Panel-based Observational Quantitative Research

Two waves are the technical minimum and allow assessment of change and simple lagged effects. However, two-wave designs have limited power to distinguish true change from measurement error, and they cannot model nonlinear trajectories. Three or more waves are strongly preferred: they enable growth-curve modeling, test for non-linear trends, and provide more robust attrition diagnostics. If only a single follow-up wave is feasible, a prospective cohort design with a baseline and one follow-up is still more informative than a cross-sectional snapshot.

Run the Hausman specification test. If the test rejects the null hypothesis (p < .05), the random-effects assumption that unit effects are uncorrelated with predictors is violated, and fixed-effects is consistent; random-effects will be biased. If the test does not reject, random-effects is more efficient and also allows estimation of time-invariant predictors, which fixed-effects cannot identify. When the theoretical interest centers on time-invariant covariates (e.g., race, gender, country), mixed models or hybrid fixed/random approaches may be required.

Attrition becomes problematic when it is non-random — that is, when participants who drop out differ systematically from those who remain on key study variables. Always compare baseline characteristics of completers versus drop-outs. If differences are found, use multiple imputation or inverse-probability weighting to correct estimates. Prevention is better than correction: maximize retention through regular contact, incentives, and easy response modes. Report attrition rates and test results transparently so readers can judge the threat to validity.

Partially. Fixed-effects estimation removes all stable unobserved unit-level confounders, which is a substantial advantage over cross-sectional designs. However, time-varying confounders remain a threat and cannot be absorbed by standard panel estimators. Stronger causal identification requires supplementary strategies such as instrumental variables, difference-in-differences with a control group, regression discontinuity, or natural experiment designs embedded within the panel. Without these, results should be described as associations with careful theoretical discussion of plausible confounders.

Sample size requirements depend on the expected effect size, the number of waves, and the analytic model. For individual-level panel regression with fixed effects, power calculations should account for the intraclass correlation (ICC) among repeated observations on the same unit: higher ICC reduces effective sample size. A priori power analysis using software such as G*Power or R packages (e.g., longpower for longitudinal designs) is strongly recommended. As a practical rule, inflate your target analytic-wave sample by 20–30% at recruitment to offset expected attrition.