glmmTMB 与不规则时间的自相关 [英] glmmTMB with autocorrelation of irregular times

问题描述

我正在组装一个 glmmTMB 模型.我每年 5 月在一个站点收集了 4 年的数据.一年内的时间分辨率可以从几分钟(甚至同一分钟)到几天不等.协方差小插图说ar1() 结构需要有规律的时间序列，但是 ou(times + 0 | group) 结构可以处理不规律的时间.也就是说 - 看起来 times 参数是一个因素 - 它如何处理不规则的时间结构??

I'm putting together a glmmTMB model. I have data collected at a single site over the course of May, every year, for 4 years. Time resolution within year can range from a few minutes (or even same minute) to days apart. The covariance vignette says that the ar1() structure requires a regular time series, but the ou(times + 0 | group) structure can handle irregular times. That said - it looks like the times argument is a factor - how does that work with irregular time structure??

那么，例如，这是对 ou() 结构的正确使用吗?

So, for example, is this a correct use of the ou() structure?

df <- structure(list(DayYear = c(234, 220, 234, 231, 243, 229, 228, 
223, 220, 218, 234, 237, 234, 231, 241, 237, 241, 241, 233, 234, 
234, 232, 218, 227, 232, 229, 220, 223, 228, 224), DateTime =     structure(c(1495477980, 
1399590540, 1495479780, 1495225920, 1464631980, 1495052760, 1463324460, 
1494525780, 1494256560, 1494088440, 1495471320, 1495730940, 1495476960, 
1495225200, 1432919940, 1495725900, 1432924200, 1432918860, 1495384020, 
1495479900, 1463848140, 1495298820, 1399420080, 1463253000, 1463692920, 
1495037040, 1494275160, 1494510780, 1463348220, 1494597180), class =     c("POSIXct", 
"POSIXt"), tzone = ""), Year = c(2017, 2014, 2017, 2017, 2016, 
2017, 2016, 2017, 2017, 2017, 2017, 2017, 2017, 2017, 2015, 2017, 
2015, 2015, 2017, 2017, 2016, 2017, 2014, 2016, 2016, 2017, 2017, 
2017, 2016, 2017), N = c(2, 2, 7, 2, 6, 4, 1, 4, 1, 3, 1, 6, 
2, 2, 2, 2, 5, 5, 3, 5, 3, 2, 4, 1, 6, 2, 2, 3, 5, 2)), row.names = c(NA, 
-30L), class = c("tbl_df", "tbl", "data.frame"))

在一年内创建采样因子

df <- df %>%
    arrange(DateTime) %>%
    group_by(Year) %>%
    mutate(times = 1:n()) %>%
    ungroup() %>%
    mutate(YearF = as.factor(Year),
            times = numFactor(times))

mod1 <- glmmTMB(N ~ DayYear + YearF + 
            ou(times + 0 | YearF),
            family = nbinom2,
            data = df)

这个特定模型运行得不太好，因为玩具数据集太小了(并且可能没有显示我需要显示的内容) - 但这是不规则时间序列下自相关结构的正确规范吗?

This particular model doesn't run too well because the toy dataset is so tiny (and probably doesn't show what I need showing) - but is that a correct specification of the autocorrelation structure under an irregular time series?

推荐答案

不，不是:您必须在 numFactor 中使用十进制时间/日期.你这样做的方式强制数据集等距.下面我使用 lubridate::decimal.date(DateTime) %% 1 来获取用作时间坐标的年份变量.

No, it's not: you have to use decimal times/dates in numFactor. The way you've done it coerces the data set to be equally spaced. Below I use lubridate::decimal.date(DateTime) %% 1 to get the fraction-of-year variable that's used as the time coordinate.

library(dplyr)
library(lubridate)
library(glmmTMB)
df2 <- (df
    %>% arrange(DateTime)
    %>% group_by(Year)
    %>% mutate(times = lubridate::decimal_date(DateTime) %% 1)
    %>% ungroup()
)

df3 <- (df2
    %>% mutate(YearF = as.factor(Year),
               times = glmmTMB::numFactor(times))
    %>% select(N, DayYear, YearF, times)
)

mod1 <- glmmTMB(N ~ DayYear + YearF + 
            ou(times + 0 | YearF),
            family = nbinom2,
            data = df3)

这篇关于glmmTMB 与不规则时间的自相关的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！