Relevant papers here

[itchyshin]

Here are some papers of relevant including my paper with colleagues

Statistics in Medicine - 2016 - Kalia.pdf : This study evaluated the accuracy of estimating ICC with time-to-event data that includes censored observations. They compared two methods to get ICC while accounting for the censoring- censoring indicators or observed event times. They found a substantial negative bias in ICC from the two methods. Removing censored data affects the overall means of the event times, and reduces variance.

RJ-2022-034.pdf : iccCounts package for quantifying repeatability of count data with Poisson, zero-inflated Poisson, negative binomial, or zero-inflated negative binomial regression. I like how the package includes goodness of fit plots and p-value estimates of the fit based on Monte Carlo simulations comparing actual vs simulated randomized quantile residuals.

Biometrical J - 2016 - Oliveira.pdf : Presents the use of generalized linear models with normal and gamma random effects to model the correlation between repeated measures and overdispersion for hierarchical count and time-to-event data. I understand the concept and rationale of these random effects, but had difficulty replicating the math.

RJ-2022-034.pdf cloglog.pdf

Austin_2017_a tutorial on multilevel survival analysis methods.pdf : Several methods to analyze time-to-event data with clustering. Cox proportional hazard models, as well as piecewise exponential (poisson) and discrete time (binomial) survival models that both involve defining intervals within which you can assess the probability that an event happened. Includes analysis of an example data set with all three methods and the variance of the random effect is approximately the same.

Crowther etal_2012_individual patient data meta-analysis of survival data using poisson regression models.pdf : Demonstrates a poisson method for analyzing survival data. They compared performance of PWE models using actual and simulated data when varying interval length and found no difference in estimates of the hazard ratio for fixed or random effects.

Hirsch etal_2016_log-normal frailty models fitted in generalized poisson regression.pdf : More in-depth mathematical look at a PWE survival model. Includes a simulation study of censored survival data and they found comparable results between their shared-frailty (PWE) model, a cox model and a Gompertz parametric model. Implemented in SAS, so this seems less helpful for R folks.

Donahue_2017_phmm package-equivalence of poisson, cox ph.pdf

[kelseybmccune]

Papers in animal behavior that use time-to-event data:

Birds Griffin, Diquelou - 2015 - Innovative problem solving in birds a cross-species comparison of two highly successful passerines.pdf : Griffin has always recognized the need for cox models in time-to-event data. This paper models innovation through cox proportional hazard models on latency to solve, as well as neophobia (latency to eat), persistence (number of attempts) and motor diversity. They determined repeatability across the 3 innovation tasks "using binary GLMMs with a logit link and multiplicative overdispersion and individual as an explanatory factor (Nakagawa & Schielzeth, 2010)". In other words, they used the rptR package to look at whether the probability of success (yes/no) was repeatable. They did not test repeatability of the other time-to-event variable, neophobia.

Vamos & Shaw_2024_consistent individual differences in caching behavior in toutouwai.pdf : They determine repeatability of multiple caching behaviors, including retrieval latency. This variable was capped at 31 seconds (almost 10% of data in this variable are censored) and modeled in rptR using the Gaussian family. Data are available that we could re-analyze with our method to see if results differ

van den Heuvel etal_2023_artificial selection for reversal learning reveals limited repeatability and no heritability of cognitive flexibility.pdf : To assess repeatability of the time to reach the reversal learning criteria (TCC; a censored variable) they "computed so-called pseudo values of the restricted mean TTC from the Kaplan–Meier survival curves (Andersen & Pohar, 2010). As a measure of TTC, pseudo values were computed for all observed trials, both censored and uncensored, using the function pseudo (Therneau, 2023). Pseudo values were evaluated at the maximum trial number reached by the individual with the highest TTC." The authors then used pseudo values as dependent variables to conduct a Poisson regression in MCMCglmm and calculated repeatability from the estimated variance components.

Other animals Peignier etal_2022_exploring links between personality traits and social environment in poison frogs.pdf : Measured personality traits in wild poison dart frogs. There were several latency and time-to-event variables that were transformed to fit a Gaussian distribution. They used rptR to quantify repeatability.

Lukas etal_2021_consistent behavioral syndrome across seasons in an invasive freshwater fish.pdf : Measured boldness as latency to emerge from a shelter, and gave a ceiling value to fish that failed to emerge (i.e., censoring is present). This variable was log transformed to calculate repeatability based on variance components from a lmm.

Briolat etal_2021_generalist camoflage can be more successful than microhabitat specialization in natural environments.pdf : The relationship between camouflage strategy and distance-to-detection was quantified using coxme. Human participant ID and location were used as grouping variables.

Holzmann & Cordoba 2024 : Howler monkey territorial male response to simulated intruder. Measured variables included closest approach distance (one censored variable where the monkey never approached the speaker), and latency to do several behaviors. Although they have repeated measures, they did not conduct repeatability analyses. Small sample size.

Villegas-Ríos et al. 2017 : Distance of cod dispersal data collected from automated telemetry arrays. Number of censored individuals not given, but Fig. A1 shows significant numbers of fish lost to follow up after a few weeks.

Brehm et al. 2019 : Seed dispersal distance by small mammals as a function of personality traits and land cover features. 35% of cached seeds were never relocated (censored data).

[kelseybmccune]

Papers in medical research that use time-to-event data

Thomson etal_2021_Effects of antibiotic resistance on bacterial pathogens related to sepsis in neonates.pdf : Used coxph and coxme (country as the grouping factor) to assess the impact of different antibiotics on survival of neonates that had sepsis.

Thompson etal_2010_Time-to-event data from epidemiological studies.pdf

[kelseybmccune]

Papers with censored data in the form of ceiling values

Rudin etal_2016_changes in dominance status affect personality.pdf : Measured cricket latency to emerge from a shelter after simulated predator attack. Trials were 10 minutes long, and if crickets didn't emerge they received a ceiling value of 600 seconds

Vamos.Shaw_2024_consistent.individual.differences.in.caching.behavior.in.toutouwai.pdf : Latency to retrieve a cached worm, where almost 10% of data were censored.

[kelseybmccune]

Ecology and Evolution papers with time-to-event data

Landes etal_2020_An introduction to event history analyses for ecologists.pdf : A review paper that explains what time-to-event (event history) data are, analysis techniques, and worked examples from animal ecology.

Wasser etal_2000_outbreeding depression in trees.pdf : Older paper, but looks at seedling emergence and survival in clustered groups of plants that varied in the spatial distance between the parent plants. Time-to-emerge, time-to-flower and survival time were the variables of interest. The authors used general linear regression on the proportion of emerged/flowering seedlings and ANOVA on log-transformed survival time with a random effect for site.

Similarly, Barak etal_2018_Cracking the case Seed traits and phylogeny predict time to germination in prairie.pdf : the authors use coxph for the analysis, but maybe should have included a random effect of species ID.

Villagomez et al. 2021 : Time to flowering of a plant after manipulating photo period and temperature. Censored data present, where some plants didn't flower, but percentage is not mentioned. Also time to initiation of brood rearing by the pollinator honey bee, with some censored data where colonies never began rearing broods. Used coxme for both analyses. Very similar paper.

Wrobel et al. 2022 : Used coxme to assess acorn dispersal distance (by jays) as a function of seed and environmental traits. The random effect was for Site ID.

Snell etal_2018_intraspecific variation in seed dispersal behavior.pdf : Review/theoretical paper that demonstrates that there is intraspecific variation in seed dispersal distances that should be accounted for when modeling plant community dynamics. The first example uses tracking data from 3 toucan species and shows distance to seed dispersal of an individual on the low end, one on the high end, as well as the population mean. ID as the random effect.

Ramírez et al. 2015 : Latency to initiate migration to the breeding grounds and return to the ocean from the breeding grounds in petrels. Also they looked at the repeatability of distance to nonbreeding grounds from the breeding sites. They used LMMs and ID as a random effect.

Brandl et al. 2019 : Latency to initiate nest building and egg laying as a function of the nest stage of nearby conspecifics. Nest box and site included as random effects in Poisson models.

Stroeymeyt et al. 2011 : Evaluated the latency of several behaviors related to ants moving to a new nest. Analyzed using glmm with colony (social group ID) as a random effect, transforming the latency variables as needed.

[itchyshin]

@kelseybmccune - these are great. You can make a table to summarise this - like in this method paper in Behavioral Ecology or MEE paper.

See table 1: https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13856

See table 1 & 2: https://academic.oup.com/beheco/article/21/5/902/198528

You might do this a bit later. I am now unsure whether there is a paper here - we do have a method that can calculate ICC, but this is not an ideal method (this is why I wrote to Josep), although I said we could start planning. I am optimistic.