library(data.table)
library(bdots)
This vignette walks through the use of the bdots package for analyzing the bootstrapped differences of time series data. The general workflow will follow three steps:
This process is represented with three main functions, bfit -> brefit -> bboot
This package is under active development. The most recent version can be installed with devtools::install_github("collinn/bdots")
.
For our example, we are going to be using eye tracking data from normal hearing individuals and those with cochlear implants using data from the Visual Word Paradigm (VWP).
head(cohort_unrelated)
#> Subject Time DB_cond Fixations LookType Group
#> <num> <int> <int> <num> <fctr> <int>
#> 1: 1 0 50 0.01136364 Cohort 50
#> 2: 1 4 50 0.01136364 Cohort 50
#> 3: 1 8 50 0.01136364 Cohort 50
#> 4: 1 12 50 0.01136364 Cohort 50
#> 5: 1 16 50 0.02272727 Cohort 50
#> 6: 1 20 50 0.02272727 Cohort 50
The bfit
function will create a curve for each unique permutation of subject
/group
variables. Here, we will let LookType
and DB_cond
be our grouping variables, though we may include as many as we wish (or only a single group assuming that it has multiple values). See ?bfit
for argument information.
bfit(data = cohort_unrelated,
fit <-subject = "Subject",
time = "Time",
y = "Fixations",
group = c("DB_cond", "LookType"),
curveFun = doubleGauss(concave = TRUE),
cores = 2)
A key thing to note here is the argument for curveType
is passed as a function call with arguments that further specify the curve. Currently within the bdots
package, the available curves are doubleGauss(concave = TRUE/FALSE),
logistic()
(no arguments), and polynomial(degree = n)
. While more curves will be added going forward, users can also specify their own curves, as shown here.
The bfit
function returns an object of class bdotsObj
, which inherits from data.table.
As such, this object can be manipulated and explored with standard data.table
syntax. In addition to the subject and the grouping columns, we also have a fit
column, containing the fit from the gnls
package, a value for R2
, a boolean indicating AR1
status, and a final column for fitCode.
The fit code is a numeric quantity representing the quality of the fit as such:
fitCode | AR1 | R2 |
---|---|---|
0 | TRUE | R2 > 0.95 |
1 | TRUE | 0.8 < R2 < 0.95 |
2 | TRUE | R2 < 0.8 |
3 | FALSE | R2 > 0.95 |
4 | FALSE | 0.8 < R2 < 0.95 |
5 | FALSE | R2 < 0.8 |
6 | NA | NA |
A fitCode
of 6 indicates that a fit was not able to be made.
In addition to plot
and summary
functions, we also have a method to return a matrix of coefficients from the model fits. Because of the data.table
syntax, we can examine subsets of this object as well
head(coef(fit))
#> mu ht sig1 sig2 base1 base2
#> [1,] 417.6899 0.1986711 145.5628 323.1882 0.01586359 0.03412371
#> [2,] 636.8447 0.2632815 306.2330 214.9787 -0.02154793 0.02858644
#> [3,] 647.5295 0.2547779 496.6745 256.4257 -0.18223561 0.01217570
#> [4,] 734.1526 0.2585742 405.6348 240.2926 -0.05751246 0.03455280
#> [5,] 501.1949 0.2258572 398.7760 158.6752 -0.16159477 0.02529158
#> [6,] 460.7152 0.3067659 382.7322 166.0833 -0.24330874 0.03992168
head(coef(fit[DB_cond == 50, ]))
#> mu ht sig1 sig2 base1 base2
#> [1,] 417.6899 0.1986711 145.5628 323.1882 0.01586359 0.03412371
#> [2,] 647.5295 0.2547779 496.6745 256.4257 -0.18223561 0.01217570
#> [3,] 501.1949 0.2258572 398.7760 158.6752 -0.16159477 0.02529158
#> [4,] 524.7172 0.2479396 287.9935 207.0812 -0.05806978 0.10493565
#> [5,] 549.6930 0.2273107 204.0645 229.6397 -0.01001132 0.02878627
#> [6,] 584.5835 0.1594028 226.3004 420.4861 0.01182188 0.00692017
The plots for this object will compare the observed data with the fitted curve. Here is an example of the first four:
plot(fit[1:4, ])
Depending on the curve type and the nature of the data, we might find that a collection of our fits aren’t very good, which may impact the quality of the bootstrapping step. Using the brefit
function, users have the option to either quickly attempt to automatically refit specified curves or to manually review each one and offer alternative starting parameters. The fitCode
argument provides a lower bound for the fit codes to attempt refitting. The default is fitCode = 1
, indicating that we wish to attempt refitting all curves that did not have fitCode == 0
. The object returned is the same as that returned by bfit
.
## Quickly auto-refit (not run)
brefit(fit, fitCode = 1L, quickRefit = TRUE)
refit <-
## Manual refit (not run)
brefit(fit, fitCode = 1L) refit <-
For whatever reason, there are some data will will not submit nicely to a curve of the specfied type. One can quickly remove all observations with a fit code equal to or greater than the one provided in bdRemove
table(fit$fitCode)
#>
#> 0 1 2
#> 20 14 2
## Remove all failed curve fits
bdRemove(fit, fitCode = 6L)
refit <-
table(refit$fitCode)
#>
#> 0 1 2
#> 20 14 2
There is an additional option, removePairs
which is TRUE
by default. This indicates that if an observation is removed, all observations for the same subject should also be removed, regardless of fit. This ensures that all subjects have their corresponding pairs in the bootstrapping function for the use of the paired t-test. If the data are not paired, this can be set to FALSE
.
The final step is the bootstrapping process, performed with bboot
. First, let’s examine the set of curves that we have available from the first step
DB_cond == 50
and DB_cond == 65
nested within either the Cohort
or Unrelated_Cohort
LookTypes
(but not both).
DB_cond
50
and 65
within the Corhort
group as \(\text{diff}_{\text{Cohort}}\) and the differences between DB_cond
50
and 65
within Unrelated_Corhort
as \(\text{diff}_{\text{UnrelatedCohort}}\). The difference of difference function will then return an analysis of \(\text{diff}_{\text{Cohort}}\) - \(\text{diff}_{\text{UnrelatedCohort}}\)
We can express the type of curve that we wish to fit with a modified formula syntax. It’s helpful to read as “the difference of LHS between elements of RHS”
For the first type, we have
## Only one grouping variable in dataset, take bootstrapped difference
~ Group1(value1, value2)
Outcome
## More than one grouping variable in difference, must specify unique value
~ Group1(value1, value2) + Group2(value3) Outcome
That is, we might read this as “difference of Outcome for value1 and value2 within Group1.”
With our working example, we would find the difference of DB_cond == 50
and DB_cond == 65
within LookType == "Cohort"
with
## Must add LookType(Cohort) to specify
~ DB_cond(50, 65) + LookType(Cohort) Fixations
For this second type of curve, we specify an “inner difference” to be the difference of groups for which we are taking the difference of. The syntax for this case uses a diffs
function in the formula:
## Difference of difference. Here, outer difference is Group1, inner is Group2
diffs(Outcome, Group2(value3, value4)) ~ Group1(value1, value2)
## Same as above if three or more grouping variables
diffs(Outcome, Group2(value3, value4)) ~ Group1(value1, value2) + Group3(value5)
For the example illustrated in (2) above, the difference \(\text{diff}_{50} - \text{diff}_{65}\) represents our inner difference, each nested within one of the values for LookType.
The “outer difference” is then difference of these between LookTypes
. The syntax here would be
diffs(Fixations, DB_cond(50, 65)) ~ LookType(Cohort, Unrelated_Cohort)
Here, we show a fit for each
bboot(formula = Fixation ~ DB_cond(50, 65) + LookType(Cohort),
boot1 <-bdObj = refit,
Niter = 1000,
alpha = 0.05,
padj = "oleson",
cores = 2)
bboot(formula = diffs(Fixation, LookType(Cohort, Unrelated_Cohort)) ~ DB_cond(50, 65),
boot2 <-bdObj = refit,
Niter = 1000,
alpha = 0.05,
padj = "oleson",
cores = 2)
#> Warning in bboot(formula = diffs(Fixation, LookType(Cohort, Unrelated_Cohort))
#> ~ : Permutation testing does not yet work for difference of difference
#> analysis. Switching to padj='oleson' instead
For each, we can then produce a model summary, as well as a plot of difference curves
summary(boot1)
#>
#> bdotsBoot Summary
#>
#> Curve Function: doubleGauss
#> Formula: Fixations ~ (Time < mu) * (exp(-1 * (Time - mu)^2/(2 * sig1^2)) * (ht - base1) + base1) + (mu <= Time) * (exp(-1 * (Time - mu)^2/(2 * sig2^2)) * (ht - base2) + base2)
#> Time Range: (0, 2000) [501 points]
#>
#> Difference of difference: FALSE
#> Paired t-test: TRUE
#> Difference: DB_cond -- 50 65
#>
#> Autocorrelation Estimate:
#> FWER adjust method: oleson
#> Alpha: 0.05
#> Adjusted alpha:
#> Significant Intervals:
#> NULL
plot(boot1)