched to the partic- ipant with missing data on a set of specified match- ing variables. Data from participants whose infants had died (n 5 2) were not estimated or included in analyses. Thus, measurement and structural model testing were based on data from 204 participants.
Descriptive Information and Correlational Results
See Table 1 for descriptive data on all measures included in model testing. All measures had adequate variance with the possible exception of Time 2 self- efficacy, which was skewed toward the positive end of the scale. See Table 2 for the correlation matrix of all model variables. These correlations were used to help clarify relationships between constructs shown in Figure 1 that were more exploratory in nature and, thus, to help finalize model testing. An alpha level of .05 was used for all statistical tests. As expected, variables within each latent construct were moderately to highly correlated with each other.
Results revealed that variables measuring representations of the infant were highly correlated with the variable measuring representations of the self as a mother, with an average correlation coefficient of .68. Thus, based on both theoretical and statistical reasons, it was decided to combine these variables to form one prenatal representations of caregiving construct. Although maternal self-efficacy was significantly related to variables measuring postnatal social support, because it had inadequate variance and was completely unrelated to infant – mother attachment, it was dropped from the model. Finally, correlations revealed that prenatal demo- graphic risks and domestic violence were more strongly related to variables measuring representations of caregiving than to infant – mother attachment and were, in fact, the highest between-construct correlations. Thus, a path from prenatal risk factors to prenatal representations of caregiving was added to the final model.
Measurement Models
Before final structural model testing, CFAs were conducted for each latent construct with more than two indicators to determine the adequacy of the hypothesized measurement models. Residual covariances were freed, as needed, when this led to a significantly improved model fit. The method of estimation was maximum likelihood (ML), which is the most frequently used and recommended esti- mation method for SEM analyses (Hoyle & Panter,
1995; McDonald & Ho, 2002).