fix(bug): correct exposure_expanded matrix calculation avoiding loops

[ntorresd]

This PR fixes the bug described in #145 . The right exposure triangular matrix with respect to the anti-diagonal is computed avoiding for loops by:

ly <- NCOL(foi_expanded)
exposure_expanded <- matrix(0, nrow = ly, ncol = ly)
exposure_expanded[apply(
  lower.tri(exposure_expanded, diag = TRUE),
  1, rev
  )] <- 1

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (044dbb8) 70.07% compared to head (b5b6002) 68.12%.

❗ Current head b5b6002 differs from pull request most recent head 22244db. Consider uploading reports for the commit 22244db to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #153      +/-   ##
==========================================
- Coverage   70.07%   68.12%   -1.95%     
==========================================
  Files          10       10              
  Lines        1751     1754       +3     
==========================================
- Hits         1227     1195      -32     
- Misses        524      559      +35     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[Bisaloo]

This issue was not detected by tests so can we add tests to ensure it doesn't pop up again?

[ntorresd]

This issue was not detected by tests so can we add tests to ensure it doesn't pop up again?

The reason why the tests did not detect this problem was because we were using incorrectly the function expect_equal(). I replaced it by a combination of expect_true and dplyr::near(), which allows for vector comparison with specific tolerance. I'm setting the tolerance as half the confidence interval size for each age as a heuristic threshold that I find reasonable. Please let me know if you can think of a better criteria. This one seems to perform well for the current models even without setting a seed and fails when using the bugged version of get_prev_expanded() for both time-varying models.

I decided to split the tests for the constant and the time-varying models into separate files in preparation for future tests that will be incorporated once we implement the additional time-varying models (#69 ) and age-vaying models (#74 ), which may be relevant for #85 .

[Bisaloo]

Thanks, this is very helpful!

For the tolerance specification, this feels more like a stats question so I'll let you double check with others if necessary.