ACF functions are used for model criticism, to test if there is structure left in the residuals. An important prerequisite is that the data is correctly ordered before running the regression models. If there is structure in the residuals of a GAMM model, an AR1 model can be included to reduce the effects of this autocorrelation.
There are basically two methods to reduce autocorrelation, of which the first one is most important:
Improve model fit. Try to capture structure in the data in the model. See the vignette on model evaluation on how to evaluate the model fit:
If no more predictors can be added, include an AR1 model. By including an AR1 model, the GAMM takes into account the structure in the residuals and reduces the confidence in the predictors accordingly.
First mark the start of each time series as TRUE, and all other data points as FALSE. For measures that develop over time this means typically marking the start of each participant-trial combination. For behavioral or response measures, such as reaction times, this means typically marking the first trial of each participant.
Determine the value for the autocorrelation coefficient
Loading the data:
In this data, for each individual subject each trial is a unique time series of at most 100 measurements. Mark the start of each time series using the function
## Group Time Trial Condition Subject Y Event start.event ## 1 Adults 0.00000 -10 -1 a01 0.7554469 a01.-10 TRUE ## 2 Adults 20.20202 -10 -1 a01 2.7834759 a01.-10 FALSE ## 3 Adults 40.40404 -10 -1 a01 1.9696963 a01.-10 FALSE ## 4 Adults 60.60606 -10 -1 a01 0.6814298 a01.-10 FALSE ## 5 Adults 80.80808 -10 -1 a01 1.6939195 a01.-10 FALSE ## 6 Adults 101.01010 -10 -1 a01 2.3651969 a01.-10 FALSE
To determine the value of
rho, we first have to run a ‘plain’ model to see how strong the residuals are correlated.
Different ways to inspect correlation, they all result in the same picture:
Determine the value of lag 1, as indicated by the red dot in the picture below:
start_value_rho basically implements the following line:
##  0.9506088
By default, the uncorrected residuals are returned. So the default ACF plots of the models
m1AR1 are the same:
Below we list a series of functions that can correct the residuals of a model with AR1 included.
Uncorrected versus corrected residuals:
Note that the value for
rho may have been too high.
By default, the data is considered as one single time series. The function can also take into account the different time series, and generate the average ACF over the time series, or examples of ACF functions over individual time series.
split_pred specifies the predictor(s) that define the time series.
par(mfrow=c(1,3), cex=1.1) acf_resid(m1AR1, split_pred = c("Subject", "Trial")) # alternatively, if the predictors are not found in the model we can use the data: acf_resid(m1AR1, split_pred=list(Subject=simdat$Subject, Trial=simdat$Trial)) # ... or the AR.start information, if provided to the model: acf_resid(m1AR1, split_pred="AR.start")
n can be used to generate n ACF plots of individual time series. By default examples are chosen that different from each other with respect to the value of lag 1. However, when
random is set to TRUE, n randomly selected time series are being plotted.
## Quantiles to be plotted: ## 0% 20% 40% 60% 80% 100% ## -0.6282108 -0.4415382 -0.3640369 -0.2972208 -0.1968810 0.1084102
The function is basically a wrapper around the functions
acf_n_plots, which are explained below.
To retrieve the corrected residuals of the model, one could use the function
Note that the function
resid_gam by default does not return NA values (similar to the function
resid). In AR1 models there are two sources of missing values:
missing values in the data result in missing values in residuals
missing values are introduced for the first element of each time series
As a result, potential problems may arise when storing the residuals in a data.frame. Below the problem and a solution is illustrated:
## Error in `$<-.data.frame`(`*tmp*`, res.m1, value = c(-1.62784861779604, : replacement has 75585 rows, data has 75600
## Error in `$<-.data.frame`(`*tmp*`, res.m1AR1, value = c(1.43657183645876, : replacement has 74829 rows, data has 75600
## Error in `$<-.data.frame`(`*tmp*`, res.m1AR1, value = c(1.43657183645876, : replacement has 74829 rows, data has 75585
acf_plot works on a vector of data, rather than a regression model object – similar to the function
acf, but different from
acf_resid. The following commands result in the same plot:
acf_plot allows for generating different time series and applies a function on the resulting ACF values.
split_by expects a list with vectors that define the time series. In contrast with the argument
split_pred of the function
acf_resid, the argument
split_by cannot handle model predictions. The following commands result in the same plot:
par(mfrow=c(1,3), cex=1.1) # when using acf_plot one need to remove missing values manually: acf_plot(resid_gam(m1AR1, incl_na = TRUE), split_by=list(Subject=simdat[!is.na(simdat$Y),]$Subject, Trial=simdat[!is.na(simdat$Y),]$Trial)) # ... acf_resid takes care of that automatically: acf_resid(m1AR1, split_pred=c("Subject", "Trial")) # ... also when using a list to identify time series: acf_resid(m1AR1, split_pred=list(Subject=simdat$Subject, Trial=simdat$Trial))
acf_plot is primarily used when the input is not a regression model object – with a regression model object
acf_resid is more convenient.
Different functions can be applied, including
sd, and custom functions.
tmp <- simdat[!is.na(simdat$Y),] # default function is mean: acf.y <- acf_plot(tmp$res.m1AR1, split_by=list(Subject=tmp$Subject, Trial=tmp$Trial), main="ACF with standard deviation") points(as.numeric(names(acf.y)),acf.y, pch=16, cex=.5) # alternatively, we could ask for SE: acf.se <- acf_plot(tmp$res.m1AR1, split_by=list(Subject=tmp$Subject, Trial=tmp$Trial), fun=sd, plot=FALSE) add_bars(as.numeric(names(acf.se)), y=acf.y+acf.se, y0=acf.y-acf.se, col=alpha('red', f=.25), border=NA, width=.5) legend('topright', legend="sd", fill=alpha('red', f=.25), border=NA, bty='n')
acf_n_plots calculates an ACF for each event, and can be used to plot multiple examples. These examples are random selected or selected as maximally different on the basis of the lag 1-values of the events.
derive_timeseries can be used to extract time series from the model, on the basis of the
AR.start information provided to the model. The function only works if the
AR.start argument is used.
## 'data.frame': 75600 obs. of 10 variables: ## $ Group : Factor w/ 2 levels "Children","Adults": 2 2 2 2 2 2 2 2 2 2 ... ## ..- attr(*, "contrasts")= num [1:2, 1] 0 1 ## .. ..- attr(*, "dimnames")=List of 2 ## .. .. ..$ : chr [1:2] "Children" "Adults" ## .. .. ..$ : chr "Adults" ## $ Time : num 0 20.2 40.4 60.6 80.8 ... ## $ Trial : int -10 -10 -10 -10 -10 -10 -10 -10 -10 -10 ... ## $ Condition : int -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 ... ## $ Subject : Factor w/ 36 levels "a01","a02","a03",..: 1 1 1 1 1 1 1 1 1 1 ... ## ..- attr(*, "contrasts")= num [1:36, 1:35] 0 1 0 0 0 0 0 0 0 0 ... ## .. ..- attr(*, "dimnames")=List of 2 ## .. .. ..$ : chr [1:36] "a01" "a02" "a03" "a04" ... ## .. .. ..$ : chr [1:35] "a02" "a03" "a04" "a05" ... ## $ Y : num 0.755 2.783 1.97 0.681 1.694 ... ## $ Event : int 1 1 1 1 1 1 1 1 1 1 ... ## $ start.event: logi TRUE FALSE FALSE FALSE FALSE FALSE ... ## $ res.m1 : num -1.6278 -0.0803 -1.3736 -3.1395 -2.6019 ... ## $ res.m1AR1 : num 1.4366 -1.3285 -1.8656 0.3502 0.0387 ...