Compute a correlation or covariance matrix from an IRW response table
Source:R/cov.R
irw_cov_compute.RdTurns a core table (the long id, item, resp layout from irw_fetch())
into an item-by-item matrix in the layout of the IRW covariance family: one
row per cell of the lower triangle, diagonal included, with columns stat
("cor" or "cov"), var1, var2, value and n (the number of people
who answered both items). irw_fetch() with source = "cov" calls this on a
core table, and irw_cov2matrix() turns the result into a square matrix.
Arguments
- df
A core IRW table in long format, with columns
id,itemand a numericresp.- method
The association measure:
"pearson"(default),"spearman","kendall","polychoric"or"tetrachoric". The last two need the psych package and give correlations only;"tetrachoric"needs every item dichotomous. Or a function of your own: it receives the wide person-by-item matrix (one column per item,NAwhere an item was not answered, afteruseis applied) and must return a square symmetric matrix with one row and column per item, in the same order. Say what it returns withtype;nis still the pairwise count.- type
"cor"(default) for a correlation matrix or"cov"for a covariance matrix."cov"is available for"pearson","spearman"and"kendall".- use
How missing responses are handled:
"pairwise"(default) uses, for each pair of items, everyone who answered both, and records that count inn;"complete"first drops anyone with a missing response, sonis the same everywhere. For polychoric and tetrachoric correlations,"pairwise"is what psych does internally.- max_items
Refuse tables with more items than this (default 500). An item bank of 27,000 items has 380 million pairs, almost all answered by nobody in common.
- table
Optional table name, recorded in the
irw_tableattribute.
Value
A tibble with columns stat, var1, var2, value and n. A
pair answered by fewer than three people in common has value NA.
Details
The result is computed, not published: its attributes irw_computed,
irw_method, irw_type, irw_use and irw_table say how it was made, and
irw_cov2matrix() carries them forward. Cite the source table, not the
matrix.
Tables with repeated measures (more than one response per
id and item, for example across waves) are refused: filter to one
occasion first, e.g. df[df$wave == 1, ]. Items with no variation have no
correlation and are dropped with a warning.
Examples
df <- data.frame(id = rep(1:50, each = 3), item = rep(c("a", "b", "c"), 50),
resp = c(replicate(50, sample(1:5, 3, replace = TRUE))))
irw_cov_compute(df)
#> # A tibble: 6 × 5
#> stat var1 var2 value n
#> <chr> <chr> <chr> <dbl> <int>
#> 1 cor a a 1 50
#> 2 cor b a 0.0629 50
#> 3 cor b b 1 50
#> 4 cor c a 0.261 50
#> 5 cor c b 0.0343 50
#> 6 cor c c 1 50
irw_cov_compute(df, method = "spearman", type = "cov")
#> # A tibble: 6 × 5
#> stat var1 var2 value n
#> <chr> <chr> <chr> <dbl> <int>
#> 1 cov a a 202. 50
#> 2 cov b a 12.8 50
#> 3 cov b b 200. 50
#> 4 cov c a 56.3 50
#> 5 cov c b 14.1 50
#> 6 cov c c 201. 50
# Your own measure: here, a correlation shrunk towards zero
shrunk <- function(x) { r <- stats::cor(x, use = "pairwise.complete.obs"); 0.9 * r + 0.1 * diag(ncol(x)) }
irw_cov_compute(df, method = shrunk)
#> # A tibble: 6 × 5
#> stat var1 var2 value n
#> <chr> <chr> <chr> <dbl> <int>
#> 1 cor a a 1 50
#> 2 cor b a 0.0566 50
#> 3 cor b b 1 50
#> 4 cor c a 0.235 50
#> 5 cor c b 0.0309 50
#> 6 cor c c 1 50