Skip to main content
. 2015 Dec 13;35(11):1880–1906. doi: 10.1002/sim.6835

Table 3.

Comparison of summarized data methods for uncorrelated IVs.

Imprecise external weights (from 5000 individuals) Precise external weights (from 50000 individuals) Oracle weights
α Median SD SE Coverage Power Median SD SE Coverage Power Median SD SE Coverage Power
Allele score method using summarized data
β U=+1 0.05 0.200 0.239 0.219 97.3 14.8 0.198 0.192 0.201 97.0 17.6 0.198 0.186 0.197 97.1 18.0
0.10 0.200 0.093 0.102 97.1 52.1 0.201 0.089 0.098 97.0 55.6 0.201 0.088 0.098 97.0 56.0
0.20 0.199 0.044 0.049 97.2 97.8 0.199 0.043 0.049 97.2 98.0 0.199 0.043 0.049 97.1 98.1
β U=−1 0.05 0.201 0.232 0.186 93.1 20.6 0.201 0.196 0.165 92.5 24.9 0.201 0.191 0.161 92.5 25.7
0.10 0.200 0.092 0.084 92.7 68.1 0.200 0.088 0.081 93.3 71.2 0.199 0.087 0.080 93.4 71.7
0.20 0.200 0.044 0.041 93.0 99.7 0.200 0.043 0.040 92.8 99.8 0.200 0.043 0.040 92.8 99.8
Summary statistic methoda
β U=+1 0.05 0.148 0.168 0.162 93.1 14.8 0.191 0.191 0.187 94.8 17.6 0.199 0.193 0.190 95.0 18.0
0.10 0.183 0.094 0.091 94.2 52.1 0.199 0.096 0.095 94.8 55.6 0.201 0.096 0.095 95.0 56.0
0.20 0.194 0.048 0.047 94.9 97.8 0.198 0.048 0.047 95.0 98.0 0.199 0.048 0.047 94.9 98.1
β U=−1 0.05 0.147 0.139 0.133 93.0 20.6 0.194 0.158 0.154 94.7 24.9 0.201 0.161 0.157 94.8 25.7
0.10 0.181 0.077 0.075 94.0 68.1 0.198 0.078 0.078 95.3 71.2 0.200 0.079 0.078 95.3 71.7
0.20 0.196 0.040 0.039 94.3 99.7 0.200 0.040 0.039 94.9 99.8 0.200 0.040 0.039 95.0 99.8
Likelihood‐based methoda
β U=+1 0.05 0.203 0.242 0.195 94.0 18.3 0.194 0.182 0.191 97.3 17.0
0.10 0.201 0.104 0.097 94.4 54.3 0.200 0.096 0.095 95.0 55.6
0.20 0.197 0.050 0.048 94.5 97.7 0.198 0.052 0.048 93.4 97.8
β U=−1 0.05 0.204 0.201 0.161 94.3 25.4 0.198 0.153 0.157 97.1 24.4
0.10 0.198 0.086 0.080 94.5 70.1 0.199 0.079 0.079 95.2 71.3
0.20 0.199 0.042 0.040 94.3 99.7 0.200 0.046 0.039 92.0 99.7

Median estimates over 10000 simulations of β X=0.2, standard deviation (SD) of estimates, median standard error (SE) of estimates, coverage (%) of nominal 95% confidence interval for the causal parameter and empirical power (%) based on nominal 95% confidence interval to detect a causal effect from simulation study with 15 uncorrelated instrumental variables (IVs) varying the direction of confounding (β U) and average strength of IV (α) using three summarized data methods: allele score, summary statistic and likelihood‐based methods, with weights taken from an external source corresponding to an independent sample of size 5000, 50000 and using the true (oracle) weights.

aThe ‘weights’ for the summary statistic and likelihood‐based methods are used as the β^Xk association estimates in equations (4) and (6). In the allele score method, the weights (w j) and the association estimates ( β^Xk) in the data under analysis are both used to provide the causal estimate.