Skip to contents

Changelog

Source: NEWS.md

MVMR 0.4.9

  • strength_mvmr() now divides the conditional Q-statistic by L - (K - 1), where L is the number of variants and K the number of exposures, as in Equation 7 of Sanderson, Spiller and Bowden (2021, https://doi.org/10.1002/sim.9133). It previously divided by L, which slightly understated the conditional F-statistics.
  • strhet_mvmr() has been removed. It was not described in the accompanying papers or the vignette. Use strength_mvmr() to assess conditional instrument strength, and qhet_mvmr() for effect estimates that are robust to weak instruments.
  • pleiotropy_mvmr() now computes its p-value on L - K degrees of freedom, as in Section 3.2 of Sanderson, Spiller and Bowden (2021). It previously used L - K - 1, which gave p-values that were too small. Its documentation now explains that the Q-statistic is evaluated at the IVW estimate and is not robust to weak instruments, so it should only be interpreted when strength_mvmr() indicates the instruments are strong.
  • qhet_mvmr() now chooses the heterogeneity parameter tau-squared so that the minimised Q-statistic equals L - K, as in Sanderson, Spiller and Bowden (2021). It previously targeted L - 2, which is only correct with two exposures. Tau-squared is now constrained to be non-negative (it was searched over -10 to 10, which could give negative weights and is poorly scaled for summary data), and is found by root finding with the Q-statistic minimised by BFGS from the IVW estimate. Effect estimates will differ from previous versions, including with two exposures.
  • qhet_mvmr() gains a CI_method argument. CI_method = "jackknife" calculates leave-one-variant-out jackknife standard errors and normal-based 95% confidence intervals, as recommended by Sanderson, Spiller and Bowden (2021); the default "bootstrap" retains the BCa bootstrap intervals. The documentation no longer refers to a non-existent se argument.
  • pleiotropy_mvmr() gains an estimator argument. The default, estimator = "ivw", evaluates the Q-statistic at the IVW estimate as before; estimator = "qmin" evaluates it at the causal effect estimates that minimise it, which are robust to weak instruments, as discussed in Section 3.2 of Sanderson, Spiller and Bowden (2021).
  • The qhet_mvmr() documentation now explains how pcor is used, and that the identity matrix should be supplied when the exposure GWAS were estimated in non-overlapping samples.

MVMR 0.4.8

  • strhet_mvmr() has been reimplemented. The previous version built a combinatorial grid of candidate coefficients with utils::combn() (which could exhaust memory with four or more exposures) and never actually minimised the Q-statistic, returning invalid conditional F-statistics. It now estimates the coefficients by iteratively reweighted least squares, giving distinct, well-identified conditional F-statistics for each exposure. Reported values will differ from previous versions.

MVMR 0.4.7

  • qhet_mvmr() now uses the estimated heterogeneity parameter (the minimiser) rather than the minimised objective value when constructing the model weights. This corrects the effect estimates, which will differ from previous versions.
  • snpcov_mvmr() now returns correct covariance matrices when the genetic instrument data are supplied as a matrix rather than a data frame.
  • strength_mvmr() and pleiotropy_mvmr() now select the covariance-matrix calculation based on whether gencov is a list, fixing a division-by-zero that occurred when a gencov list for exactly two variants was supplied.
  • ivw_mvmr() no longer emits a spurious warning about fixing the covariance at zero; the gencov argument does not affect the IVW estimates and is retained only for interface consistency.
  • mvmr() no longer fits the same weighted regression twice.
  • Tidied format_mvmr() internal column naming.

MVMR 0.4.6

  • Add vignette on estimating phenotypic correlations.

MVMR 0.4.5

  • Bump roxygen2 to 8.0.0 and add package level helpfile.

MVMR 0.4.4

  • Implement some optimizations.

MVMR 0.4.3

  • The snpcov_mvmr() function accidentally omitted intercepts from its regressions of exposure on genotype. This has been fixed.

MVMR 0.4.2

  • The qhet_mvmr() function with argument CI = TRUE is now faster as it can use multiple processor cores (except on Windows) and it calculates the bootstrap confidence interval limits more efficiently as an unnecessarily repeated function call has been removed (thanks @nickhir).

MVMR 0.4.1