ARCHIVED - Early primary school outcomes associated with maternal use of alcohol and tobacco during pregnancy and with exposure to parent alcohol and tobacco use postnatally

4. Data Analysis

4.1 Data Analysis Part 1: ANCOVA

Analyses of the 79 child outcome measures were carried out using Analyses of Covariance (ANCOVA) with the mother’s drinking alcohol and smoking during pregnancy as the two independent variables. The general idea of ANCOVA is to use statistical methods to create a level playing field for the comparisons of groups. The measures used as covariates in the ANCOVA analyses were those that might bias the results due to factors other than smoking or drinking during pregnancy. By including these measures in the analyses, statistical controls were used to remove any bias these variables may have had on the differences between groups. A complete list of the measures used as covariates in the analyses appears in Appendix 2, and includes measures of family income, maternal education, immigrant status, home language and single-parent status.

With the four prenatal smoking and drinking groups statistically equated on the covariate variables, we could then compare the average results for different groups with increased confidence. We compared the averages in a statistical manner so that we were less tempted to seize on a false result that favours our hypotheses.

4.1.1 Statistical Significance

It is standard practice to report the results of statistical analyses in terms of significance levels or p values. In the current analyses, the significance level (p value) is the probability that the difference between groups on any given measure is due to chance factors alone. If the p value is low (i.e. .01 or less), we can conclude that differences between groups are likely to be due to differences in whether or not mothers drank or smoked during pregnancy rather than being due to chance. If the p value is .01 or less, we conclude that the group differences are statistically significant. Statistically significant results allow us to say something similar to the phrase used in consumer polls; we will be right at least 99 times out of 100 whenever we say that the averages of two groups are in fact different, and not due to chance. We chose to use a conservative p value of .01 because of the large number of tests we were reporting.

4.1.2 Effect Size

The effect size reflects how large average differences are across different variables in a standardized manner. One of the problems with using many different measures is that the numbers used mean different things from one measure to another. A difference of 10 points means one thing in a depression score and another in an IQ score. In an effort to produce numbers that mean the same thing from measure to measure, we calculated a statistic called an effect size (more specifically a d statistic). When we compare two groups of children, the d statistic allows us to express the difference between the two groups in units determined by the variability of the children within their groups. This gives a common metric across measures and effectively allows us to compare “apples to oranges.”

4.1.3 Analysis Process

Each of the child outcome measures was analyzed three ways. First, outcomes for children of the smoking mothers were compared with children of the non-smoking mothers. This allowed for a comparison of children exposed to any prenatal tobacco to those exposed to none. A second analysis compared outcomes for children of the high-risk drinking mothers to those of low-risk drinking mothers. A third analysis compared children of mothers who were both high-risk drinkers and smokers to those who were low-risk drinkers and non-smokers. This comparison allowed for an assessment of the outcomes of children exposed to both prenatal tobacco and high levels of alcohol.

4.2 Data Analysis Part 2: Structural Equation Modelling

For the Structural Equation Modelling (SEM) analysis of the relationship between measures of maternal tobacco and alcohol consumption and later internalizing and externalizing behaviours, we limited the analysis to the four measures of smoking and drinking behaviour (and two arithmetic products of those measures), four measures of externalizing behaviours and two measures of internalizing behaviours collected from the teachers of the children and the same from the parents of the children.

The earliest measures were the tobacco use and alcohol use measures collected from mothers when the children were 3 months old. These are the same measures described in Table 1, and were coded dichotomously (i.e. higher-risk drinking versus no high-risk drinking; any smoking during pregnancy versus no smoking). The product of the two measures was used as a third variable, sensitive to an interaction between tobacco and alcohol use. Note that the product gave the non-drinkers and non-smokers the same value as the  non-drinking smokers and the non-smoking drinkers. Thus, these “one substance only” cases were included in the SEM analysis while they were omitted from the ANCOVA analyses described above.

When the children were 33 months old, we also collected data on maternal alcohol use and smoking in the home (as an indication of exposure to second-hand smoke). These were coded dichotomously and the product computed. This gave us a total of six measures of smoking and drinking behaviour, three collected at each of two times.

When the child was in Grade 3, we collected a large array of measures of child behaviour and social and emotional functioning, as described above. From that list of measures, we chose six parent report measures and six teacher measures that would allow us to estimate externalizing behaviour and internalizing behaviour. The measures are listed below.

4.2.1 Parent Measures

4.2.2 Teacher Measures

4.2.3 Procedure

We controlled for the same covariates described above by computing the covariate adjusted residuals for our Grade 3 parent and teacher variables (12 measures).

The data were analyzed using AMOS 17.0 (SPSS; Levesque 2007). Although AMOS does not have an option for selecting list-wise/pair-wise deletion, it can handle missing cases using a method called “Full Information Maximum Likelihood” (FIML, also known as “Raw Maximum Likelihood”), which is the technique that we used to deal with missing cases. This technique leads to indefinite sample sizes because while 502 people contributed data, only 177 have complete data for all 18 variables. Classical list-wise deletion would have limited the analysis to the information provided by the 177 subjects with complete data. The FIML procedure uses all the information available from the 502 subjects while assessing for bias imposed by the procedure. Given that the missing data were randomly missing, this technique is more efficient.

The SEM analysis was broken into segments to simplify the process. One segment modelled the relationships among the alcohol and tobacco measures, another tackled the internalizing and externalizing measures.