What is happening?
Summaries, rates, trends, and visualizations establish the pattern before a model tries to explain it.
Output: an accurate picture of the observed data.Methods education
Understand the method. Use the evidence wisely.
MP2 teaches leaders and practitioners how to ask what a model can answer, how it was tested, and where context and human judgment must enter.
01 Choose the question
The method follows the decision.
Start with the problem leaders need to understand. Then choose the simplest design that can answer it responsibly.
What is more likely next?
Regression and machine learning combine prior signals to estimate a probability, count, or future value.
Output: decision support, not causal proof.Did an intervention change outcomes?
Causal designs compare observed change with a credible estimate of what would have happened otherwise.
Output: an effect estimate with assumptions and uncertainty.02 Learn the toolkit
Six methods in the evidence lineage.
Each tool has a job. None eliminates the need to inspect data quality, assumptions, errors, and the consequences of acting.
01 Association
Linear regression
Estimates how a numeric outcome changes as one or more factors change.- Use it for
- Continuous outcomes such as scores, rates, or readiness measures.
- Read carefully
- A coefficient describes an adjusted association. It does not establish why the relationship exists.
02 Probability
Logistic regression
Estimates the probability of a yes-or-no outcome from a set of observed factors.- In the URFT
- The public base model uses logistic regression to estimate unit-level incident probability from recent patterns and selected unit factors.
- Read carefully
- Probability is not certainty, and a unit signal is not an assessment of any person.
03 Counts
Poisson regression
Models how many events may occur during a defined period.- Use it for
- Incident counts, when the outcome is more informative than a simple yes-or-no flag.
- Read carefully
- Count data can be clustered or unusually variable; the model's assumptions must be checked.
04 Machine learning
XGBoost & tree ensembles
Learn nonlinear patterns and interactions that simpler equations may miss.- Use it for
- Prediction when relationships may involve thresholds, interactions, or complex combinations.
- Read carefully
- Test on data the model did not train on. High sensitivity can still produce too many false positives.
05 Causal evaluation
Difference-in-differences
Compares change over time in an exposed group with change in a comparison group.- Use it for
- Estimating an intervention's effect when random assignment is not available.
- Read carefully
- The estimate depends on a credible comparison and assumptions about how the groups would otherwise have changed.
06 Counterfactuals
Panel-data estimators
Use patterns across units and time to estimate the outcome that was not observed.- Use it for
- Testing whether an apparent effect persists across alternative comparison strategies.
- Read carefully
- A sophisticated estimator cannot repair unclear treatment timing, missing context, or weak source data.
03 Read a model
Six questions before action.
A useful model is more than an accuracy score. It is a transparent chain from a decision need to an accountable human response.
- 01What is the unit of analysis?
A person, a team, a unit, or a week are not interchangeable.
- 02How is the outcome defined?
Recorded incidents are organizational records, not a complete measure of underlying harm.
- 03Was performance tested honestly?
Separate training and test data, compare with a useful baseline, and inspect calibration.
- 04Which errors matter?
False positives and false negatives create different burdens, risks, and obligations.
- 05What assumptions remain?
Confounding, missingness, timing, and local context can change the meaning of a result.
- 06How will people use the signal?
Define the human review, protective action, governance, and learning loop before deployment.
The essential boundary
A model can surface a signal. It cannot diagnose a person.
MP2 uses organizational evidence to focus inquiry, strengthen care, and improve systems. It does not support individual surveillance, automated personnel action, or clinical decisions.
04 Continue learning
See the methods in context.
URFT base-model working paperLogistic regression and unit-level prediction ↗Protecting Soldiers, Preventing Harmful Behaviors, and Boosting Combat Readiness—with DataLogistic, Poisson, and XGBoost models in the public synthesis ↗Taking a Data-Centric Approach to Unit ReadinessA practical research-and-analytics workflow ↗