Session Objectives
In this workshop, we will discuss how rigor in applied epidemiological methods, combined with qualitative stakeholder involvement, can generate useful and realistic policy insights.
The Policymaker’s Dilemma
Imagine you are a policymaker with a $1M public health budget.
What kind of evidence would you demand before investing?
A. Associative Evidence “Sugar and dental decay go hand in hand.”
B. Causal Evidence “If we reduce sugar consumption by 20g per day, we avert 15% of all new dental disease.”
You need causal answers, not just correlations.
The Problem: Association ≠ Causation
🔍 Association (Comparing Subgroups)
- The Approach: We partition the observed data.
- The Contrast: We compare the dental outcomes of the “high sugar” fraction directly against the “low sugar” fraction.
- The Focus: Contrasting different subsets of people.
🎯 Causation (Contrasting the Whole Population)
- The Approach: We construct counterfactual scenarios.
- The Contrast: We compare the entire population’s actual outcomes against the entire population’s expected outcomes under a new policy.
- The Focus: Contrasting the exact same population under different conditions.
What Can Regression Tell Us?
| Regression CAN tell us: | Regression CANNOT tell us: |
|---|---|
| “People who eat more sugar tend to have more cavities” | “If we reduce sugar, fewer people will get cavities” |
| Variables are related | What would happen under a POLICY |
| Prediction | Causation |
Key message: Regression models isolate robust predictive patterns. However, unmeasured structural confounding ensures these statistical associations frequently fail to translate into causal policy impacts.
What is Exchangeability?
Exchangeability To make valid causal claims, the groups we compare must be fundamentally similar across all relevant dimensions before the intervention. We require a true “apples to apples” comparison.
Confounding When structural factors such as baseline income or education differ systematically between groups, exchangeability breaks down. The comparison becomes “apples to oranges,” (direct contrasts are biased).
Structure of Confounding
Lower-income people consume more sugar AND have more tooth decay — for reasons beyond sugar alone
Static vs Modified Policy
Static (bad for continuous): > “Set everyone to exactly 40g sugar” — unrealistic, no data for many people
Modified Treatment Policy (good): > “For people eating more than 50g, reduce by 20g. For others, keep the same.”
Visual:
“This is like a REALISTIC policy — we only ask heavy eaters to cut back, not everyone.”
\[d_1(a_t, h_t) = \begin{cases} a_t - 20 & \text{if } a_t > 50 \\ a_t & \text{otherwise} \end{cases}\]
What Does the Intervention Look Like?
- Mirrors a realistic public health intervention (e.g., sugar tax, labelling policy)
- Stays within the support of the observed data ✅
- Respects individual baseline levels ✅
- Avoids positivity violations ✅
Defining the Shift in R
# Policy: Reduce sugar by 20 g/day for those consuming > 50 g/day
policy_reduce_20 <- function(data, trt) {
a <- data[[trt]]
ifelse(a > 50, a - 20, a)
}
# Preview what it does to a few observations
dental_data |>
select(sugar_consumption) |>
mutate(
sugar_policy = policy_reduce_20(dental_data, "sugar_consumption")
) |>
slice_sample(n = 7)# A tibble: 7 × 2
sugar_consumption sugar_policy
<dbl> <dbl>
1 75.6 55.6
2 56.6 36.6
3 25.6 25.6
4 61.0 41.0
5 44.5 44.5
6 45.9 45.9
7 68.0 48.0