Estimates of the weight function parameters.

Standard errors of the estimates, calculated from a numerically approximated information matrix, are given in parentheses. Parameter w(I) described bias in initial studies, w(E)1 and w(E)2 describe bias in early replication studies in Model 3 and the Proteus model (Fig. 1 and Methods), and w(S) describes bias in subsequent studies. In Model 1, w(S) describes bias in all studies, in Model 2, all but initial studies. Further details are given in Fig. 1 and Methods. The AIC is given by 2k-2L, where L is the maximized value of the log likelihood function and k is the number of parameters. The values given in the table are differences to the values for the unbiased model. Model 1 shows clear indication for selection bias. Model 2 shows that bias is larger for initial studies on a marker, compared to subsequent ones. The Proteus model indicates that non-significant studies opposing the initial result tend to be more likely published than non-significant studies confirming it. Note that the two parameters log w(E)1 and log w(E)2 are estimated relative to the outer intervals, and therefore the errors in the estimates are correlated. The difference between the two parameters is log w(E)1 - log w(E)2 = 0.32. The standard error of the difference can be calculated from the variances and co-variances between the two estimates as determined by the information matrix and is given by sqrt{var(log w(E)1)+var(log w(E)2) -2 covar (log w(E)1, log w(E)2)} = 0.14. Thus non-significant studies confirming the initial result are published with a probability of 73% relative to non-significant studies opposing the initial result, with a confidence interval ranging from 55% to 96%. Model 3 shows that the direction of the second study does not matter per se. Selection bias for the early replication studies falls in between bias for initial and for subsequent results.

估计量的标准误（standard errors）通过数值近似的信息矩阵（information matrix）计算得到，结果置于括号内。参数w(I)用以表征初始研究中的偏倚，w(E)₁与w(E)₂分别对应模型3及Proteus模型（图1及方法部分）中早期重复研究的偏倚，w(S)则表征后续研究中的偏倚。在模型1中，w(S)指代所有研究的偏倚；在模型2中，其指代除初始研究外的全部研究的偏倚。更多细节参见图1及方法部分。
赤池信息准则（Akaike Information Criterion, AIC）的计算公式为2k-2L，其中L为对数似然函数的极大值，k为参数个数。表格中给出的数值为其与无偏模型对应值的差值。
模型1明确显示存在选择偏倚。模型2结果表明，针对某一标记物的初始研究偏倚程度大于后续研究。Proteus模型结果显示，与初始研究结果相悖的非显著性研究，相较于验证初始结果的非显著性研究，更易被发表。需注意，参数log w(E)₁与log w(E)₂是相对于外区间进行估计的，因此二者的估计误差存在相关性。两参数的差值为log w(E)₁ − log w(E)₂ = 0.32。该差值的标准误可通过信息矩阵得到的两估计值的方差与协方差计算得出，公式为√[var(log w(E)₁)+var(log w(E)₂) - 2covar(log w(E)₁, log w(E)₂)] = 0.14。据此，验证初始结果的非显著性研究的发表概率，相较于与初始结果相悖的非显著性研究，约为73%，其置信区间为55%至96%。
模型3显示，第二项研究的方向本身并不会对偏倚产生影响。早期重复研究的选择偏倚程度介于初始研究与后续研究之间。