Connection across places: does one indicator rise where another does?

For two indicators published for the same places (regions, provinces, Inuit regions), their rank correlation across places, an exact permutation p-value, and an interval that carries the sampling error of the published estimates: for connection to land and wellness, community belonging and mental health, education and health.

Usage

connassoc(
  data,
  place,
  x,
  y,
  xlo = NULL,
  xup = NULL,
  ylo = NULL,
  yup = NULL,
  reps = 10000,
  seed = 1,
  conflvl = 0.95
)

Arguments

data: A data frame, one row per place.
place: Column of place names.
x, y: Columns of the two indicators.
xlo, xup, ylo, yup: Optional columns of their published limits (for the interval).
reps: Draws for the interval and, past 8 places, for the p-value (default 10,000).
seed: Seed for the draws (default 1).
conflvl: Confidence level of rho's drawn interval (default 0.95).

Value

A list of class rmbl_connection: data (the places used), rho, p_value, exact, rholo, rhoup, n.

Details

Mashford-Pringle and Stewart (2019) describe the land as “an instrumental part of all aspects of Indigenous life including health and well-being”; Mashford-Pringle (2016) describes Indigenous health and education as interconnected. Across places, that connection has a testable shadow: where more people are on the land (or belong, or finish school), do more report wellness? This function measures it by Spearman's \(\rho\). Its p-value is exact when there are at most 8 places (over all \(n!\) orderings) and a Monte Carlo estimate otherwise (\((b + 1)/(R + 1)\) for \(b\) of \(R\) = reps random orderings at least as extreme). Each estimate carries sampling error, so rholo and rhoup are the quantiles of \(\rho\) over reps draws of both indicators from normal distributions with their published standard errors. That is a sensitivity band, not a confidence interval for a correlation across places: it shows how much the published sampling error alone moves \(\rho\) for these places, it is pulled toward 0 (noise is added to estimates that are already noisy), and it ignores the sampling of the places themselves. Quality flags are not carried: filter suppressed rows out first.

This is an association across places (an ecological one). It cannot say that connection to land makes a person well, which needs data on persons; a handful of places gives little power; and a correlation is not a measure of the relationship the authors describe. It is a description of how the published indicators move together.

Acknowledgement

This estimator puts to quantitative use the work of Professor Angela Mashford-Pringle (Timiskaming First Nation) and Professor Suzanne Stewart (Yellowknives Dene First Nation), both of the University of Toronto.

See the acknowledgements of the first-nations-metis-inuit-health vignette.

References

Mashford-Pringle A, Stewart SL (2019). Akiikaa (it is the land): exploring land-based experiences with university students in Ontario. Global Health Promotion 26(3_suppl):64-72. doi:10.1177/1757975919828722

Mashford-Pringle A (2016). Aboriginal health and Aboriginal education are interconnected. Research in the Sociology of Health Care 34:85-96. doi:10.1108/s0275-495920160000034005

Examples

# across regions, 2022 (table 41-10-0067-01): where First Nations people
# report excellent or very good mental health, do Metis people living in the
# same region too?
mh <- indiex("mentalhealth")
mh <- mh[mh$rating == "Excellent or very good" & mh$region != "Canada", ]
fn <- mh[mh$group == "First Nations", ]
me <- mh[mh$group == "M\u00e9tis", ]
d <- merge(fn, me, by = "region", suffixes = c("fn", "me"))
ca <- connassoc(
  d,
  "region",
  "estimatefn",
  "estimateme",
  "lowerfn",
  "upperfn",
  "lowerme",
  "upperme"
)
ca
#> ╭── Connection across places ──────────────────────────────────────────────╮
#> │   Places            11                                                   │
#> │   Indicators        estimatefn and estimateme                            │
#> │   Rank correlation  rho = 0.52                                           │
#> │   p-value           0.106 (permutation)                                  │
#> │   Interval          95% [-0.19, 0.77], with the published sampling error │
#> │                                                                          │
#> │   Rank correlation across 11 places: rho = 0.52, permutation p = 0.106   │
#> │   Across these places the two rank moderately together.                  │
#> │                                                                          │
#> │   An association across places only:                                     │
#> │   it does not show that one causes the other for any person.             │
#> ╰──────────────────────────────────────────────────────────────────────────╯
plot(ca)


# your own places: made-up shares in six regions, without limits
own <- data.frame(
  region = c("North", "South", "East", "West", "Coast", "Interior"),
  language = c(41, 18, 27, 33, 22, 38),
  wellbeing = c(68, 55, 60, 66, 57, 63)
)
connassoc(own, "region", "language", "wellbeing", reps = 2000)
#> ╭── Connection across places ──────────────────────────────────────╮
#> │   Places            6                                            │
#> │   Indicators        language and wellbeing                       │
#> │   Rank correlation  rho = 0.94                                   │
#> │   p-value           0.017 (exact, every permutation)             │
#> │                                                                  │
#> │   Rank correlation across 6 places: rho = 0.94, exact p = 0.0167 │
#> │   Across these places the two rank strongly together.            │
#> │                                                                  │
#> │   An association across places only:                             │
#> │   it does not show that one causes the other for any person.     │
#> ╰──────────────────────────────────────────────────────────────────╯