Measurement error in income and schooling, and the bias for linear estimators, with Paul Bingley, Journal of Labor Economics 35, no. 4 (2017): 1117-1148.

We propose a general framework for determining the extent of measurement error bias in OLS and IV estimators of linear models, while allowing for measurement error in the validation source. We apply this method by validating Survey of Health, Ageing and Retirement in Europe (SHARE) data with Danish administrative registers. Contrary to most validation studies, we find measurement error in income is classical, once we account for imperfect validation data. We find non-classical measurement error in schooling, causing a 38 percent amplification bias in IV estimators of the returns, with important implications for the program evaluation literature.

Mental retirement and schooling, with Paul Bingley, European Economic Review, 63:292-298, 2013

We assess the validity of differences in eligibility ages for early and old age pension benefits as instruments for estimating the effect of retirement on cognitive functioning. Because differences in eligibility ages across country and gender are correlated with differences in years of schooling, which affect cognitive functioning at old ages, they are invalid as instruments without controlling for schooling. We show by means of simulation and a replication study that unless the model incorporates schooling, the estimated effect of retirement is negatively biased. This explains a large part of the “mental retirement” effects which have recently been found.

Færdigheder og forventninger, with Paul Bingley and Kristian B. Karlson, In Mai Heidi Ottosen, editor, 15-åriges hverdagsliv og udfordringer, chapter 9, pages 217-242. SFI - Det Nationale Forskningcenter for Velfærd, 2012

Working papers

Placement Optimization in Refugee Resettlement, with Andrew C. Trapp, Alexander Teytelboym, Tommy Andersson, and Narges Ahani

Every year thousands of refugees are resettled to dozens of host countries. While there is growing evidence that the initial placement of refugee families profoundly affects their lifetime outcomes, there have been few attempts to optimize resettlement destinations. We integrate machine learning and integer optimization technologies into an innovative software tool that assists a resettlement agency in the United States with matching refugees to their initial placements. Our software suggests optimal placements while giving substantial autonomy for the resettlement staff to fine-tune recommended matches. Initial back-testing indicates that Annie can improve short-run employment outcomes by 22%-37%. We discuss several directions for future work such as incorporating multiple objectives from additional integration outcomes, dealing with equity concerns, evaluating potential new locations for resettlement, managing quota in a dynamic fashion, and eliciting refugee preferences.


In the media: Dagens Nyheter (Swedish), SVT (Swedish)

Dynamic Refugee Matching, with Tommy Andersson and Lars Ehlers

Only one percent of the 17.2 million asylum seekers in 2016 was part of international resettlement programs: The remaining 99 percent arrived directly to their host countries without assistance from resettlement agencies. These asylum seekers are assigned to a locality directly upon arrival based on some type of dynamic matching system, which is often uninformed and does not take the background of the asylum seekers into consideration. This paper proposes an informed, intuitive, easy-to-implement and computationally efficient dynamic mechanism for matching asylum seekers to localities. This mechanism can be adopted in any dynamic refugee matching problem given locality-specific quotas and that asylum seekers can be classified into specific types. We demonstrate that any matching selected by the proposed mechanism is Pareto efficient and that envy between localities is bounded by a single asylum seeker. We evaluate the performance of the proposed mechanism in settings resembling the US and the Swedish situations, and show that our mechanism outperforms uninformed mechanisms even in presence of severe misclassification error in the estimation of asylum seeker types. With realistic misclassification error (24 percent), the proposed matching mechanism increases efficiency up to 75 percent, and guarantees a reduction in envy of between 17 and 50 percent.

The Effects of Schooling on Wealth Accumulation Approaching Retirement, with Paul Bingley

Education and wealth are positively correlated for individuals approaching retirement, but the direction of the causal relationship is ambiguous in theory and has not been identified in practice. We combine administrative data on individual total wealth with a reform expanding access to lower secondary school in Denmark in the 1950s, finding that schooling increases pension annuity claims but reduces the non-pension wealth of men in their 50’s. These effects grow stronger as normal retirement age approaches. Labour market mechanisms are key, with schooling reducing self-employment, increasing job mobility and employment in the public sector, and improving occupational pension benefits.

Long-Run Saving Dynamics: Evidence from Unexpected Inheritances, with Jeppe Druedahl

Long-run saving dynamics are a crucial component of consumption-saving behavior. This paper makes two contributions to the consumption literature. First, we exploit inheritance episodes to provide novel causal evidence on the long-run effects of a large financial windfall on saving behavior. For identification, we combine a longitudinal panel of administrative wealth reports with variation in the timing of sudden, unexpected parental deaths. We show that after inheritance net worth converges towards the path established before parental death, with only a third of the initial windfall remaining after a decade. These dynamics are qualitatively consistent with convergence to a buffer-stock target. Second, we analyze our findings through the lens of a generalized consumption-saving framework, and show that life-cycle consumption models can replicate this behavior, but only if the precautionary saving motive is stronger than usually assumed. This result also holds for two-asset models, which imply a high marginal propensity to consume.

Paper Appendix Slides
Does Liquidity Substitute for Unemployment Insurance? Evidence from the Introduction of Home Equity Loans in Denmark, with Kristoffer Markwardt and László Sandór

Would the value of unemployment insurance fall if more people had a buffer stock of liquid savings? Using quasi-experimental evidence from the unexpected introduction of home equity loans in Denmark, where public unemployment insurance is voluntary, we find that liquidity and insurance are substitutes. A Danish reform provided less levered homeowners with more liquidity. Using a ten-year-long panel dataset drawn from administrative registries, we find that people who obtained access to extra liquidity were less likely to sign up for unemployment insurance. The effect is concentrated among those for whom insurance has negative expected value. In this group, extra liquidity from housing equity worth one year’s income decreases insurance up-take by as much as a 0.3 percentage point fall in the risk of unemployment. Placebo tests for earlier years show no differential trends by leverage before the natural experiment. This implies that the liquidity of financial assets influences unemployment insurance uptake in the absence of public provision of insurance.

Measurement error in the Survey of Health, Ageing and Retirement in Europe: A validation study with administrative data for education level, income and employment, with Paul Bingley

We link the Survey of Health, Ageing and Retirement in Europe (SHARE) to Danish Administrative Registers, comparing schooling, retirement status and income. We are able to retrieve administrative records for 1670 out of the original 1707 respondents from the first survey wave in 2004. We compare individual linked records in an analysis of measurement error. Overall, we find only minor non-random misclassification of schooling, but otherwise SHARE provides reliable data for socio-economic analysis of schooling, income and retirement. SHARE Denmark overestimates the proportion of individuals with higher education: the probability of misclassification is higher for lower educated, richer individuals. Labour market status is precisely reported, and misclassification probability decreases with age. Average gross household income is not statistically different in SHARE and register data, and we show that measurement error is classical.

SHARE working paper 16:2014