Attrition and Non-Response in Panel Data: The Case of the Canadian Survey of Labor and Income Dynamics

This paper provides an analysis of the effects of attrition and non-response on employment and wages using the Canadian Survey of Labour and Income Dynamics. We consider a structural model composed of three freely correlated equations for nonattrition/response, employment and wages. The model is estimated using microdata from 22,990 individuals who provided sufficient information in the first wave of the 1996-2001 panel. The main findings of the paper are that attrition is not random. Attritors and non-respondents likely are less attached to employment and come from low-income population. The correlation between non-attrition and employment is positive and statistically significant, though small. Also, wage estimates are biased upwards. Observed wages are on average higher than wages that would be observed if all the individuals initially selected in the panel remained in the sample.

💡 Research Summary

This paper investigates how attrition and non‑response in a longitudinal labour‑market survey bias estimates of employment status and wages. Using the Canadian Survey of Labour and Income Dynamics (SLID) for the period 1996‑2001, the author constructs a three‑equation structural model in which (i) a selection equation predicts the probability that an individual remains in the panel (i.e., does not attrit or become a non‑respondent), (ii) an employment equation determines whether the individual is employed in a given wave, and (iii) a wage equation models the observed wage conditional on employment. The three equations are allowed to be freely correlated, thereby capturing the joint dependence of panel retention, labour‑force attachment, and earnings.

The sample consists of 22,990 respondents who supplied sufficient baseline information in the first wave. Because attrition creates missing observations across subsequent waves, the model is estimated by full‑information maximum likelihood (FIML), which yields consistent parameter estimates even in the presence of non‑random missing data. Identification is achieved by including variables that affect panel retention but not directly the employment or wage outcomes (e.g., initial willingness to participate, geographic mobility, household composition) in the selection equation, while the employment and wage equations contain the usual labour‑market covariates such as age, education, experience, industry, and regional labour‑market conditions.

The empirical findings are threefold. First, attrition is decidedly non‑random. Individuals with lower household income, lower educational attainment, and a history of precarious or non‑standard employment are significantly more likely to drop out of the panel or to become non‑respondents. For example, respondents in the lowest quintile of income have an attrition odds ratio of roughly 1.8 relative to the median‑income group. Second, the correlation between the non‑attrition (response) equation and the employment equation is positive and statistically significant, but its magnitude is modest (approximately 0.09). This indicates that while staying in the survey is associated with a slightly higher probability of being employed, the effect is economically small. Third, the wage estimates derived from the observed sample are upwardly biased. After accounting for the selection process, the model predicts that the average wage for the original cohort (had everyone remained) would be 5–7 % lower than the wage measured in the observed data. This bias arises because attritors and non‑respondents tend to be low‑wage workers, and their exclusion skews the observed wage distribution upward.

The paper underscores the policy relevance of these results. Analyses that ignore attrition and non‑response will overstate employment rates marginally and overstate average wages more substantially, potentially leading to misguided conclusions about labour‑market health and the effectiveness of employment policies. The author recommends two complementary approaches: (1) improving survey design to reduce attrition (e.g., incentives, more intensive follow‑up, and adaptive contact strategies) and (2) routinely applying selection‑model techniques such as the presented structural system to correct for bias in empirical work.

In sum, the study provides a rigorous methodological framework for quantifying selection bias in panel labour‑market data and demonstrates that attrition systematically excludes low‑income, less‑attached workers, thereby inflating observed employment and wage statistics. These insights are valuable for researchers conducting longitudinal analyses and for policymakers who rely on such data to evaluate labour‑market interventions.