Social Context of Welfare in Manitoba, Canada

Abstract

Social stratification and life course approaches are enlisted to study the effects of health and social events on early adult welfare use. Given the strong link between individual and parental economic disadvantage, the mechanisms by which social context affects welfare use are examined. This unique approach is made possible by the linkage of several administrative databases in Manitoba, Canada, allowing for the follow-up of a large population (n = 42,598) and subpopulation of siblings (n = 7920) from birth to age 26. Gradients of inequality were found for many of the predictors. Regardless of background, improved educational achievement and better childhood and adolescent mental health seem likely to decrease the use of welfare in early adulthood. Addressing the risk factors identified in this study would reduce inequities and lower the need for welfare in early adulthood.

Appendices

Appendix 1: Covariates

Covariates	Neighborhood income quintile at birth
	1	2	3	4	5
Moved
Ages 0–3	3284	3410	3368	3936	3683
Ages 4–8	3923	3888	3891	4607	4737
Ages 9–13	3271	2773	2696	2833	2980
Ages 14–17	2928	2212	1915	1723	1949
Major mental health conditions
Ages 0–3	17	11	12	17	11
Ages 4–8	35	19	20	32	26
Ages 9–13	71	64	61	72	78
Ages 14–17	306	242	259	241	257
Minor mental health conditions
Ages 0–3	689	765	807	1062	1049
Ages 4–8	808	940	1041	1232	1140
Ages 9–13	1015	1076	1235	1341	1301
Ages 14–17	1773	1819	1990	2169	2195
Major injuries
Ages 0–3	2838	3230	3539	4180	4179
Ages 4–8	2731	3040	3344	3985	4086
Ages 9–13	2461	2756	3109	3672	3936
Ages 14–17	2220	2376	2688	3140	3313

Appendix 2: Model Validation

2.1 Cross Validation

We used tenfold cross validation to check on over-fitting. The following steps used to cross validate our sample are taken from Sainani (2013).

• 1. Randomly divide your data into 10 pieces, 1 through k.

• 2. Treat the 1st tenth of the data as the test dataset. Fit the model to the other nine-tenths of the data (which are now the training data).

• 3. Apply the model to the test data (e.g., for logistic regression, calculate predicted probabilities of the test observations).

• 4. Repeat this procedure for all 10 tenths of the data.

• 5. Calculate statistics of model accuracy and fit (e.g., ROC curves) from the test data only.

Table 7 shows the C-statistics (the measure of fit) for the full model (as presented in the paper) and those obtained after cross-validation. The fit stat

Table 7 C-statistics and confidence intervals of full model and cross-validated model (neighborhood income quintile at age 18)

istics were not significantly different, indicating our models were robust.

2.2 Bootstrapping

Bootstrapping was done to determine robustness of the standard errors associated with our estimates. Unrestricted random sampling with replacement was done at the individual level; each outcome was modelled 500 times with different randomly selected samples. Table 8 highlights the similarities between the confidence intervals from the original models and the bootstrapped confidence intervals; the significant predictors remain the same between the two models.

Table 8 Model and bootstrapped confidence intervals, neighborhood income quintile models

Appendix 3: Specific Mental Health Conditions

Specific mental health conditions are defined using ICD codes from physician claims and hospitalizations. All medical records used ICD-9-CM codes during the relevant period; however, hospital discharge codes switched from ICD-9-CM to ICD-10-CA/CII in 2004/05. These ICD-10 files were converted to ICD-9-CM files using an existing crosswalk (MCHP 2006). The ICD-9-CM system classifies mental, behavioral and neurodevelopmental disorders under codes 290–319 (CDC 2012). Only diagnoses with at least 6 individuals in each subgroup are included to protect the privacy of individuals.

Among welfare recipients, the percentage of individuals in each neighborhood income quintile with a mental health diagnosis grew as neighborhood affluence increased (with the exception of ‘hyperkinetic syndrome of childhood’ and ‘nondependent abuse of drugs’) (Fig. 2). Access to services does not seem to be an issue; the utilization of mental health services is relatively constant across quintiles for those not using welfare.

Fig. 2

Mental health conditions of individuals who did and did not receive welfare in early adulthood, by neighborhood income quintile at age 18

Appendix 4: Gradient in Odds Ratios across Neighborhood Income Quintiles

For each covariate, estimates from each of the five logistic regressions were modelled as a general linear model with income in the corresponding quintile (1–5) as the outcome. Since estimates are from different populations, bootstrapping was done to determine the standard error of the slope (with 500 replications) (Efron and Tibshirani 1998). Confidence intervals were constructed from the bootstrapped standard errors; if the confidence interval did not contain zero then the slope (or gradient) was deemed significant. See Table 9.

Table 9 Odds ratios for models including family instability variables (N = 32,743)

Appendix 5: CFS Models and Variable Correlations

See Table 10.

Table 10 Pearson correlation coefficients for CFS variables