Minimal example
The main function is fit_cumhist() that takes a data
frame with time-series as a first argument. In addition, you need to
specify the name of the column that codes the perceptual state
(state argument) and a column that holds either dominance
phase duration (duration) or its onset
(onset). The code below fits data using Gamma distribution
(default family) for a single run of a single participant. By default,
the function fits cumulative history time constant but uses default
fixed mixed state value (mixed_state = 0.5) and initial
history values (history_init = 0).
library(bistablehistory)
data(br_singleblock)
gamma_fit <- fit_cumhist(br_singleblock,
state="State",
duration="Duration",
refresh=0)Alternatively, you specify onset of individual dominance phases that will be used to compute their duration.
gamma_fit <- fit_cumhist(br_singleblock,
state="State",
onset="Time")You can look at the fitted value for history time constant using
history_tau()
history_tau(gamma_fit)
#> # A tibble: 1 x 3
#> Estimate `5.5%` `94.5%`
#> <dbl> <dbl> <dbl>
#> 1 0.986 0.783 1.23and main effect of history for both parameters of gamma distribution
historyef(gamma_fit)
#> # A tibble: 2 x 4
#> DistributionParameter Estimate `5.5%` `94.5%`
#> <fct> <dbl> <dbl> <dbl>
#> 1 shape 1.05 0.120 1.92
#> 2 scale 0.269 -0.676 1.31The following model is fitted for the example above, see also companion vignette for details on cumulative history computation. \[Duration[i] \sim Gamma(shape[i], rate[i]) \\ log(shape[i]) = \alpha^{shape} + \beta^{shape}_H \cdot \Delta h[i] \\ log(rate[i]) = \alpha^{rate} + \beta^{rate}_H \cdot \Delta h[i] \\ \Delta h[i] = \text{cumulative_history}(\tau, \text{history_init})\\ \alpha^{shape}, \alpha^{rate} \sim Normal(log(3), 5) \\ \beta^{shape}_H, \beta^{rate}_H \sim Normal(0, 1) \\ \tau \sim Normal(log(1), 0.15)\]
Passing Stan control parameters
You can pass Stan control parameters via control
argument, e.g.,
gamma_fit <- fit_cumhist(br_singleblock,
state="State",
duration="Duration",
control=list(max_treedepth = 15,
adapt_delta = 0.99))See Stan documentation for details (Carpenter et al. 2017).
Run
By default, fit_cumhist() function assumes that the
time-series represent a single run, so that history states are
initialized only once at the very beginning. You can use
run argument to pass the name of a column that specifies
individual runs. In this case, history is initialized at the beginning
of every run to avoid spill-over effects.
gamma_fit <- fit_cumhist(br_single_subject,
state="State",
onset="Time",
run="Block")Experimental session
Experimental session specifies which time-series were measured
together and is used to compute an average dominance phase duration
that, in turn, is used when computing cumulative history: \(\tau_H = \tau \cdot <D>\), where
\(\tau\) is normalized time constant
and \(<D>\) is the mean dominance
phase duration. This can be used to account for changes in overall
alternation rate between different sessions (days), as, for example,
participants new to the stimuli tend to “speed up” over the course of
days (Suzuki and Grabowecky 2007). If you
do not specify session parameter then a single
mean dominance phase duration is computed for all runs of a single
subject.
Random effect
The random_effect argument allows you to specify a name
of the column that codes for a random effect, e.g., participant
identity, bistable display (if different displays were used for a single
participant), etc. If specified, it is used to fit a hierarchical model
with random slopes for the history effect (\(\beta_H\)). Note that we if random
independent intercepts are used as prior research suggest large
differences in overall alternation rate between participants (Brascamp et al. 2019).
Here, is the R code that specifies participants as random effect
gamma_fit <- fit_cumhist(kde_two_observers,
state="State",
duration="Duration",
random_effect="Observer",
run="Block")And here is the corresponding model, specified for the shape parameter only as identical formulas are used for the rate parameter as well. Here, \(R_i\) codes for a random effect level (participant identity) and a non-centered parametrization is used for the pooled random slopes.
\[Duration[i] \sim Gamma(shape[i], rate[i]) \\ log(shape[i]) = \alpha[R_i] + \beta_H[R_i] \cdot \Delta h[i] \\ \Delta H[i] = \text{cumulative_history}(\tau, \text{history_init})\\ \alpha[R_i] \sim Normal(log(3), 5) \\ \beta_H[R_i] = \beta^{pop}_H + \beta^{z}_H[R_i] \cdot \sigma^{pop}_H\\ \beta^{pop}_H \sim Normal(0, 1) \\ \beta^{z}_H[R_i] \sim Normal(0, 1) \\ \sigma^{pop}_H \sim Exponential(1) \\ \tau \sim Normal(log(1), 0.15)\]
Identical approach is take for \(\tau\), if tau=' "1|random"'
was specified and same holds for mixed_state=' "1|random"'
argument, see below.
Fixed effects
fit_cumhist() functions allows you to specify multiple
fixed effect terms as a vector of strings. The implementation is
restricted to:
- Only continuous (metric) independent variables should be used.
- A single value is fitted for each main effect, irrespective of whether a random effect was specified.
- You cannot specify an interaction either between fixed effects or between a fixed effect and cumulative history variable.
Although this limits usability of the fixed effects, these restrictions allowed for both a simpler model specification and a simpler underlying code. If you do require more complex models, please refer to companion vignette that provides an example on writing model using Stan directly.
You can specify custom priors (a mean and a standard deviation of a
prior normal distribution) via history_effect_prior and
fixed_effects_priors arguments. The former accepts a vector
with mean and standard deviation, whereas the latter takes a named list
in format .
Once fitted, you can use fixef() function to extract a
posterior distribution or its summary for each effect.
Cumulative history parameters
fit_cumhist() function takes three parameters for
cumulative history computation (see also companion vignette):
-
tau: a normalized time constant in units of mean dominance phase duration. -
mixed_state: value used for mixed/transition state phases, defaults to0.5. -
history_init: an initial value for cumulative history at the onset of each run. Defaults to0.
Note that although history_init accepts only fixed
values either a single value used for both states or a vector of two. In
contrast, both fixed and fitted values can be used for the other three
parameters. Here are possible function argument values
- a single positive number for
tauor single number within [0, 1] range formixed_state. In this case, the value is used directly for the cumulative history computation, which is default option formixed_state. -
NULL: a single value is fitted and used for all participants and runs. This is a default fortau. -
'random': an independent tau is fitted for each random cluster (participant, displays, etc.).random_effectargument must be specified. -
'1|random': values for individual random cluster are sampled from a fitted population distribution (pooled values).random_effectargument must be specified.
You can specify custom priors for each cumulative history parameter
via history_priors argument by specifying mean and standard
deviation of a prior normal distribution. The
history_priors argument must be a named list, , e.g.,
history_priors = list("tau"=c(1, 0.15)).
Once fitted, you can use history_tau() and
history_mixed_state()functions to obtain a posterior
distribution or its summary for each parameter.
Distribution family
fit_cumhist currently supports three distributions:
'gamma', 'lognormal', and
'normal'.
Gamma
\[Duration[i] \sim Gamma(shape[i], rate[i])\] For Gamma distribution independent linear models with a log link function are fitted for both shape and rate parameter. Priors for intercepts for both parameters are \(\alpha ~ Normal(log(3), 5)\).
Model comparison
Models fits can be compared via information criteria. Specifically,
the log likelihood is stored in a log_lik parameter that
can be directly using loo::extract_log_lik() function (see
package (@ loo?)) or used
to compute either a leave-one-out cross-validation (via
loo() convenience function) or WAIC (via
waic()). These are information criteria that can be used
for model comparison the same way as Akaike (AIC), Bayesian (BIC), or
deviance (DIC) information criteria. The latter can also be computed
from log likelihood, however, WAIC and LOOCV are both preferred for
multi-level models, see (Vehtari, Gelman, and
Gabry 2017). The model comparison itself can be performed via
loo::loo_compare() function of the loo
package.
Predicted values
You can predict durations for individual dominance phases via
predict() function. You have an option of getting a summary
(an average expected duration plus an optional credible interval) or
computing predicted durations for every sample.
For summary statistics with 89% credible interval.
Predictions for every sample for full length time-series
(invalid samples are filled with NA):
prediction_samples <- predict(gam_fit, summary = FALSE, full_length = TRUE)Predictions for every sample only for valid samples:
prediction_samples <- predict(gam_fit, summary = FALSE, full_length = FALSE)Computing and using cumulative history
If you are interested in the cumulative history itself, you can
extract from the fitted object via predict_history()
function. Note that there are five different history types you can
extract:
-
"1": cumulative history for the first perceptual state, i.e., state with index of 1. -
"2": cumulative history for the second perceptual state, i.e., state with index of 2. -
"dominant": for the state that is dominant during the following phase. -
"suppressed": for the state that is suppressed during the following phase. -
"difference": difference between cumulative histories (\(\Delta h = h_{suppressed} - h{dominant}\)), which is used in linear models.
H <- predict_history(gam_fit, "difference")Alternatively, you can skip fitting and compute history directly
using predefined values via compute_history().
df <- compute_history(br_singleblock,
state="State",
duration="Duration",
tau=1,
mixed_state=0.5,
history_init=0)