changed name of Olival functions script

Author: Liam Shaw

scripts/Olival-functions.R

@@ -0,0 +1,296 @@
+# This code was originally written by Olival et al. (2017)
+# and was adapted (lightly) by Liam Shaw 2019 (liam.philip.shaw at gmail dot com)
+# for this project.
+
+# See: https://zenodo.org/record/807517 for the original code repository this code was sourced from
+# I am grateful to Olival et al. for making their code available under an MIT License.
+# https://opensource.org/licenses/MIT
+# Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+# The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+fit_all_gams <- function(data_set, outcome_variable, terms) {
+
+  fit_gam = function(frm) {
+    try(gam(formula=as.formula(frm), family="binomial", data_set, select=TRUE), silent=TRUE)
+  }
+
+  terms_grid = do.call(expand.grid, terms)
+
+  #Create model forumulas from the grid
+  formulas = apply(as.matrix(terms_grid), 1, function(row) paste(row, collapse = " + ")) %>%
+    stri_replace_all_regex("\\s[\\+\\s]+\\s", " + ") %>%
+    {paste(outcome_variable, "~", .)} %>%
+    rearrange_formula %>%
+    unique
+
+  models = data_frame(formula = formulas)
+
+  n_cores = detectCores()
+  n_cores_use = round(nrow(models) / ceiling(nrow(models) / (n_cores - 1)))
+  options(mc.cores = n_cores_use)
+  message("Using ", n_cores_use, " cores to fit ", nrow(models), " models")
+
+  models_vec = mclapply(models$formula, fit_gam)
+
+  models = models %>%
+    mutate(model = models_vec)
+  #print(models)
+
+  # Calculate models
+  models = models %>%
+    filter(map_lgl(model, ~ !("try-error" %in% class(.) | is.null(.)))) %>%
+    mutate(aic = map_dbl(model, MuMIn::AICc),
+           daic = aic - min(aic),
+           weight = exp(-daic/2)) %>%
+    arrange(aic)
+  #print("Calculated models")
+  #print(models)
+
+  # Remove unused terms from models and reduce to unique ones
+  models_reduced = models %>%
+    dplyr::select(model) %>%
+    mutate(formula = map_chr(model, ~rearrange_formula(rm_low_edf(.)))) %>%
+    distinct(formula, .keep_all = TRUE)
+  #print("Remove unused terms")
+  #print(models)
+
+  n_cores_use = round(nrow(models_reduced) / ceiling(nrow(models_reduced) / (n_cores - 1)))
+  options(mc.cores = n_cores_use)
+  message("Using ", n_cores_use, " cores to fit ", nrow(models_reduced), " reduced models")
+
+  # Reduce the remaining models
+  models_reduced = models_reduced %>%
+    mutate(model = mclapply(model, reduce_model))
+  #print("Reduced remaining models")
+  #print(models_reduced)
+
+  models_reduced = models_reduced %>%
+    filter(map_lgl(model, ~ !("try-error" %in% class(.) | is.null(.)))) %>%
+    mutate(aic = map_dbl(model, MuMIn::AICc)) %>%
+    mutate(formula = map_chr(model, ~rearrange_formula(rm_low_edf(.)))) %>%
+    distinct(formula, .keep_all=TRUE) %>%
+    arrange(aic) %>%
+    mutate(daic = aic - min(aic),
+           weight = exp(-daic/2),
+           terms = shortform(map(model, ~ rearrange_formula(.$formula))),
+           relweight = ifelse(daic > 2, 0, weight/sum(weight[daic < 2])),
+           relweight_all = weight/sum(weight),
+           cumweight = cumsum(relweight_all))
+  #print("Final reduction")
+  #print(models_reduced)
+
+  return(models_reduced)
+
+}
+
+fit_all_gams_poisson <- function(data_set, outcome_variable, terms) {
+  
+  fit_gam = function(frm) {
+    try(gam(formula=as.formula(frm), family="poisson", data_set, select=TRUE), silent=TRUE)
+  }
+  
+  terms_grid = do.call(expand.grid, terms)
+  
+  #Create model forumulas from the grid
+  formulas = apply(as.matrix(terms_grid), 1, function(row) paste(row, collapse = " + ")) %>%
+    stri_replace_all_regex("\\s[\\+\\s]+\\s", " + ") %>%
+    {paste(outcome_variable, "~", .)} %>%
+    rearrange_formula %>%
+    unique
+  
+  models = data_frame(formula = formulas)
+  
+  n_cores = detectCores()
+  n_cores_use = round(nrow(models) / ceiling(nrow(models) / (n_cores - 1)))
+  options(mc.cores = n_cores_use)
+  message("Using ", n_cores_use, " cores to fit ", nrow(models), " models")
+  
+  models_vec = mclapply(models$formula, fit_gam)
+  
+  models = models %>%
+    mutate(model = models_vec)
+  #print(models)
+  
+  # Calculate models
+  models = models %>%
+    filter(map_lgl(model, ~ !("try-error" %in% class(.) | is.null(.)))) %>%
+    mutate(aic = map_dbl(model, MuMIn::AICc),
+           daic = aic - min(aic),
+           weight = exp(-daic/2)) %>%
+    arrange(aic)
+  #print("Calculated models")
+  #print(models)
+  
+  # Remove unused terms from models and reduce to unique ones
+  models_reduced = models %>%
+    dplyr::select(model) %>%
+    mutate(formula = map_chr(model, ~rearrange_formula(rm_low_edf(.)))) %>%
+    distinct(formula, .keep_all = TRUE)
+  #print("Remove unused terms")
+  #print(models)
+  
+  n_cores_use = round(nrow(models_reduced) / ceiling(nrow(models_reduced) / (n_cores - 1)))
+  options(mc.cores = n_cores_use)
+  message("Using ", n_cores_use, " cores to fit ", nrow(models_reduced), " reduced models")
+  
+  # Reduce the remaining models
+  models_reduced = models_reduced %>%
+    mutate(model = mclapply(model, reduce_model))
+  #print("Reduced remaining models")
+  #print(models_reduced)
+  
+  models_reduced = models_reduced %>%
+    filter(map_lgl(model, ~ !("try-error" %in% class(.) | is.null(.)))) %>%
+    mutate(aic = map_dbl(model, MuMIn::AICc)) %>%
+    mutate(formula = map_chr(model, ~rearrange_formula(rm_low_edf(.)))) %>%
+    distinct(formula, .keep_all=TRUE) %>%
+    arrange(aic) %>%
+    mutate(daic = aic - min(aic),
+           weight = exp(-daic/2),
+           terms = shortform(map(model, ~ rearrange_formula(.$formula))),
+           relweight = ifelse(daic > 2, 0, weight/sum(weight[daic < 2])),
+           relweight_all = weight/sum(weight),
+           cumweight = cumsum(relweight_all))
+  #print("Final reduction")
+  #print(models_reduced)
+  
+  return(models_reduced)
+  
+}
+
+
+# Returns a model formula from a GAM with the low_edf terms removed
+rm_low_edf <- function(mod, edf_cutoff = 0.001) {
+  fr = as.character(formula(mod))
+  lhs = fr[2]
+  rhs = fr[3]
+  edfs = pen.edf(mod)
+  low_edfs = edfs[edfs < edf_cutoff]
+  vars_to_remove = stri_replace_all_fixed(unique(stri_extract_first_regex(names(low_edfs), "(?<=s\\()[^\\)]+(?=\\))")), ",",", ")
+  vars_regex = paste0("(", paste(vars_to_remove, collapse="|"), ")")
+  new_rhs = stri_replace_all_regex(rhs, paste0("\\s*s\\(", vars_regex, "\\,[^\\)]+\\)\\s*\\+?"), "")
+  new_rhs= stri_replace_all_fixed(new_rhs, "+, k = 7) ", "")
+  new_rhs= stri_replace_all_fixed(new_rhs, "+ +s", "+ s")
+  new_formula = paste(lhs, "~", new_rhs)
+  new_formula = stri_replace_all_regex(new_formula, "[\\s\\n]+", " ")
+  new_formula = stri_replace_all_regex(new_formula, "[+\\s]*$", "")
+  return(new_formula)
+}
+
+rm_terms <- function(mod, terms) {
+  fr = as.character(formula(mod))
+  lhs = fr[2]
+  rhs = fr[3]
+  vars_regex = paste0("(", paste(terms, collapse="|"), ")")
+  new_rhs = stri_replace_all_regex(rhs, paste0("\\s*s\\(", vars_regex, "\\,[^\\)]+\\)\\s*\\+?"), "")
+  new_rhs= stri_replace_all_fixed(new_rhs, "+, k = 7) ", "")
+  new_rhs= stri_replace_all_fixed(new_rhs, "+ +s", "+ s")
+  new_formula = paste(lhs, "~", new_rhs)
+  new_formula = stri_replace_all_regex(new_formula, "[\\s\\n]+", " ")
+  new_formula = stri_replace_all_regex(new_formula, "[+\\s]*$", "")
+  return(new_formula)
+}
+
+#' Alphabetizes the right-hand side of a formula so as to compare formulas across models
+rearrange_formula = function(formula) {
+  if(class(formula) == "formula") {
+    formula = as.character(formula)
+    formula = paste(formula[2], "~", formula[3], collapse=" ")
+  }
+  lhs = stri_extract_first_regex(formula, "^[^\\s~]+")
+  rhs = stri_replace_first_regex(formula, "[^~]+~\\s+", "")
+  terms = stri_split_regex(rhs, "[\\s\\n]+\\+[\\s\\n]+")
+  terms = lapply(terms, sort)
+  new_formula = mapply(function(lhs, terms) {paste(lhs, "~", paste(terms, collapse = " + "))}, lhs, terms)
+  new_formula = stri_replace_all_regex(new_formula, "[\\s\\n]+", " ")
+  new_formula = stri_replace_all_fixed(new_formula, "+ +s", "+ s")
+  new_formula = stri_replace_all_fixed(new_formula, "+ +", "+")
+  names(new_formula) <- NULL
+  return(stri_trim(new_formula))
+}
+
+# Re-fits a gam model, dropping terms that have been selected out
+reduce_model <- function(mod, edf_cutoff = 0.001, recursive=TRUE) {
+  low_edf_vars = any(pen.edf(mod) < edf_cutoff)
+  if(recursive) {
+    while(low_edf_vars) {
+      mod = update(mod, formula = as.formula(rm_low_edf(mod, edf_cutoff)))
+      low_edf_vars = any(pen.edf(mod) < edf_cutoff)
+    }
+  } else {
+    if(low_edf_vars) {
+      mod = update(mod, formula = as.formula(rm_low_edf(mod, edf_cutoff)))
+    }
+  }
+  return(mod)
+}
+
+# Makes a reduced version of the RHS of a formula
+shortform = function(formula) {
+  rhs = stri_replace_first_regex(formula, "[^~]+~\\s+", "")
+  rhs = stri_replace_all_regex(rhs, "s\\(([^\\,]+)\\,[^\\)]+\\)", "s($1)")
+  rhs= stri_replace_all_fixed(rhs, "+ +s", "+ s")
+  stri_replace_all_fixed(rhs, "(1 + | + 1)", "")
+}
+
+check_vals <- function(x) {
+  w = which(is.nan(x) | is.na(x) | is.infinite(x))
+  z = x[w]
+  names(z) = w
+  return(z)
+}
+
+
+get_relative_contribs <- function(gam_model) {
+
+  terms <- attributes(gam_model$terms)
+  offset <- attributes(gam_model$terms)$offset
+
+  formula_chr <- as.character(formula(gam_model))
+  model_data <- gam_model$model
+  offset_name = stri_replace_first_regex(names(model_data)[terms$offset], "offset\\(([^\\)]+)\\)", "$1")
+  names(model_data) <- stri_replace_all_regex(names(model_data), "offset\\(([^\\)]+)\\)", "$1")
+  gam_model = gam(formula(gam_model), data=model_data, family=gam_model$family, select=FALSE)
+
+  y = model_data[,terms$response]
+  preds = predict(gam_model, type="iterms")
+  intercept = attributes(preds)$constant
+  smooth_pars  <- gam_model$sp
+  rhs <- formula_chr[3]
+  lhs <- formula_chr[2]
+  devs <- map_dbl(terms$term.labels, function(term) {
+    # sub_preds <- rowSums(preds[, !stri_detect_fixed(colnames(preds), term), drop=FALSE]) + intercept
+    # if(length(offset_name) ==1 ) sub_preds = sub_preds + model_data[, offset]
+    # sub_preds <- gam_model$family$linkinv(sub_preds)
+    # res <- gam_model$family$dev.resids(y, sub_preds, 1)
+    # s <- attr(res, "sign")
+    # if (is.null(s))
+    #   s <- sign(y - sub_preds)
+    # res <- sqrt(pmax(res, 0, na.rm=TRUE)) * s
+    # dev <- sum(res^2)
+    term = c(terms$term.labels[terms$term.labels != term], offset_name)
+    term_regex= paste0("(", paste(term, collapse="|"), ")")
+    new_formula <- stri_extract_all_regex(rhs, paste0("(s|offset)\\(", term_regex, "[^\\)]*\\)"))[[1]] %>%
+      paste(collapse = " + ") %>%
+      {paste(lhs, "~", .)} %>%
+      as.formula()
+    smooth_par <- smooth_pars[stri_detect_regex(names(smooth_pars), term_regex)]
+    smooth_par <- smooth_par[!stri_detect_regex(names(smooth_par), "\\)[23456789]+$")]
+    new_model <- gam(formula = new_formula, sp=smooth_par, data=model_data, family=gam_model$family, select=FALSE)
+    dev = deviance(new_model)
+    return(dev)
+  })
+  null_formula = paste(lhs, "~ 1")
+  if(length(offset_name) ==1 ) null_formula = paste0(null_formula, "+ offset(", offset_name, ")")
+  null_model = gam(formula = as.formula(null_formula), data=model_data, family=gam_model$family, select=FALSE)
+  null_model_dev = deviance(null_model)
+  orig_dev<- deviance(gam_model)
+  dev_expl = (devs - orig_dev)
+  dev_expl = dev_expl/sum(dev_expl)
+  data_frame(term = terms$term.labels, rel_deviance_explained = dev_expl)
+
+}
+
+
+