predict_delta_comps function

Get predictions from compositional ilr multiple linear regression model

Provided the data (containing outcome, compositional components and covariates), fit a ilr multiple linear regression model and provide predictions from reallocating compositional values pairwise amunsnst the components model.

predict_delta_comps(
  dataf,
  y,
  comps,
  covars = NULL,
  deltas = c(0, 10, 20)/(24 * 60),
  comparisons = c("prop-realloc", "one-v-one")[1],
  alpha = 0.05
)

Arguments

• dataf: A data.frame containing data
• y: Name (as string/character vector of length 1) of outcome variable in dataf
• comps: Character vector of names of compositions in dataf. See details for more information.
• covars: Optional. Character vector of covariates names (non-comp variables) in dataf. Defaults to NULL.
• deltas: A vector of time-component changes (as proportions of compositions , i.e., values between -1 and 1). Optional. Changes in compositions to be computed pairwise. Defaults to 0, 10 and 20 minutes as a proportion of the 1440 minutes in a day (i.e., approximately 0.000, 0.007 and 0.014).
• comparisons: Currently two choices: "one-v-one" or "prop-realloc" (default). Please see details for explanation of these methods.
• alpha: Optional. Level of significance. Defaults to 0.05.

Returns

Messages are printed to the console as the function tests the inputs, produces the isometric log ratios (ilrs), fits the linear model and produces the redistributed time-use predictions (with confidence intervals).

Returns a data.frame of the time-use redistribution predictions (and 95% confidence intervals) with the following columns:

• comp+: the compositional variable with the addition of the delta value
• comp-: the compositional variable with the subtraction of the delta value
• delta: the time-use redistribution value
• alpha: significance level for the 100(1-alpha)% confidence interval
• delta_pred: the predicted mean change in the outcome variable
• ci_lo: the lower limit of 100(1-alpha)% confidence interval corresponding to delta_pred
• ci_up: the upper limit of 100(1-alpha)% confidence interval corresponding to delta_pred
• sig: "*" if the delta_pred is significantly different from 0 at the alpha level (empty string otherwise)

The data.frame has a class of deltacomp_obj which denotes there are additional attributes of the returned object accessible using attr(*, "attribute_name").

The possible values for "attribute_name" are:

• dataf: a data.frame of the predictors (covariates and ilrs)
• y: a vector of the outcome variable
• comps: a character vector of the time-use composition names
• lm: the lm object of the multiple linear regression fit (using y and dataf from above)
• deltas: the redistributed time-use values used in the predictions
• comparisons: "one-v-one" or "prop-realloc" provided as the comparisons argument
• alpha: significance level for the 100(1-alpha)% confidence intervals
• ilr_basis: the ilr change of basis matrix V
• mean_pred: a single row data.frame with the predicted mean outcome (fit column) value from the "average" set of predictors

Details

Values in the comps columns must be strictly greater than zero. These compositional values are NOT assumed to be constrained to (0, 1) values as the function normalises the compositions row-wise to sum to 1 in part of it's processing of the dataset before analysis.

Please see the deltacomp package README.md file for examples and explanation of the comparisons = "prop-realloc" and comparisons = "one-v-one" options.

Examples

predict_delta_comps(
  dataf = fat_data,
  y = "fat",
  comps = c("sl", "sb", "lpa", "mvpa"),
  covars = c("sibs", "parents", "ed"),
  deltas = seq(-60, 60, by = 5) / (24 * 60),
  comparisons = "one-v-one",
  alpha = 0.05
)

delta_comp_out <- predict_delta_comps(
  dataf = fat_data,
  y = "fat",
  comps = c("sl", "sb", "lpa", "mvpa"),
  covars = NULL,
  deltas = seq(-60, 60, by = 5) / (24 * 60),
  comparisons = "prop-realloc",
  alpha = 0.05
)

# get the mean prediction from the returned object
attr(delta_comp_out, "mean_pred")

Author(s)

Ty Stanford tystan@gmail.com