---
title: "Post-hoc credible-set filtering without an LD reference"
author: "Alex McCreight"
date: "`r Sys.Date()`"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Post-hoc credible-set filtering without an LD reference}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(collapse = TRUE, comment = "#", fig.width = 5,
                      fig.height = 3, fig.align = "center", dpi = 120)
```

`susie_get_cs_attainable()` is a post-hoc credible-set filter for the
case when neither a genotype matrix nor an LD reference is available,
so the usual `min_abs_corr` purity filter in `susie_get_cs()` cannot
be used. It uses only the fitted `alpha` matrix. For each effect, it
computes an *attainable coverage*, an idea introduced by
[SparsePro (Zhang et al. 2023)](https://doi.org/10.1371/journal.pgen.1011104),
by zeroing every entry of `alpha` that is not the column maximum and
then summing the rows, which gives the coverage that effect could
attain if it did not have to share probability mass with other
effects. A second filter caps the entropy
of each effect's attainable mass, dropping effects whose mass is
spread too thinly across many variables.

## Example

```{r}
library(susieR)
set.seed(1)
data(N3finemapping)
X <- N3finemapping$X
y <- N3finemapping$Y[, 1]
```

Fit SuSiE with a fixed prior so noise effects are not shrunken away:

```{r}
fitted <- susie(X, y, L = 20,
                estimate_prior_variance = FALSE,
                scaled_prior_variance = 0.05,
                verbose = FALSE)
```

With the genotype matrix available, `susie_get_cs()` filters by
purity:

```{r}
sets_purity <- susie_get_cs(fitted, X = X, coverage = 0.95)
length(sets_purity$cs)
```

Without `X` or `Xcorr`, several poorly localized effects slip through:

```{r}
sets_noLD <- suppressMessages(susie_get_cs(fitted, coverage = 0.95))
length(sets_noLD$cs)
sapply(sets_noLD$cs, length)
```

`susie_get_cs_attainable()` recovers the purity-filtered result
without using `X`:

```{r}
sets_attain <- susie_get_cs_attainable(fitted, coverage = 0.95)
length(sets_attain$cs)
```

The two filters select the same effects:

```{r}
sets_purity$cs
sets_attain$cs
```

## References

Zhang, W., Najafabadi, H., Li, Y. (2023). SparsePro: An efficient
fine-mapping method integrating summary statistics and functional
annotations. *PLOS Genetics* **19**, e1011104.
doi:10.1371/journal.pgen.1011104.

## Session information

```{r}
sessionInfo()
```