TABLE 1.
Regression coefficients (and 95% SIs) for the effect of sugars on fasting plasma glucose estimated using 4 different causal estimand scenarios and adjustment approaches1
| Model number | Model | Model name | Estimand | True estimate | Model estimate (95% SI), no confounding | Model estimate (95% SI), with confounding |
|---|---|---|---|---|---|---|
| 0 |
|
The unadjusted model | Total causal effect | 5.002 | 5.002 (3.80, 6.21) | 8.222 (7.09, 9.35) |
| 1 |
|
The energy partition model | Total causal effect | 5.002 | 5.002 (3.95, 6.06) | 5.492 (4.53, 6.45) |
| 2 |
|
The standard model | Average relative causal effect | 2.002 | 1.942 (0.83, 3.04) | 2.282 (1.22, 3.34) |
| 3a |
|
The nutrient density model | Obscure3 | 0.404 | 0.144 (0.11, 0.40) | 0.474,5 (0.16, 0.75) |
| 3b |
|
The multivariable nutrient density model | Obscure3 | 0.404 | 0.354 (0.14, 0.56) | 0.394,5 (0.18, 0.58) |
| 4 |
|
The residual model | Average relative causal effect | 2.002 | 1.942 (0.83, 3.04) | 2.282 (1.22, 3.34) |
| 5 |
|
The all-components model | Total causal effect (
|
5.002 | 5.002 (3.95, 6.05) | 5.002 (3.95, 6.05) |
Average relative causal effect ( ) |
2.002 | 2.002 (0.87, 3.13) | 2.002 (0.88, 3.11) |
1
PRO, proteinSI, simulation interval.
2
Values are expressed as mg/dL/100 kcal.
3
The nutrient density model evaluates an obscure estimand, but it is conceptually closest to the average relative causal effect rescaled as a proportion of total energy.
4
Values are expressed as mg/dL/1%.
5
The confounded estimates are closer to the true estimates than expected by chance because of the direction of confounding.