Parents with an Unemployed Adult Child: Consumption, Income, and Savings Effects

From the full PSID sample, we create an analytical sample of 38,004 person-year observations of mothers matched to independently living adult children between 1985-2013. Details of the matching process, including matching rates and a comparison of dependent and covariate variables across matched and unmatched mothers, are discussed in Appendix B. Our analysis relies on mother-child pairs only. We are able to match mothers at a much higher rate than fathers, and most of the matched fathers were in households with previously matched mothers. Hence, the only excluded parents from our sample are single, custodial fathers and divorced, non-custodial fathers. The former makes up a very small share of the populatio

The U.S. Census Bureau estimates that from 1970 - 1990, when the adult children in our sample were minors, the share of households with children under 18 that were headed by a single father averaged just under 2 percent (Annual Social and Economic Supplement, 2016).

Table 1 summarizes all of the matched mothers in our sample, together and separated into three age groups: working age (less than 62 years old), the retirement windo

This is defined by Social Security, the early retirement age is 62 years old; and after age 70, there is no longer a financial incentive to delay claiming.

Figure 1

Indeed, we find basic demographic differences between the groups beyond age. Our sample is 77.2 percent white, and just under half have any college experience, at 42.3 percent. But working-age mothers have higher educational attainment than retiring or retired mothers, at 45.3 percent with some college attendance, compared to 25.6 percent and 16.6 percent, respectively. In addition, working-age mothers have a higher share divorced (at 24 percent) compared to older mothers (19.7 percent and 10.5 percent) and a lower share self-reporting to be a housewife (12.1 percent, compared to 14.2 and 17.2 percent). Clearly, it is important to keep in mind the latent cohort differences when interpreting the subgroup results, and important to control for the year of birth of the mother.

The bottom half of Table 1 summarizes the dependent variables that we will use in our analysis

Appendix A, figures A2 - A13 show the distribution of all of these variables by age of mother.

The second group of variables describes usual household food consumption, which combines the monetary value of food stamp benefits with the reported spending on food consumed at and away from home. It is the only consistent measure of consumption in the survey. Total household food consumption declines with age, from $8060 to $5060, though per capita food consumption (the household total divided by family size) rises and then falls, from $4738 for working-age mothers, to $5361 and then $4430 for the oldest mothers. These variations could reflect economies of scale in larger households, changes in preferences towards or ability to afford more expensive food, as well as reductions in food consumption overall. However, it likely also reflects a measurement issue. Usual household food consumption is measured over a year, but family size is measured at the time of survey. If the family size changed over the course of the year, the members contributed to consumption, but may not be included in family size, depending on when they exited or entered. For this reason, we show per capita consumption in this summary, but will not use it as a dependent variable in our analysis.

The third group of variables summarizes sources of income, including the total weeks worked in the year, the share of mothers whose household has any Social Security Income

The PSID varies in its collection of information regarding Social Security income, asking about individual receipt in some years and family receipt in others. We use household-level summaries in all years for consistency, but that means we cannot identify the recipient within the household for most years of the survey.

Finally, the PSID added in 1999 regular measures of wealth, including individual contributions to pensions or retirement savings accounts, the value of total family wealth and its components. We examine value of vehicles and value of IRAs. 12.2 percent ofworking-age mothers are contributing to a pension, and contribute an unconditional average of just under 1 percent of their income, while comparatively few retired mothers are currently contributing. However, the course of retirement savings is visible in our sample. Younger mothers, still contributing to their IRA have lower IRA value of $47,000, mothers around the time they retire have roughly $112,000 in their IRAs, and retired mothers, who have likely drawn down on their savings, have $73,000.

In summary, our sample straddles mothers of varying financial resources and flexibility, from younger mothers who are working and contributing to savings, older mothers who are on Social Security income and drawing down their savings, and the mothers who are transitioning between the two states.

We define an unemployment spell as a calendar year in which the child was unemployed for at least one week

We use the term spell for parsimony; it could in fact be multiple spells within the annual total.

Figure 2

The demographic characteristics of unemployed adult children matched to mothers in our sample, as well as characteristics of the spell, are summarized in Table 2 for all unemployed adult children, and grouped by the age of their mother. The average age of unemployed children is 32.2, with a nearly even split of men and women, though the child’s average age increases with the age of the mothers, from 28.8 years old to 44.9. Similar to mother’s age predicting the sources and scope of financial resources available to her, child’s age is indicative of the needs and constraints of her household. For example, although marriage rates are similar for each set of unemployed children, the share with children of their own at home drops from 67.6 percent to 49.3 percent. In addition, 50.2 percent of unemployed children of working-age mothers have more education than their mother, but this increases to 63.1 percent for the oldest mothers. Hence, the relative position of mothers to children is different among the threegroups.

We also examine events that are concurrent with the year of the child’s unemployment spell that could lead to financial assistance from parents. Self-reporting of disability joint with an unemployment spell is rare, with less than one percent of spells for children or working-age and retiring mothers, and 1.9 percent of spells for children of the oldest mothers. More common is the concurrent reporting of a new child in the unemployed household in the year of a spell, averaging 11.1 percent of spells for children of working-age mothers and 4.1 percent for children of oldest mothers. Also more common is for a child’s unemployment to coincide with that child’s divorce, averaging 10.1 percent of spells, and fairly consistent across the age groups. Due to the timing of the survey window, we can only identify that both events (unemployment and birth/divorce) occurred within the same calendar year and not which preceded the other, and it could be the case that the unemployment spell is prompted by the birth of a new child or divorce. Finally, 30.8 percent of unemployment spells occur within the year that a sibling is also unemployed; this share varies less by age of mom, from 30.9 percent, to 31.7 percent, to 28.5 percent.

To understand more about the nature of the child’s unemployment, we also examine the characteristics of the spell itself. First, we show the reason given for the spell. The PSID varies in who was administered the question eliciting the reason for unemployment; in certain years the head had to be unemployed at the time of the survey, in other years not. Hence, the categories include that the spell was associated with displacement, either layoff or firm closing; the spell was voluntary; or, the nature of the spell is not known. Of the known spells, a roughly even amount are associated with displacement (27.7 percent) and voluntary (26.1 percent), however, nearly half are not known (47.7 percent). Moreover, displacement rates are similar whatever the age of mom, but voluntary unemployment is less likely among older adults of older mothers, falling to 15.1 percent, and not known rising to 59.5 percent. Because such a large number of spells are not associated with a known reason, we do not feature these divisions in the main results, although displacement has a much stronger exogeneity argument. Instead, we discuss them in our subgroup analysis.

Recall that our definition of unemployment excludes spells that are associated with school exit. In the bottom of table 2, we show whether the individual was employed in full-year work (40 or more weeks employed) in the year prior to the spell. The majority of unemployed children were employed more than 40 weeks the year prior to the spell (59.2 percent), which rises with the age of the worker to 64.3 percent for the oldest group. Once unemployed, the majority experience a spell of less than 26 weeks, the maximum number of weeks that unemployment insurance benefits are available. The proportion with shorter unemployment spells falls with age, from 74.1 percent of spells for younger to 66.1 percent for older. We are also able to observe labor force participation in the year following the spell. For all age groups, labor force attachment is high; 75.8 percent remain in the labor force in the year following an spell of unemployment. Hence, the majority of unemployed children in our sample are full-time, full-year workers who experience a short spell and then return to work in the following year.In summary, unemployed children have similar demographic variation as their mothers do, but their spells and labor force attachment is comparatively much more uniform.

We first estimate equation (1) when Y_pt are measures of financial assistance from mother to child. Prior literature has established that parents send financial assistance to their children in certain instances, including unemployment (Cox and Way, 2011, Edwards, 2017); we want to confirm that finding within our sample before examining other outcomes. We use two measures of transfers in Y_pt, presented in Table 3: the first as a dummy variable for any transfer sending, and the second as the real dollar amount of the transfer. The estimated γ when transfers are a dummy variable in column 1 is 0.018 (.004), or a 1.8 percentage-point increase off of a mean rate of 7.0 percent, a 25 percent increase in the probability of sending a transfer. In column 2, we show similar estimates when Y_pt is the dollar amount of the transfer sent; the results are comparable to the first regression. There is a large and precisely estimated coefficient for the total population of mothers, at $77.844 (25.561). Off of a mean of $329, this is equivalent to a 22 percent increase in the amount of transfers sent to a child in the year of a child’s unemployment spell.

In Figure 3, we show the coefficient estimates for the parental transfer dependent variable when UC_ct is the time-path before and after unemployment, the event study from equation (2). The markers are the point estimate of the regression coefficient and the bars are the standard errors; three years, the omitted variable, is zero. The transfer amount dependent variable shows a spike in the year of the spell, but the years outside the spell are not positive. Hence, there is little evidence that parents pre-finance spells or may continue to assist even after the spell is completed

The exact coefficients (and errors) are: two years prior, 10.3 (24.1); one year prior, -5.8 (28.6); year of, 76.8 (25.0); one year following, 4.6 (27.4); two years following, 26.5 (25.0); three years following, -0.9 (28.8). The dummy transfer amount event study looks similar; those coefficient estimates can be found in Appendix Table 3.

Figure 3

To further understand the financial flows between parents and children, we also examine whether unemployed children effected the probability of mothers receiving a transfer, rather than sending one. There was no significantly estimated effect (neither decrease or increase), for either measure of transfers

Results in Appendix Table 1.

Our finding that parents are more likely to send a transfer to a child in the year they experience an unemployment spell corroborates similar results in previous studies. One shortcoming of the results presented here is that they are likely not exhaustiveof the total financial assistance sent from parents to children; the transfer themselves may be an underreporting of cash assistance, and by definition exclude any cohabiting or any in-kind transfers. Regardless, they do establish that parents financially assist unemployed children and that that assistance does not precede spells, providing support for exogeneity in the unemployment spell. We can turn to our research question, how this assistance is financed, and if we can observe any change in behavior.

Consumption is the most basic measure of welfare, and it is also an adjustment that all individuals, even those on a fixed income, can make in response to an income shock. Most accounting of consumption and household expenditures are not captured in the PSID and, in the years when more consumption measures are available, they are not measured frequently or consistently, save annual household spending on food. Although this it only measures part of total household consumption, food consumption has been used in prior studies to proxy for overall consumption, notably in the Gruber (1997) study of the consumption smoothing effects of unemployment insurance.

In the third and fourth columns of Table 3, we estimate equation (1) when the dependent variable Y_pt is the annual food consumption in 2014 dollars and in log. For both forms of the dependent variable, the main effect in negative and precisely estimated, at -$254.958 (49.384) in dollars, and -.041 (.008) in log. Given mean household annual food consumption of $7196, this is an estimated decline in consumption of 3.5 percent.

Figure 4 shows the event study coefficients for the log dependent variable. Consumption declines in the year of the spell; the decline persists for the two years following before regaining its former level

The exact coefficients (and errors) are: three years prior, 0.01 (0.011); two years prior, -0.016 (0.008); year prior omitted; year of, -0.039 (0.008); one year following, -0.020 (0.01); two years following, -0.028 (0.008); three years following, -0.006 (0.011). The dollar amount event studylooks similar; those coefficient estimates can be found in Appendix Table 3.

Figure 4

There are two measures of income behavior in the PSID that we can examine through the empirical framework in equation (1). Labor supply, as measured through the weeks worked per year and hours worked per week; and program participation in Social Security and SNAP. Asset and dividends, such as from the sale of a car, we consider dis-saving, rather than income, and discuss in that section.

The coefficient estimates from equation (1) when Y_pt is mothers’ total weeks employed in a year are presented in Table 3. As noted previously, an average of 49.2 percent of mothers are working in any given year. In the fifth column of Table 3, we test the individual reported weeks employed by the mother. The coefficient estimate of γ on child’s unemployment is a 0.381 increase in the number of weeks worked, on a base mean of 25.0 weeks, or 1.5 percent. This is a small though precisely estimated increase in labor supply in the year a child experiences unemployment. However, we do not find that there is an increase in the usual hours worked per week, in the subsequent column, which shows an imprecise point estimate of 0.058 (0.147).

We offer a few explanations of why weeks would adjust upward, but hours would not. First, they are measured differently. The survey asks for the number of weeks employed in the prior year, as well as the usual hours worked per week. The former is an accounting of time while the latter is not. Second, there may be less of a margin to adjust hours per week, either because it is not compensated or it is not feasible. A salaried employee, for example, may increase their earned income by taking less time off, but likely would see no earnings return to working longer hours. Workers in our sample are mostly at full-time hours. Conditional on working, 42 percent of mothers in our sample report working exactly forty hours per week and an addition 21 percent reportworking more than 40 hours. By comparison, conditional on working, half of mothers report working fewer than 48 weeks in the year prior

A histogram of usual hours per week for working mothers can be found in figure A7 of the Appendix. A histogram of weeks employed in the prior year can be found in A6.

In Figure 5, we present the coefficients from the event study in equation (2) for the total weeks employed of mothers to determine if the increase in labor supply precedes the unemployment spell or persists following. There is no increase in weeks employed in the years leading up to a child’s unemployment spell, a large increase in the year of the spell, and noisy estimates of the years following, though they weakly suggest that weeks worked remains high relative to the years prior to the child’s spell

The exact coefficients (and errors) are: three years prior, -0.045 (0.226); two years prior, -0.044 (0..172); year prior omitted; year of, 0.359 (0.161); one year following, 0.200 (0.201); two years following, 0.226 (0.166); three years following, -0.191 (0.219). The coefficient estimates for hours per week can be found in Appendix Table 3.

Figure 5

The second income behavior we examine is from program participation in Social Security or SNAP. The margins by which mothers and their households can alter their behavior in order to increase income via either program is very small, given the eligibility constraints and observed claiming behavior. Individuals may not claim Social Security retirement benefits until age 62, the Early Retirement Age (ERA), and most claim by age 70. Individuals of any age can apply for Social Security disability benefits, if they demonstrate they have previously worked but can no longer. In our sample, we measure household Social Security income, which would capture receipt for retirement or disability of any household member. The coefficient estimates from equation (1) when Y_pt is a dummy for any Social Security income are presented in the final column of Table 3. The result is positive, but imprecise, at 0.005 (0.004).

The other program, SNAP, has income-based eligibility requirements. To increase SNAP in the year of a child’s unemployment, a mother would have to have been eligible but not claiming household benefits until the year of the child’s spell, a very small margin. In addition, average SNAP receipt in our sample is low, at 7.0 percent. We do not find any significant results

Results presented in Appendix Table 1.

The last set of behaviors that we examine is savings. Savings variables were measured in each wave beginning in 1999, and take two forms. First, there is a measure of retirement savings contributions; the PSID asks what percent of income is put towards a pension. Second, there are measures of asset values of components of total wealth: home equity, debt, stock, farm or business income, other real estate, IRAs, and vehicles. For these regressions, we add to equation (1) a five-knot linear spline controlling for total household wealth.

To start, Table 4 shows the coefficient estimates when Y_pt is the unconditional percentage of income contributed to retirement savings. We find there is an imprecisely estimated decrease in savings rates of -0.068 (0.047), off a sample mean of 0.6 percent. In Figure 6, we present the coefficients from the event study in equation (2) for the percentage of income contributed to retirement savings. None of the coefficients are precise, though they show a pattern year-of-spell only decline

The exact coefficients (and errors) are: six year prior, 0.040 (0.057); four years prior, -0.024 (0.047); two years prior omitted; year of spell, -0.061 (0.046); two years following, 0.021 (0.081); four years following, 0.014 (0.056); six years following 0.002 (0.062).

Next, we examine asset values of mothers’ households in the year of child’s unemployment spell. Although the PSID has seven asset measures, we present results for two: vehicle value and IRA value. Very few mothers have any farm/business income or real estate outside of a primary home; the distribution of both are highly skewed. The results for stock value, debt, and home equity were imprecise and insignificant. They also may not logically be expected to respond to unemployment of a child. Stocks are reasonably liquid, but few mothers report any stock assets, and may be constrained in selling based on the value at the time relative to when purchased. Mothers may take on debt if they are not borrowing constrained and can afford the interest. And home equityis not a reasonably liquid asset, especially relative to a vehicle or an IRA

Results for stock, debt, home equity, as well as a dummy for any retirement saving, are presented in Appendix Table 2.

The coefficient estimates when Y_pt is vehicle value are presented in the second column of Table 4. We find an imprecisely estimated decline in vehicle value of -$324.56 (262.568). However, in Figure 7, we show the coefficient estimates from the event study of equation (2); the decline in vehicle value does not precede a child’s unemployment, but decreases, significantly estimated at -$567.436 (275.268), in the two years following, before returning to the baseline level

The exact coefficients (and errors) are: six years prior, 453.998 (316.475); four years prior, 145.156 (270.109); two years prior omitted; year of spell, -309.049 (277.631); two years following, -567.436 (275.268); four years following, -23.906 (289.482); six years following, 87.327 (291.876).

Figure 6

Figure 7

In the previous sections of results, we estimate the effect of having an unemployed child on various dependent variables for the full sample of mothers. In this section of results, we examine those previously estimated effects across key subgroups of unemployed children and their mothers. For both, we introduce the division in the treatment effect by substituting the single dummy U_ct of child’s unemployment with a vector UC_ct. For example, to investigate how the main effect varies by the reason of unemployment, we execute:

In the top half of Tables 5 and 6, we present estimates when UC_ct spans reason for the child’s unemployment, the labor supply before the spell, the duration of the spell, and the labor supply after the spell. We show this for each of the key dependent variables.