Poverty, Social Capital, Parenting and Child Outcomes in Canada

4. Statistical Modeling

4.1 Overview of the Analysis

The present study uses extensions of linear regression models to predict indicators of family stress, parenting style and child outcomes. We chose the SAS system for data management and statistical analysis because it includes routines for the estimation of regression models using generalized estimating equations (Diggle, Liang and Zeger, 1994) as well as conventional ordinary least squares (OLS). SAS also includes routines for estimating models with complex random effects and these can be used for estimating the coefficients of hierarchical models (Littell et al. 1996; Singer 1998).

We test the basic model that the experience of chronic and, to a lesser extent, temporary poverty brings about symptoms of depression in parents as well as conflict between spouses and that these, in turn, cause poor parenting practices, resulting in the following child outcomes: poorer physical health²⁶, greater risk hyperactivity-inattention disorder, and lower mathematics test scores. We test the modified Family Stress Model by estimating the coefficients of a logical sequence of nested multiple regression models.

4.2 The National Sample and the Neighbourhoods Sample

We report two distinct groups of results. The first is based on data from up to 15,000 "longitudinal children" for whom data are available at Cycle 3. We use this "national sample" to predict indicators of family stress and parenting behaviour, as well as child outcomes in the domains of physical health, emotional adjustment and cognitive development. In this first series of analyses, the census indicators are measures of the enumeration area in which the child was living at the time, but the survey-based measures of neighbourhoods are usually obtained from the child's mother who is often the same person giving information about child outcomes. Since there are never more than two longitudinal children in a family at Cycle 3 and since most Enumeration Areas (EAs) contain no more than one family with a longitudinal child, it was impossible to estimate hierarchical models at with the national sample. Instead, we used OLS and GEE methods to estimate conventional multiple regression models in which we predicted parent and child outcomes from their history of exposure to long-term poverty and other characteristics.

The second group of results is based on a much smaller sample of children selected because they were from neighbourhoods containing more than five children at Cycle 3, thus allowing us to estimate "random slopes" models in a hierarchical modeling framework. This "neighbourhoods sample" includes some 200 different EAs (130 in analyses when observations from only Cycle 1 and Cycle 3 are included) and is further restricted in that the measures of neighbourhood social capital are restricted survey-based information about neighbourhood that has been averaged over at least six adult respondents. We discuss this sample in more detail below, but first we report the results of analyses using the national sample.

4.3 Results from the National Sample

These results are in three sections: (1) the prediction of family stress from long-term poverty and other measures including indicators of neighbourhood social capital; (2) the prediction of parenting style from long-term poverty, family stress and other measures as before; and (3) the prediction of selected child outcomes from long-term poverty, family stress, parenting style, etc. Since our theoretical model specifies that measures of family stress and parenting style mediate in the causal chain between long-term poverty and child outcomes, they appear as dependent variables in some tables and as independent variables in others.

4.3.1 Determinants of Family Stress (Data from Longitudinal Children in the National Sample at Cycle 3)

As discussed in the introduction, the Family Stress Model proposes a causal path leading from long-term poverty through depression and dissatisfaction in the family to inadequate parenting and unsatisfactory child outcomes. A first step in testing this model is to establish the causes of family stress. We use long-term low income, change in family income and neighbourhood average income along with a battery of other measures in order to predict scale scores on family dysfunction, PMK depression, and the PMK response to a single question about her degree of satisfaction with her spouse (All things considered, how satisfied or dissatisfied are you with your marriage or relationship with your partner? Which number comes the closest to how you feel, where 1 is completely dissatisfied and 11 is completely satisfied?).

Results from regression analyses with data from children who were followed up to Cycle 3 are shown in the table. Each indicator of family stress is predicted from Model 1 (poverty indicators only) and Model 2 (poverty indicators along with all other explanatory variables). The results from Model 1 and Model 2 support the family stress hypothesis in that family dysfunction has significant partial regression coefficients for all three income indicators, while depression and satisfaction with one's partner have significant associations with long-term low income and income change respectively. If the effects of poverty upon family stress were mediated by neighbourhood and family variables we would expect that the coefficients for income measures in Model 2 would be smaller than in Model 1. This is true for the measures of family dysfunction and depression though only in the case of depression do we see a coefficient become non-significant in Model 2. Lone-parents (on average poorer) have higher depression scores and higher dysfunction as compared to role-specialized couples. All three indicators of family stress have significant relationships in the expected directions with the census indicator of neighbourhoods with larger households, as well as with the individual measure of social support. They differ markedly in the degree to which they can be predicted; family dysfunction having a squared multiple correlation of 0.22 while the value for PMK depression is 0.10 and that for partner satisfaction is only 0.04. We conclude that the NLSCY measure of the quality of conjugal relationships does not support the Family Stress Model but that depression and family dysfunction behave roughly as would be expected.

Table 5
National Sample regressions predicting family stress indicators
Explanatory Variables		Dysfunction Model 1	Dysfunction Model 2	Depression Model 1	Depression Model 2	Satisfaction Model 1	Satisfaction Model 2
1.	Income Indicators
	Long-Term Low Income	0.116**	0.047**	0.186**	0.032*	0.009^ns	0.000^ns
	Change in income to needs	-0.048**	-0.034**	-0.042**	-0.018^ns	0.031**	0.028**
	EA Average Income	-0.036**	0.027*	-0.046**	-0.010^ns	-0.031**	-0.039**
2.	Neighbourhood Indicators
	EA Female labor force participation rate		0.035**		0.012^ns		-0.031**
	EA Persons per economic family		-0.054**		-0.033**		0.029*
	EA Non traditional households		-0.060**		-0.029*		0.029^ns
	EA New Canadians		0.043*		0.020^ns		-0.013^ns
	EA Recent movers		-0.009^ns		0.000^ns		0.009^ns
	EA Asian visible minorities		-0.021^ns		0.012^ns		0.014^ns
	Perceived Collective Efficacy		-0.112**		-0.020^ns		0.019^ns
	Perceived Social Support		-0.366**		-0.089**		0.114**
	Perceived Not a good place to bring up children’		0.006^ns		0.039**		-0.042**
	Volunteer work (No)		0.000^ns		-0.003^ns		-0.013^ns
	Infrequent Religious Attendance		-0.006^ns		0.055**		-0.044**
3.	Family Indicators
	PMK Years of Formal Education		-0.073**		-0.073**		0.007^ns
	Renting rather than owning home		-0.024*		0.025**		0.009^ns
	Recent Immigrant		0.000^ns		-0.001^ns		0.009^ns
	Stepfamily		0.050**		0.051**		-0.015^ns
	Parental Work/Study Status
	Both work or study		0.015^ns		-0.033**		-0.043**
	Neither works/studies		.003^ns		0.028**		0.020^ns
	Lone Parent works/studies		0.071**		0.134**		-
	Lone Parent not working nor studying		0.058**		0.124**		-
	Gender of respondent (female)		-0.025**		0.022*		0.017^ns
	Adjusted R-Squared	0.020	0.216	0.043	0.102	0.002	0.044
	Number of Observations	10,188	10,188	10,171	10,171	8,554	8,554
^*=p<0.01, ^=p<0.05, =0.10 (two-tailed tests of statistical significance), ns = not significant Family dysfunction and PMK depression are measured as multiple item scales. Satisfaction with the spousal relationship (for those with a partner only) is measured by the PMK responses to a single question. The family work/study variable encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. These coefficients are estimated from least squares regressions over child records. Since some families contribute two children, the sample is effectively clustered, thus biasing the standard errors though not regression coefficients except insofar as the estimates are pulled towards the parameters in the population of adults with more than one child.

Table 6
National Sample: models predicting two indicators of parenting style
Predicting the Child's Experience of Two Parenting Style Measures at Cycle 3.
Based on responses linked to longitudinal children aged 4 to 11 at Cycle 3
Explanatory Variables		Positive Parenting Model 1	Positive Parenting Model 2	Hostile Parenting Model 1	Hostile Parenting Model 2
	Intercept	13.04	17.88	8.242	8.245
1.	Income Indicators
	Long-Term Low Income	0.034^ns	0.064^ns	0.060^ns	-0.122^ns
	EA Average Income	0.011^ns	0.003^ns	0.102^**	0.109^**
	Change in family income to needs	-0.099^**	-0.105^**	-0.010^ns	-0.010^ns
2.	Family Stress Indicators
	Family Dysfunction		-0.088^**		0.110^**
	PMK Depression		0.005^ns		0.125^**
3.	Neighbourhood Indicators
	EA Female labor force participation		-0.021^ns		0.172^**
	EA Persons per economic family		0.417^*		-0.597^**
	EA Non traditional households		0.098^*		-0.065^ns
	EA New Canadians		-0.249^*		0.343^*
	EA Recent movers		-0.003^ns		0.047^ns
	EA Asian visible minorities		0.366^ns		-2.11^ns
	Perceived Collective Efficacy		0.087^**		-0.012^ns
	Perceived Social Support		0.007^ns		0.0^ns
	"Not a good place to bring up children"		-0.050^ns		0.337^**
	Volunteer work (No)		-0.251^*		-0.010^ns
	Infrequent Religious Attendance		-0.131^**		0.069^ns
4.	Family Indicators
	PMK Years of Formal Education		0.032^ns		0.052*
	Renting rather than owning home		0.107^ns		-0.208^ns
	Recent Immigrant		0.198^ns		-0.523^ns
	Stepfamily		0.128^ns		0.056^ns
	Parental Work/Study Status
	Both work or study		-0.306^**		0.252^ns
	Neither works/studies		0.863^**		-0.274^ns
	Lone Parent works/studies		-0.423^**		0.378^ns
	Lone Parent not working nor studying		0.212^ns		-0.247^ns
	Gender of respondent (female)		0.041^ns		0.542^ns
	Child's number of siblings		-0.491^**		0.269^**
	Between-child correlation	0.601	0.674	0.553	0.512
	Deviance/df	7.989	6.216	12.925	11.751
	Adjusted R-Squared	0.002	0.221	0.002	0.086
	Number of Observations	8,423	7,004	8,384	6,845
^*=p<0.01, ^=p<0.05, =0.10 (two-tailed tests of statistical significance), ns = not significant The EA female labour force participation rate is the rate for females with children over six years old only. The family work/study variable encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. Maternal age and child gender as well as child age, age squared, birth weight and birth order, are included in the full model but the coefficients are not shown. Empirical standard estimates are used (with the "sandwich" estimator.)

4.3.2 Determinants of Parenting Styles (Longitudinal Children in the National Sample at Cycle 3)

Even within the same family, parents treat their children differentially depending on their age, birth order, gender and temperament. Mothers often parent differently from fathers and biological parents differently from stepparents. Our questions about parenting flow from the Family Stress Model and relate to differences between better off families and those exposed to poverty, and/or between well-functioning families and those under stress. From this point of view, the effects of child age and gender upon how they are parented are not of primary interest, so although we control for them in our analyses, we do not present them in our results and we minimize our discussion of them.

Table 6 displays the results of fitting models predicting two measures of parenting style: (1) positive parenting; and (2) hostile-aggressive parenting. As in the previous table, we present restricted and full models (Model 1 and Model 2) for each of these parenting style measures. The results for Model 1 establish that neither of them is linked with long-term poverty, thus failing to support the predictions of the Family Stress Model. On the other hand both these parenting measures have highly significant relationships with Family Dysfunction. Parents in more dysfunctional families give less positive parenting and more hostile-aggressive parenting. Both measures of parenting style are also linked to family size: larger families being associated with less positive parenting and more of the hostile-aggressive style. The relationships are not always consistent: hostile-aggressive parenting is affected by PMK depression as well as by the perception that the neighbourhood is not a good place to bring up children, while positive parenting is not. Some patterns of association unique to positive parenting are that it is higher when the PMK has a high degree of trust in her neighbours (Collective Efficacy) and lower when both parents, or a lone-parent, is engaged in work or study.

There are also some interesting linkages with objective Census measures of neighbourhoods: both parenting styles are affected by the EA's average size of economic families, which seems to increase positive parenting and to decrease hostile-aggressive parenting. Both the EA average income level for economic families and the local level of labour force participation by mothers of children over 6 increase the amount of hostile-ineffective parenting, but these two Census measures of neighbourhood have no relationship with positive parenting style.

4.3.3 Determinants of Selected Child Outcomes (Longitudinal Children in the National Sample at Cycle 3)

As before, we report a logical sequence of regression models. In this case, we focus upon the child outcomes of physical health (as judged by the PMK in her response to two global questions about the child's health status), hyperactivity (as reported upon by the PMK in an eight-item scale) and mathematics test scores, as administered in school.

The analysis of physical health is shown in Table 7; of hyperactivity-inattention in Table 8; and of mathematics test scores in Table 9.

Model 1 shows the effect of long-term low income and it is highly significant for all three child outcomes. Model 4 shows the effect of including family stress and parenting indicators and Model 5 adds family indicators. All models include classic indicators of disadvantage such as child's birth weight, number of siblings and number of older siblings, each of which (except birth weight which becomes non-significant in the full model for mathematics test scores) is a highly significant predictor for all three child outcomes. One surprise here is that children with more siblings tend to do better in mathematics than children from smaller families.

Maternal age is a highly significant predictor in all models for hyperactivity-inattention and the mathematics test score, but is unrelated to physical health status except in Model 5 after many other predictors have been held constant. Children in stepfamilies are more hyperactive and do worse in the mathematics test, but are no different from other children in terms of their general physical health. On the other hand, the children of lone-parents who work or study are more hyperactive than children in role-specialized families and have worse physical health, but are no different in mathematics test scores. Family dysfunction is associated with worse physical health and lower mathematics scores, but has no effect on hyperactivity-inattention. The PMK level of depression affects the child's physical health status in both models in which it appears as a predictor, but has inconsistent effects upon hyperactivity-inattention and mathematics scores. Completing the complex pattern of results, hostile-aggressive parenting has highly significant relationships with child hyperactivity-inattention and the child's physical health status but no significant relationship with mathematics test scores.

Measures capturing aspects of the neighbourhood provide a similarly complex set of results. The objective Census characteristic of Non-Traditional Households, which some might think of as a form of disadvantage is in fact associated with better mathematics test scores in both Models 4 and 5. The same explanatory variable is unrelated to general physical health status or to hyperactivity-inattention. PMK judgements about the neighbourhood being a bad place to raise children are associated with the child's physical health status as well as her mathematics test scores, but not with hyperactivity-inattention.

Table 7
National Sample: models predicting the child's physical health status
		Metric Partial Regression Coefficients
Explanatory Variables		Model 1	Model 4	Model 5
1.	Intercept	3.265	3.211	3.687
2.	Income Indicators
	Long-Term Low Income	0.089^**	0.043^**	0.046^**
	Change in family income to needs			0.000^ns
	EA Average Income for 2+ pers holds	-0.030^**	-0.014	-0.008^ns
3.	Family Stress Indicators
	Family Dysfunction Score		0.017^**	0.016^**
	PMK Level of Depression		0.023^**	0.024^**
4.	Parenting Indicators
	Positive Parenting 2-11		-0.001^ns	-0.004^ns
	Hostile-Ineffective Parenting 2-11		0.011^**	0.011^**
	Consistent Parenting 2-11		-0.001^*	-0.006^ns
	Punitive-Aversive Parenting 2-11		-0.006^ns	-0.009^ns
5.	Neighbourhood Indicators
	EA Female labor force participation		0.002^ns	0.008^ns
	EA Persons per economic family		-0.077^ns	-0.114^*
	EA Non traditional households		0.001^ns	0.017^ns
	EA New Canadians		0.004^ns	0.01^ns
	EA Recent movers		-0.011^ns	-0.017^*
	EA Asian visible minorities		0.062^ns	0.037^ns
	Perceived Collective Efficacy		0.002^ns	0.000^ns
	Perceived Social Support		-0.010^ns	-0.009^ns
	"Not a good place to bring up children"		0.082^**	0.063^*
	Volunteer work (No)		-0.016^ns	-0.029^ns
	Infrequent Religious Attendance		0.001^ns	-0.004^ns
6.	Family Indicators
	PMK Years of Formal Education			-0.030^**
	Renting rather than owning home			0.038^ns
	Recent Immigrant			0.022^ns
	Stepfamily			0.099^ns
	Parental Work/Study Status
	Both work or study			-0.005^ns
	Neither works/studies			-0.136^ns
	Lone Parent works/studies			-0.115^*
	Lone Parent not working nor studying			-0.069^ns
	Gender of the PMK (female)	0.027^ns	0.021^ns	0.013^ns
7.	Child Characteristics
	Child gender (female)	-0.047^*	-0.025^ns	-0.028^ns
	Child age	-0.019^ns	-0.019^ns	-0.026^ns
	Child age squared	0.001^ns	0.001^ns	0.001^ns
	Maternal age	0.001^ns	0.003^ns	0.007^*
	Birth weight	-0.089^**	-0.075^**	-0.077^**
	Number of siblings	-0.075^**	-0.067^**	-0.062^**
	Number of older siblings	0.066^**	0.067^**	0.051^**
8.	Summary Statistics
	Between-child correlation	0.255	0.221	0.220
	Deviance/df from GEE	1.138	1.093	1.095
	Number of Observations	11,857	10,734	9,539
	Number of Households at Cycle 3	8,336	7,403	6,576
^*=p<0.01, ^=p<0.05, =0.10 (two-tailed tests of statistical significance), ns = not significant The measure of poor health is the result of adding numerical scores from two questions asked of the child's main caregiver (the PMK.) Higher scores mean poorer health status. The EA female labour force participation rate is the rate for females with children over six years old only. For children aged 12-15 we used the most recent parenting scores. The family work/study variable also encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. Maternal age and child gender as well as child age, age squared, birth order, and PMK age are included as predictors in all models. Empirical standard estimates are used (with the "sandwich" estimator.)

Table 8
National Sample: models predicting the child's hyperactivity-inattention scores
		Metric Partial Regression Coefficients
Explanatory Variables		Model 1	Model 4	Model 5
1.	Intercept	9.418	1.60	1.596
2.	Income Indicators
	Long-Term Low Income	0.217**	1.71**	0.079^ns
	Change in family income to needs	0.001^ns	-0.001^ns	0.001^ns
	EA Average Income for 2+ pers holds	-0.018^ns	-0.015^ns	-0.014^ns
3.	Family Stress Indicators
	Family Dysfunction Score		0.017^ns	0.014^ns
	PMK Level of Depression		0.073**	0.068**
4.	Parenting Indicators
	Positive Parenting 2-11		0.016^ns	0.018^ns
	Hostile-Ineffective Parenting 2-11		0.352**	0.348**
	Consistent Parenting 2-11		-0.037	-0.035
	Punitive-Aversive Parenting 2-11		0.018^ns	0.026^ns
5.	Neighbourhood Indicators
	EA Female labor force participation		0.022^ns	0.014^ns
	EA Persons per economic family		0.082^ns	0.110^ns
	EA Non traditional households		0.031^ns	0.022^ns
	EA New Canadians		0.133^ns	0.107^ns
	EA Recent movers		-0.040^ns	-0.056^ns
	EA Asian visible minorities		-1.171^ns	-0.991^ns
	Perceived Collective Efficacy		-0.038^ns	-0.037^ns
	Perceived Social Support		0.021^ns	0.019^ns
	"Not a good place to bring up children"		0.135^ns	0.106^ns
	Volunteer work (No)		0.286	0.228^ns
	Infrequent Religious Attendance		0.092	0.082
6.	Family Indicators
	PMK Years of Formal Education			-0.016^ns
	Renting rather than owning home			0.299
	Recent Immigrant			-0.158^ns
	Stepfamily			0.641**
	Parental Work/Study Status
	Both work or study			0.048^ns
	Neither works/studies			0.037^ns
	Lone Parent works/studies			0.640**
	Lone Parent not working nor studying			-0.083^ns
	Gender of the PMK (female)	0.190^ns	-0.166^ns	-0.216^ns
7.	Child Characteristics
	Child gender (female)	-1.339**	-1.201**	-0.120**
	Child age	-0.041^ns	0.520^ns	0.411^ns
	Child age squared	-0.006^ns	-0.036	-0.029^ns
	Maternal age	-0.076**	-0.040**	-0.030*
	Birth weight	-0.367**	-0.367**	-0.349**
	Number of siblings	-0.442**	-0.482**	-0.424**
	Number of older siblings	0.399**	0.529**	0.465**
8.	Summary Statistics
	Between-child correlation	0.201	0.151	0.146
	Deviance/df from GEE	11.688	10.638	10.563
	Number of Observations	7,994	4,546	4,464
	Number of Households at Cycle 3	6,132	3,930	3,861
^*=p<0.01, ^=p<0.05, =0.10 (two-tailed tests of statistical significance), ns = not significant The measure of child hyperactivity is a scale formed of questions responded to by the PMK for children aged 4-11 only. The EA female labour force participation rate is the rate for females with children over six years old only. The family work/study variable also encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. Maternal age and child gender as well as child age, age squared, birth order, and PMK age are included as predictors in all models. Empirical standard estimates are used (with the "sandwich" estimator.)

Table 9
National Sample: models predicting the child's mathematics test scores
		Metric Partial Regression Coefficients
Explanatory Variables		Model 1	Model 4	Model 5
1.	Intercept	-81.09	4.364	-31.861
2.	Income Indicators
	Long-Term Low Income	-8.292**	-8.766**	-5.794**
	Change in family income to needs			0.001^ns
	EA Average Income for 2+ pers holds	1.087^ns	2.155**	1.251^ns
3.	Family Stress Indicators
	Family Dysfunction Score		-0.729**	-0.595*
	PMK Level of Depression		-0.340**	-0.228^ns
4.	Parenting Indicators
	Positive Parenting 2-11		-1.037*	-0.905*
	Hostile-Ineffective Parenting 2-11		-0.453^ns	-0.662^ns
	Consistent Parenting 2-11		0.262^ns	0.123^ns
	Punitive-Aversive Parenting 2-11		-0.858^ns	-0.668^ns
5.	Neighbourhood Indicators
	EA Female labor force participation		-0.785^ns	-1.224^ns
	EA Persons per economic family		-3.748^ns	-6.104^ns
	EA Non traditional households		5.843**	5.125**
	EA New Canadians		0.837^ns	-2.362^ns
	EA Recent movers		-2.364*	-1.951^ns
	EA Asian visible minorities		14.07^ns	26.071^ns
	Perceived Collective Efficacy		-0.650^ns	-0.887^ns
	Perceived Social Support		0.269^ns	0.237^ns
	"Not a good place to bring up children"		-9.444**	-8.013**
	Volunteer work (No)		1.005^ns	3.422^ns
	Infrequent Religious Attendance		-1.384^ns	-0.724^ns
6.	Family Indicators
	PMK Years of Formal Education			4.686**
	Renting rather than owning home			-7.36ns
	Recent Immigrant			10.068^ns
	Stepfamily			-8.996*
	Parental Work/Study Status
	Both work or study			2.926^ns
	Neither works/studies			-0.943^ns
	Lone Parent works/studies			-3.074^ns
	Lone Parent not working nor studying			2.341^ns
	Gender of the PMK (female)	-1.330^ns	1.225^ns	4.702^ns
7.	Child Characteristics
	Child gender (female)	4.000	2.007ns	2.967^ns
	Maternal age	1.834**	1.633**	1.140**
	Birth weight	4.832**	4.763*	2.830^ns
	Number of siblings	4.548**	5.060**	4.34*
	Number of older siblings	-8.525**	-8.502**	-7.97**
8.	Summary Statistics
	Between-child correlation	0.200	0.200	0.166
	Deviance/df from GEE	4612.0	4453.8	4408.3
	Number of Observations	4,391		3,510
^*=p<0.01,^=p<0.05, =0.10 (two-tailed tests of statistical significance), ns = not significant The mathematics scale score is constructed in such a way that it has a built-in correlation with age. Child age and age squared are included in all models. The measure of poor health is the result of adding numerical scores from two questions asked of the child's main caregiver (the PMK) The EA female labour force participation rate is the rate for females with children over six years old only. For children aged 12-15 we used the most recent parenting scores. The family work/study variable also encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. Maternal age and child gender as well as child age, age squared, birth order, and PMK age are included as predictors in all models. Empirical standard estimates are used (with the "sandwich" estimator.)

This analysis of child outcomes using the national sample at Cycle 3 of the NLSCY confirms that long-term low income has negative outcomes for children. In the case of hyperactivity-inattention, the introduction of variables which can be considered to mediate the impact of long-term poverty reduces its effect to non-significance. However, in the case of physical health and mathematics scores, the effects of long-term low income persist, even after adding a large number of family level and neighbourhood level variables to the battery of predictors.

Our analysis shows that some neighbourhood measures are significantly associated with child outcomes and with indicators of parenting style. However, we consider this part of the analysis to be inadequate as regards the survey-based measures of neighbourhoods since they are in most cases based on the same PMK reporting both on the neighbourhood and on her child. Furthermore, many children in the national sample are the only child in their enumeration area, so it is impossible to find out whether poverty has different effects from one neighbourhood to another. For this reason we decided to focus on a smaller number of children clustered so that there are several per enumeration area and we now turn to analysis of this "neighbourhoods sample".

4.4 Results from the Neighbourhoods Sample

4.4.1 The Sub-Sample for Analyzing Neighbourhood Effects

Numbers are reduced in the neighbourhoods sample by the requirement that most families should include more than one child and that each neighbourhood should contain at least a reasonable number of families²⁷. This requirement stems from the logic of asking multilevel questions and also for the technical reason that while mixed models can produce unbiased estimates from unbalanced layouts where randomly occurring missing observations are present, data which depart too much from a balanced layout present too great a computational challenge. After some experimentation we set the selection condition that child-occasions must be from EAs which each contained more than ten child-occasions from children about whom we had longitudinal data. This resulted in our analyses being carried out on up to 5,250 child-occasions when from Cycles 1, 2 or 3 and 3,455 child-occasions when we restricted attention to Cycles 1 and 3 only. The numbers are reduced when predicting child outcomes that are only measured for children in a limited age range and sometimes by the presence of missing data on certain explanatory variables.

4.4.2 Clustered Observations and Longitudinal Analysis

Children in the NLSCY sample are to some degree "clustered" within households and enumeration areas. (In addition, the longitudinal aspect of the data produces a kind of "clustering" of observations over time.) So far as children within households and within enumeration areas are concerned, most such clusters are of size one, while a small percentage are of size two, three or more²⁸. Children at Cycle 3 are to some degree clustered within families, though more than half were the only child sampled from their household. Families, in turn, are to some degree clustered within census EAs, though again, a considerable proportion of families are the only ones sampled in their neighbourhoods. Census tracts are larger scale geographical units but even so, less than 100 of them contain more than 20 households from the NLSCY sample (Foster et al. 2001.)

There are two general approaches to the analysis of longitudinal data of the kind collected in the NLSCY. The first is to focus upon outcomes at the most recent data collection and to use as much as possible of the history of each child in explanatory models. The second is to restructure the data on each child's history into a set of child-occasions (see for example Allison 1999) and use mixed models to carry out hierarchical modeling.

Mixed models are very general in that they can treat each child's occasions (wave by wave observations) as clustered within the child who is clustered within a family etc. When the data are imperfect, as is almost inevitable with longitudinal studies, mixed models are more forgiving than standard approaches to repeated measures analysis since, in principle, they can obtain unbiased estimates even with unbalanced and randomly missing data (Littell et al. 1996 : 115.) Hierarchical mixed models typically show that most of the variation in child outcomes is between children (often associated with age or gender) and after that between families, with quite small proportions of variation lying between EAs, census tracts or larger units (Boyle and Lipman 1998; Foster 2001.) While all approaches can estimate cross-level interaction effects (for example, the interaction between neighbourhood characteristics and parenting styles in affecting child outcomes) mixed models can estimate the variances of random slopes, thus testing whether or not family-level relationships are similar in different neighbourhoods (Snijders and Bosker 1999 : 67-85.)

4.4.3 The Multi-Level Modeling Framework

Variation in child outcomes that relates to family income, ethnicity, parenting practices and other parent characteristics is "between-family" and in the simplest theoretical model "between-family" differences should affect all children in the same family in the same way. In a more sophisticated "cross-level interaction" formulation a family level event, such as the experience of bereavement, divorce or transient poverty, may affect younger children differently from older ones. Of course, children brought up within the same household often have different outcomes. Such differences may be related to early health deficits, gender, birth order, differential parenting or other factors, but they are all "within-household". In the language of hierarchical modeling, "within-family" variation in child outcomes is nested within "between-family" variation. In turn, families are located within neighbourhoods; so we can also talk about "between-family" variation in child outcomes being nested within "between-neighbourhood" variation. We could also discuss "between-family" variation in parenting practices being nested within "between-neighbourhood" variation. When we have measured the same outcome repeatedly for each child as in a longitudinal survey, we also have "within-child" variation.

Thinking about effects that may occur at different levels of analysis leads us to the classical discussions of "ecological" and "individualistic" fallacies. These so-called fallacies arise because partial regression coefficients estimated at one level of analysis do not necessarily have the same size, or even sign, as the corresponding coefficients estimated at a lower or higher level of analysis. When this happens it is usually because different social processes are operating at the various levels of analysis. For example, a correlation over census tracts between the percent of lone-parent families and the percent of children who drop out of high school might reflect relatively less effective parenting by lone parents, but it might equally represent lone parents tending to live in low income areas, where all children are at higher risk of dropout. A virtue of hierarchical linear modeling is that it explicitly models the "levels of analysis" issue and produces best linear unbiased estimates of model parameters.

Hierarchical linear models use "mixed model" data analysis that estimates not only the "fixed" part of the standard multiple regression model, but also the variances of random intercepts at different levels of the model. It is also possible to estimate the variances of random slopes, also at different levels. In the "random intercepts model" we estimate the variance of random intercepts summarizing differences between EAs, _{Neighbourhoods}, as well as the variance of random intercepts representing differences between families, _Families, and the variance of random intercepts summarizing differences between children, _Children. A residual variance, ² can be estimated from within-child variation between waves of the longitudinal survey. One useful result that can be obtained from hierarchical modeling is a statistical test of the hypothesis that the variance of random intercepts at some level of the model is zero. For example, if the random intercepts for neighbourhoods are essentially the same value (zero variance) then the neighbourhood level explanatory variables in the predictive model can be said to explain neighbourhood differences in child outcomes. In general, as more independent variables are added to the predictive model, the unexplained variation at neighbourhood, family and child levels of the model is reduced and each can be reported separately. The "random slopes model" contains the same terms as the random intercepts model but adds a term for the variance of the slope parameter that is hypothesized to vary over clusters. We present such a model below for the slope of the relationship between long-term family poverty and child health and we test the hypothesis that this slope varies over neighbourhoods.

4.4.4 Hierarchical Linear Models for Predicting Child Outcomes

In this section we report on the results of fitting hierarchical linear models for the prediction of two child outcomes and one parenting style. We chose the child's physical health as rated by the person most knowledgeable about the child (the PMK) in Cycles 1, 2 and 3 as our first dependent variable and the child's level of hyperactivity-inattention (again as rated by the PMK) in Cycles 1 and 3 as our second. We were unable to analyse the mathematics test scores in this way, because of deficiencies in their administration at Cycle 1. The final analysis in this section is the prediction of hostile-aggressive parenting within the framework of the same multilevel model.

4.4.4.1 Hierarchical Linear Models for Predicting Child's Physical Health

The child's physical health is measured by the PMK's responses to two global questions. Higher scores indicate poorer health, as perceived by the PMK. Table 10 shows the prediction of PMK-rated child health in five analytical steps. All five models include the basic predictors of child gender, age and age squared, as well as child birth-weight, maternal age, PMK gender, the total number of siblings, the number of older siblings, and the wave of the survey. A summary of a sixth model (the "empty" or "null" model) is given in the bottom left of the table. Model 1 includes the exogenous variables of long term poverty and neighbourhood average income. Models 2, 3 and 4 add the variables that mediate and modify associations: family dysfunction and PMK depression (Model 2), parenting styles (Model 3) and neighbourhood social capital indicators (Model 4.) Model 5 includes further time-varying family-level measures, including the work/study status of the PMK and partner (if any), a measure that also encodes lone parent status. Further measures incorporated in Model 5 include whether the family is a stepfamily, immigration status, home ownership and the number of years of formal education of the PMK.

Table 10
Neighbourhoods Sample Random Slope and Intercepts Models of the Child's Physical Health at Cycles 1, 2 and 3
Explanatory Variables			Model 1	Model 2	Model 3	Model 4	Model 5
1.	Intercept		3.675	3.430	3.353	3.778	3.905
2.	Poverty Indicators
	Long-Term Low Income (Ltpoverty)		0.113**	0.085**	0.074*	0.072*	0.012^ns
	EA Average Income for 2+ person hse-holds		-0.003^ns	-0.001^ns	-0.004^ns	0.013^ns	0.013^ns
	Change in family income to needs						-0.001^ns
3.	Family Stress Indicators
	Family Dysfunction Score			0.015**	0.014**	0.012**	0.012**
	PMK Level of Depression			0.015**	0.015^ns	0.015**	0.013**
4.	Parenting Indicators
	Positive Parenting 2-11				0.009^ns	0.008^ns	0.008^ns
	Hostile-Ineffective Parenting 2-11				0.017**	0.018**	0.019**
	Consistent Parenting 2-11				-0.016**	-0.017**	-0.014*
	Punitive-Aversive Parenting 2-11				-0.004^ns	-0.006^ns	-0.003^ns
5.	Neighbourhood Indicators
	EA Female labor force participation rate					0.015^ns	0.019^ns
	EA Persons per economic family					-0.088^ns	-0.079^ns
	EA Non traditional households					-0.002^ns	-0.004^ns
	EA Change in Non traditional households						-0.256**
	EA New Canadians					0.022^ns	0.031^ns
	EA Recent movers					-0.008^ns	-0.011^ns
	EA Asian visible minorities					-1.356	-1.297
	EA Collective Efficacy					0.009	0.0^ns
	EA Social Support					-0.003	-0.020^ns
	EA "Not a good place to bring up kids"					0.126*	0.108*
	EA Volunteer work					-0.084^ns	-0.057^ns
	EA Infrequent Religious Attendance					-0.013^ns	-0.022^ns
6.	Family Indicators
	PMK Years of Formal Education						-0.032**
	Renting rather than owning home						0.139*
	Recent Immigrant						-0.476
	Stepfamily						0.086^ns
	Parental Work/Study Status
	Both work or study						-0.071*
	Neither works/studies						-0.077^ns
	Lone Parent works/studies						-0.070^ns
	Lone Parent not working nor studying						0.063^ns
	Child Birth Weight		-0.164**	-0.153**	-0.148**	-0.152**	-0.154**
	Number of siblings in the household		-0.059*	-0.054	-0.060*	-0.061*	-0.067*
	Likelihood Ratio Chi Square	14862.6	14744.8	14617.2	14343.4	14247.2	13431.1
	Variance Components Empty Model
	Residual ²	0.617**	0.615**	0.615**	0.609**	0.611**	0.599**
	_Children	0.372**	0.364**	0.361**	0.354**	0.357**	0.355**
	_Families	0.191**	0.167**	0.148**	0.148**	0.141**	0.119**
	_{LTpoverty* Poor Health}		0.026**	0.025**	0.027**	0.026**	0.026**
	_{Neighbourhoods}	0.024*	0.008^ns	0.004^ns	0.002^ns	0.0^ns	0.0^ns
	Observations Used	5,250	5,235	5,208	5,131	5,097	4,858
^*=p<0.01, ^=p<0.05 , =0.10 (two-tailed tests of statistical significance), ns = not significant Analysis of child-occasions data from 2271 children, 1451 households and 215 enumeration areas, each enumeration area containing at least 20 child-occasions. The measure of poor health is the result of adding numerical scores from two questions asked of the child's main caregiver (the PMK.) Higher scores mean poorer health status. The EA female labour force participation rate is the rate for females with children over six years old only. For children aged 12-15 we used the most recent parenting scores. The family work/study variable also encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. Maternal age and child gender as well as child age, age squared, birth order, and PMK age are included as predictors in all models.

We begin this section by reporting estimates of the fixed effects in the mixed model. PMK ratings of their children's physical health are associated with long-term poverty, though not with neighbourhood average income (Model 1). The sign of the regression coefficient is positive, indicating that longer-term poverty is associated with poorer child health. After adjustment for family stress in Model 2, this coefficient is considerably reduced, but remains statistically significant. Further adjustment, this time for parenting styles (Model 3) has no effect on the coefficient for long term poverty and the addition of neighbourhood measures (Model 4) reduces the coefficient for long term poverty by only a small amount. The introduction of the final set of family measures reduces the coefficient for long term poverty to less than one third of the value it had in Model 1 and renders it statistically insignificant. We conclude that the effect of long-term poverty on children's physical health is entirely mediated by the combination of family stress measures introduced in Model 2 and the remaining family measures added in Model 5. Parenting styles affect children's physical health but contrary to the family stress model, they do not mediate the effects of long-term poverty.

Our estimates of random effects from the mixed model are equally informative. These include _{Neighbourhoods} as well as _Families as the variances of random intercepts and _{LTpoverty*Poor Health} which is the variance of the slope of the relationship between long-term poverty and poor health. As in our analysis of how estimates of fixed effects change when more explanatory variables are added to the model, we can observe how the different variance components from one model to the next.

Between-neighbourhoods variation includes both _{Neighbourhoods} and _{LTpoverty*Poor Health}. In an earlier analysis (not shown) we had left that variance undivided and had been able to show that between neighbourhoods variation was significantly different from zero in the empty model as well as in Model 1, but that by the time we got to Model 5, it had been reduced to less than half its original size and was not significantly different from zero. This showed that there were indeed between-neighbourhood differences in child health and that they could be accounted for by controlling for measures of poverty, family stress and neighbourhood social capital.

The empty model summarized in the bottom left of Table 10 summarizes the results of predicting physical health from intercepts only. This model has no random slope included and the between-neighbourhoods variance is significantly different from zero. Model 1 introduces the fixed effects of long-term poverty and other predictors. It also introduces a random slope for the relationship between long-term poverty and the child's physical health, a step that immediately reduces _{Neighbourhoods} to non-significance, while almost all of its variation goes into _{LTpoverty*Poor Health}, which expresses the variability of this relationship over neighbourhoods. The variance of the random slope remains highly significant in Models 2 through 5 and we have not yet found a cross-level interaction in the fixed effects which would reduce its size. Thus, we have shown that the effect of long-term poverty on child health has a different slope depending on the neighbourhood in which children were living at Cycle 3 of the NLSCY. Neighbourhoods do indeed modify the effects of long-term poverty. This could be because of a variety of social processes, ranging from the differing degree of supportiveness of neighbourhoods to differential migration between neighbourhoods by the children. Precisely how it occurs must be the subject for future research.

Between-families variation in children's physical health is significantly different from zero in all models. This means that significant between-family variation remains even after all explanatory variables have been entered. Comparisons of how _Families changes as we move from the empty model to Model 5 shows that the introduction of poverty measures (Model 1) reduces it by only a small amount, while adding measures of family stress (Model 2) contributes a further small reduction and adding parenting measures only produces a further slight per cent drop. It makes sense that the addition of neighbourhood measures has no effect on _Families, while adding the remaining family measures reduces it by a further ten per cent or so.

Returning to description of the fixed effects of explanatory variables upon child health, there are interesting results regarding the effects of family stress, parenting measures and certain neighbourhood indicators. PMK depression and hostile-aggressive parenting have highly significant effects that are not much affected by adding any other explanatory variables. Consistent parenting has significant effects upon child health and this effect is essentially unchanged by adding neighbourhood measures in Model 4, but disappears after the addition of family level measures in Model 5. These family measures clearly mediate the effects of consistent parenting, or the lack of it, upon children's physical health. Family dysfunction displays a similar pattern, though its effect in Model 5 remains statistically significant, but somewhat attenuated. Neighbourhoods with higher scores on census indicators of non-traditional households are associated with poorer child health in Models 4 and 5, though the direction of the partial association changes after the introduction of family-level indicators of non-traditional households in Model 5. Neighbourhoods with higher numbers of Asian visible minorities are associated with better child health in both Model 4 and Model 5. Among aggregated neighbourhood measures there are linkages to child health both from averaged social support and averaged responses to the question about whether the neighbourhood was a good or a poor place to bring up children. The effect of averaged social support is significant in Model 4 but not in Model 5, so we conclude that it is mediated by family-level measures.

4.4.4.2 Hierarchical Linear Models for Predicting Child's Hyperactivity-Inattention

Table 11 shows the same five nested models as reported in the previous table but this time for prediction of the hyperactivity-inattention scale for children aged 4-11 at Cycles 1 and 3. We do not estimate a random slope of the kind shown in the previous table. We excluded observations from Cycle 2 in order to make the aggregated survey measures of neighbourhoods more closely linked with the child record²⁹. Because of this and since the measure of hyperactivity-inattention is only available for children up to age 11, the number of child occasions for analysis is at most 2,742 and is reduced to 2,529 in the final model due to missing data in some explanatory variables.

Table 11
Neighbourhoods Sample¹ Random Intercepts Models Predicting Hyperactivity-inattention at Cycles 1 and 3
Explanatory Variables			Model 1	Model 2	Model 3	Model 4	Model 5
1.	Intercept		9.369	8.439	4.995	1.034	2.858
2.	Poverty Indicators
	Long-Term Low Income		0.233**	0.114**	0.189*	0.193*	0.042^ns
	EA Average Income		-0.091^ns	-0.080^ns	-0.136*	-0.128^ns	-0.090^ns
3.	Family Stress Indicators
	Family Dysfunction Score			0.039**	0.004^ns	0.014^ns	0.011^ns
	PMK Level of Depression			0.098**	0.055**	0.052**	0.048**
4.	Parenting Indicators
	Positive Parenting 2-11				0.041	0.040^ns	0.033^ns
	Hostile-Ineffective Parenting 2-11				0.353**	0.352**	0.348**
	Consistent Parenting 2-11				-0.037	-0.035	-0.029^ns
	Punitive-Aversive Parenting 2-11				0.084*	0.079*	0.081*
5.	Neighbourhood Indicators (EA)
	EA Female labor force participation rate					0.039^ns	0.027^ns
	EA Persons per economic family					0.264^ns	0.2557^ns
	EA Non traditional households					0.130^ns	0.130^ns
	EA Change in neighbourhood (dquart)						-0.201^ns
	EA New Canadians					0.495^ns	0.744^ns
	EA Recent movers					-0.071^ns	-0.059^ns
	EA Asian visible minorities					-2.645^ns	-4.26^ns
	EA Collective Efficacy					0.012^ns	-0.021^ns
	EA Social Support					0.114*	0.124*
	EA "Not a good place to bring up kids"					0.226^ns	0.092^ns
	EA Volunteer work					0.732*	0.661*
	EA Infrequent Religious Attendance					-0.024^ns	-0.059^ns
7.	Family Indicators
	PMK Years of Formal Education						-0.136**
	Renting rather than owning home						0.183^ns
	Recent Immigrant						-1.331^ns
	Stepfamily						0.328^ns
	Parental Work/Study Status
	Both work or study						-0.039^ns
	Neither works/studies						0.270^ns
	Lone Parent works/studies						0.496^ns
	Lone Parent not working nor studying						0.251^ns
	Child Birth Weight		-0.517**	-0.470**	-0.438**	-0.420**	-0.363**
	Maternal age		-0.044*	-0.036*	-0.027^ns	-0.029^ns	-0.013^ns
	Number of siblings		-0.347**	-0.299**	-0.440**	-0.457**	-0.449**
	Number of older siblings		0.338**	0.302**	0.419**	0.426**	0.329**
	Likelihood Ratio Chi Square		14205.1	14077.4	13602.2	13533.9	12634.7
	Variance Components: Empty Model
	Residual ²	5.649**	5.543**	5.530**	5.556**	5.536**	5.568**
	_Children	4.287**	3.526**	3.575**	2.093**	2.120**	2.009**
	_Families	1.795**	1.895**	1.569**	1.289**	1.291**	1.217**
	_{Neighbourhoods}	0.616**	0.647**	0.526**	0.397**	0.341**	0.330**
	Observations Used		2,736	2,724	2,713	2.701	2,529
^*=p<0.01, ^=p<0.05, =0.10 (two-tailed tests of statistical significance), ns = not significant The family work/study variable encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. Birth weight, maternal age and child gender as well as child age, age squared, number of siblings and birth order, are included in all models. Child gender (not shown) has a highly significant effect on hyperactivity-inattention in all models with boys being more hyperactive. ¹N=3,455 child-occasions at Cycles 1 and 3.

As in our analysis of the national sample (Table 8) there is a highly significant relationship between long-term poverty and hyperactivity-inattention in Model 1. This relationship is attenuated as further explanatory variables are controlled and becomes non-significant in Model 5. As predicted by the Family Stress model the coefficient for long-term poverty is reduced by half when measures of family dysfunction and PMK depression are introduced in Model 2. The addition of parenting style measures in Model 3 does not further reduce the estimated effect of long-term poverty but reduces the coefficient for family dysfunction to non-significance while having no effect on the coefficient for PMK depression. These results are consistent with the idea that family stress mediates the effect of long-term poverty upon the child's hyperactivity-inattention and that some measures of parenting style in turn mediate the effect of family dysfunction. The effects of PMK depression are not mediated by any of our parenting measures and remain highly significant in all models up to and including Model 5. Hostile-ineffective parenting has a highly significant relationship with hyperactivity-inattention in all models in which it appears and the same is true for consistent parenting, though at a lower level of statistical significance. Neighbourhood-level measures of social support and the prevalence of volunteer work are positively associated with children's hyperactivity-inattention in models 4 and 5. Analysis of the national sample had shown that being from a stepfamily or a lone parent family where the parent worked or studied made children more hyperactive, but these results are not confirmed in analysis of the neighbourhoods sample.

We show estimates for the random effects part of the model at the bottom of Table 11. These include _{Neighbourhoods} as well as _Families and _Children as the variances of random intercepts at different levels of the hierarchy. As in our analysis of how estimates of fixed effects change when more explanatory variables are added to the model, we can observe how the different variance components change from one model to the next. Starting with the model containing only intercepts as predictors (the "empty model") we observe that between-neighbourhoods variation in child hyperactivity is significantly different from zero, though about one seventh of that which lies between children and about one third of the between-family variance. These estimates can be contrasted with those reported by Boyle and Lipman (1998), though they dealt with data from Cycle 1 only, used a much larger sample of children and employed a different estimation method. Comparing the random effects in Models 1 through 5, it is clear that adding successive groups of explanatory variables gradually reduces the between-neighbourhoods variation until by Model 5, it is almost half its original value though still significantly different from zero. This means that there are persistent neighbourhood differences in child hyperactivity, even after all child-level, family-level and neighbourhood level explanatory variables have been controlled. Further analysis of these data will explore the effects of including data from Cycle 2 and fitting random slope terms of the kind reported in the multilevel modelling of children's physical health status.

4.4.4.3 Hierarchical Linear Models for Predicting Child's Experience of Hostile-Ineffective Parenting

We report a series of multi-level random intercepts models predicting hostile-aggressive parenting in Table 12. As in Table 11 the data are from the neighbourhood sample and only include waves 1 and 3. Because of this and also because of the fact that this parenting scale is only available for children up to age 11, the number of child occasions for analysis is at most 2,742 and is reduced to 2,545 in the final model due to missing data in some explanatory variables. We focus on the sources of hostile-aggressive parenting because our analyses of the national sample confirm previous studies using the NLSCY, which have shown that this style is associated with negative child outcomes.

Table 12 shows four models labelled Model 1, Model 2, Model 4 and Model 5, the absent Model 3 being the one that would have included parenting variables as predictors had we been building models of child outcomes. The fixed effects estimated for the various models show that long-term poverty initially has no effect on hostile-ineffective parenting, a finding that confirms results from the national sample. As other explanatory variables are introduced in Models 2 through 5 the relationship becomes statistically significant, though in the opposite direction to the one expected under our interpretation of the Family Stress Model: that is, net of other factors the experience of long-term poverty is associated with reduced levels of hostile-aggressive parenting. Both indicators of family stress have highly significant effects on hostile-aggressive parenting in all the models and these effects are in the expected direction: family dysfunction and caregiver depression are associated with higher levels of hostile-aggressive parenting. Our aggregated survey measures of Collective Efficacy, Social Support etc. have no association with hostile-ineffective parenting. However, objective indicators of neighbourhood characteristics show very interesting relationships. Higher levels of labour force participation by mothers of children over six has a highly significant impact upon hostile-aggressive parenting (b = 0.249 in the final model). This is the same result as we obtained in the national sample, though the regression coefficient was smaller (b = 0.173). The Census indicator of New Canadians shows a significant relationship with hostile-ineffective parenting such that the higher the incidence of New Canadians, the greater the hostile-ineffective parenting score. Our interpretation of this relationship must take into account the further finding that the presence of East Asian Visible Minorities in the enumeration area is associated with a marked reduction in hostile-ineffective parenting. Overall this style of parenting is affected by family stress and by several objective indicators of neighbourhood type.

Table 12
Neighbourhoods Sample Random Intercepts Models of the Child's Experience of Hostile-Ineffective Parenting at Cycles 1 and 3
Explanatory Variables			Model 1	Model 2	Model 4	Model 5
1.	Intercept		8.563	7.200	8.145	8.223
2.	Poverty Indicators
	Long-Term Low Income (Ltpoverty)		-0.059^ns	-0.217*	-0.222*	-0.218
	EA Average Income for 2+ person hse-holds		0.165*	0.176*	0.028^ns	0.032^ns
	Change in family income to needs					0.001^ns
3.	Family Stress Indicators
	Family Dysfunction Score			0.081**	0.087**	0.089**
	PMK Level of Depression			0.107**	0.106**	0.105**
4.	Neighbourhood Indicators
	EA Female labor force participation rate				0.238**	0.249**
	EA Persons per economic family				-1.034*	-0.985*
	EA Non traditional households				-0.230	-0.251
	EA Change in Non traditional households					0.027^ns
	EA New Canadians				1.317*	1.528*
	EA Recent movers				0.136^ns	0.090^ns
	EA Asian visible minorities				-8.833*	-10.12**
	EA Collective Efficacy				-0.115ns	-0.100^ns
	EA Social Support				0.116	0.099^ns
	EA "Not a good place to bring up kids"				0.371^ns	0.406^ns
	EA Volunteer work				-0.511^ns	-0.474^ns
	EA Infrequent Religious Attendance				0.160^ns	0.125^ns
5.	Family Indicators
	PMK Years of Formal Education					-0.016^ns
	Renting rather than owning home					-0.044^ns
	Recent Immigrant					-0.268^ns
	Stepfamily					0.556
	Parental Work/Study Status
	Both work or study					0.181^ns
	Neither works/studies					0.063^ns
	Lone Parent works/studies					0.233^ns
	Lone Parent not working nor studying					-0.281^ns
	Gender of PMK (female)		1.298**	1.117**	1.050**	0.884*
	Number of siblings in the household		0.307**	0.369**	0.360**	0.322**
	Number of children older than this child		-0.203	0.263*	-0.268*	-0.269*
	Likelihood Ratio Chi Square	14380.8	14282.4	14122.1	14016.6	13084.0
	Variance Components: Empty Model
	Residual ²	6.413^**	6.335**	6.482**	6.439**	6.470**
	_Children	0.251^ns	0.155^ns	0.078^ns	0.092^ns	0.059^ns
	_Families	5.288^**	5.147^**	4.486^**	4.446^**	4.154^**
	_{Neighbourhoods}	1.267**	1.250**	1.009**	0.719**	0.716**
	Observations Used	2,742	2,742	2,729	2,717	2,545
^*=p<0.01, ^=p<0.05 , =0.10 (two-tailed tests of statistical significance), ns = not significant The EA female labour force participation rate is the rate for females with children over six years old only. The family work/study variable also encodes lone parent status: the reference category is a role-specialized couple where one partner works or studies and the other does neither. Maternal age and child gender as well as child age, age squared, birth order, and PMK age are included as predictors in all models.

The random effects shown in Table 12 display a completely different pattern from that typical when predicting child outcomes. Instead of having most of the non-residual variation between children, as almost always happens when modeling child outcomes (Boyle and Lipman, 1998), Table 12 shows that the largest random intercept variance is for _Families. Furthermore the between neighbourhoods variance _{Neighbourhoods} is considerably larger as a proportion of the residual variance when predicting hostile-aggressive parenting than when predicting child outcomes such as physical health. The variance component for children declines in size as we move from the empty model to Model 5, though it is non-significant in all models. The intercept variance for families declines only slightly (from around 5.2 to around 4.1) as further explanatory variables are introduced and remains highly significant in Model 5. The variance component for neighbourhoods is of course smaller and exhibits a steeper decline until by Model 5 it is at around 60% of its value in the empty model, though still highly significant. It appears then that while parents certainly adopt different styles when socializing older or younger, male or female children hostile-aggressive parenting, as measured in the NLSCY has little between-child variation and very significant between-family and between-neighbourhood variation.

²⁶Physical health is measured by PMK responses to questions about the child's general health status hlcq01 and hlcq02. The Health Utility Index might have been preferable but was not asked of all longitudinal children in Cycle 3.
²⁷As discussed above, we have aggregated respondents' judgments about their neighbourhoods to the level of enumeration areas and our analyses of the neighbourhood sample use only those aggregated neighbourhood variables that are based on data from six or more respondents. These respondents do not have to be parents of the children whose outcomes are being predicted in any particular analysis. It is quite possible for example that some the PMKs who provided judgments of a given enumeration area were mothers of very young children while we would be predicting outcomes for seven year olds. The upshot of this focus upon neighbourhoods about which we have aggregated measures is that we drop children from enumeration areas about which we have insufficient information
²⁸While many of the sampled households contribute only one child to the survey, the sampling design for cycle 1 allowed up to four children to be sampled in any economic family. In cycles 2 and 3 the maximum number of children per newly selected economic family was limited to two and while in principle all cycle 1 children were to have been followed up, it appears that the definition of a "longitudinal child" excludes those children who were one of three or four children sampled from the same economic family in cycle 1. The essential point here is that while cycle 1 of the NLSCY has a certain amount of clustering of children sampled from the same economic family, the definition of "longitudinal children" is such as to minimize the consequences that this might have for respondent burden and accurate estimation of standard errors. The maximum number of longitudinal children per household at Cycle 3 is two. Cross sectional estimates from cycle 1 data are of considerable interest when estimating models which focus upon the effects of household structure upon child outcomes but longitudinal children are of much less use in this respect.
²⁹No questions had been asked about neighbourhoods in Cycle 2. While our method of aggregating PMK responses about neighbourhoods to the EA level should allow us to use aggregated neighbourhood measures for Cycle 2 observations, we took a conservative approach here and excluded them.

Contents

Poverty, Social Capital, Parenting and Child Outcomes in Canada - March 2002