From Analysis to Decision-Making

Published

Jun 2026

  • ID: DS-L16
  • Type: Premium
  • Audience: Intermediate to Advanced
  • Theme: Translating analytical results into real-world decisions

So far, we have built a complete analytical pipeline:

data → model → evaluation → interpretation → communication

Now we take the final step:

How do we move from analysis to action?

This is where applied data science becomes more than modeling.

A model may produce predictions.

A report may communicate findings clearly.

But a real-world decision still requires context, judgment, trade-offs, and responsibility.

How to Run This Lesson

Run the supporting script from the project root:

python scripts/python/16a_build_decision_summary.py

This creates the expected outputs in the reports/ directory:

reports/diabetes-decision-summary.md
reports/diabetes-decision-framing-table.csv

Then render the Quarto site:

quarto render

You can also read this chapter as a conceptual bridge from analytical evidence to decision support.

The script-based workflow is preferred because it creates reusable decision-support artifacts that can be reviewed, committed, and reused in the end-to-end case study.

The Final Gap

Even well-executed analyses can fail at the last step.

This usually happens when analytical results are not connected clearly to decisions.

Common problems include:

  • results are presented without a decision context
  • uncertainty is ignored or minimized
  • stakeholders do not know what action is reasonable
  • the model is treated as if it makes the decision by itself
  • limitations are not translated into practical risk

This lesson focuses on bridging that final gap.

The goal is not to turn model outputs into automatic instructions.

The goal is to turn model outputs into useful decision support.

What Decisions Require

A decision is not based only on:

  • a model
  • a metric
  • a plot
  • a coefficient
  • a prediction

A decision requires:

  • a goal
  • a decision-maker
  • a context
  • available options
  • acceptable risk
  • consequences of being wrong
  • domain knowledge
  • ethical and operational constraints

Models inform decisions.

They do not make them.

This distinction is central to responsible applied data science.

From Prediction to Action

Suppose a model predicts higher disease progression for some individuals.

That prediction does not automatically imply:

  • immediate intervention
  • treatment change
  • policy change
  • individual-level clinical action

Instead, it raises a decision question:

What should be done given this information, uncertainty, and context?

The model contributes evidence.

The decision process determines how that evidence should be used.

Load the Communication Summary

In the previous lesson, we created a communication summary.

This chapter builds on that output by translating findings into a decision-support format.

from pathlib import Path

summary_path = Path("reports/diabetes-communication-summary.md")

if summary_path.exists():
    communication_summary = summary_path.read_text()
    print(communication_summary[:800])
else:
    print("Communication summary not found. Run Lesson 15 first.")

This reinforces an important applied workflow principle:

decision support should build from documented analytical evidence, not from memory.

Decision Framing

A useful decision-support structure asks five questions.

Question Purpose
What is the goal? Defines the purpose of action
What does the model provide? Clarifies the evidence available
What are the possible actions? Identifies decision options
What are the risks? Makes uncertainty practical
What should not be concluded? Prevents overreach

This structure helps prevent a common mistake:

jumping from model output directly to action.

Build a Decision-Framing Table

The table below translates the diabetes modeling workflow into decision-support language.

import pandas as pd

framing_df = pd.DataFrame({
    "decision_component": [
        "Goal",
        "Model evidence",
        "Possible use",
        "Main uncertainty",
        "Reasonable action",
        "Unsupported action"
    ],
    "decision_framing": [
        "Identify individuals who may have higher predicted disease progression.",
        "The model shows moderate predictive ability and uses features such as BMI in prediction.",
        "Use predictions as one input for prioritizing closer monitoring or further review.",
        "Predictions have error and are based on associations in this dataset.",
        "Flag higher-risk cases for review while combining model output with domain expertise.",
        "Do not use the model alone to prescribe treatment or claim causation."
    ]
})

framing_df

This kind of table is useful because it makes the decision logic explicit.

It separates what the model says from what the decision process may do with it.

Save the Decision-Framing Table

Decision-support outputs should be saved as project artifacts.

from pathlib import Path

Path("reports").mkdir(exist_ok=True)

framing_df.to_csv(
    "reports/diabetes-decision-framing-table.csv",
    index=False
)

framing_df

This creates:

reports/diabetes-decision-framing-table.csv

The output can be reviewed by analysts, collaborators, or decision-makers.

Example: Translating Results Into Decision Support

Model output

Higher BMI is associated with higher predicted disease progression in this model.

Unsupported action

“All high-BMI patients must receive treatment.”

This is too strong because the model does not establish causation and is not sufficient for clinical decision-making on its own.

Better decision framing

“Patients with higher predicted progression may be prioritized for closer monitoring or additional review, depending on clinical context and available resources.”

This statement is better because it:

  • uses the model as decision support
  • avoids causal overreach
  • leaves room for domain expertise
  • acknowledges that action depends on context

The Role of Uncertainty

No model is perfect.

Decisions must consider:

  • prediction error
  • variability across data splits
  • limitations of the dataset
  • missing variables
  • measurement error
  • consequences of acting incorrectly

A useful decision mindset is:

decisions should be robust to uncertainty, not dependent on perfect predictions.

A model can reduce uncertainty.

It cannot eliminate it.

CDI Insight

Models reduce uncertainty.

They do not remove judgment.

Responsible decision-making requires combining model evidence with context, domain expertise, and careful consideration of consequences.

Decision Thresholds

Many decisions require thresholds.

For example:

flag a case if predicted progression exceeds a certain value.

However, thresholds are not determined by the model alone.

They depend on:

  • cost of false positives
  • cost of false negatives
  • available resources
  • urgency of the problem
  • ethical and operational constraints
  • stakeholder priorities

A model can help estimate risk.

A decision process must decide how much risk is acceptable.

Trade-Offs

Every decision involves trade-offs.

Examples include:

  • sensitivity versus specificity
  • early intervention versus unnecessary action
  • simplicity versus complexity
  • automation versus human review
  • speed versus certainty

The “best” decision depends on priorities.

For example, in some contexts it may be better to flag more cases and accept more false positives.

In other contexts, false positives may be costly or harmful.

The model does not resolve that trade-off by itself.

Communicating for Decisions

When presenting results for decision-making, include:

  • what the model suggests
  • how reliable the evidence is
  • what uncertainty remains
  • what options are available
  • what actions are not supported
  • where human judgment is needed

Avoid:

  • presenting predictions as instructions
  • using causal language without causal evidence
  • hiding uncertainty
  • recommending action without context
  • treating a model score as a final answer

Clear decision communication makes the boundary between evidence and action visible.

Build a Decision Summary

A short decision summary can help stakeholders understand how to use the analysis.

decision_summary = """# Diabetes Model Decision Summary

## Objective
Identify individuals who may have higher predicted disease progression using clinical features.

## Analytical Finding
The model shows moderate predictive ability and reasonably stable performance across evaluation splits. Features such as BMI are used by the model when estimating disease progression.

## Interpretation
BMI and related features are useful predictors in this dataset, but the model describes associations and prediction patterns rather than causal mechanisms.

## Decision Context
Predictions may support prioritization for closer monitoring, further review, or additional data collection. The model should not be used as the sole basis for treatment decisions.

## Key Limitation
The model has prediction error and was trained on a specific dataset. Decisions should account for uncertainty, clinical context, and consequences of false positives or false negatives.

## Recommended Use
Use the model as a decision-support tool, not as an automatic decision-maker.
"""

Path("reports/diabetes-decision-summary.md").write_text(decision_summary)

print(decision_summary)

This creates:

reports/diabetes-decision-summary.md

This output will also be useful in the end-to-end case study.

From Insight to Action

A responsible decision process follows a sequence:

  1. Understand the analytical objective
  2. Build and evaluate the model
  3. Interpret the results carefully
  4. Communicate the evidence clearly
  5. Identify possible decisions
  6. Evaluate risks and trade-offs
  7. Decide within context
  8. Monitor outcomes after action

The final step is especially important.

A decision is not the end of the analytical system.

In real-world systems, decisions produce outcomes, and those outcomes should inform future learning.

What This Means in Practice

The model can help answer:

  • Who may have higher predicted progression?
  • Which features are useful for prediction?
  • How stable is the model performance?
  • How much uncertainty remains?

The model cannot answer by itself:

  • What treatment should be given?
  • What threshold is ethically acceptable?
  • What action is best under resource constraints?
  • Whether a feature causes the outcome?

Those questions require decision context.

Key Insight

Analysis supports decisions.

It does not replace judgment.

The purpose of applied data science is not to automate responsibility away.

It is to make evidence clearer, uncertainty more visible, and decisions better informed.

Summary

In this lesson, we moved from communicated results to decision support.

We learned that:

  • models inform decisions, but do not make them
  • decisions require context, risk assessment, and domain knowledge
  • uncertainty must be acknowledged
  • thresholds depend on trade-offs, not metrics alone
  • clear communication enables responsible action
  • decision-support outputs should be saved as reusable artifacts

The central lesson is:

a model is useful when it improves the quality of a decision without hiding uncertainty or replacing judgment.

What Comes Next

In the next lesson, we bring the full workflow together.

We will move from isolated lessons into an end-to-end applied data science case study:

data → features → model → evaluation → interpretation → communication → decision support

→ End-to-end project: from data to decision