Children and Familial Economic Welfare: The Effect of Income on Child Development

3. Methodological Considerations and Hypotheses

3.1 Sample Information and Descriptive Data Selection

The data used in this study are obtained from the share file of the NLSCY which is a nationally representative sample of children collected in 1994-95 and 1996-97. The unit of analysis in the survey is the child. All information and analyses, therefore, must be interpreted from the standpoint of the child rather than that of the family or parent. The total sample of children surveyed in 1996-97 was 20,025, ranging in age from 0 to 13 years. Our analysis in this paper focuses on the share file cohort of children aged from 2 to 13 years old in 1996 (ages 0 to 11 in 1994), yielding a sample of 15,266 children. Our reason for choosing this sample was to represent as wide an age range as possible and also thereby obtain a broad number of developmental outcomes to investigate. Children one year or younger were excluded because information for them was not available for both cycles. As in other research using the NLSCY (Lefebvre and Merrigan, 1998) we decided to concentrate our analysis on children in two parent and lone mother parent situations, and where the person most knowledgeable (PMK) was either the mother or father (including biological, adoptive and step).

The analyses use longitudinal and cross-sectional weights where appropriate to obtain the results and estimates, which were provided on the NLSCY database. Statistics Canada release guidelines for data quality have been followed for this analysis. Where any tests of significance were necessary we constructed a new "sample" weighting variable for the sample population of children ages 2 to 13 years in 1996. This new sample weighting variable was constructed by dividing each respondent's existing longitudinal weight by the mean of the overall longitudinal weights. The new sample weight has a mean of one, but avoids overestimations for tests of significance while maintaining the relative positioning or distribution of the original variables being tested.

For the descriptive data analysis we constructed a number of income, family structure and wage earner change variables, as well as recoding the child outcomes. Income change is measured in two ways: using the proportional change in household income from 1994 to 1996, and using a derived variable of whether children's household income changed relative to Statistics Canada's low-income cut-off (LICO) from 1994 to 1996. Family structure change is measured as a derived variable that observed whether children were in the same family type (lone-parent or two-parent) in 1996 compared to 1994. Wage earner change is calculated as the change in the number of parents participating in the paid labour market from 1994 to 1996.

Our child outcome variables were recoded in two different ways. For the Peabody Picture Vocabulary Test (PPVT) and Motor and Social Development (MSD) standardized scales we used pre-existing cutoffs to differentiate delayed, normal and advanced development. We then calculated whether children had stayed or moved from one category to the other from 1994 to 1996. For example, if they moved from advanced to normal or delayed development on one of the scales, then they were considered to have a worse score in 1996 than in 1994. For the behavioural variables we used a technique that Offord and Lipman (1996) developed. This technique views the top 10 per cent of the distribution on a scale (such as the property offences scale, for example) as exhibiting the behaviour being studied. Applying the 1994 scale score as the cut-off for 1996, we then established whether children were still in the same category in 1996 as in 1994. If they were, then they had not changed. A child who went from the top 10 per cent in 1994 to below the top 10 per cent in 1996 were viewed as improving, and vice versa meant the child's behaviour scored had worsened.

3.2 Regression Model Data Selection

The most common manner in which the effect of income security on children's outcomes is studied, uses a reduced form OLS regression analysis (Blau, 1999; Duncan et al., 1998; Mayer, 1997). As Mayer (1997) explains, the reduced form model does not try to identify all of the possible mechanisms through which family income affects children's outcomes. Instead, this strategy wants to observe what the direct effect would be of an increase in family income on children's outcomes. For example, how much of an effect would an increase of $10,000 have on a particular child outcome. Conventionally a reduced form regression model controls for those characteristics that affect the relationship between household income and children's outcomes, but which are causally prior to the income variable. This is not quite as straight forward as it may seem.

The causal relationship between income and the independent control variables is important to bear in mind when selecting variables for inclusion in the analysis. We do not want to include in our analysis any variables which may be a result of income, except for children's developmental outcomes. By controlling the effect of these variables in our regression we effectively reduce the impact that income will have on the child outcomes. Mayer (1997) for example, argues that the inclusion of a marital status control reduces the effect of income on children's outcomes. Marital status is in part a result of income (low-income men and women are less likely to marry when they have a child than men and women with higher incomes, and when low-income men and women do marry they are more likely to separate or divorce), but it is also the case that marital status causes income (lone-parent families are relatively poorer due to have only one earner). There is no way to estimate just how large the underestimation of the income effect would be in this case. Mayer (1997) admits that excluding marital status variables leads to income effects that are over estimated. However, other studies do include marital status including Duncan et al. (1998), to account for the effect that reduced incomes can have on children in these families. As we discuss below, our study relies on previous research to select the control variables, although in some cases their inclusion may be debatable.

Once we have controlled for the effect of background characteristics then the direct effect of income on child outcomes can be observed and discussed. Unfortunately we are unable to replicate the study by Mayer (1997) to obtain the "true" effect of income, or the study by Duncan et al. (1998) investigating the effect of siblings, because the NLSCY database does not contain the appropriate variables. We will instead rely on the reduced form OLS regression approach to estimate the effect of income on our dependent child outcomes.

3.3 Regression Variable Description and Construction

3.3.1 Child Outcomes: Dependent Variables

For our analysis we wish to use a wide variety of outcome measures. The literature review demonstrates that the effect of income depends on the outcome variable used and the age of the child. We use two cognitive development measures from the 1996 NLSCY, Motor and Social Development (MSD) and the Peabody Picture Vocabulary Test (PPVT), which cover children in the age ranges from 2 to 3, and 4 to 7 years respectively. As well, we investigate the influence of income on six behavioural scales measuring children's hyperactivity-inattention, prosocial behaviour, emotional disorder-anxiety, aggression, indirect aggression and property offences, for children aged 4 to 11 years in 1996. We chose the behaviour scales as dependent variables because American research has shown the Behaviour Problems Index (BPI) in the NLSY to be related to income (Mayer, 1997; Hanson et al., 1997). Lastly, we observe the effect of income on two academic measures - Reading and Math - for children aged 10 to 13 years in 1996. Following Blau (1999) we divide each of our dependent variables by its standard deviation. This will allow us to express any changes in our regression coefficients in terms of standard deviation units for each dependent variable.

3.3.2 Home Environment Dependent Variable

The importance of the home environment for children's development is recognized by social researchers (Blau, 1999; Lefebvre and Merrigan, 1998; Duncan and Brooks-Gunn, 1997; Hanson et al., 1997). This importance is reflected in the fact that data gathered for the NLSY in America allows for the construction of the Home Observation for Measurement of the Environment (HOME) scale composed of cognitive, social and physical environment variables (Jekielek et al., 1998). This index, or a shortened version of it, has been used in a number of American studies that investigate the effect of economic security and child outcomes (Blau, 1999; Jekielek et al., 1998; Smith et al., 1997). Evidence from these studies suggests that the HOME scale is associated, although somewhat modestly, with children's developmental outcomes. As well, Blau (1999) demonstrates that family income is quite strongly related with the HOME variable. Given the evidence from other studies pointing to the connection between income and the HOME scale, as well as between the HOME scale and children's developmental outcomes, we have chosen to try and capture the extent of the children's home environment for our sample of children by constructing our own index variable to serve as a substitute for the HOME environment variable.

In a recent Canadian study, Lefebvre and Merrigan (1998) used the extent of literacy activity in children's homes as a measure of the cognitive environment. While this recognizes the importance of the home environment for children, it doesn't capture enough of the facets included in the HOME scale. Our proxy home environment variable will include literacy activity as well as some other variables that approximate the variables used in the HOME scale to capture the social and physical environment of children's homes. Since many of our child well-being measures, as well as many of the possible variables to measure the home environment, differ by children's age, we have constructed four separate home environment proxies — reflecting the age groups of the children to which they apply.

The HOME scale was developed by Caldwell and Bradley as a 45 minute inventory designed to assess particular characteristics of young children's environment (Desai et al., 1990). The HOME measure is obtained by means of a self-report section by the child's mother and an interviewer evaluation. Two major subscales compose the measure, the cognitive stimulation available to the child and the emotional support provided by the mother (Desai et al., 1990). The original scale is considered to be a very reliable measure but most of the recent research has used smaller sub-scales of the original, including Blau (1999), Jekielek et al. (1998) and Desai et al. (1990).

The construction of our proxy variables attempts to reconstruct the HOME scale using the item list from Blau (1999), Jekielek et al. (1998) and Desai et al. (1990). A problem with this is that the NLSCY does not contain assessment data from the interviewer in cycle 2. Our proxy variables, therefore, will rely solely on information provided by the PMK. Variables from the NLSCY are chosen which we feel are close in content and meaning to those items listed in the studies mentioned. As described in Blau (1999), each of the items composing the HOME environment index are dichotomized so that a score of one indicates a better home environment. We have followed this procedure as well. Our proxy variables are obtained from the dichtomous components by simply summing across the various components. The final proxy variables have scores ranging from 0 to 4, where a higher score indicates a better home environment.

The first proxy used in our analysis - Home Environment Proxy 1 — measures the home environment for children ages 2 to 3 years and consists of the following four components:

Whether or not the PMK or others ever read to the child. Those who responded "yes" were assigned a value of 1, while those who responded "no" were assigned a value of 0.
A second cognitive component focus on "How often does the child play with pencils or markers doing real or pretend writing?". Children who did so "once a week" or less were assigned a 0, those who did it "a few times a week" or more were given a 1.
We also included a third cognitive component that again focuses on the parent-child interaction and learning, specifically "How often do you help or encourage the child to write or pretend to write?". This approximates the HOME variables regarding whether the parent helps their child to learn the alphabet, numbers, colours and shapes. Those parents who helped their children "once a week" or less were assigned a 0, those who did it "a few times a week" or more were given a 1.
A family involvement variable is included that addresses the extent to which children are exposed to a lack of emotional support. The original HOME variable uses information over two components regarding physical punishment (such as spanking) and the anger of the child's parents. We have used a parenting question "When "child" breaks rules or does things that they are not supposed to, how often does the parent use physical punishment?". We have coded children whose parents "always", "often", or "sometimes" do this 0, while parents who "rarely", or "never" do this were given a 1. PMK's participation and interaction with the child. This used the question "How often do you and [child] talk or play with each other, focussing attention on each other for five minutes or more, just for fun?". Those who responded "never", "about once a week or less", or "a few times a week" were assigned a value of 0, while those who responded "one or two times a day", or "many times each day" were assigned a value of 1.

Our second home environment proxy (Home Environment Proxy 2) , covers children 4 to 7 years of age corresponding to the age range for the Peabody Picture-Vocabulary Test - Revised (PPVT-R). This proxy consists of four components also. We use both the parental help with reading, or reading alone for the older children (7 years), and the parental physical punishment components used in Home Environment Proxy 1 and four additional variables described below:

A variable addressing the child's television and home video viewing was included. The original HOME scale measured whether the TV was on 4 hours or less per day. Using this cutoff we constructed a variable on children's TV or videos habits at home. Those who watched 4 hours a day or less were assigned a value of 1, while children who watched 5 hours or more per day of TV were assigned a value of 0.
The fourth component of the Home Environment Proxy 2 variable was regarding children's structured recreational activities. The original HOME scale measured whether children obtained special lessons or had access to a musical instrument. We have used the structured recreational activities to approximate this observing whether children participated in organized sports, took dance, gymnastics or martial arts, took lessons in music, art or was involved in a community activity such as guides, cubs or a church group. Those who participated once a week or more frequently were assigned a value of 1, those who participated less frequently or not at all were assigned a value of 0.

Our third and fourth home environment variables (Home Environment Proxy 3 and Home Environment Proxy 4) are used for children ages 4 to 11 years and 7 to 11 years, corresponding to the age ranges for the six behavioural outcomes and the reading/math scores, respectively. The same four components that were used in the Home Environment Proxy 2 are included in the last two proxies. The differences between the three are with regard to the relevant age related questions used to generate the proxy. For example, the structured recreational activity questions used in the Home Environment Proxy 2 were answered by the parents and refer to children aged 4 to 9 years. Older children self-respond to similarly worded questions regarding their participation in various structured recreational activities, their TV watching, and their reading habits. The parental question regarding physical punishment is answered by the parents for children of all ages.

Given the lack of interviewer information and low correspondence between the NLSCY questions and those found in the original HOME scale, our four components for composing each proxy variable is somewhat less than the 11 to 13 components in the HOME sub-scales used by Blau (1999), Jekielek et al. (1998) and Desai et al. (1990). The distribution and coverage of our four proxy variables are therefore limited compared to the original HOME scale (Table 1 - see Appendix for all Tables). Nevertheless, they provide us with some measure of the extent to which the home environment is affected by income. A reliability test of the four home environment proxies demonstrates the limits of these variables, providing low to very low Cronbach's alpha scores (Table 2). However, a correlation performed with two of the income variables we have constructed and each of our four home environment proxies shows that the proxies are definitely correlated with income, although we observe these are range from very low to low/moderate correlations (Table 2).

3.3.3 Independent Variables

Income Variables

The exact choice of which variables to include in our model is not straightforward. A variety of income measures have been used in past research including current income, permanent income, logarithmic income, categorised income, income-to-needs ratio, and poverty. Current income has been shown by Mayer (1997) and Blau (1999) to be a less than adequate measure of a family's income over time. Fluctuations in income associated with specific time periods are considered to be transitory in nature and not to have much influence on children's development because families maintain their consumption patterns by borrowing against future earnings (Mayer, 1997: 63). Therefore a more accurate representation of income is to use an average of the family's income over a set time period, that is "permanent" income. Since we only have two cycles we will use an average of the household income using information from the 1994 and 1996 NLSCY. Again following Blau (1999), we express our "permanent" income measure in units of $10,000. As well, we calculate a logarithmic "permanent" income measure to adjust for income's skewed distribution, which is expressed in natural log units.

We also adopt a categorical income variable, following from Duncan et al. (1998). This allows us to observe the effect of income transitions across discrete income breaks. We do not include an income-to-needs measure in our analysis because Bane and Ellwood (1985) argue against the use of income-to-needs ratios. They contend that while we may talk about "permanent" and "transitory' components of family earnings it is not clear that this same logic can be applied to the income-to-needs concept. As well, using an income-to-needs ratio would not allow us to estimate the separate effects of income and needs, such as family size. Lastly, we decided to include in our model a dichotomous poverty variable, derived from the 1996 LICO variable in the NSLCY database. It is coded 0 for non-poor and 1 for poor families. This will provide important information about the effect that movement out of poverty may have on children's developmental outcomes.

Other Control Variables

Mayer (1997) argues for a very limited number of control variables: age and race of the child, household size, mother's age at child's birth, mother's measured intelligence, and mother's education. Blau (1999) on the other hand excludes variables such as parental labour supply, household structure and parental education, since these are potentially jointly chosen variables with income. That is, income may influence your choice of whether and when to have children, or get more education or enter the labour market. Blau does control for other variables such as gender of the child, ethnicity/race, age of child, mother's education (present date), number of siblings (older and younger), marital status of the mother, mother's age and measured intelligence, and variables relating to the child's grandparents and geographic location where mother was raised. Duncan et al. (1998) use household structure and labour participation as control variables as well as the race and gender of the child, the total number of siblings of the child, mothers age at child's birth, mother's years of schooling, residential mobility and where the mother was raised.

Since many of the geographic and extended family variables used in these other studies are not available on the NSLCY we cannot use them in our research. Instead of using race or ethnicity in our study we will use, as do Lefebvre and Merrigan (1998), a variable indicating whether the mother of the child was an immigrant to Canada. We use this variable as a dichotomous variable (yes or no, mother was an immigrant) while Lefebvre and Merrigan categorise the time since immigration. We include family structural variables as controls in keeping with Duncan et al. (1998) and because of the research literature showing the connection. The family structure variables consist of a dichotomous variable representing lone-parent family membership in 1996, and a longitudinal variable measuring the change in family structure from 1994 to 1996 (whether the family structure stayed the same, either two parent or lone parent, or changed from a two parent to lone parent, or vice versa.

Following the lead of Blau (1999) and Duncan et al. (1998), we include family variables such as the size of the family, the gender of the child, and the number of siblings, in our analysis. As well, characteristics of the mother are integrated into the analysis, including mother's age at birth of child, mother's education, and mother's employment. We also integrate a geographic variable to control for background characteristics of urban versus rural dwellers. We divided the urban dwellers between large urban centres (100,000 people or more), small urban centres (less than 100,000 people) and then rural dwellers.

3.4 Hypotheses

The main hypothesis to be tested in this study is that a family's economic circumstances - as measured by income level and income change - will affect child outcomes. We expect that children from families with lower income security are more likely to have negative behavioural outcome scores and lower cognitive development scores (as measured in the NLSCY). However, children from families with higher income security are more likely to have positive behavioural outcome scores and higher cognitive development scores. We also believe that changes in familial circumstances such as divorce and separation, as well as labour market changes such as a decrease in the number of earners in a family, will have a negative impact on the income security of children's families.

Contents

Children and Familial Economic Welfare: The Effect of Income on Child Development - April 2001