从gamm模型错误产生的随机效应预测：无法在'newdata'上评估组合期望水平

Question

我试图使用newdata参数从gamm模型（来自mgcv数据包）生成预测。我想对模型的lme部分做出预测，以便预测包括随机效应。但是，我认为，由于模型系数的命名方式，我正在遇到问题。

我的问题是，如何结构/命名newdata参数以允许预测。谢谢。

甲兆瓦

mod <- gamm(outcome ~ s(time) + predvar, data=d, 
         random=list(groupvar=~1), 
         correlation = corARMA(form=~1|groupvar, p = 1))  
# okay 
pred <- predict(mod$lme) 

# Not okay 
pred <- predict(mod$lme, newdata=d) 

产生错误

Error in predict.lme(mod$lme, newdata = d) : cannot evaluate groups for desired levels on 'newdata'

---

如果我跑在nlme模型没有花键而言，newdata执行没有问题

mod2 <- lme(outcome ~ time + predvar, data=d, 
         random=list(groupvar=~1), 
         correlation = corARMA(form=~1|groupvar, p = 1))  
# okay 
pred2 <- predict(mod2, newdata=d) 

---

d <- structure(list(time = c(0, 1, 2, 3, 4, 5, 6, 3, 4, 5, 6, 7, 8, 
9, 10, 11, 12, 13, 14, 15, 16, 17, 5, 6, 7, 8, 9, 10, 11, 12, 
13, 14), outcome = c(-1.85, -1.57, -1.38, -1.22, -1.27, -1.63, 
-2.07, -1.36, -0.33, 0.08, 0.3, 0.44, 0.78, 1.03, 1.13, 1.14, 
1.05, 0.94, 0.73, 0.51, 0.08, 0.01, 0.42, 0.59, 0.71, 0.79, 0.87, 
0.75, 0.6, 0.38, 0.01, -0.63), predvar = c(-1.83, -1.77, -1.7, 
-1.84, -1.84, -1.72, -1.69, 0.01, -0.07, 0.16, -0.04, 0.04, 0.25, 
0.19, 0.17, 0.22, 0.34, 0.54, 0.7, 0.81, 0.92, 1.12, 0.58, 0.63, 
0.63, 0.68, 0.62, 0.56, 0.61, 0.73, 0.92, 1.07), groupvar = structure(c(1L, 
1L, 1L, 1L, 1L, 1L, 1L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 2L, 
2L, 2L, 2L, 2L, 2L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L), .Label = c("a", 
"b", "c"), class = "factor")), .Names = c("time", "outcome", 
"predvar", "groupvar"), row.names = c(NA, -32L), class = "data.frame") 

info：我没有指定随机效应作为样条（s（。 ，bs =“re”）），因为我的RE比上面的例子更复杂。

Answer 1

单向预测新数据，如果需要随机效应，则对模型的gam部分进行预测，并添加随机效应。

使用上面的例子，

library(mgcv) 

mod <- gamm(outcome ~ s(time) + predvar, data=d, 
         random=list(groupvar=~1), 
         correlation = corARMA(form=~1|groupvar, p = 1))  
# For comparison: predict with RE: we cant use the newdata arg here 
pred <- predict(mod$lme) 

# Extract the random effects from the model and match with the relevant observation 
re <- coef(mod$lme)[ncol(coef(mod$lme))] 
pred_ref <- re[[1]][match(d$groupvar, gsub(".*/", "", rownames(re)))] 

# Predict on gam part of model and adjust for RE 
pred2 <- as.vector(predict(mod$gam, data=d) - pred_ref) 

# Compare 
all.equal(pred, pred2, check.attributes = F, use.names = F)