Balancing a panel • panel990

panel_balance() trims a panel to a balanced rectangle – a block of years in which every retained organization is observed in every retained year. It is the complement of panel_complete(): completion fills gaps to reach a full rectangle, balancing drops organizations and years to reach one.

panel_balance() trimming a mixed panel to its largest fully observed block of organizations and years

df <- data.frame(
  EIN2     = c(rep("A", 4), rep("B", 4), rep("C", 3), rep("D", 2)),
  TAX_YEAR = c(2019:2022,   2019:2022,   2019:2021,   c(2019, 2020))
)

A and B are observed all four years; C misses 2022; D only appears in 2019–2020.

The year window is the trade-off

A wider window keeps more years but fewer organizations; a narrower one keeps more organizations. By default panel_balance() uses the full observed range – the strictest balance:

full <- panel_balance(df)
unique(full$EIN2)          # only orgs present in all of 2019-2022
#> [1] "A" "B"
attr(full, "balance")
#> $years
#> [1] 2019 2020 2021 2022
#> 
#> $n_years
#> [1] 4
#> 
#> $orgs_kept
#> [1] 2
#> 
#> $orgs_dropped
#> [1] 2

Restrict the window and more organizations qualify:

win <- panel_balance(df, years = 2019:2021)
unique(win$EIN2)           # C now qualifies; D still does not
#> [1] "A" "B" "C"

Let it find the largest block

If you would rather have the package pick the window that retains the most data (organizations x years), use strategy = "max_rectangle":

best <- panel_balance(df, strategy = "max_rectangle")
attr(best, "balance")[c("years", "orgs_kept")]
#> $years
#> [1] 2019 2020 2021
#> 
#> $orgs_kept
#> [1] 3

Here the 3-year x 3-organization block (2019–2021, A/B/C = 9 observations) beats the full 4-year x 2-organization block (8), so that is what it returns. min_years sets a floor on how short a balanced window may be.

Notes

The "balance" attribute always records the retained years and the keep/drop counts. If no organization spans the window, the result is simply empty (zero rows) rather than an error – a useful signal that your window is too demanding. And because balancing only selects rows, run deduplicate() first if an organization can have more than one filing per year, so each cell of the rectangle is a single record.