NEWS

---

sffdr 1.1.4

• The density window in sffdr() now spans the documented ~5,000 neighbours. nn was taken as a fraction of all SNPs but applied to the rows the density is fitted on (the tail plus 100,000 downsampled null SNPs), which left about 600 neighbours at one million SNPs. Results at up to ~100,000 SNPs are unchanged; above that, calls at fq <= 0.05 move by up to 2% (fewer at one million SNPs) at the same ranking and the same type I error with independent SNPs.
• Perfect-LD copies (SNPs with identical p-values) now contribute one point each to the density fit, with the density evaluated for every SNP. With many copies the narrow window inflated functional p-values below 1e-3 by up to 2.4-fold at one million SNPs; it is now within the level of the same number of independent SNPs. Ties from rounded p-values, which sit almost only in the bulk of the distribution, are recognised by that and left alone.
• fpvalues() gives tests with identical lfdr and p the same values (the set's), so the result no longer depends on the input order; negative lfdr is an error and all-zero lfdr gives functional p-values of 0 instead of NaN.
• kernelEstimator() no longer crashes R on tied or near-tied values in the univariate case (tied values are collapsed, the fit uses at most 20 iterations and at least 10 distinct neighbours). Its maxk default is 5,000 instead of 500,000 (35 MB instead of 2.7 GB of evaluation space), and the fit is repeated with more space if locfit runs out; the fit is unchanged. ... is passed to locfit(), as it always was, and the arguments set by the estimator cannot be passed through it. The error after a failed cascade reports locfit's last message.
• pi0_model() and fpi0est() accept indep_snps as a logical vector, row indices or SNP ids (row names). Row indices were summed as if logical in pi0_model() (far too many SNPs per knot) and were recycled in fpi0est().
• pi0_model() stops on non-numeric input (an rsID column was ranked as text and modelled) and warns when a column is not a p-value column or qvalue() fails, instead of silently using the fallback knot.
• fpi0est() checks the lengths of p, weights and indep_snps, requires lambda in (0, 1), warns when the chosen model did not converge and when weights above 1 are clamped (LD scores passed in place of inverse LD scores gave an unweighted fit). Tests with a missing p-value or covariate get fpi0 = 1 and the floor of 0.01 are documented. tableLambda no longer holds the 19 fitted models (~1 GB per million SNPs); the chosen model is returned as model.
• sffdr() stops on negative p-values (a log10 p-value column became the strongest discovery) and clamps exact zeros only.
• plot.sffdr() showed the wrong SNPs' fpi0 in panel (a) when some p-values were missing; also removed a duplicated scale and a deprecated legend position.
• The C++ smoothers check their input lengths and ordering.
• Documentation: monotone smoothing has been the default since 1.1.2; the bivariate kernelEstimator() takes the covariate first and the p-values second (the example and tests had them swapped).
• tests/testthat.R added, so R CMD check runs the test suite (it did not before).

sffdr 1.1.3

• decorrelate_informative() now removes the primary's noise from the informative studies exactly, (z_y - rho z_x) / sqrt(1 - rho^2), instead of scaling the correction by the primary's local FDR, which left part of the overlap in the covariates and inflated false discoveries at larger overlap. It takes several informative studies at once and, without rho_o, estimates the overlap with the new estimate_overlap(). p_x, p_y and lfdr_x are no longer used.
• New estimate_overlap(): the null correlation of z-scores between studies from their products near zero.
• overlap_null_density(), run_empirical_overlap_correction(), estimate_rho_overlap() and discover_empirical_nulls() are defunct.
• sffdr() no longer uses weights (a warning is given). Its density is estimated over all tested SNPs, the set the functional p-values count errors in; reweighting it by inverse LD scores biased the local FDR, because associated SNPs have larger LD scores (conservative with inverse-LD weights, anti-conservative with weights larger at associated SNPs). LD scores belong in the functional pi0 model as a covariate; inverse-LD weights remain available in pi0_model() and fpi0est() in place of LD pruning. Results without weights are unchanged.

sffdr 1.1.2 (2026-03-31)

• Added monotonicity option (running min or PAVA)
• Better handling of weights in kernel density estimation
• function docorrelate_informative to handle overlapping samples

sffdr 1.1.1

• Added a weights argument which allows users to input inverse LD scores and fit to all SNPs.

sffdr 1.1.0 (2026-02-11)

Major Changes

• Now uses fastglm exclusively for significantly faster GLM fitting.

• Simplified pi0_model() function:

  • Removed knots parameter. Knots are now selected adaptively based on FDR thresholds and signal strength.
  • Added n_knots parameter (default: 5) for target knot count.
  • Added min_snps_per_knot parameter (default: 100) to ensure stable spline fitting.
  • Added max_knots parameter (default: 5) to prevent overfitting.

• Removed indep_snps functionality from sffdr(). Kernel density estimation now always uses all SNPs with appropriate nn smoothing to handle LD.

• Improved speed of locfit by subsampling the null region of the distribution (accounted for using the weights argument).

• Added 98 unit tests covering kernelEstimator, pi0_model, fpi0est, sffdr, and fpvalues.

sffdr 1.0.0 (2024-12-02)

• Initial CRAN submission.