Survey Research — Survey Research Design

Required sample size depends on the expected effect size, desired statistical power (conventionally 0.80), significance level (typically 0.05), and the complexity of planned analyses. For a simple two-group comparison detecting a medium effect (d = 0.50) with 80% power you need roughly 64 per group. Structural equation models typically require at least 200 complete cases. Use an a priori power analysis (e.g., G*Power) rather than rules of thumb, and add 20–30% to account for anticipated incomplete responses.

A low response rate is a threat, not automatically a fatal flaw. What matters is whether respondents differ systematically from non-respondents on study-relevant variables. Conduct a non-response bias analysis — compare early and late respondents on key demographics, or compare your sample profile with known population benchmarks — and report the findings. Acknowledge remaining uncertainty about generalisation even if the analysis is reassuring.

Common-method variance (CMV) is spurious covariance among variables caused by the fact that they were measured with the same instrument at the same time — inflating observed correlations. Procedural remedies include temporal separation of predictor and outcome measurement, using different response formats for different constructs, and including a marker variable. Statistical checks such as Harman's single-factor test or the unmeasured latent factor approach can detect but not fully correct CMV.

Online surveys are cost-effective, fast, and appropriate when the target population has reliable internet access — most working-age adults in developed countries. Postal surveys remain valuable for older populations or those without internet access, at the cost of slower turnaround and higher per-response cost. Telephone surveys have declining response rates but allow complex branching logic and reach populations without internet. Mixed-mode designs (offering multiple response channels) typically achieve higher response rates than any single mode.

A single cross-sectional survey cannot establish causation because it lacks temporal ordering and random assignment. Longitudinal panel designs that measure the same respondents at multiple time points can establish temporal precedence, which is a necessary (but not sufficient) condition for causation. Cross-lagged panel models and random-intercept cross-lagged panel models can control for stable individual differences, strengthening causal inference, but unmeasured confounds remain a threat. For unambiguous causal claims, experimental or quasi-experimental designs are required.