Poverty and Child Well-Being in Canada and the United States:Does it Matter How We Measure Poverty? - September 2000

3. Multivariate Analysis of the Probability of Child Poverty Using SCF versus NLSCY data and Alternative Poverty Lines

Previous

Contents

Next

Given these rather striking differences in the estimated incidence of poverty, it seems appropriate to proceed to a multivariate analysis of the factors associated with an increase in the probability of a child living in poverty using the different data sets and poverty lines. Tables 3a and 3b report probit models of the probability that a child is observed to live in poverty, for Canada and the US, respectively.¹³ For both Canadian data sets as well as for the US Current Population Survey (CPS), we focus on the sample of all children aged 0 to 11 years, regardless of mother age, but also report one example of a regression using the restricted mother-age sample. Estimates for the National Longitudinal Survey of Youth (NLSY) can obviously only be for the restricted age sample.

The specification employed for the probit models is extremely parsimonious both because overlap in survey content across 4 different data sets was necessary for this exercise and because the intent of the regressions is not to 'explain' poverty, but merely to check basic correlation patterns. Thus, control variables are limited to number of children aged less than 18 years and a series of dummies indicating less than high-school level of education for the mother,¹⁴ that the child is aged 7 to 11 years, that the child lives in a lone-parent family, and that the mother is aged 35 or more.

Appendix Table A1 provides sample means for the estimating samples. Note, first, that the two Canadian samples for all children appear very similar, except that a slightly larger percentage of mothers have less than high-school education in the Survey of Consumer Finance (SCF) (20.3% versus 16.2%). If we compare the two mother age-restricted samples for the US, sample means are quite similar. But, how do the characteristics of the age-restricted sample differ from those of the full sample? If we compare the two sets of means for the CPS, it is first, of course, true that a smaller percentage of mothers are older than 35 in the age-restricted sample (31.6% versus 44.6%). It is also true that a smaller percentage of mothers in the age-restricted sample have less than high-school education (13.3% versus 18.5%). Children are slightly older in the age restricted sample (44.1% versus 41.4%) and fewer live in lone-parent households (22.4% versus 26.7%). Overall, it does seem important to keep in mind that NLSY data set is not entirely representative of all US children of a particular age.

This point is reinforced by examining the differences between the age-restricted and full samples for the Canadian data sets, where most of the same patterns are apparent (e.g., mothers are less likely to have low education in the age-restricted sample and children are less likely to be living in a lone-parent household). One cross-country difference is that for both Canadian data sets, children with mothers aged 29 to 37 are slightly less likely to be aged 7 to 11 (e.g., 38.9% versus 41.7% using the NLSCY) while the opposite is true for the US data.

Turning to the multivariate results, we consider first the two Canadian probability of poverty equations estimated using the Low-Income Cutoffs (LICO's) with the SCF and NLSCY, respectively (see Table 3a). While most estimated coefficients are somewhat larger using the NLSCY than using the SCF (e.g., low education, lone-parent, number of children), the key difference between the two estimated equations is that the age of the child is not statistically significant in the case of the SCF, but being aged 7 to 11 years is associated with a lower probability of poverty when the full sample of NLSCY data is used. This pattern holds regardless of the poverty measure employed and seems reasonable since school-aged children are presumably less of an impediment to labour force participation. Note that one important difference between the SCF and NLSCY which may be relevant for this finding is that the NLSCY weighting system provides more accurate estimates of the distribution of children across age categories than the SCF (see footnote 13).

If we focus instead upon a comparison across measures of poverty rather than across data sets, one key difference is that the number of children living in the family is associated with larger increases in poverty when using the OECD equivalence scale than, for example, when the LICO's are employed. This is not surprising, since the LICO's assume larger economies of scale are available to individuals who live together. Thus, at the same income, larger families are more likely to be classified as poor using the OECD than the LICO approach. This point is valid regardless of the data set employed.

Finally, if we compare estimates obtained using the mother-age restricted sample with the full sample (using a 50% of median equivalent income poverty line calculated with a LIS equivalence scale), the dummy variable for older children being less likely to be poor drops to insignificance with the NLSCY.¹⁵

For the United States, if we compare coefficient estimates obtained with different data sets but the same measure of poverty (e.g., the US official poverty lines), it is again true that the most notable difference is for the estimated association between child age and child poverty. Using the full sample of CPS data, children aged 7 to 11 years are less likely to be poor (as was true using the NLSCY but not the SCF); using the age restricted sample and the LIS equivalence scale, there is no significant difference in the probability of poverty for older and younger children (and this was the same finding with the age-restricted Canadian samples). However, using the NLSY, older children are more likely to be poor. It is not obvious to us why this should be the case.

If we compare across poverty lines using the same data set for the US, we again find that the estimated coefficient on number of children living in the family is largest when OECD equivalence scales (assuming smaller economies of scale) are employed.

Finally, it is interesting to compare Canadian and US estimates. While it is obvious that the absolute value of coefficients involved in the comparison will be sensitive to the poverty measure chosen, in fact most of the qualitative points are true regardless of poverty measure. We focus first on the SCF and CPS estimates, using the unrestricted mother age samples. Low education of the mother and lone parenthood are associated with higher rates of poverty in both countries, though the magnitude of both effects is much larger in the United States. Additional siblings are also associated with higher probabilities of poverty in the United States than in Canada. If the mother is aged 35 or higher, the associated probability of child poverty is lower, but this is more dramatically the case for the US than for Canada. An important difference is that older children are less likely to be poor using the CPS data, but this effect is not observed using the SCF.

If we compare the age-restricted NLSCY estimates with the NLSY estimates (for the 'LIS' poverty lines), most of the points made above remain valid. That is, children living with lone parents or whose mothers have low levels of education are more likely than others to be poor, but the magnitudes of these associations are much larger in the US. Additional siblings are associated with higher probabilities of poverty in both countries, but the association is stronger in the US; children with older mothers are less likely to be poor, but more so in the US. Controlling for mother's age, there is no association between child age and poverty in the mother-age -restricted sample of the NLSCY; older children are more likely than others to be poor in the US.

Two themes have been emphasized in the paper thus far. The first is that estimates of child poverty are not the same using the NLSCY and the SCF for Canada or using the NLSY and the CPS for the United States. The SCF is generally believed to produce the best estimates of income and poverty in Canada (e.g., the SCF is the data set used by Statistics Canada to produce estimates of income distribution and low-income in Canada -- see Statistics Canada, 1997 a or, b). Thus, there is reason to be more comfortable with the SCF than the NLSCY estimates. The troubling feature for the over-all purposes of this research is that to understand the links which exist between child health and poverty, we are forced to use the income information available in the NLSCY. Perhaps future waves of data collection could focus upon improving the income content of the survey. Until that time, it remains possible that one reason for some of the current findings (e.g., of the relative unimportance of low-income status) is limited information about income. Until better income information is available in data sets concerning child well-being, this will continue to be an issue.

The second theme is that choice of poverty line can affect our estimates of the incidence and correlates of child poverty. A central question of this paper is then whether choice of poverty line may also influence our perception of whether or by how much poverty matters for child well-being and whether this is consistent for different dimensions of well-being? This is the topic of the next section.

 

Table 3a Probit Estimate of the Probability of Being Poor. Canadian Children, Aged 0-11 in 1994

	Survey of Consumer Finance					National Longitudinal Survey of Children and Youth
	LICO1	US Official  (in Can $)	OECD2	LIS3		LICO1	US Official  (in Can $)	OECD2	LIS3
				ALL	PMK 29-37				ALL	PMK 29-37
Dummy = 1 if PMK < High School Education	0.699* (0.027)	0.558* (0.030)	0.730* (0.027)	0.710* (0.027)	0.712* (0.039)	0.768* (0.026)	0.751* (0.027)	0.758* (0.026)	0.808* (0.026)	0.821* (0.037)
Dummy = 1 if Child 7-11 years old	-0.032 (0.026)	-0.039 (0.030)	-0.029 (0.025)	-0.035 (0.026)	0.011 (0.035)	-0.143* (0.023)	-0.124* (0.026)	-0.139* (0.022)	-0.139* (0.023)	-0.025 (0.030)
Dummy = 1 if Lone parent  Family	1.372* (0.029)	1.103* (0.031)	1.222* (0.029)	1.464* (0.030)	1.553* (0.043)	1.470* (0.027)	1.269* (0.026)	1.352* (0.027)	1.529* (0.027)	1.620* (0.039)
Number of children < 18 years in  household	0.148* (0.011)	0.176* (0.012)	0.275* (0.011)	0.192* (0.011)	0.256* (0.016)	0.196* (0.010)	0.251* (0.011)	0.338* (0.010)	0.242* (0.010)	0.242* (0.014)
Dummy = 1 if  PMK>= 35 years  old.	-0.343* (0.026)	-0.235* (0.030)	-0.338* (0.026)	-0.384* (0.026)	-0.173* (0.037)	-0.375* (0.023)	-0.239* (0.026)	-0.439* (0.022)	-0.402* (0.023)	-0.188* (0.032)
Intercept	-1.474* (0.031)	-1.961 (0.036)	-1.715* (0.031)	-1.572* (0.031)	-1.931* (0.047)	-1.370* (0.027)	-1.993* (0.032)	-1.577* (0.028)	-1.509* (0.028)	-1.764* (0.041)
*significance with 99% confidence. 1. Low-Income Cut-off. 2. Organization of Economic Cooperation and Development. 3. Luxemburg Income Study.

 

Table 3b Probit Estimate of the Probability of Being Poor. US Children, Aged 0-11 in 1994

	Current Population Survey					National Longitudinal Survey of Youth-Children
	US Official	OECD1	½ Median Canadian  OECD Scale (US $)	LIS2		US Official	OECD1	½ Median Canadian  OECD Scale (US $)	LIS2
				ALL	PMK 29-37
Dummy = 1 if PMK < High School Education	1.463* (0.040)	1.761* (0.039)	1.771* (0.039)	1.684* (0.039)	1.964* (0.065)	1.364* (0.126)	1.411* (0.120)	1.410* (0.120)	1.370* (0.123)
Dummy = 1 if Child 7-11 years old	-0.246* (0.038)	-0.200* (0.033)	-0.208* (0.033)	-0.171* (0.034)	0.022 (0.049)	0.227** (0.108)	0.342* (0.093)	0.343* (0.093)	0.389* (0.096)
Dummy = 1 if Lone parent  Family	2.368* (0.037)	2.153* (0.035)	2.161* (0.035)	2.296* (0.035)	2.574* (0.054)	2.963* (0.113)	2.811* (0.102)	2.809* (0.102)	3.008* (0.103)
Number of children < 18 years in  household	0.465* (0.014)	0.574* (0.014)	0.573* (0.014)	0.401* (0.013)	0.491* (0.020)	0.539* (0.048)	0.676* (0.046)	0.675* (0.046)	0.549* (0.046)
Dummy = 1 if  PMK>= 35 years  old	-0.792* (0.039)	-0.871* (0.034)	-0.869* (0.034)	-0.976* (0.035)	-0.511* (0.055)	-0.409* (0.123)	-0.636* (0.106)	-0.633* (0.106)	-0.608* (0.110)
Intercept	-3.470* (0.050)	-2.809* (0.044)	-2.792* (0.044)	-2.509* (0.042)	-3.331* (0.070)	-4.555* (0.177)	-3.958* (0.154)	-3.956* (0.153)	-3.872* (0.155)
*significance with 99% confidence. **significance with 95% confidence. 1. Organization of Economic Cooperation and Development. 2. Luxemburg Income Study.

• ¹³Sample weights are employed for all regressions.
• ¹⁴Observations were excluded in the event of non-response to any of the variables used in the estimating model. One difference which exists between the SCF and NLSCY samples, but which should make little difference to the results reported here, is that in the NLSCY, extensive use has been made of the 'persons most knowledgeable' (pmk) about the child. In the regressions, we use age and education of the pmk as control variables. Most, but not all pmk's are mothers. Hence, in the SCF data, we use age and education of the mother, unless the mother is not present, in which case appropriate values for father or other care-giver are substituted. While these procedures are not identical, the coefficients on the 'mother' variables are remarkably similar across data sets. For both US data sets, we use information about the mother, where available.
• ¹⁵We estimated probits for all measures of poverty with the restricted age sample and the same conclusion was valid regardless of poverty line. We report only one set of coefficients to save space.