Crowd-out, Education, and Employer Contributions to Workplace Pensions: Evidence from Canadian Tax Records* Derek Messacar„

Abstract This study assesses whether workplace pensions help individuals overcome knowledge barriers to saving for retirement. Using administrative data from Canada and exploiting unique features of the pension system, I find compelling evidence that each $1 contributed to workplace pensions partially crowds out other retirement saving by approximately $0.50—among interior savers—in a regression kink design, centering on unionized

workers for methodological reasons. Further analysis indicates active versus passive decisions are influenced by education, exploiting compulsory schooling reforms for identification. I conclude by showing pension and education reform are both viable mechanisms for boosting retirement saving from a life-cycle perspective. JEL Codes: E21, H30, I21, J32 Keywords: Workplace pension, retirement saving, crowd-out, educational attainment, compulsory schooling, regression kink design, instrumental variables

*I

am grateful to Robert McMillan, Kory Kroft, and David Seim for their guidance on this study. I would also like to thank Winnie Chan, Marc Frenette, Adam Lavecchia, Pierre-Carl Michaud, Philip Oreopoulos, Yuri Ostrovsky, Laura Turner, Michael Veall, seminar participants from the Bank of Canada, Brock University, Department of Finance Canada, Employment and Social Development Canada, Federal Reserve Board, McGill University, Statistics Canada, Universit´e du Qu´ebec `a Montr´eal, University of Toronto, National Tax Association conference 2014, and Canadian Economics Association conference 2015, as well as Brigitte Madrian and two anonymous referees for helpful comments. I am indebted to Ren´e Morissette, Brian Murphy, Hung Pham, Paul Roberts, and Grant Schellenberg at Statistics Canada for their assistance. Financial support from the Canadian Labour Market and Skills Researcher Network and Royal Bank Graduate Fellowship in Public and Economic Policy is gratefully acknowledged. Any errors are my own. „ Research Analyst, Social Analysis and Modelling Division, Statistics Canada; and Adjunct Professor, Department of Economics, Memorial University of Newfoundland. The views and opinions expressed in this study are solely those of the author and do not necessarily reflect the views of Statistics Canada or the Government of Canada.

1

Introduction

Changes in the economic landscape over the past several decades have led to reductions in the generosities of public and employer-sponsored pensions in many OECD countries. Governments have increased retirement ages and strengthened work incentives in order to improve the sustainability of their pension systems amid pressures from increasing life expectancies, aging populations, poor investment performance, and greater economic uncertainty (OECD, 2012). Employers have also responded by moving from traditional, defined benefit pensions toward defined contribution plans to mitigate the costs and investment risks of providing such programs (KPMG, 2011; OECD, 2013). These trends imply the onus to save for retirement is increasingly being left to individuals. Alongside these changes, savings rates in some countries have declined sharply in recent years, raising concerns about the future retirement prospects of today’s workers (de Serres and Pelgrin, 2003). The standard approach of offering tax incentives on assets held in designated accounts is often regarded as an ineffective way to boost retirement wealth (Attanasio et al., 2004). Instead, many economists now advocate greater use of nonfinancial mechanisms for assisting or prescribing individuals to save. Pension features such as automatic enrollment, active decisions, simplification, and commitment devices have been found to raise saving in workplace accounts (Madrian and Shea, 2001; Choi et al., 2004; Thaler and Benartzi, 2004; Carroll et al., 2009; Madrian, 2012). Despite these innovations to the approaches used for changing pension plan outcomes, the effect of random variation in workplace saving on total wealth accumulation remains unresolved. While the illiquidity of pensions implies these plans raise the overall saving of individuals who face borrowing constraints (Hubbard, 1986), for others the effect is ambiguous. Workplace pensions may induce individuals to actively reduce non-pension wealth, as less is needed to hit a target level of consumption in retirement. As annuities, pensions may reduce saving by providing insurance against an uncertain lifespan (Hubbard, 1987).

1

Workplace pensions likely raise the overall saving of individuals who save for non-retirement purposes (Gale, 1998). If savers are passive (Chetty et al., 2014), pension saving passes through into greater total wealth accumulation for such reasons as mental accounting, rule-ofthumb saving, or bounded rationality (Thaler, 1990; Rubinstein, 1998; Benartzi and Thaler, 2001; Card and Ransom, 2011; Binswanger and Carman, 2012). Passive behavior may reflect a wider problem that individuals do not understand how to save adequately on their own. In this case, the life-cycle effect of shifting consumption from working years to retirement is unclear, since the costs of inducing some workers to over-save can be as great as the consumption losses to others in retirement from myopia (Whitehouse, 2013). The objective of this study is to investigate the interconnections between how individuals respond to changes in their workplace saving and how they save for retirement. The study makes three contributions. First, I assess the extent to which employer pension contributions increase total wealth accumulation or crowd out saving in other retirement accounts. This analysis adds to a large literature examining the effects of public and workplace pensions on saving outcomes (Bernheim, 2002). I provide credible new insight into this unresolved empirical issue by estimating the effect of random variation in saving arising from a unique institutional feature of pensions in Canada. Specifically, the identification exploits the fact that employer contribution rates increase discontinuously on earnings above the average industrial wage. The magnitude of the change in employer contributions and its displacement effect on other retirement saving are jointly estimated in a two-stage regression kink design (RKD). The results indicates that, among unionized workers who save in both types of plans but below their contribution limits, each $1 of workplace saving partially displaces other retirement saving by approximately $0.50, on average.1 The second contribution of this study is to provide new insight into the factors behind active versus passive choice. Specifically, I investigate the extent to which education affects such decisions. To this end, I obtain schooling measures from unique datasets that link tax records for nearly 800,000 individuals to their 1991 or 2006 census responses. As in the

2

returns-to-education literature, compulsory schooling reforms are used to obtain exogenous variation in education to identify this effect (Acemoglu and Angrist, 2000; Oreopoulos, 2006b). The results show that workers with high education respond to a change in workplace pension contributions by reducing other saving, while those with low schooling are passive. Hence, active versus passive choice in part results from human capital traits that are amenable to change through education policy. To the extent that education affects the knowledge-based costs of active decision-making as these results suggest, workers with low adjustment costs are indeed more responsive than workers with high adjustment costs. The third contribution is to consider the direct relationship between active versus passive choice and life-cycle saving outcomes. Since passive behavior is to some extent explained by low levels of schooling, workers with low education must also be found to save less than those with high education for any intervention that shifts consumption from working years to retirement to be desirable from a life-cycle perspective. I examine whether this condition likely holds in practice by estimating the effect of educational attainment on individuals’ savings rates in tax-preferred accounts. The results show that compelling individuals to complete high school raises savings rates by 3 to 6 percentage points annually, an effect that persists remarkably throughout normal working years. This finding is robust to controlling for a wide set of channels through which education indirectly affects saving, including family composition, employment, permanent income, home equity, and health. Taken together, the findings indicate that many individuals respond to changes in workplace saving, but that total wealth accumulation increases among those who likely benefit the most. However, education reform may be a substitute for interventions that directly target saving: more schooling reduces the need for, and effectiveness of, such policies as workers learn to save on their own, a finding that warrants consideration when designing socially optimal programs for encouraging individuals to save. This study contributes to three related literatures. The first addresses the longstanding issue of whether workplace pensions raise or redistribute total saving. Whereas several studies

3

find these plans create new saving (Poterba et al., 1994; Venti and Wise, 1996; Gelber, 2011; Chetty et al., 2014), others find they tend to crowd out contributions to other accounts (Engen and Gale, 1994; Veall, 2001; Benjamin, 2003). The lack of consensus likely stems from a lack of reliable data and credible research designs, which are problems that beset this topic of research (Bernheim, 2002). For example, measurement error of reported wealth in survey data can lead to over-estimates of the extent to which workplace pensions create new saving (Engelhardt and Kumar, 2011). Since workers sort into firms based on pension coverage, and firms offer benefits based on the demands of their workers (Ippolito, 1997), there is a need for quasi-experimental variation in workplace saving for credible identification. This study addresses these issues by using administrative data and exploiting a novel design feature of workplace pensions in Canada in the identification. Previous research on the effects of pensions on overall saving tend to focus on changes in pension wealth arising from wide-spanning events, including: changes in eligibility (Engen and Gale, 1994; Poterba et al., 1994; Venti and Wise, 1996; Gelber, 2011; Beshears et al., 2014); worker mobility (Chetty et al., 2014); and mandatory pension reform (Euwals, 2000; Arnberg and Barslund, 2013). In contrast, I analyze a nuanced change in workplace saving that only affects workers who are already pension members, irrespective of age or job tenure. In this sense, the second set of literature to which this study relates concerns how nonfinancial mechanisms and ‘nudges’ affect retirement saving (Madrian and Shea, 2001; Choi et al., 2004; Thaler and Benartzi, 2004; Carroll et al., 2009). Whereas rational agency predicts that individuals respond to changes in workplace saving by adjusting contributions to other plans, under-responses may occur due to behavioral factors including inertia or procrastination. Third, this study adds to large literatures on the returns to schooling and financial literacy (Angrist and Krueger, 1991; Bernheim and Garrett, 2009; Oreopoulos, 2007; Lusardi and Mitchell, 2014; Venti and Wise, 2014), as well as on the effects of limited rationality on saving outcomes (Choi et al., 2011; Beshears et al., 2013). The study proceeds as follows: the next section presents a simple conceptual model to

4

guide the empirical analysis; Section 3 reviews the institutional details of Canada’s retirement income system; Section 4 describes the data and sample selection; Sections 5, 6, and 7 present the results for crowd-out, effects of education on saving responses, and returns to education, respectively; and Section 8 concludes.

2

Conceptual Framework

This section presents a stylized model of life-cycle saving to help guide the empirical analysis. I extend models of active versus passive choice by Chetty et al. (2014) and Bernheim et al. (2015) by positing that the costs of active choice depend in part on human capital, providing a testable prediction of the factors behind such behavior. 2.1

Model Setup

There are I agents, i ∈ {1, 2, ..., I}, who live for two periods, t ∈ {1, 2}. In the first period, each agent supplies one unit of labor in exchange for earnings E and a pension contribution P . Each agent can then invest in taxable and tax-preferred saving plans, Si and Ri respectively, to purchase consumption {Ci,1 , Ci,2 } to maximize utility Ui (Ci,1 , Ci,2 ), subject to:

Ci,1 ≡ c1 (E, Si , Ri , P, r0 ) = E − (1 + φ)Si − (Ri + P )

(1)

Ci,2 ≡ c2 (E, Si , Ri , P, r0 ) = r0 (Si + Ri + P )

(2)

Si ≥ S¯

(3)

Ri ≥ 0

(4)

Ri + P ≤ L

(5)

Preferences satisfy the standard assumptions for risk aversion. For each i: ∂Ui (·)/∂Ci,t > 0; 2 ∂ 2 Ui (·)/∂Ci,t < 0; and ∂ 2 Ui (·)/∂Ci,t ∂Ci,l ≥ 0, for each t 6= l. Consistent with tax regulations,

L is a contribution limit. Hence, Ci,1 equals income minus savings and Ci,2 equals wealth accumulated from the first period. The term φ ∈ (0, 1) ensures that Ri and P yield a higher 5

rate of return than Si ; the minimum level of taxable saving, S¯ ∈ (0, E), is imposed as a money liquidity requirement on the basis that assets held in taxable accounts are generally more liquid than those held in pensions or retirement saving accounts.2 In this section, I assume r0 = 1 is known with certainty, although this is later relaxed. For ease of notation, denote utility as a function of the choice variables as Ui (Si , Ri ) and, similarly, the marginal utility of consumption as Ui,t (Si , Ri ) = ∂Ui (·)/Ci,t , for each t. 2.2

Optimal Behavior

Given that retirement saving strictly dominates taxable saving in this stylized model, agents prefer to save in the former over the latter in absence of the liquidity and contribution limit constraints. In what follows, I distinguish between three types of savers. ¯ L − P ) ≤ Ui,2 (S, ¯ L − P ). Shifting consumption First, an agent is a ‘limit’ saver if Ui,1 (S, from the first period to the second via Ri is (weakly) desired, but not possible due to the contribution limit, L; the optimal saving is Si? = arg max Ui (Si , L − P ) and Ri? = L − P , Si

where

Si?

¯ 0) ≥ Ui,2 (S, ¯ 0). The ¯ Second, an agent is financially ‘constrained’ if Ui,1 (S, ≥ S.

liquidity restriction leads to excess saving, and borrowing from Ri is (weakly) desired but not possible due to the non-negativity constraint; the optimal saving is Si? = S¯ and Ri? = 0. Limit and constrained savers are both at corner solutions in Ri . Third, an agent is an ‘interior’ saver if neither the contribution limit nor the non-negativity constraint binds; the optimal ¯ Ri ), where R? ∈ (0, L − P ). saving is Si? = S¯ and Ri? = arg max Ui (S, i Ri

˜ i }, may differ from the optimal choices The empirically-observed levels of saving, {S˜i , R for various reasons, discussed in Section 7. As such, programs for increasing total wealth accumulation may be desirable under certain conditions.

2.3

Comparative Statics

I consider how each type of agent is expected to respond to a (small) change in employer pension contributions, dP . For limit and interior savers, the model predicts dSi? = 0 and

6

dRi? = −dP . This result is a mechanical effect of the binding contribution limit constraint if the agent is a limit saver and dP > 0. Otherwise, this result arises from Ri and P being perfect substitutes under the tax code, whereas Si and P are not due to φ. For constrained savers, if dP > 0, first-period consumption falls and total wealth accumulation necessarily rises by this full amount because these agents are at a lower corner solution in saving and are unable to respond, dSi? = 0 and dRi? = 0. However, if dP < 0, first-period consumption may increase by this full amount or some additional saving may be made in the retirement account, dSi? = 0 and dRi? > 0, depending on preferences. An agent will under-responds to dP if the costs outweigh the associated benefits. The gain from adjusting saving is: Gi ≡ Gi (Si? , Ri? |dP ) = Ui (Si? − dSi? , Ri? − dRi? |dP ) − Ui (Si? , Ri? |dP ) + Vi (Ri? + P + dP − L). This reflects the difference in utility from re-optimizing saving relative to not responding, conditional on dP . The last term is an opportunity cost of the penalty from not adhering to the contribution limit, subject to: Vi (x) = 0, ∀x ≤ 0; and Vi (x) > 0 and dVi (x)/dx > 0, ∀x > 0. I posit that the adjustment cost depends on human capital hi , which ¯ The total cost is ki ≡ ki (hi , qi ; ζ) = ζf (hi ) + (1 − ζ)qi , I assume is dichotomous: hi ∈ {0, h}. where ζ ∈ [0, 1]. The human capital element reflects costs of attentiveness and learning, ¯ other costs are reflected by qi , where qi > 0 and qi ⊥ hi , for each i. An agent f (0) > f (h); crowds out dP if and only if Gi ≥ ki , which leads to the definition (Chetty et al., 2014) of active versus passive choice: Definition 1 Given dP , an agent is an ‘active’ saver if the taxable and tax-preferred saving chosen are {Si? − dSi? , Ri? − dRi? }, and is a ‘passive’ saver if {Si? , Ri? } remain chosen. This model gives two predictions that form the basis of the first two sections of the empirical analysis. First, the baseline prediction is dRi? = −dP for interior and limit savers, but the aggregate response may be weaker due to adjustment costs, which is testable if dP is well-identified. Limit savers are the most likely to exhibit perfect crowd-out because of the penalty on over-contributions. Second, crowd-out is expected to be dichotomous based on hi (unless ζ = 0), which is testable if exogenous variation in human capital exists. 7

3

Institutional Details

This section briefly reviews the institutional features of Canada’s retirement income system, focusing on the details most relevant to the empirical analysis. The system can be described as comprising three tiers (Baker et al., 2007).

3.1

Public Pensions

The first two tiers constitute Canada’s public pension system. First, there is a citizenshipbased pension available to most Canadians aged 65 or older who meet residency requirements called the Old Age Security (OAS), funded out of tax revenues. As of December 2010, the maximum monthly entitlement was $522. Benefits are linked to inflation and are taxable as income. For low-income recipients, a supplement can be available of $658 and $435 per person for single and married individuals, respectively, as of December 2010. The second tier comprises income-tested pensions called the Canada Pension Plan (CPP) and Qu´ebec Pension Plan (QPP). The CPP operates across Canada except in the province of Qu´ebec, where the QPP provides similar benefits. The CPP and QPP are designed to replace approximately 25 percent of workers’ mean lifetime earnings up to the average industrial wage. In 2010, the maximum monthly benefit payment was $934. Benefits are determined by a complex function of earning histories, lengths of time spent contributing, ages at which benefits start to be collected, and other factors such as time spent child-rearing (Baker and Benjamin, 1999). These programs are funded out of matching employer and employee payroll deductions, where both contribute a fraction of earnings (4.95 percent in 2010) between a basic exemption ($3,500 in 2010) and the year’s maximum pensionable earnings (YMPE), which is indexed to the average industrial wage ($47,200 in 2010). On earnings above this amount, the marginal contribution rates of both employers and employees fall to zero, a program feature that I exploit in the crowd-out analysis.

8

3.2

Private Pensions and Voluntary Retirement Accounts

Private pensions and retirement saving accounts constitute the third tier of this system. Using these plans is important for middle- and high-income households to prevent significant drops in living standards at retirement (Ostrovsky and Schellenberg, 2009). Workplace pensions, called Registered Pension Plans (RPP), have historically played an important role in this process. These plans can be defined benefit, defined contribution, or hybrid arrangements to provide income to retired employees in the form of periodic payments. Employers offering workplace pensions must contribute at least 1 percent of earnings annually, and employees also make positive contributions around 75 percent of the time. Contributions are non-taxable, investment income accumulates tax-free, and assets lock in after a short vesting period. Income from these plans are subject to income taxes at retirement. For the sample of taxfilers used in this study, Table 1 compares demographic, income, and saving characteristics by whether they have workplace pension and union coverage. Notably, workers with pensions earn and save more than those who do not have such coverage. A unique feature of workplace pensions in Canada that I exploit in the crowd-out analysis is that most plans are integrated with the contribution schedules of the second-tier public pensions. When these programs were introduced in 1966, they imposed additional labor costs on employers already sponsoring workplace plans. This led many employers to amend their workplace pensions in recognition of the additional costs and the fact that the public pensions somewhat duplicated existing coverage (Frenken, 1996). For example, “the pension per year of service may be 1.3% of earnings up to the YMPE [the average industrial wage] and 2.0% of earnings over the YMPE, with members being able to make contributions of 4.8% of their earnings up to the YMPE and 7.5% of earnings above it” (Statistics Canada, 2003, p.3). The same principle applies to defined benefit plans, where benefits accrue at lower rates on earnings up to the average industrial wage than on earnings above it (Baldwin, 2007). This feature has persisted over the last 50 years; in 1994, for example, 80 percent of workplace pension members had integrated plans (Statistics Canada, 1996, 2003). 9

Integration results in a typical workplace pension contribution schedule that kinks upward at the average industrial wage, which gives plausibly exogenous variation for identifying crowd-out since this feature is determined at the firm level. While employers’ total labor costs remain relatively constant around the average industrial wage, the change in overall saving from workers’ perspective is more discernible. This is because workers’ eventual benefit entitlements in retirement from the public pensions are not directly tied to contributions; in 1997, for example, reforms were enacted that gradually increased the combined employeremployee contribution rate to public pensions from 5.85 percent to 9.9 percent in 2003 and beyond, whereas benefit entitlements were reduced (Pesando, 2001). Marginal deviations in earnings around the average industrial wage have larger effects on workplace saving than on future income from public pensions; in the empirical analysis, I perform a placebo test of whether workers appear to respond in some way to the public pensions. Individuals may also save for retirement in voluntary, defined-contribution accounts called Registered Retirement Savings Plans (RRSP), which are set up and maintained through financial institutions. Contributions to these plans are tax-deductible, the income taxes being owed when funds are withdrawn. The maximum amount that may be contributed annually into these accounts is the lesser of 18 percent of income and a pre-set threshold ($22,000 in 2010), but any unused contribution room has carried forward indefinitely since 1991 so most taxfilers to not contribute close to their limits (Messacar, 2015). Individuals are permitted to over-contribute a cumulative lifetime amount of $2,000 to these accounts but excess contributions incur a penalty of 1 percent per month. For savers with workplace pensions or deferred profit-sharing plans, contribution room is reduced dollar-for-dollar by the amount saved in these other plans.

4

Data and Sample Selections

I use several related administrative tax databases from Statistics Canada in this study. These data provide rich, longitudinal information on Canadian taxfilers’ saving behavior in 10

workplace pensions and voluntary retirement accounts. The datasets and sample selections are described in this section. First, the Longitudinal Administrative Databank (LAD) is a panel file comprising a 20-percent representative sample of Canadian tax records, derived from the central incometax register. The LAD contains information about demographics, earnings, income, taxes, credits and allowances, transfers, and saving for the individuals represented and their census families. Although information on taxable saving or other wealth, such as home equity, are not available as this information is not reported on tax forms, the LAD is one of the most comprehensive databases for studying retirement saving in Canada. The following sample restrictions are imposed. I condition on the years 1991-2010, which coincides with the first year that data on workplace pension coverage is available and which runs up to the last year of data availability when this study began. I restrict to individuals between the ages of 25 and 49 in 1991 (44 and 68 in 2010). The upper age limit (68 in 2010) accounts for the fact that individuals are required to start collecting from their workplace pensions by the time they turn 69 years of age. I focus the analysis on individuals satisfying these criteria who are observed filing their taxes on time in at least 18 out of the 20 years. While relaxing this assumption does not change the main results, this restriction allows me to exploit the longitudinal component of the data in later stages of the analysis. Finally, I omit observations where individuals are observed collecting public or private pension income to center the analysis on pre-retirement saving behavior. Taken together, the sample comprises approximately 34 million observations on 1.8 million taxfilers. The LAD does not provide schooling information, since this is not collected by tax authorities. For this, I turn to newly-created datasets that link individuals’ 1991 or 2006 census of Canada responses to panels of their tax records (hereafter, the ‘Census-Tax Linkages’ file). This file is limited in that it does not contain the full breadth of information or sample size needed to carry out every crowd-out analysis that I perform with the LAD. However, replicating the baseline results is possible and, in addition, the census has information on

11

respondents’ reported highest level of education. I impose the same sample restrictions in the linkages file as before, except that I condition on individuals aged 25-68 at any time over the period 1991-2010 and do not specify a minimum number of repeat occurrences to increase the sample size for the RKD analysis. I also restrict to individuals who were born in a Canadian province in order to assign compulsory schooling laws, which vary by province and cohort, to individuals based on their province of birth as reported in the census. In all, the Census-Tax Linkages file comprises observations on nearly 800,000 individuals.

5

Crowd-out Analysis

This section estimates whether an exogenous change in employer pension contributions raises overall saving or induces workers to reduce contributions to voluntary retirement accounts. I first describe the empirical method and then present results and robustness checks.

5.1

Empirical Strategy

The identification exploits a change in employer pension contributions arising from the integration feature of these plans. The local average change in workplace saving at the kink (the average industrial wage), and its resulting displacement effect on voluntary retirement saving, are jointly estimated using a two-stage RKD:

Pit = α1 + β1 (Eit − Tt ) + γ1 (Eit − Tt )Dit + Xit0 δ + υit

(6)

Rit = α2 + β2 (Eit − Tt ) + γ2 (Eit − Tt )Dit + Xit0 θ + ξit

(7)

conditional on (Eit − Tt ) ∈ [−B, B], where B is the estimation bandwidth. The variables Eit , Pit , and Rit , are labor earnings, workplace pension contributions, and voluntary retirement savings for individual i at time t, respectively. Denote Xit as a vector of observable covariates and Dit = 1(Eit ≥ Tt ) as an indicator of whether earnings exceeds the kink point Tt in the reference year. Individuals are constrained, interior, or limit savers if Rit = 0, Rit +Pit ∈ (0, Lt ), 12

or Rit + Pit ≥ Lt , respectively. The empirical analysis focuses primarily on interior savers, although all three types are considered. Equation (6) estimates the ‘first-stage’ effect of workplace pensions being integrated with the public pensions, whereas equation (7) estimates the ‘second-stage’ effect on voluntary retirement saving, and crowd-out is γ =

γ2 . γ1

I estimate both equations jointly as seemingly

unrelated regressions (SUR) and obtain standard errors for crowd-out using the Delta method.3 The standard errors are clustered at the individual level to account for unit-specific correlations of the residuals, as recommended in Lee and Lemieux (2010). The estimator makes two identifying assumptions. The first is that the change in voluntary retirement saving comes only from the effect of integration on workplace pensions and not from a direct response by workers to the public pension discontinuity, an assumption that likely holds in practice for reasons discussed in Section 3.2. Second, the estimator provides a test of substitution between workplace and non-workplace retirement saving holding constant total compensation given that the running variable is employment income. However, the RKD assumes workers’ earnings are randomly assigned around the kink, which I assess using the density test of McCrary (2008). These results (in Figure A-1 of the Online Appendix) indicate sorting occurs in the full sample, but that the effect is driven by non-unionized workers, perhaps since collective bargaining makes it difficult to control earnings at the individual level. This leads me to condition this analysis on unionized workers (67.7 percent of all workplace pension members).4

5.2

Salience of the Treatment Effect

I consider two issues regarding the salience of workplace saving: how to interpret contributions to defined benefit plans; and how such information is disseminated to employees. The measure of annual contributions to workplace pensions available in tax records is called the ‘pension adjustment’ (PA). The PA reflects saving in traditional pensions, deferred profit-sharing plans (DPSPs), and some unregistered accounts.5 For defined contribution

13

plans, the PA is the value of total contributions in a year. For defined benefit plans, the PA translates the accrued pension compensation from the past year of service into a dollar equivalent, and is interpreted as the net present value of saving. Information about workplace saving is disseminated in two ways. First, workers receive a statement of remuneration from their employers each year during tax season stating their PA. Second, taxfilers receive a deduction limit statement each year from the central tax authority stating their PA and the amount they can contribute to voluntary retirement accounts without exceeding their limits. Thus, workers need not have a deep understanding of how their pensions operate (Luchak and Gunderson (2000) show this is often the case) to know roughly how much they saved in these plans in the past year.

5.3

Results

The graphical inspection and regression estimates of the primary RKD analysis, given by equations (6) and (7), are shown in Figure 1. This analysis centers on interior savers so that the results are not affected by those who are unable to respond or a mechanical effect of contribution limits.6 Note that only approximately 50 percent of workplace pension members also contribute to voluntary retirement accounts in a given year. These workers have higher total income ($54,750 versus $43,850), are less likely to be unionized (64.1 percent versus 71.9 percent), to collect unemployment insurance (8.1 percent versus 14.8 percent), and to make employee pension contributions (72.1 percent versus 79.1 percent) compared to those not saving in voluntary retirement accounts. The results for other savers are also considered to test the validity of the model’s predictions for these groups. The results indicate, first, that employer contributions increase by an average of $0.025 per $1 of employment income beyond the kink among interior savers, which represents a significant rise in saving of 36.2 percent from the base (pre-treatment) rate of $0.069 per $1. The magnitude of this effect is consistent with expectations: since employers’ contribution rate to the public pensions averaged $0.032 per $1 over the time period 1991-2010, and 80

14

percent of workplace pension members had integrated plans in 1994, the public pensions displace employer contributions to workplace pensions by $0.025 ÷ ($0.032 × 80%) = $0.977, nearly a dollar-for-dollar response. Second, no effect is observed for the employees’ share of contributions. While Frenken (1996) notes that employee contributions may be integrated, such an arrangement is not pronounced enough to be discerned empirically. The corresponding decrease in voluntary retirement saving at the kink is estimated to be $0.015 per $1 of employment income, a 33.3 percent decline relative to the base savings rate of $0.045 per $1 earned. As column (1) of of Table 2 shows, these RKD results jointly imply that

each $1 contributed to workplace pensions displaces other retirement saving by an average of $0.60, while the remaining $0.40 passes through into greater total wealth accumulation.7 These results control for a wide set of demographic, income, and other characteristics observed in the tax data, listed in the table’s notes. The p-values reported from the t-tests of whether perfect pass-through (γ = 0) and perfect crowd-out (γ = −1) are rejected empirically indicate this result is consistent with partial displacement. To further explore the robustness of this result, I augment the primary RKD analysis in two ways. First, I re-estimate the model controlling for individual and year fixed effects (FE), forcing the estimator to exploit variation in saving within individuals who move around the (non-stationary) kink over time, similar to ‘bracket creep’ (Saez, 2003).8 Second, to check that the response is driven entirely by the effect of interest, and is not confounded by the public pensions, the procedure is reapplied on workers who save in voluntary retirement accounts but do not have a workplace pension. As Figure 1 shows, this ‘placebo’ test does not detect any change in saving at the kink, consistent with expectations. I augment the statistical model to a difference-in-differences (DD) framework using savers from this placebo test as a control group. The second-stage estimating equation becomes:

Rit = α ˜ 2 + χIit + β˜2 (Eit − Tt )Iit + γ˜2 (Eit − Tt )Dit Iit + λ(Eit − Tt ) + ψ(Eit − Tt )Dit + Iit Xit0 θ˜ + Xit0 ω + ξ˜it (8) 15

where Iit = 1(Pit > 0) is an indicator of workplace pension coverage, and the sample is extended to include all unionized workers who meet the selection criteria. Hence, γ˜2 measures the change in the savings rate in voluntary retirement accounts net of any detectable response from the control group given by ψ. The RKD-FE and RKD-DD results, shown in columns (2) and (3) of Table 2, are very similar to the findings from the primary model specification, and suggest that each $1 contributed to workplace pensions crowds out other retirement saving by $0.464 and $0.643, respectively. The latter result is not significantly different from unity, although the point estimate is consistent with the results of the primary model and the estimator’s precision is weakened by the DD augmentation. In the Online Appendix, I also show that these findings are robust to standard RKD robustness checks, including: bandwidth sensitivity; polynomial order; optimal bandwidth selection (Calonico et al., 2014); and the permutation test (Ganong and J¨ager, 2016). The remaining columns of Table 2 extend the analysis to different types of savers to test predictions from Section 2 for these groups. For example, among limit savers, each $1 contributed to workplace pensions is estimated to crowd out other retirement saving by an average of $1.405; this effect is significantly different from zero but not from the prediction of perfect crowd-out from the model. Limit savers also have a higher pre-treatment savings rate of $0.059 per $1 earned in voluntary retirement accounts compared to interior savers, which is to be expected since they are saving at, not below, their limits. Among all savers, the estimated crowd-out, of $0.273 per $1, is approximately half of the response observed among unconstrained (interior and limit) savers, of $0.616 per $1. This result follows from the fact that only around half of workplace pension members also save in voluntary retirement accounts, as shown in Table 1. Among constrained savers, employer pension contributions pass through into greater total wealth accumulation notwithstanding any adjustments in taxable saving or other forms of wealth not observed in the data. Overall, a $1 increase in employer pension contributions crowds out other retirement saving by approximately $0.50 among interior savers, depending on the model specification.

16

This finding raises the question of why dollar-for-dollar substitution was not detected based on the model’s prediction. There are several reasons why workplace pensions and voluntary retirement accounts may be imperfect substitutes. Due to the lock-in provisions of workplace pensions, such assets are intended for retirement whereas the Canadian tax code does not penalize pre-retirement withdrawals from the voluntary accounts.9 Thus, assets held in voluntary retirement accounts may reflect saving for such reasons as income-smoothing, precautionary purposes (Mawani and Paquette, 2011), home equity, and retirement. I expect crowd-out to be larger for individuals who only use these plans for retirement. While administrative tax records do not reveal how savers intend to use voluntary retirement accounts, the longitudinal data show how they are actually used. Table 3 assesses how crowd-out varies based on withdrawal behavior. In particular, column (1) uses net saving (contributions less withdrawals in the reference year) in voluntary accounts as the dependent variable in the second-stage regression, and columns (2) and (3) condition on whether savers were ever observed making pre-retirement withdrawals. These results show that employer pension contributions partially displace other retirement saving in each case; non-retirement saving motives do not likely affect the substitutability of these plans in practice. Lastly, to consider whether adjustments occur along other margins, column (6) uses investment income as the second-stage dependent variable to proxy for taxable saving; consistent with the model’s expectations, no change at the kink is detected in this case.

6

Active versus Passive Choice and Education

In this section, I investigate the extent to which exogenous variation in educational attainment affects active versus passive choice, predicated on the hypothesis that the costs of re-optimizing saving depend at least in part on human capital acquisition. To motivate this analysis, Figure 2 shows that the primary crowd-out result of approximately $0.50 per $1 for interior savers can be decomposed based on workers’ propensities to save in voluntary retirement accounts. Exploiting the longitudinal data, Panel (a) shows that 17

workers who contributed frequently to voluntary retirement accounts in the past are more likely to crowd out workplace saving than non-frequent savers. In addition, Panel (b) shows that crowd-out approaches zero as the sample is conditioned progressively on workers with larger unused contribution room. As discussed in Section 3, unused contribution room carries forward such that, holding everything else constant, individuals with more room have saved less in these plans over their lifetimes. Therefore, workers who save more regularly appear to understand how to substitute adequately between the two saving plans, which could occur due to gradual learning from repeated interactions. In columns (2) and (3) of Table 4, crowd-out is decomposed by workers’ reported highest level of educational attainment. Specifically, individuals are sorted according to whether they have at least a high school diploma (‘high education’) or not (‘low education’).10 The results indicate, first, that the treatment effect is homogeneous across groups. Second, workers with low education under-respond compared to workers with comparatively high education, which suggests human capital partly determines active versus passive choice. To control for the possibility that saving adjustments are endogenous with education, columns (4) and (5) condition on a predicted measure of education that exploits compulsory schooling reforms. I estimate the incidence of completing high school, given by: 0 0 0 1(HSipc = 1) = $ + ϑMpc + X¯ ipc Λ + Z¯ipc Ω + Kpc Φ + εipc

(9)

where HSipc indicates high school completion and Mpc is the mandatory years of schooling required by province-of-birth p and cohort c.11 The regression (shown in column (3) of Table 5 and discussed in the next section) controls for covariates Xipc , and other factors Zipc from the Census-Tax Linkages file listed in the table’s notes. While education is only observed ¯ ipc from 1991 or 2006 census responses, the tax dataset is longitudinal. Hence, I construct X and Z¯ipc as inflation-adjusted averages of Xipc and Zipc for each individual across all observed years. To control for endogeneity between schooling reforms and other institutional factors

18

affecting education (Stephens Jr. and Yang, 2014), I include education policy covariates Kpc listed in the table’s notes, province-of-birth and cohort fixed effects, and province-of-birth specific cohort trends. Using the results of equation (9), I separate individuals into low- and high-education groups to keep sample sizes similar to actual attainment, which yields nearly an 85 percent success rate. Overall, the results using predicted education are similar to the baseline findings. Columns (6) and (7) condition on workers with high levels of schooling and, as expected, crowd-out is the largest for these groups. These findings indicate workers with high education are more active savers than those with low education, suggesting that active versus passive choice to some extent results from knowledge-based traits that are amenable to change through educational attainment. Finally, I assess whether underlying differences in observable characteristics drive these results. Figure 3 shows how earnings and saving differ by education. Since the distributions of voluntary retirement saving perfectly overlap across groups, the results are not likely due to low-education workers being unable to cut back their saving by as much because they save less. Second, the characteristics of workers in the RKD sample are comparable: low- and high-education workers are about the same age (45.0 versus 44.5, respectively), equally likely to be married (68.2 versus 66.5 percent), and have similar average incomes from non-employment ($2,750 versus $3,000). These similarities arise because the RKD conditions on workers localized around the average industrial wage, imposing a degree of homogeneity irrespective of education.

7

Saving Behavior and Education

This section assesses the life-cycle implications of shifting consumption from working years to retirement via workplace pensions. While previous studies subsume a role for redistribution, (Carroll et al., 2009; Chetty et al., 2014; Bernheim et al., 2015), these frameworks are extended here to endogenize this role. In particular, I consider the case in which saving adequacy is affected by human capital, motivated by the previous result that education affects active 19

versus passive choice and the growing literature that finds education has wide-spanning effects on: financial literacy; stock market participation and returns; diversification; and the use of financial planners (Calvet et al., 2009; Lusardi and Mitchell, 2010; Mullock and Turcotte, 2012; Lusardi et al., 2013). I then empirically assess the relationship between active versus passive choice and saving adequacy using the model to help guide this analysis.

7.1

Extensions of the Conceptual Framework

The conceptual framework is extended to account for two classes of models in which actual saving may not be optimal: (1) agents are present-biased; and (2) characterization failure occurs. I show that a role for raising saving exists only under certain conditions when these issues are determined by human capital, and that the welfare justification is different in each ˜ i (hi )}. case. Denote actual saving as a function of human capital as {S˜i (hi ), R 7.1.1

Present-biased Agents

Suppose agents are present-biased for such reasons as inertia, hyperbolic discounting or rational temptation (Laibson, 1997; Gul and Pesendorfer, 2001). I model present-bias in a reduced-form approach: Ui (Ci,1 , Ci,2 ) = ui (Ci,1 , Ci,2 ) + bi wi (Ci,1 , Ci,2 ), where ui (·) is forwardlooking utility and wi (·) is a present-biased component (Krusell et al., 2010). Both components satisfy the assumptions for risk aversion, and it is assumed ui,1 (·)/ui,2 (·) < wi,1 (·)/wi,2 (·) so that wi (·) tilts preferences toward the first period. The term bi affects the level of present-bias, ¯ = S ? and R ¯ = R? , ˜ i (h) and depends on human capital according to bi = 1(hi = 0). Thus, S˜i (h) i i since high-education agents are forward-looking. For low-education agents, S˜i (0) = S¯ and ˜ i (0) ≤ Ri? (with strict inequality if Ri? > 0), unless the contribution limit binds in which R ˜ i (0) = L − P , since myopia leads to under-saving. case S˜i (0) < S ? and R In this setting, there may be a role for dP > 0 to help uninformed agents if the social planner is paternalistic and over-value the forward-looking component of utility relative to the present-biased component compared to agents’ valuations. From a planner’s perspective,

20

agents are heterogeneous along two dimensions: (1) active versus passive; and (2) forward˜ i (hi )) as the forward-looking component of looking versus present-biased. Denote ui (S˜i (hi ), R utility as a function of the choice variables. Provided that a planner only values forwardlooking utility, the aggregate welfare effect of dP > 0 is:

A=

Xn

1(Gi < ki ) ui (S˜i (hi ), R˜ i (hi ) + dP ) − ui (S˜i (hi ), R˜ i (hi )) − 1(Gi ≥ ki )ki 



o

(10)

i

This effect is strictly negative for active agents, since adjustment costs are incurred and consumption is unaffected. For passive savers, the effect depends on how they save before treatment: (1) forward-looking agents are worse off since they save optimally on their own, ¯ R ¯ ¯ R ¯ which follows from S˜i (h) ¯ = S ? and R ¯ = R? being ˜ i (h) ˜ i (h)+dP ˜ i (h)), ui (S˜i (h), ) < ui (S˜i (h), i i the solution to the maximization problem in ui (·); (2) present-biased non-limit savers stand ˜ i (0) + dP ) ≥ ui (S˜i (0), R ˜ i (0)) (with strict inequality if Ri? > 0); and (3) to benefit, ui (S˜i (0), R present-biased limit savers are worse off, provided that the penalty on excess contributions exceeds the gain from redistribution. Thus, A > 0 only if there are sufficiently many non-limit savers who are both passive and present-biased.

7.1.2

Characterization Failure

Even if all agents are forward-looking, bi = 0 for each i, saving inadequacy may stem from an imperfect understanding of financial systems. This problem could occur due to bounded rationality or financial illiteracy, such as incomplete knowledge of how inflation, interest compounding, taxation, or diversification affect savings (Lusardi and Mitchell, 2010; Choi et al., 2011; Beshears et al., 2013). In the spirit of (Ambuehl et al., 2014), these issues can be modeled in a ‘complexly-framed’ setting where agents have different information about how investment instruments map into future consumption. ¯ but agents Suppose that r0 is known to both the social planner and agents with hi = h, with hi = 0 are uncertain about this mapping and have beliefs about r0 given by the random

21

¯ = S ? and R ¯ = R? . For ˜ i (h) variable r, subject to r > 0. Thus, for informed agents, S˜i (h) i i ˜ i (0)} ∈ arg max Er [Ui (Si , Ri )] subject to equations (1) to (5), uninformed agents, {S˜i (0), R {Si ,Ri }

except Ci,t ≡ ct (E, Si , Ri , P, r) for each t. The aggregate welfare effect of dP > 0 continues to be that of equation (10) in this case; however, uninformed agents may under- or over-save resulting from the competing substitution and income effects of the interest rate uncertainty (Sandmo, 1970). Thus, A > 0 only if there are sufficiently many non-limit savers who are passive, uninformed, and under-save on their own. In this setting, agents’ decisions are based on incomplete information about opportunities to save, which affects utility only through its effect on consumption, hence characterization failure occurs (Chetty et al., 2009; Ambuehl et al., 2014; Bernheim et al., 2015). Given the information asymmetry about r0 between some agents and the planner, there is a possible role for the redistribution of saving to correct this problem, the value of which can be assessed through the framework for welfare analysis of Bernheim and Rangel (2009) without imposing paternalistic judgment on agents’ preferences.

7.2

Empirical Analysis

The welfare justification for shifting consumption from working years to retirement depends on why saving inadequacy occurs. Based on the model’s implications and the previous finding that education affects active versus passive choice, a necessary empirical condition for dP > 0 ˜ i (hi )/dhi > 0. to target individuals who stand to benefit the most is dR To test this condition, I estimate the effect of compelling individuals to complete high school on their savings rates in tax-preferred accounts over the life-cycle in an instrumental variables (IV) framework that exploits the compulsory schooling reforms. The first-stage regression is given by equation (9), and the second-stage effect of high school completion on savings rates is: 0 0 0 ¯ ipc sr ¯ ipc = ι + η 1(HSipc = 1) + X Γ + Z¯ipc Π + Kpc Ψ + πipc

22

(11)

Using the average values of saving, income, and other variables from the longitudinal tax data helps to control for permanent-income effects that would otherwise be omitted if only the 1991 and 2006 cross-sectional data were used. Saving adequacy is assessed using exogenous variation in education on the basis that more schooling should not influence savings rates if individuals were already acting optimally, controlling for observed channels through which education indirectly affects such behavior. The first- and second-stage results are shown in Table 5. These findings indicate that high school completion induces individuals to raise savings rates in tax-preferred accounts by 3 to 6 percentage points. The inclusion of additional income and demographic controls does not alter the regression estimates, suggesting that income is the primary channel through which education indirectly affects saving and that other omitted variables are not a significant concern. In Figure 4, the effect of high school completion on the savings rates is illustrated. Those with high education save the most in tax-preferred accounts over the life-cycle up to the point of retirement when savings rates fall significantly for all groups. The finding that education raises individuals’ savings rates even at early ages suggests there are large cumulative returns to schooling on wealth in retirement. Overall, the data are consistent with a necessary condition for programs that shift consumption from working years to retirement via workplace pensions to be desired. In particular, passive savers likely stand to benefit the most and active savers re-optimize at low (human capital) cost. It is important to recognize, however, that many individuals with high education may over-save for retirement (Scholz et al., 2006). The extent to which saving inadequacy occurs in Canada is a contended issue. The first two tiers of the retirement income system, which include a demogrant, provide high replacement rates to current retirees who were in the bottom of the income distribution during working years, especially by international standards (Ostrovsky and Schellenberg, 2009, 2011; Whitehouse, 2009). However, depending on the dataset and empirical methodology used, and the assumptions over what constitutes an acceptable replacement rate, existing studies find that between 17 and 50 percent of

23

middle-income working Canadians will experience meaningful drops in living standards when they retire (LaRochelle Cot´e et al., 2008; Moore et al., 2010; Horner, 2011; Wolfson, 2011; Liu et al., 2013; Company, 2015). Additional saving in workplace pensions would, therefore, benefit at least a substantial minority of workers who do not prepare financially for retirement on their own, especially among constrained savers who tend to contribute very little to non-workplace pensions and are the focus of much of policy.

8

Conclusion

This study provided new insight into the extent to which employer pension contributions increase total wealth accumulation versus simply induce workers to reduce how much they contribute to other retirement saving accounts. Using administrative tax records from Canada and exploiting unique institutional features of the retirement income system, the results indicated that each $1 contributed by employers to workplace pensions partially displaces other retirement saving by approximately $0.50, among interior savers. This finding was consistent across a range of robustness checks and model extensions. Therefore, programs that shift consumption from working years to retirement via workplace pensions appear to be viable policy levers for boosting aggregate private saving, although many individuals are very responsive to such distortions. These results raise the question of why the saving responsiveness observed here is greater than in related literature. For example, several studies in public finance estimate that pensions increase overall saving (Poterba et al., 1994; Venti and Wise, 1996; Gelber, 2011; Chetty et al., 2014), but that the effect diminishes as broader measures of non-pension wealth are used (Gale, 1998). In contrast, this analysis detected crowd-out focusing only on wealth accumulation in similar types of retirement saving plans. Recent lessons from behavioral economics also suggest most workers remain passive to workplace saving distortions (Madrian and Shea, 2001; Choi et al., 2004). Those studies tend to analyze the effects of default enrollment or other nudges on the saving outcomes of new employees within firms. However, 24

firm switches are significant events from the workers’ perspective; in such cases, passive decisions may arise from: too much choice (Sethi-Iyengar et al., 2004; Choi et al., 2009); lack of knowledge about new pensions (Mitchell, 1988; Luchak and Gunderson, 2000; Gustman et al., 2009); or a view that the program is implicit financial advice. New hires are also younger and have less experience with pensions and saving than the typical worker. In contrast, this study estimated the effect of a change in workplace saving arising from a permanent feature of the retirement income system, with uniform effects across most workers irrespective of firm tenure. The way in which contribution limits to voluntary retirement accounts are linked to the amounts saved in workplace pensions may also increase the salience of the treatment effect. While the analysis is limited in that the results are local estimates for unionized workers near the average industrial wage for methodological reasons, this group is an interesting point of focus from a policy perspective.12 Further, in this study I showed that programs that shift consumption from working years to retirement likely help individuals save who stand to benefit the most without disadvantaging others who save adequately on their own. However, the welfare justification for such interventions depends on the reason why saving inadequacy occurs. Workers with low education save less than those with high education but remain passive to such a distortion, whereas workers with high education re-optimize saving across accounts at low cost. Thus, active versus passive choice to some extent depends on human capital. The finding that education affects saving suggests programs aimed at improving schooling outcomes or financial literacy may be imperfect substitutes for programs that directly target savings rates. The efficient mix of policies that assist or prescribe individuals to save, educate, and simplify choice (Choi et al., 2005, 2011; Beshears et al., 2013) remains an unresolved issue with implications for the future design of retirement income systems.

25

Notes 1

The crowd-out analysis is restricted to unionized workers (approximately two-thirds of

all workplace pension members) for methodological reasons, discussed below. 2

Other approaches to model the money liquidity requirement are to do so explicitly in

utility (Chetty et al., 2014) or in a three-period setting with short positions or lock-in pension provisions (Gale and Scholz, 1994; Milligan, 2002). As in Chetty et al. (2014), the approach used here can be viewed as a reduced-form version of a three-period model. 3

This is analogous to estimating the effect of Pit on Rit , conditional on (Eit − Tt ) and Xit ,

in a two-stage least squares approach using (Eit − Tt )Dit as the excluded instrument. 4

Restricting the analysis to unionized workers is necessary to ensure that an underlying

assumption of the RKD is satisfied, but introduces the possibility of sample selection bias. This issue must be recognized given that the statistical inferences are based on standard errors computed from the restricted sample. This issue is to some extent mitigated by clustering the standard errors at the individual level. 5

The PA slightly over-estimates pension coverage due to DPSPs (Morissette and Ostrovsky,

2006). However, changes in the PA are expected to be driven by the effect of interest. I exclude individuals with pension contributions or voluntary retirement saving above the 99th percentile, which tend to far exceed the contribution limits, to exclude individuals with very large DPSPs or erroneous saving information in the empirical analysis. 6

By restricting the analysis to interior savers, this potentially introduces bias from selection

of the sample based on the endogenous variable. The crowd-out results for other types of savers, including the unrestricted sample, are also presented below. 7

Total contributions is used in the first-stage regressions to remain consistent with sub-

sequent analyses using the Census-Tax Linkages dataset, which does not separately report employer and employee contributions. Point estimates of the changes in the employees’ share of contributions range from 0.000 to 0.003 and are statistically insignificant, hence the

26

first-stage effect operates through the employers’ share of contributions. 8

Individual-specific FEs are not necessary for credible identification in the regression

discontinuity design with panel data provided that treatment is random, but their inclusion should not meaningfully affect the results (Lee and Lemieux, 2010). Indeed, the finding of partial crowd-out continues to hold when the RKD-FE model specification is used. 9

For example, Individual Retirement Accounts (IRAs) in the United States impose a 10

percent penalty on withdrawals made before the age of 59.5. No equivalent penalty exists on withdrawals from voluntary retirement accounts in Canada, and assets held in these accounts may be used to finance a first home (Steele, 2007) or higher education under the Home Buyers’ Plan (HBP) and Lifelong Learning Plan (LLP), respectively. 10

The Census-Tax Linkages file does not provide data on respondents’ total number of

years of schooling, but indicates the highest level of attainment. Since this analysis exploits exogenous variation in education from compulsory schooling, which affects attainment at lower levels of schooling, the groups were separated by high school completion. 11

I match individuals to the entry age that was in effect in their province of birth at 6 years

old and to the exit age in effect at 14 years old. The Census-Tax Linkage sample consists of individuals who were 14 years of age from 1938 to 1995, so the reforms exploited herein span this time period. See Oreopoulos (2006a) for further information. 12

In Canada and other countries, public pensions provide adequate replacement incomes in

retirement to low-income groups (Ostrovsky and Schellenberg, 2009, 2011), whereas highincome earners have sufficient resources to save adequately. For those in between, public pensions do not provide high replacement rates but saving inadequacy may arise due to financial constraints or other factors.

27

References Acemoglu, D. and Angrist, J. (2000). How large are human-capital externalities? evidence from compulsory schooling laws. NBER Macroeconomics Annual, 15:9–59. Ambuehl, S., Bernheim, B. D., and Lusardi, A. (2014). Financial education, financial competence, and consumer welfare. Working Paper n. 20618, National Bureau of Economic Research. Angrist, J. D. and Krueger, A. B. (1991). Does compulsory school attendance affect schooling and earnings? Quarterly Journal of Economics, 106(4):979–1014. Arnberg, S. and Barslund, M. (2013). The crowding-out effect of mandatory labour market pension schemes on private savings: Evidence from renters in Denmark. Discussion Papers 11/2013-054, Network for Studies on Pensions, Aging and Retirement. Attanasio, O. P., Banks, J., and Wakefield, M. (2004). Effectiveness of tax incentives to boost (retirement) saving: Theoretical motivation and empirical evidence. OECD Economic Studies, 39(2):145–172. Baker, M. and Benjamin, D. (1999). Early retirement provisions and the labor force behavior of older men: Evidence from Canada. Journal of Labor Economics, 17(4):724–756. Baker, M., Gruber, J., and Milligan, K. (2007). Simulating the response to reform of Canada’s income security program. In Gruber, J. and Wise, D. A., editors, Social Security Programs and Retirement Around the World: Fiscal Implications of Reform, pages 83–118. University of Chicago Press. Baldwin, B. (2007). Registered pension plans in Canada – their evolution, structure and challenges. Report prepared for the Strategic Analysis, Audit and Evaluation Branch, Human Resources and Social Development Canada, Informetrica Limited.

28

Benartzi, S. and Thaler, R. H. (2001). Naive diversification strategies in defined contribution saving plans. American Economic Review, 91(1):79–98. Benjamin, D. J. (2003). Does 401(k) eligibility increase saving? evidence from propensity score subclassification. Journal of Public Economics, 87(56):1259–1290. Bernheim, B. D. (2002). Taxation and saving. In Auerbach, A. J. and Feldstein, M., editors, Handbook of Public Economics, volume 3, pages 1173–1249. Elsevier. Bernheim, B. D., Fradkin, A., and Popov, I. (2015). The welfare economics of default options in 401(k) plans. Unpublished Manuscript. Bernheim, B. D. and Garrett, D. M. (2009). The effects of financial education in the workplace: Evidence from a survey of households. Journal of Public Economics, 124(1):51–104. Bernheim, B. D. and Rangel, A. (2009). Beyond revealed preference: Choice-theoretic foundations for behavioral welfare econnomics. Quarterly Journal of Economics, 87(78):1487–1519. Beshears, J., Choi, J. J., Laibson, D., and Madrian, B. C. (2013). Simplification and saving. Journal of Economic Behavior & Organization, 95(C):130–145. Beshears, J., Choi, J. J., Laibson, D., and Madrian, B. C. (2014). Does front-loading taxation increase savings? evidence from Roth 401(k) introductions. Working Paper n. 20738, National Bureau of Economic Research. Binswanger, J. and Carman, K. G. (2012). How real people make long-term decisions: The case of retirement preparation. Journal of Economic Behavior & Organization, 81(1):39–60. Calonico, S., Cattaneo, M. D., and Titiunik, R. (2014). Robust data-driven inference in the regression-discontinuity design. The Stata Journal, 14(4):909–946. Calvet, L., Campbell, J., and Sodini, P. (2009). Measuring the financial sophistication of households. American Economic Review, 99(2):393–398. 29

Card, D. and Ransom, M. (2011). Pension plan characteristics and framing effects in employee savings behavior. Review of Economics and Statistics, 93(1):228–243. Carroll, G. D., Choi, J. J., Laibson, D., Madrian, B. C., and Metrick, A. (2009). Optimal defaults and active decisions. Quarterly Journal of Economics, 124(4):1639–1674. Chetty, R., Friedman, J. N., Leth-Petersen, S., Nielsen, T., and Olsen, T. (2014). Active vs. passive decisions and crowdout in retirement savings accounts: Evidence from Denmark. Quarterly Journal of Economics, 129(3):1141–1219. Chetty, R., Looney, A., and Kroft, K. (2009). Salience and taxation: Theory and evidence. American Economic Review, 99(4):1145–1177. Choi, J. J., Laibson, D., and Madrian, B. C. (2009). Reducing the complexity costs of 401(k) participation through quick enrollment. In Wise, D. A., editor, Developments in the Economics of Aging, pages 57–82. University of Chicago Press. Choi, J. J., Laibson, D., and Madrian, B. C. (2011). $100 bills on the sidewalk: Suboptimal investment in 401(k) plans. Review of Economics and Statistics, 93(3):748–763. Choi, J. J., Laibson, D., Madrian, B. C., and Metrick, A. (2004). For better or for worse: Default effects and 401(k) savings behavior. In Wise, D. A., editor, Perspectives on the Economics of Aging, pages 81–126. University of Chicago Press. Choi, J. J., Laibson, D., Madrian, B. C., and Metrick, A. (2005). Passive decisions and potent defaults. In Wise, D. A., editor, Analyses in the Economics of Aging, pages 59–78. University of Chicago Press. Company, M. . (2015). Building on canada’s strong retirement readiness. Technical report. de Serres, A. and Pelgrin, F. (2003). The decline in private saving rates in the 1990s in OECD countries: How much can be explained by non-wealth determinants? Economic Studies 36, Organisation for Economic Co-operation and Development (OECD). 30

Engelhardt, G. V. and Kumar, A. (2011). Pensions and household wealth accumulation. Journal of Human Resources, 46(1):203–236. Engen, E. M. and Gale, W. G. (1994). Do saving incentives work? Brookings Papers on Economic Activity, 1994(1):85–180. Euwals, R. (2000). Do mandatory pensions decrease household saving? evidence from the Netherlands. De Economist, 148(5):643–670. Frenken, H. (1996). The impact of changes in the Canada Pension Plan on private pensions. Canadian Business Economics, 4(4):65–72. Gale, W. G. (1998). The effects of pensions on household wealth: A reevaluation of theory and evidence. Journal of Political Economy, 106(4):706–723. Gale, W. G. and Scholz, J. K. (1994). IRAs and household saving. American Economic Review, 84(5):1233–1260. Ganong, P. and J¨ager, S. (2016). A permutation test for the regression kink design. Unpublished Manuscript. Gelber, A. M. (2011). How do 401(k)s affect saving? evidence from changes in 401(k) eligibility. American Economic Journal: Economic Policy, 3(4):103–122. Gul, F. and Pesendorfer, W. (2001). Temptation and self-control. Econometrica, 69(6):1403– 1435. Gustman, A. L., Steinmeier, T. L., and Tabatabai, N. (2009). Do workers know about their pension plan type? comparing workers’ and employers’ pension information. In Lusardi, A., editor, Overcoming the Saving Slump, pages 47–81. University of Chicago Press. Horner, K. (2011). A new pension plan for Canadians: Assessing the options. IRPP Study 18, Institute for Research on Public Policy. 31

Hubbard, R. G. (1986). Pension wealth and individual saving: Some new evidence. Journal of Money, Credit and Banking, 18(2):167–178. Hubbard, R. G. (1987). Uncertain lifetimes, pensions, and individual saving. In Bodie, Z., Shoven, J. B., and Wise, D. A., editors, Issues in Pension Economics, pages 175–210. University of Chicago Press. Ippolito, R. A. (1997). Pension Plans and Employee Performance: Evidence, analysis, and policy. University of Chicago Press. KPMG (2011). Defined contribution pensions: A global solution to pension funding challenges. Frontiers in Finance Supplement November, KPMG. Krusell, P., Kuru¸sc¸u, B., and Smith Jr., A. A. (2010). Temptation and taxation. Econometrica, 78(6):2063–2084. Laibson, D. (1997). Golden eggs and hyperbolic discounting. Quarterly Journal of Economics, 112(2):443–478. LaRochelle Cot´e, S., Myles, J., and Picot, G. (2008). Income security and stability during retirement in Canada. Analytical Studies Branch Research Paper Series, no. 306. Statistics Canada Catalogue no. 11F0019M, Ottawa: Statistics Canada. Lee, D. S. and Lemieux, T. (2010). Regression discontinuity designs in economics. Journal of Economic Literature, 48(2):281–355. Liu, H., Ostrovsky, Y., and Zhou, J. (2013). Saving and wealth: The adequacy of household saving in Canada. The Canadian Economy in Transition Research Paper Series, no. 029. Statistics Canada Catalogue no. 11-622-M, Ottawa: Statistics Canada. Luchak, A. A. and Gunderson, M. (2000). What do employees know about their pension plan? Industrial Relations: A Journal of Economy and Society, 39(4):646–670.

32

Lusardi, A., Michaud, P.-C., and Mitchell, O. S. (2013). Optimal financial knowledge and wealth inequality. Working Paper n. 18669, National Bureau of Economic Research. Lusardi, A. and Mitchell, O. S. (2010). How ordinary consumers make complex economic decisions: Financial literacy and retirement readiness. Working Paper 2010/11, Center for Financial Studies. Lusardi, A. and Mitchell, O. S. (2014). The economic importance of financial literacy: Theory and evidence. Journal of Economic Literature, 52(1):5–44. Madrian, B. C. (2012). Matching contributions and savings outcomes: A behavioral economics perspective. Working Paper n. 18220, National Bureau of Economic Research. Madrian, B. C. and Shea, D. F. (2001). The power of suggestion: Inertia in 401(k) participation and savings behavior. Quarterly Journal of Economics, 116(4):1149–1187. Mawani, A. and Paquette, S. (2011). Pre-retirement RRSP withdrawals. Canadian Tax Journal, 59(2):183–219. McCrary, J. (2008). Manipulation of the running variable in the regression discontinuity design: A density test. Journal of Econometrics, 142(2):698–714. Messacar, D. (2015). Do workplace pensions crowd out other retirement savings? evidence from Canadian tax records. Analytical Studies Branch Research Paper Series, no. 371. Statistics Canada Catalogue no. 11F0019M, Ottawa: Statistics Canada. Milligan, K. (2002). Tax-preferred savings accounts and marginal tax rates: Evidence from RRSP participation. Canadian Journal of Economics, 35(3):436–456. Mitchell, O. S. (1988). Worker knowledge of pension provisions. Journal of Labor Economics, 6(1):21–39.

33

Moore, K. D., Robson, W., and Laurin, A. (2010). Canada’s looming retirement challenge: Will future retirees be able to maintain their living standards upon retirement? Pension Paper n. 317, C.D. Howe Institute. Morissette, R. and Ostrovsky, Y. (2006). Pension coverage and retirement savings of Canadian families, 1986 to 2003. Analytical Studies Branch Research Paper Series, no. 286. Statistics Canada Catalogue no. 11F0019MIE, Ottawa: Statistics Canada. Mullock, K. and Turcotte, J. (2012). Financial literacy and retirement savings. Working Paper 2012-01, Department of Finance. OECD (2012). OECD pensions outlook 2012, Organisation for Economic Co-operation and Development (OECD). OECD (2013). Pension markets in focus, Organisation for Economic Co-operation and Development (OECD). Oreopoulos, P. (2006a). The compelling effects of compulsory schooling: Evidence from Canada. Canadian Journal of Economics, 39(1):22–52. Oreopoulos, P. (2006b). Estimating average and local average treatment effects of education when compulsory schooling laws really matter. American Economic Review, 96(1):152–175. Oreopoulos, P. (2007). Do dropouts drop out too soon? wealth, health and happiness from compulsory schooling. Journal of Public Economics, 91(11-12):2213–2229. Ostrovsky, Y. and Schellenberg, G. (2009). Pension coverage, retirement status, and earnings replacement rates among a cohort of Canadian seniors. Analytical Studies Branch Research Paper Series, no. 321. Statistics Canada Catalogue no. 11F0019M, Ottawa: Statistics Canada. Ostrovsky, Y. and Schellenberg, G. (2011). A note on pension coverage and earnings replacement rates of retired men: A closer look at distributions. Analytical Studies Branch 34

Research Paper Series, no. 326. Statistics Canada Catalogue no. 11F0019M, Ottawa: Statistics Canada. Pesando, J. (2001). The Canada Pension Plan: Looking back at the recent reforms. In Grady, P. and Sharpe, A., editors, The State of Economics in Canada: Festschrift in honour of David Slater, pages 137–150. Centre for the Study of Living Standards. Poterba, J. M., Venti, S. F., and Wise, D. A. (1994). 401(k) plans and tax-deferred saving. In Wise, D. A., editor, Studies in the Economics of Aging, pages 105–142. University of Chicago Press. Rubinstein, A. (1998). Modeling Bounded Rationality. MIT Press. Saez, E. (2003). The effect of marginal tax rates on income: A panel study of ‘bracket creep’. Journal of Public Economics, 87(5-6):1231–1258. Sandmo, A. (1970). The effect of uncertainty on saving decisions. The Review of Economic Studies, 37(3):353–360. Scholz, J. K., Seshadri, A., and Khitatrakun, S. (2006). Are Americans saving “optimally” for retirement? Journal of Political Economy, 114(4):607–643. Sethi-Iyengar, S., Huberman, G., and Jiang, G. (2004). How much choice is too much? contributions to 401(k) retirement plans. In Mitchell, O. S. and Utkus, S. P., editors, Pension Design and Structure: New lessons from behavioral finance, pages 157–184. Oxford University Press. Statistics Canada (1996). Canada’s Retirement Income Programs: A statistical overview, Statistics Canada Catalogue no. 75-507-XIE. Statistics Canada (2003). Canada’s Retirement Income Programs: A statistical overview (1990-2000), Statistics Canada Catalogue no. 75-507-XIE.

35

Steele, M. (2007). The Canadian Home Buyers’ Plan: Tax benefits, tax expenditure, and policy assessment. Canadian Tax Journal, 55(1):1–30. Stephens Jr., M. and Yang, D.-Y. (2014). Compulsory education and the benefits of schooling. American Economic Review, 104(6):1777–1792. Thaler, R. H. (1990). Anomalies: Saving, fungibility, and mental accounts. Journal of Economic Perspectives, 4(1):193–205. Thaler, R. H. and Benartzi, S. (2004). Save More TomorrowTM : Using behavioral economics to increase employee saving. Journal of Political Economy, 112(S1):S164–S187. Veall, M. R. (2001). Did tax flattening affect RRSP contributions? Canadian Journal of Economics, 34(1):120–131. Venti, S. F. and Wise, D. A. (1996). The wealth of cohorts: Retirement saving and the changing assets of older Americans. Working Paper n. 5609, National Bureau of Economic Research. Venti, S. F. and Wise, D. A. (2014). The long reach of education: Early retirement. Working Paper n. 20740, National Bureau of Economic Research. Whitehouse, E. (2009). Canada’s retirement-income provision: An international perspective. Technical report, Organisation for Economic Co-operation and Development (OECD). Whitehouse, E. (2013). Policies to encourage private pension savings: Evidence from OECD countries. In Hinz, R., Holzmann, R., Tuesta, D., and Takayama, N., editors, Matching Contributions for Pensions: A review of international experience, pages 27–50. International Bank for Reconstruction and Development/The World Bank. Wolfson, M. C. (2011). Projecting the adequacy of Canadians’ retirement incomes: Current prospects and possible reform options. IRPP Study 17, Institute for Research on Public Policy. 36

Table 1: Summary statistics, by workplace pension coverage (LAD) Has workplace pension No workplace pension All Unionized All Unionized (1) (2) (3) (4) Demographics Age 44.5 44.8 44.0 43.1 Female 48.6 48.0 56.0 50.0 Married 79.1 78.3 77.2 77.3 Employment Employed 99.7 99.9 63.3 94.2 Self-employed 2.9 3.0 15.0 9.0 Unionized 67.7 100.0 12.5 100.0 Has unemployment insurance 11.2 13.6 16.0 27.7 Conditional income Employment income, gross 47,300 46,500 26,850 28,900 Self-employment income, net 1,700 1,000 19,550 15,200 Other income 3,600 3,500 12,450 8,750 Total income 49,700 49,000 28,600 33,750 Savings participation Has workplace pension 100.0 100.0 0.0 0.0 Has employee pension contributions 75.3 81.3 0.8 3.7 Contributes to voluntary accounts 53.6 50.8 29.0 43.5 Withdraws from voluntary accounts 6.4 5.8 5.4 7.0 Has unused contribution room 87.6 87.5 82.5 91.3 Conditional savings Total workplace pension contributions 4,700 4,850 0 0 Employee pension contributions 2,350 2,450 1,400 1,550 Contributions to voluntary accounts 3,150 2,950 4,350 3,950 Withdrawals from voluntary accounts 2,000 1,900 2,200 2,150 Unused room in voluntary accounts 21,350 20,400 21,150 23,350 Conditional total savings Savings 6,250 6,250 3,500 3,350 Savings rate 15.9 12.7 8.4 7.3 Mean values are based on the LAD data for the years 1991-2010 inclusive. The ‘conditional’ statistics refer to the mean values conditional on those values being non-zero. The savings rate is calculated as net saving in tax-preferred accounts (both workplace pensions and voluntary retirement accounts) relative to total income. The income and saving values (rounded to the nearest $50) are expressed in nominal dollars.

37

Primary (1) Slope Kink

Slope Kink

38

Crowd-out

0.118*** (0.001) 0.026*** (0.002) 0.045*** (0.002) -0.015*** (0.003)

H0 : γ = 0 H0 : γ = −1

0.600 (0.124) [.000]*** [.001]***

Individuals Observations

225,270 735,682

Table 2: RKD baseline results (LAD) Interior Limit Unconstrained Fixed Difference-in effects differences Primary Primary (2) (3) (4) (5) PANEL A: Workplace pensions 0.077*** 0.118*** 0.121*** 0.119*** (0.001) (0.001) (0.002) (0.001) 0.017*** 0.026*** 0.022*** 0.025*** (0.002) (0.002) (0.004) (0.002) PANEL B: Voluntary retirement accounts 0.037*** 0.045*** 0.059*** 0.051*** (0.002) (0.004) (0.005) (0.002) -0.008*** -0.016*** -0.031*** -0.015*** (0.004) (0.008) (0.008) (0.003) PANEL C: Crowd-out 0.464 0.643 1.405 0.616 (0.175) (0.308) (0.416) (0.126) [.008]*** [.037]** [.001]*** [.000]*** [.002]*** [.247] [.330] [.002]*** 225,270 735,682

283,289 949,804

89,260 185,863

255,342 921,545

Non-limit

All

Primary (6)

Primary (7)

0.119*** (0.001) 0.026*** (0.002)

0.120** (0.001) 0.024*** (0.001)

0.041*** (0.001) -0.008*** (0.002)

0.049*** (0.001) -0.007*** (0.002)

0.297 (0.073) [.000]*** [.000]***

0.273 (0.080) [.001]*** [.000]***

367,901 1,660,048

403,129 1,945,197

‘All’ includes constrained, interior, and limit savers; ‘Non-limit’ includes constrained and interior savers; and ‘Unconstrained’ includes interior and limit savers. The ‘Primary’ model is given by equations (6) and (7). The bandwidth used is B = 6, 000. Results are for the years 1991-2010 inclusive. The slope estimates refer to the slope of the savings function before the kink. The difference-in-differences slope estimates and kink estimates correspond to the effects for workplace pension members net of the effects for those not covered by a workplace pension. The following covariates are included in every regression: female, married, province of residence, has unemployment insurance, has self-employment income, age, age-squared, disability and medical expense tax allowances, and other income (total income before taxes less employment income). Standard errors (in parentheses) are clustered by individual. The p-values are reported separately for the t-tests of the null hypothesis (H0 ) that there is perfect pass-through (γ = 0) or perfect crowd-out (γ = −1) of the change in workplace pension contributions. *** and ** denote significance at the 1% and 5% levels, respectively.

Table 3: RKD robustness checks for interior savers (LAD) Pre-retirement withdrawal? Investment Net saving No Yes income (1) (2) (3) (4) PANEL A: Workplace pensions Slope 0.118*** 0.120*** 0.115*** 0.105*** (0.001) (0.002) (0.002) (0.003) Kink 0.025*** 0.027*** 0.024*** 0.045*** (0.002) (0.003) (0.003) (0.006) PANEL B: Voluntary retirement accounts or investment income Slope 0.048*** 0.045*** 0.044*** -0.001 (0.002) (0.002) (0.003) (0.002) Kink -0.014*** -0.017*** -0.013*** 0.000 (0.003) (0.004) (0.004) (0.004) PANEL C: Crowd-out Crowd-out 0.550 0.645 0.545 -0.004 (0.139) (0.162) (0.193) (0.081) H0 : γ = 0 [.000]*** [.000]*** [.005]*** [.964] H0 : γ = −1 [.001]*** [.029]** [.018]** [.000]*** Individuals Observations

239,084 792,901

111,799 391,358

113,471 344,324

202,357 593,721

The dependent variable in the second-stage regression (in Panel B) of column (1) is net saving (contributions less withdrawals in the reference year) in voluntary retirement accounts. Columns (2) and (3) exploit the longitudinal nature of the LAD data to condition the analysis on individuals who are not, or who are, ever observed withdrawing from their voluntary retirement accounts before retirement. Column (4) uses investment income as the dependent variable in the second-stage regression; in this case, the sample is restricted to ‘interior’ savers, but defined as individuals who have both a workplace pension and strictly positive investment income. See Table 2 for more information about the regression specifications. *** and ** denote significance at the 1% and 5% levels, respectively.

39

Slope Kink

Slope Kink

40

Crowd-out H0 : γ = 0 H0 : γ = −1 Individuals Observations

Table 4: RKD results for interior savers, by education (Census-Tax Linkages) Higher levels of Unconditional Actual education Predicted education actual education Some post- University Full sample Low High Low High secondary graduate (1) (2) (3) (4) (5) (6) (7) PANEL A: Workplace pensions 0.120*** 0.101*** 0.120*** 0.099*** 0.122*** 0.118*** 0.105*** (0.002) (0.006) (0.002) (0.006) (0.002) (0.003) (0.004) 0.034*** 0.047*** 0.033*** 0.023** 0.034*** 0.024*** 0.049*** (0.003) (0.010) (0.003) (0.011) (0.003) (0.005) (0.006) PANEL B: Voluntary retirement accounts 0.044*** 0.052*** 0.042*** 0.041*** 0.045*** 0.038*** 0.041*** (0.002) (0.008) (0.003) (0.007) (0.002) (0.004) (0.005) -0.017*** -0.009 -0.018*** 0.001 -0.019*** -0.017** -0.032*** (0.004) (0.013) (0.004) (0.013) (0.004) (0.007) (0.008) PANEL C: Crowd-out 0.510 0.184 0.555 -0.033 0.569 0.684 0.651 (0.123) (0.273) (0.135) (0.566) (0.131) (0.284) (0.174) [.000]*** [.499] [.000]*** [.953] [.000]*** [.016]** [.000]*** [.001]*** [.003]*** [.001]*** [.068]* [.001]*** [.267] [.045]** 58,992 217,714

6,092 23,819

52,943 193,895

6,765 24,113

52,297 193,601

20,862 82,592

15,838 45,138

Column (1) replicates the RKD analysis from the LAD on the full sample from the Census-Tax Linkages data. Columns (2) and (3) separate individuals based on actual level of education, whereas columns (4) and (5) separate based on predicted attainment from the first-stage instrumental variables regression from column (3) of Table 5. Columns (6) and (7) condition the analysis on individuals with some post-secondary education and a bachelor’s degree or higher, respectively. The same set of covariates as those used in the LAD data are included here, except for the following changes due to data availability: has unemployment insurance, has self-employment, and medical expense allowances are not available, but home equity values reported on the 1991 or 2006 census and sector of employment indicators (1-digit NAICS code) are included. The bandwidth is B = 9, 000 to give more power to the estimator given the small sample size compared to the LAD. See Table 2 for more information about the regression specifications. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

Table 5: Returns to education on life-cycle savings rates (Census-Tax Linkages) First stage Second stage Dep. var.: Completed high school Dep. var.: Savings rate (1) (2) (3) (4) (5) (6) (7) (8) Completed high school 0.060* 0.062* 0.061** 0.032** (0.034) (0.033) (0.030) (0.016) Entry age -0.009*** (0.002) Dropout age 0.005*** (0.002) Mandatory years of schooling 0.007*** 0.006*** 0.006*** (0.001) (0.002) (0.002)

41

Additional controls Total income polynomial Earnings and job characteristics Education covariates Birth province and cohort effects Birth province × cohort trends Linear Quadratic Observations

X X

X X

X X X X X

X

X

X

492,765

492,765

374,625

X X X X X

X 374,625

X X X

X X X X

X X X X X

X

X

X

370,598

368,570

368,570

X X X X X

X 368,570

The number of mandatory years of schooling is the difference between the legal age at which students may exit high school and the legal age at which they must begin elementary school. The additional control variables include: indicators of employment status, union coverage, and sector of employment (1-digit NAICS code); a 7th-order polynomial in total income; and the value of home equity as reported in the 1991 or 2006 census. The education covariates Kpc are: child labor age; an indicator of whether a restrictive labor law was in place; an indicator of whether exemptions existed; the log of school expenditure; and the number of schools and teachers per capita (see Oreopoulos (2006a) for more details). Province-of-birth polynomial cohort trends are used to control for unobserved factors across provinces that may have changed over time that affected educational attainment. The data permit me to exploit reforms to compulsory schooling laws that were enacted between 1938 and 1995. See Table 2 for more information about the regression specifications. ***, **, and * denote significance at the 1%, 5%, and 10% levels, respectively.

Figure 1: RKD primary results for interior savers (LAD) (a) Employer contributions

(b) Employee contributions

Contributions

Contributions

2,600

1,900

2,300

1,700

2,000

1,500

1,700

1,300

1,400 -6,000

-3,000

0

3,000

6,000

1,100 -6,000

-3,000

0

3,000

6,000

Employment income relative to the kink

Employment income relative to the kink

Slope = 0.069*** (0.001) Kink = 0.025*** (0.002)

Slope = 0.049*** (0.001) Kink = 0.001 (0.002)

(c) Voluntary saving, has workplace pension

(d) Voluntary saving, no workplace pension

Contributions

Contributions

2,600

4,000

2,450

3,600

2,300

3,200

2,150

2,800

2,000 -6,000

-3,000

0

3,000

6,000

2,400 -6,000

-3,000

0

3,000

Employment income relative to the kink

Employment income relative to the kink

Slope = 0.045*** (0.002) Kink = -0.015*** (0.003)

Slope = 0.101*** (0.003) Kink = 0.001 (0.006)

6,000

Each graph shows contributions as a function of employment income relative to the kink (the average industrial wage). Each data point represents the average contribution within a bin of width $400. The corresponding linear regression results are also reported. See Table 2 for more information about the regression specifications. *** denotes significance at the 1% level.

42

Figure 2: RKD results for interior savers, by contribution history and unused room in voluntary retirement accounts (LAD) (a) Observed history of saving in voluntary retirement accounts Crowd-out

(b) Unused contribution room in voluntary retirement accounts Kink in workplace pensions

0.8

0.6

0.4

0.2

0.04

1.6

0.03

1.2

0.02

0.8

0.01

0.4

0.00

0.0

0.0 No

Yes

Saved last year?

No

Crowd-out

Unused contribution room

Yes

Kink

Ever saved before?

Crowd-out

Graph (a) shows how crowd-out varies according to whether individuals were observed saving in a voluntary retirement account last year (left) or ever in the past (right). Graph (b) shows the effect of integration on workplace pension contributions (left vertical axis) and crowd-out (right vertical axis) conditional on workers with progressively larger unused contribution room in voluntary retirement accounts. Holding everything else constant, individuals with larger unused contribution room have saved less over their lifetimes in these accounts. See the text for more information about this institutional characteristic. See Table 2 for more information about the regression specifications.

43

Figure 3: Distribution of employment income and tax-preferred saving among workers in the RKD sample, by educational attainment (Census-Tax Linkages) (a) Employment income

(b) Workplace pension contributions

Percent

Percent

2.5

6.0

2.0

5.0 4.0

1.5 3.0 1.0

2.0 0.5

1.0

0.0 -9,000

0.0 -4,500

0

4,500

0

9,000

5,000

7,500

10,000

Employment income relative to the kink

Employment income relative to the kink Low

2,500

High

Low

High

(c) Voluntary retirement saving Percent 12.0 10.0 8.0 6.0

4.0 2.0 0.0 0

2,500

5,000

7,500

10,000

Employment income relative to the kink Low

High

These graphs plot the distributions of employment income, workplace pension contributions, and voluntary retirement saving by education conditional on workers who appear in the RKD analysis. Employment income is expressed relative to the kink (the average industrial wage). Education levels are based on actual reported highest level of schooling attained.

44

Figure 4: Effect of education on life-cycle savings rates (Census-Tax Linkages) (a) Low versus high actual education

(b) Low versus high predicted education

Savings rate

Savings rate

14

14

10

10

6

6

2

2

-2

-2 20

30

40

50

60

70

20

30

40

Age Low

50

60

70

Age High

Low

High

(c) Various levels of actual education Savings rate 14 10 6 2 -2

20

30

40

50

60

70

Age

High school dropout

High school or trades

Some college

University graduate

These graphs plot the average savings rate by age, for individuals aged 25-67. The savings rate is calculated as net saving in tax-preferred accounts (both workplace pensions and voluntary retirement accounts) relative to total income. Graph (a) separates individuals based on actual level of education, whereas Graph (b) separates based on predicted attainment from the first-stage instrumental variables regression from column (3) of Table 5. Graph (c) conditions the analysis on individuals in four categories of educational attainment: did not complete high school (‘dropout’); has a high school diploma or trades certificate; completed some college but did not graduate; and has a university bachelor’s degree or higher. Some recipients of trades certificates may not have completed high school; separating these individuals based on whether they graduated from high school is not possible given that the census questionnaire ranks highest level of educational attainment using a market-based ranking. High school graduates and trades certificate recipients were grouped for illustrative purposes given that their saving profiles nearly perfectly overlap.

45

Online Appendix Table A-1: RKD tests of bandwidth and polynomial order for interior savers (LAD)

Slope Kink

Slope Kink

Crowd-out H0 : γ = 0 H0 : γ = −1 Individuals Observations

Linear Cubic B = 5,000 B = 7,000 B = 9,000 B = 9,000 (1) (2) (3) (4) PANEL A: Workplace pensions 0.116*** 0.117*** 0.117*** 0.119*** (0.002) (0.001) (0.001) (0.002) 0.025*** 0.028*** 0.028*** 0.021*** (0.003) (0.002) (0.001) (0.005) PANEL B: Voluntary retirement saving 0.044*** 0.044*** 0.043*** 0.044*** (0.002) (0.001) (0.001) (0.003) -0.012*** -0.014*** -0.014*** -0.012* (0.004) (0.002) (0.002) (0.006) PANEL C: Crowd-out 0.456 0.493 0.494 0.559 (0.156) (0.093) (0.066) (0.323) [.003]*** [.000]*** [.000]*** [.084]* [.001]*** [.000]*** [.000]*** [.172] 202,516 614,553

246,303 855,249

283,918 1,083,783

283,918 1,083,783

The Akaike Information Criterion (AIC) and Baysian Information Criterion (BIC) tend to indicate that the linear model is preferred, although the point estimates do not meaningfully change when a cubic polynomial in the running variable is used. Specifically, the AIC and BIC statistics from columns (3) and (4) are as follows. Panel A, Linear: AIC = 17.204; BIC = 3,585,320.661. Panel A, Cubic: AIC = 17.204; BIC = 3,585,330.611. The difference in the BIC of 9.950 provides very strong support for the linear model. Panel B, Linear: AIC = 17.957; BIC = 4,401,234.755. Panel B, Cubic: AIC = 17.957; BIC = 4,401,256.574. The difference in the BIC of 21.819 provides very strong support for the linear model. See Table 2 for more information about the regression specifications. *** and * denote significance at the 1% and 10% levels, respectively.

46

Figure A-1: Distribution of employment income relative to the kink, by type of worker (LAD) (a) Full sample

(b) Has workplace pension

Percent

Percent

2.5

2.5

2.0

2.0

1.5

1.5

1.0

1.0

0.5

0.5

0.0 -6,000

-3,000

0

3,000

6,000

0.0 -6,000

-3,000

0

3,000

Employment income relative to the kink

Employment income relative to the kink

McCrary estimate = 0.050*** (0.010)

McCrary estimate = 0.013 (0.013)

(c) Unionized, no workplace pension

(d) Non-unionized, no workplace pension

Percent

Percent

2.5

2.5

2.0

2.0

1.5

1.5

1.0

1.0

0.5

0.5

0.0 -6,000

-3,000

0

3,000

6,000

6,000

0.0 -6,000

-3,000

0

3,000

6,000

Employment income relative to the kink

Employment income relative to the kink

McCrary estimate = 0.028 (0.034)

McCrary estimate = 0.118*** (0.017)

Each graph shows the distribution of employment income relative to the kink (the average industrial wage), in bins of width $200. Results are for the years 1991-2010 inclusive. The McCrary discontinuity test estimates the extent to which a bunching response at the kink point is statistically significant using the optimal kernel density bandwidth. Standard errors are in parentheses. *** denotes significance at the 1% level.

47

Permutation Test In this note, I describe the results of the permutation test of Ganong and J¨ager (2016), which complements the standard RKD inference. This test constructs a distribution of placebo estimates in regions with and without the kink in the running variable, and uses this distribution to gauge statistical significance. This test requires that a range of ‘plausible’ values for the kink is determined, from which the placebo kinks are randomly drawn. To determine this range, it is helpful to notice that the actual kink in this study corresponds to the average industrial wage on the basis that Canada’s means-tested public pensions are designed to provide adequate replacement income in retirement up to this amount (see Section 3 for details). Similar thresholds that could have been chosen include the median industrial wage or the average national wage. A recent reform of Canada’s public pension system provides additional insight: it is currently expected that the actual kink will increase significantly to $82,700 by 2025. I conduct three separate permutation tests using different regions from which the placebo kinks are drawn, each from a uniform probability distribution: (1) 95-135 percent of the actual kink (‘narrow region’); (2) 90-150 percent of the actual kink (‘moderate region’); and 50-200 percent of the actual kink (‘wide region’). Given the economic interpretation of the kink point and recent reforms, these regions imply it is relatively unlikely the kink point would have been set much lower than its actual value, although the wide region allows for this possiblity. In each case, I draw 1,000 placebo kinks from the relevant region. The two-sided p-values from these purmutation tests are: .076 using the narrow region; .054 using the moderate region; and .038 using the wide region. These results support the RKD methodology. While the test’s two-sided p-value is the largest using a narrow region, this is not surprising since the placebo kinks are only being drawn from values that closely resemble the actual kink in this case. The p-value decreases as less-realistic placebo kinks are drawn.

48

Optimal Bandwidth In this note, I describe my implementation of the bandwidth selection procedure of Calonico et al. (2014) applied to the RKD results for interior savers. Specifically, the Stata program ‘rdbwselect.ado’ is used here. This analysis is implemented using Census-Tax Linkages data, the smaller of the two datasets used in this study. First, I restrict the sample to workers with employment income within a large range of the kink. Second, from this restricted sample, I conduct the bandwidth selection procedure. Given the large sizes of the administrative datasets used in this study, it was not possible to implement this procedure using the LAD or without restricting the initial sample from the Census-Tax Linkages data in some way. The kink point (the average industrial wage) was $30,500 in 1991 and $47,200 in 2010, hence this procedure is carried out separately for workers with employment income: (1) from $0 to $61,000 (‘input sample A’); and (2) from $0 to $94,400 (‘input sample B’). The results are as follows. Based on input sample A, the optimal bandwidths are $7,988 and $5,389 for the regressions on workplace pension contributions and voluntary retirement saving, respectively. Based on input sample B, the optimal bandwidths are $9,790 and $11,314 for the regressions on workplace pension contributions and voluntary retirement

saving, respectively. The actual bandwidth used of $9,000 with the Census-Tax Linkages data is a good approximation of the suggested bandwidths from this procedure. As Table A-1 shows, the results do not vary meaningfully as the bandwidth changes.

49