metafor中SMCR效应量聚合模型与多层模型估计结果不一致问题

我正在遵循metafor的研究层面数据聚合指南开展meta分析，作者Viechtbauer提到：

以这种方式聚合估计值时，对这些合并估计值进行meta分析得到的结果与上述多层模型的结果完全一致

但我的结果并不匹配。我猜测差异源于教程用了escalc(measure="SMD")，而我用的是SMCR。

效应量计算代码 #

df <- escalc(measure="SMCR", 
             m1i=me_post_int, 
             m2i=me_pre_int, 
             sd1i=sd_pre_int, 
             ni=num_int,
             ri=r_int_imp,
             data=df, 
             append=TRUE,  
             replace = F,
             var.names = c("yi_int", "vi_int")) 

df <- escalc(measure="SMCR", 
             m1i  = me_post_c,
             m2i  = me_pre_c, 
             sd1i = sd_pre_c, 
             ni   = num_c, 
             ri   = r_c_imp, 
             data = df, 
             append=TRUE, 
             replace = F,
             var.names = c("yi_c", "vi_c")) 

# 反转负向结局的效应量
df$yi_int <- ifelse(df$direction == "neg", yes = df$yi_int*(-1), no = df$yi_int)
df$yi_c <- ifelse(df$direction == "neg", yes = df$yi_c*(-1), no = df$yi_c)

df$yi <- df$yi_int - df$yi_c
df$vi <- df$vi_int + df$vi_c

多层模型代码及输出（df_pain为df子集） #

pain_3lv<- rma.mv(yi, vi, random = ~ 1 | study_id/Unique_ID_ES, 
                  method="REML", data=df_pain)
pain_3lv

输出结果：

Multivariate Meta-Analysis Model (k = 50; method: REML)

Variance Components:

            estim    sqrt  nlvls  fixed                 factor 
sigma^2.1  0.2641  0.5139     14     no               study_id 
sigma^2.2  0.0764  0.2764     50     no  study_id/Unique_ID_ES 

Test for Heterogeneity:
Q(df = 49) = 271.0640, p-val < .0001

Model Results:

estimate      se    zval    pval   ci.lb   ci.ub     
  0.4517  0.1581  2.8568  0.0043  0.1418  0.7616  ** 

---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

聚合模型代码及输出 #

agg <- aggregate(df_pain, cluster=study_id, V=vcov(pain_3lv, type="obs"),  addk=TRUE)

res <- rma(yi, vi, method="EE", data=agg, slab = study_id)
res

输出结果：

Equal-Effects Model (k = 14)

I^2 (total heterogeneity / total variability):   79.70%
H^2 (total variability / sampling variability):  4.93

Test for Heterogeneity:
Q(df = 13) = 64.0479, p-val < .0001

Model Results:

estimate      se    zval    pval   ci.lb   ci.ub      
  0.3016  0.0475  6.3448  <.0001  0.2084  0.3948  *** 

---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

我能找到的唯一代码差异就是效应量测量方法（SMD vs SMCR），其余完全复制教程代码。我按照指南用了"EE"（固定效应模型），但即使换成REML模型，估计结果还是和原多层模型不一致。

内容的提问来源于stack exchange，提问作者kinbeat