Changelog • deltapif

deltapif 0.5.0

Shared parameters are declared explicitly. p and beta in pif() and paf() can be given as a named list with one element, e.g. beta = list("smoking_dementia" = 0.46) or p = list("447" = c(0.12, 0.87)). Fractions using the same name (identifier) share the estimate and are correlated; using one identifier with different values or variances is an error. Parameters given as plain numbers are now treated as independent. Previously, fractions whose beta and var_beta (or p and var_p) were numerically identical were silently assumed to share the estimate, which missed shared estimates that were rounded differently and invented sharing between independent estimates that happened to coincide. Code that relied on this should give the shared parameters a common identifier.
weighted_adjusted_paf() and weighted_adjusted_pif() build their confidence intervals on the log scale by default (new argument link), consistent with the log-scale variance of the adjusted fractions; use link = "identity" for the previous symmetric intervals.
The dementiacov dataset is replaced by dementiacor. dementiacov was built from the pairwise tetrachoric correlations between the dementia risk factors (exposures) in ARIC (eTable 2 of Lee et al.) multiplied by the standard errors of the log relative risks, so it was not a covariance between relative-risk estimates. dementiacor contains the published correlations themselves, documented as correlations between exposures.

pif() accepts the counterfactual prevalence p_cft as a function of p (e.g. function(p) 0.85 * p) for scenarios defined relative to the current exposure. The uncertainty in p is then propagated through the counterfactual (numerically or with the new argument p_cft_deriv). Numeric p_cft values are still treated as fixed targets.
pif_total(), pif_ensemble(), their paf versions and the weighted adjusted fractions accept var_p and var_beta: covariance matrices between different estimates whose row and column names are the identifiers of the parameters (e.g. relative risks from the same meta-analysis).
New variance_components() splits the variance of a fraction into the parts due to the prevalences (p), the relative risks (beta) and the weights.
Compatibility with the development version of S7 (> 0.2.2): constructors of the fraction classes build their parent instances, and the package uses S7’s load, unload and build hooks when available (@t-kalinowski, #2).

pif_total(), paf_total(), pif_ensemble(), paf_ensemble() (and hence weighted_adjusted_pif() and weighted_adjusted_paf()) used the inverse of the link instead of its derivative, producing confidence intervals that were too narrow. Point estimates and variances were not affected.
weighted_adjusted_pif() and weighted_adjusted_paf() now pass pif_total_link and pif_ensemble_link (and their inverses and derivatives) instead of ignoring them.

dementiarisk gains a communality column (%, Table 2 of Lee et al.). Its documentation now states that the prevalence columns are percentages and that sdlog is the standard error of the log relative risk.

The vignettes used the ARIC exposure prevalences of smoking, excessive alcohol and physical inactivity (5.9%, 3.8% and 31.1%) as if they were communalities; they now use the published communalities from dementiarisk$communality, with weights 1 - communality in the same order as the fractions.
Fixes the variance of the number of cases derived from its confidence interval and the alcohol example, which passed relative risks instead of log relative risks.
The README and the Introduction vignette now state that Lee et al. obtained their intervals by substituting the confidence limits of the relative risk, and the vignettes show relative counterfactuals, variance_components() and shared identifiers.

Fixes bug in printing standard deviation of attributable_cases() and averted_cases() that multiplied by 100.