Fig. 2
From: Robust estimation of the effect of an exposure on the change in a continuous outcome
Performance of the REM and cprobit model in estimating the linear effect in simulation studies. Mean and standard error (panel a), coverage (panel b) and type I error and power (panel c) of the estimated linear effect under null and strong effects from the random effects model (REM) and the conditional probit (cprobit) model when applied to the scenarios where no transformation was required (“None”) and Box-Cox transformation was considered (\( \lambda =0,\frac{1}{3},1 \)), with normal and skewed intercept terms, small and large sample sizes (n = 300, 1200). Solid vertical grey lines indicate the true effect sizes in panel a, and the nominal value of the coverage and type I error in panel b and c. Dashed vertical grey lines indicate a 10% bias in the estimate under the strong effect in panel A, and ±1% deviation from the nominal values in panel B and C. (Note: Under strong effect, the coverage of the REM with skewed intercepts was 34.2% or lower for \( \lambda =0,\frac{1}{3},1 \) and beyond the plot range for panel b)