A methodologically principled and scientifically responsible process translating data into meaning is the craft of GoodScience.

Do Good Science.

FRAMEWORK

The GoodScience framework is designed to generate meaningful and reliable evidence from data and support high-quality decisions.

Nature
Nature
The complex nexus of causes that produce the phenomena of our world that are available for empirical study. The underlying causal structure of nature is often abstruse or inscrutable.
  • Underlying causal reality
  • Beginning of the data generating process
Population
Population
All of the objects (existing, extant and/or possible) in the category of interest for study. The population is the realization of causal process in nature. The Population is the expression of the 'long run' probabilistic tendencies in nature's causes. The population is also the primary object of study and inference.
  • Importance of study / experimental design
  • Risk of selection bias
  • Confounding by indication
Sample
Sample
The subset of the population available for study and observed.
  • Omitted variables
  • Missing data
  • Measurement issues
  • Information bias
Data
Data
The actual observations made and recorded on the sample. Not all observations/variables of all possible variables from the sample are collected. Measurements are made imperfectly and recorded with errors. The particular instance of the data (out of many possible instances) are the source of the statistical likelihood on which the analysis is predicated.
Uncertainties
  • Model specification
  • Model selection
  • Assumptions re: distributions
Analysis
Analysis
The mathematical procedures that account for both the structure and randomness of the data. Typically a model is used or is at least implicit. All analyses require assumptions (both strong and weak).
Analytic bias
  • Model selection
  • Model misspecification
  • Over-fitting
  • Residual confounding
  • Arbitrary categorization
  • Collider bias
Inference
& Belief
Inference & Belief
The conclusions drawn from the analysis of the data (and in combination with any external information), including whether any associations observed are causal in nature and likely reproducible effects in independent data. Belief depends on the strength of the findings and the research process, coherence with existing knowledge, and numerous cognitive and psychological factors including biases, intentions and motivations.
  • Association vs. Causation
  • Cognition / psychology
  • Intentions & motivations
Decisions
& Actions
Decisions & Actions
The consequences, if any, of the research activities. The impact of the research will depend in part on the strength of the belief resulting from the inference, and the relevance for problems faced by others. Consequences include clinical behavior and medical decision making; and scientific behavior including confirmatory reproduction of research, and motivation of additional research.
Hover over nodes and arrows for more detailed information.

ETHOS

Any science needs to faithfully connect its observations and measures with a sciences' systematic implications.

Analysis methods are a filter and a lens in the process leading from observations to knowledge and value.

In the communication between the data generating process and knowledge, analysis methods can be low- or high-fidelity. The details of how this connection is performed matter.

All but the most trivial of scientific questions will require modeling. The numerous decisions and assumptions made in the process of regression modeling and estimation determine the accuracy of any signal or pattern that is the target of scientific interest.

Modeling is not simply the plumbing between data and conclusions. Most observational data in health science are generated by largely nonrandom and poorly understood mechanisms of subject selection, exposure assignment, measurement error and missing data; and are analyzed to evaluate effects that are latent under these mechanisms.

Modeling well is a craft: much more than just facility with a toolkit of regression techniques. Rigorous understanding of the specific purposes of the research, the processes that generated the data, the specific analytic strategy and modeling tools employed, and a thorough understanding of the raw material (data and assumptions) that are fed to the model, are all part of a successful program of scientific modeling. 

Modeling well begins with asking good questions.

A principled, well integrated and coherent, consistent, and reliable process translating data into meaning is the craft of good science.

PERSPECTIVES

PUBLICATIONS

Drew Levy is the founder of GoodScience, where he focuses on statistical modeling, prediction, causal inference, and evidence quality in healthcare. His work examines how evidence is generated, how signal is distinguished from noise, and how analytic methods support reliable decision-making under uncertainty. He practices the craft of analysis with a principled framework and modern applied methods.

ABOUT

CONTACT

drew@dogoodscience.com