Determine priors for time-varying hazards and hazard ratios

Computes consequences of priors chosen for the parameters hsd and hrsd in a flexible hazard model survextrap on an interpretable scale. This can be used to calibrate Gamma priors for these parameters to match interpretable beliefs.

Usage

prior_haz_sd(
  mspline,
  coefs_mean = NULL,
  prior_hsd = p_gamma(2, 1),
  prior_hscale = p_normal(0, 20),
  smooth_model = "exchangeable",
  prior_loghr = NULL,
  formula = NULL,
  cure = NULL,
  nonprop = NULL,
  newdata = NULL,
  prior_hrsd = NULL,
  tmin = 0,
  tmax = NULL,
  nsim = 1000,
  hq = c(0.1, 0.9),
  quantiles = c(0.025, 0.5, 0.975)
)

prior_hr_sd(
  mspline,
  coefs_mean = NULL,
  prior_hsd = p_gamma(2, 1),
  prior_hscale = p_normal(0, 20),
  smooth_model = "exchangeable",
  prior_loghr = NULL,
  formula = NULL,
  cure = NULL,
  nonprop = NULL,
  newdata = NULL,
  newdata0 = NULL,
  prior_hrsd = NULL,
  tmin = 0,
  tmax = 10,
  nsim = 100,
  hq = c(0.1, 0.9),
  quantiles = c(0.025, 0.5, 0.975)
)

Arguments

mspline

A list of control parameters defining the spline model.

knots: Spline knots. If this is not supplied, then the number of knots is taken from df, and their location is taken from equally-spaced quantiles of the observed event times in the individual-level data.

add_knots: This is intended to be used when there are external data included in the model. External data are typically outside the time period covered by the individual data. add_knots would then be chosen to span the time period covered by the external data, so that the hazard trajectory can vary over that time.

If there are external data, and both knots and add_knots are omitted, then a default set of knots is chosen to span both the individual and external data, by taking the quantiles of a vector defined by concatenating the individual-level event times with the start and stop times in the external data.

df: Degrees of freedom, i.e. the number of parameters (or basis terms) intended to result from choosing knots based on quantiles of the data. The total number of parameters will then be df plus the number of additional knots specified in add_knots. df defaults to 10. This does not necessarily overfit, because the function is smoothed through the prior.

degree: Polynomial degree used for the basis function. The default is 3, giving a cubic. This can only be changed from 3 if bsmooth is FALSE.

bsmooth: If TRUE (on by default) the spline is smoother at the highest knot, by defining the derivative and second derivative at this point to be zero.

coefs_mean

Spline basis coefficients that define the prior mean for the hazard function. By default, these are set to values that define a constant hazard function (see mspline_constant_coefs). They are normalised to sum to 1 internally (if they do not already).

prior_hsd

Gamma prior for the standard deviation that controls the variability over time (or smoothness) of the hazard function. This should be a call to p_gamma(). The default is p_gamma(2,1). See prior_haz_sd for a way to calibrate this to represent a meaningful belief.

prior_hscale

Prior for the baseline log hazard scale parameter (alpha or log(eta)). This should be a call to a prior constructor function, such as p_normal(0,1) or p_t(0,2,2). Supported prior distribution families are normal (parameters mean and SD) and t distributions (parameters location, scale and degrees of freedom). The default is a normal distribution with mean 0 and standard deviation 20.

Note that eta is not in itself a hazard, but it is proportional to the hazard (see the vignette for the full model specification).

"Baseline" is defined by the continuous covariates taking a value of zero and factor covariates taking their reference level. To use a different baseline, the data should be transformed appropriately beforehand, so that a value of zero has a different meaning. For continuous covariates, it helps for both computation and interpretation to define the value of zero to denote a typical value in the data, e.g. the mean.

smooth_model

The default "exchangeable" uses independent logistic priors on the multinomial-logit spline coefficients, conditionally on a common smoothing variance parameter.

The alternative, "random_walk", specifies a second-order random walk for the multinomial-logit spline coefficients, based on logistic distributions with a common smoothing variance. See the methods vignette for full details.

In non-proportional hazards models, setting smooth_model also determines whether an exchangeable or random walk model is used for the non-proportionality parameters (\(\delta\)).

prior_loghr

Priors for log hazard ratios. This should be a call to p_normal() or p_t(). A list of calls can also be provided, to give different priors to different coefficients, where the name of each list component matches the name of the coefficient, e.g. list("age45-59" = p_normal(0,1), "age60+" = p_t(0,2,3))

The default is p_normal(0,2.5) for all coefficients.

formula

A survival formula in standard R formula syntax, with a call to Surv() on the left hand side.

Covariates included on the right hand side of the formula with be modelled with proportional hazards, or if nonprop is TRUE then a non-proportional hazards is used.

If data is omitted, so that the model is being fitted to external aggregate data alone, without individual data, then the formula should not include a Surv() call. The left-hand side of the formula will then be empty, and the right hand side specifies the covariates as usual. For example, formula = ~1 if there are no covariates.

cure

If TRUE, a mixture cure model is used, where the "uncured" survival is defined by the M-spline model, and the cure probability is estimated.

nonprop

Non-proportional hazards model specification. This is achieved by modelling the spline basis coefficients in terms of the covariates. See the methods vignette for more details.

If TRUE, then all covariates are modelled with non-proportional hazards, using the same model formula as formula.

If this is a formula, then this is assumed to define a model for the dependence of the basis coefficients on the covariates.

IF this is NULL or FALSE (the default) then any covariates are modelled with proportional hazards.

newdata

A data frame with one row, containing variables in the model formulae. Samples will then be drawn, for any covariate-dependent parameters, with covariates set to the values given here.

prior_hrsd

Prior for the standard deviation parameters that smooth the non-proportionality effects over time in non-proportional hazards models. This should be a call to p_gamma() or a list of calls to p_gamma() with one component per covariate, as in prior_loghr. See prior_hr_sd for a way to calibrate this to represent a meaningful belief.

tmin

Minimum plotting time. Defaults to zero.

tmax

Maximum plotting time. Defaults to the highest knot.

nsim

Number of simulations to draw

hq

Quantiles which define the "low" and "high" values of a time-varying quantity (hazard in prior_haz_sd and the hazard ratio in prior_hr_sd). The ratio between the high and low values will be summarised, as a measure of time-dependence. By default, this is c(0.1, 0.9), so that the 10% and 90% quantiles are used respectively.

quantiles

Quantiles used to summarise the implied prior distributions of the simulated quantities.

newdata0

A data frame with one row, containing "reference" values of variables in the model formulae. The hazard ratio between the hazards at newdata and newdata0 will be returned.

Value

A data frame with columns sd_haz (SD of the hazard), sd_mean (SD of the inverse hazard) and hr (ratio between high/low hazards) (for prior_haz_sd), and rows giving prior quantiles of these.

In prior_hr_sd, sd_hr is the SD of hazard ratios over time, and hrr is the ratio between high/low hazard ratios.

Details

The spline model in survextrap allows the hazard to change over time in an arbitrarily flexible manner. The prior distributions on the parameters of this model have implications for how much we expect the hazard to plausibly vary over time. These priors are hard to interpret directly, but this function can be used to compute their implications on a more easily-understandable scale.

This is done by:

(1) simulating a set of parameters from their prior distributions

(2) computing the hazard at a fine grid of equally-spaced points spanning the boundary knots

(3) calculating the empirical standard deviation of the set of hazards at these points

(4) repeatedly performing steps 1-3, and summarising the distribution of the resulting standard deviations. This is the implied prior for the hazard variability.

prior_haz_sd computes the SD of the hazard, and the SD of the inverse hazard is also computed. The inverse hazard at time t is the expected time to the event given survival to t. The hazard ratio between a high and low value (defined by quantiles of values at different times) is also computed.

prior_hr_sd computes the SD of the hazard ratio between two covariate values supplied by the user.

All of these SDs refer to the variability over time, e.g. a SD of 0 indicates that the hazard (or inverse hazard, or hazard ratio) is constant with time.