When estimating the effect of a covariate in grouped data, investigators may choose "fixed effects" (FE) approaches with specialized standard errors (e.g., cluster-robust standard errors, CRSE), or multilevel models employing "random effects" (MLM/RE). From over 100 articles in political science, sociology, and education, we find that MLM/RE is often chosen on grounds of greater efficiency and predictive accuracy, simultaneously allowing the inclusion of group-level variables and varying intercepts or slopes, and claims that it ensures correct standard errors for grouped data. Yet, other traditions have favored the FE approach because, unlike MLM/RE, it fully accounts for group-level confounding. We first dissect how these models compare in three analytical pieces: (i) random effects are simply "regularized" fixed effects; (ii) bias is often problematic in MLM/RE, but has a longstanding solution; and (iii) MLM standard errors rely upon narrow assumptions, which can be relaxed. These lessons are known, but we find they are neglected in practice. We then describe how to address the bias and standard error issues within MLM. Fortunately, the resulting approach produces coefficients and standard error estimates exactly equal to the analogous FE estimates with CRSEs, resolving debate about which framework is appropriate.
Figures - uploaded by Chad Hazlett
All figure content in this area was uploaded by Chad Hazlett
Content may be subject to copyright.