Draft pool_means() and pool_comparisons()

[strengejacke]

Fixes #151

@DominiqueMakowski WDYT? pool-function when dealing with missing data. Suggestions for other names, or even one function name (where inside the function, based on the class, automatically deals with pooled means or contrasts)

estimate_means

library(modelbased)
data("nhanes2", package = "mice")

# regular
m <- lm(bmi ~ age + hyp + chl, data = nhanes2)
estimate_means(m, "age")
#> Estimated Marginal Means
#> 
#> age   |  Mean |   SE |         95% CI |  t(8)
#> ---------------------------------------------
#> 20-39 | 31.44 | 1.85 | [27.17, 35.71] | 16.98
#> 40-59 | 24.82 | 1.50 | [21.36, 28.28] | 16.55
#> 60-99 | 20.26 | 2.62 | [14.21, 26.31] |  7.72
#> 
#> Variable predicted: bmi
#> Predictors modulated: age
#> Predictors averaged: hyp, chl (1.9e+02)

# imputed and pooled
imp <- mice::mice(nhanes2, printFlag = FALSE)
predictions <- lapply(1:5, function(i) {
  m <- lm(bmi ~ age + hyp + chl, data = mice::complete(imp, action = i))
  estimate_means(m, "age")
})
pool_means(predictions)
#> Estimated Marginal Means
#> 
#> age   |  Mean |   SE |         95% CI | t(20)
#> ---------------------------------------------
#> 20-39 | 30.69 | 1.30 | [14.12, 47.26] | 23.54
#> 40-59 | 24.74 | 1.42 | [ 6.65, 42.82] | 17.31
#> 60-99 | 23.35 | 1.68 | [ 1.97, 44.74] | 14.98
#> 
#> Variable predicted: bmi
#> Predictors modulated: age
#> Predictors averaged: hyp, chl (1.9e+02)

estimate_contrasts

library(modelbased)
data("nhanes2", package = "mice")

# regular
m <- lm(bmi ~ age + hyp + chl, data = nhanes2)
estimate_contrasts(m, "age")
#> Marginal Contrasts Analysis
#> 
#> Level1 | Level2 | Difference |   SE |          95% CI |  t(8) |     p
#> ---------------------------------------------------------------------
#> 40-59  | 20-39  |      -6.62 | 2.30 | [-11.93, -1.31] | -2.88 | 0.021
#> 60-99  | 20-39  |     -11.18 | 3.36 | [-18.93, -3.43] | -3.33 | 0.010
#> 60-99  | 40-59  |      -4.56 | 2.89 | [-11.22,  2.10] | -1.58 | 0.153
#> 
#> Variable predicted: bmi
#> Predictors contrasted: age
#> Predictors averaged: hyp, chl (1.9e+02)
#> p-values are uncorrected.

# imputed and pooled
imp <- mice::mice(nhanes2, printFlag = FALSE)
predictions <- lapply(1:5, function(i) {
  m <- lm(bmi ~ age + hyp + chl, data = mice::complete(imp, action = i))
  estimate_contrasts(m, "age")
})
pool_contrasts(predictions)
#> Marginal Contrasts Analysis
#> 
#> Level1 | Level2 | Difference |   SE |          95% CI | t(20) |      p
#> ----------------------------------------------------------------------
#> 40-59  | 20-39  |      -5.22 | 2.02 | [ -9.43, -1.01] | -2.00 |  0.059
#> 60-99  | 20-39  |      -7.49 | 2.49 | [-12.67, -2.30] | -4.33 | < .001
#> 60-99  | 40-59  |      -2.27 | 3.00 | [ -8.51,  3.98] | -2.35 |  0.029
#> 
#> Variable predicted: bmi
#> Predictors contrasted: age
#> Predictors averaged: hyp, chl (2e+02)
#> p-values are uncorrected.

^{Created on 2025-02-05 with reprex v2.1.1}

[strengejacke]

Now also works for the prediction functions:

library(modelbased)
data("nhanes2", package = "mice")

# regular
m <- lm(bmi ~ age + hyp + chl, data = nhanes2)
estimate_expectation(m, by = "age")
#> Model-based Predictions
#> 
#> age   | Predicted |   SE |         95% CI
#> -----------------------------------------
#> 20-39 |     30.26 | 1.47 | [26.87, 33.65]
#> 40-59 |     23.64 | 1.75 | [19.59, 27.69]
#> 60-99 |     19.08 | 2.80 | [12.63, 25.53]
#> 
#> Variable predicted: bmi
#> Predictors modulated: age
#> Predictors controlled: hyp (no), chl (1.9e+02)

# imputed and pooled
imp <- mice::mice(nhanes2, printFlag = FALSE)
predictions <- lapply(1:5, function(i) {
  m <- lm(bmi ~ age + hyp + chl, data = mice::complete(imp, action = i))
  estimate_expectation(m, by = "age")
})
pool_means(predictions)
#> Model-based Predictions
#> 
#> age   | Predicted |   SE |         95% CI
#> -----------------------------------------
#> 20-39 |     29.01 | 1.65 | [25.78, 32.24]
#> 40-59 |     23.34 | 1.54 | [20.31, 26.36]
#> 60-99 |     23.13 | 2.11 | [18.99, 27.26]
#> 
#> Variable predicted: bmi
#> Predictors modulated: age
#> Predictors controlled: hyp (no), chl (1.9e+02)

@DominiqueMakowski WDYT? I think this is quite useful, since imputing missing data is increasingly common, as well as using marginal means/adjusted predictions, and now being able to combine both is really powerful.

[strengejacke]

I'm not sure pool_contrasts(), and especially pool_means(), are good names. Maybe a better name for pool_means()?

[strengejacke]

Renamed pool_means() into pool_predictions()