The problem
Choosing the analysis after seeing the results. Trying several models and reporting the one that worked inflates false positives, even when nobody intends it. A plan written in advance is the simplest protection.
Data rarely arrives ready to analyze. Variables are coded three different ways, records are duplicated, and the missing values are not missing at random. The cleaning is tedious, undocumented, and often where the biggest decisions in the whole project are quietly made.
Then comes the choice of method. Software makes it easy to run a model and hard to know whether it was the right one. Assumptions go unchecked, clustered data is treated as independent, and a p-value becomes the whole story.
Finally, someone has to explain the result. A table of coefficients is not an answer for a program director, a funder or a journal reader. Without a clear interpretation, careful analysis still fails to inform anything.
Anscombe's quartet (Anscombe, 1973). All four data sets have the same means, the same correlation (0.82) and the same fitted line, shown in red. Only a plot reveals the curve, the outlier and the single point driving the fit.
Who this is for
- Researchers with data in hand who want a second statistician to plan, run or check the analysis.
- Labs and groups whose analysis lives in fragile scripts that one person understands.
- Nonprofits and organizations with program or survey data they need to understand and report.
- Authors responding to statistical comments from reviewers.
What we do
Reproducible analysis means someone else can take your raw data and your code and get the same numbers, figures and tables. It usually needs version control, scripted cleaning rather than manual edits, and a record of software versions.
- Write an analysis plan. Before touching the outcome data where possible: the questions, the variables, the models, how missing data will be handled, and what counts as a sensitivity analysis.
- Clean and prepare data. Scripted, documented cleaning that turns raw exports into an analysis dataset, with a log of every decision that changes a value or drops a record.
- Choose appropriate methods. Methods that match the design and the data, from descriptive statistics and regression to mixed models, survival analysis and qualitative coding support, with assumptions checked and reported.
- Build reproducible workflows. Analysis scripted in a standard statistical language, organized so it can be rerun from raw data to final tables, with version control and a short guide for your team.
- Make figures and tables. Clear, honest graphics and tables formatted for your journal, report or slide deck, showing uncertainty rather than hiding it.
- Interpret in plain language. What the results mean, what they do not mean, and how confident anyone should be, written for the audience that will act on them.
- Review an existing analysis. A careful second look at code and output before submission, or in response to reviewers.
How it works
A confidence interval tells you far more than a p-value: the size of the effect and how precisely it was estimated. Many journals now expect intervals; check the current author guidelines for yours.
Discovery conversation
We discuss the questions, how the data was collected, its size and format, and what the results need to support.
Data handling agreement
Before any data moves, we agree how it will be shared, stored and deleted, and whether it needs to be de-identified first.
Written scope and plan
You receive a proposal and, where appropriate, a draft analysis plan for your review before analysis starts.
Analysis with check-ins
We clean, analyze and share interim results at agreed points, so surprises are discussed early rather than discovered in the final report.
Delivery and handover
Final outputs, code and documentation, plus a walkthrough so your team can rerun and extend the work.
What you receive
- A written analysis plan, agreed before the main analysis.
- Documented cleaning scripts and a data dictionary for the analysis dataset.
- Analysis code organized to run from raw data to final outputs.
- Publication-ready figures and tables in the formats you need.
- A results summary in plain language, with the main limitations stated.
- Suggested text for the statistical methods section of a paper or report.
How we can work together
- Project
- A defined analysis, from plan to final outputs, for a paper, report or evaluation.
- Retainer
- A standing statistical colleague for a lab or organization that produces analyses regularly.
- Hourly review
- A focused check of an analysis plan, code or results, or help answering statistical comments from reviewers.
Timelines
Timelines depend on scope, and above all on the state of the data. Clean, well-documented data moves quickly; data assembled from several sources takes longer to prepare than to analyze. Typical ranges are agreed in the proposal, after we have seen a sample of the data or its documentation.
Questions
Usually widely used open tools such as R or Python, or the environment your team already works in where that is practical. We agree this at the start, because code your team cannot run after we leave is not much of a deliverable.
Carefully, and preferably not at all. Where possible we work with de-identified data. Where identifiable data is necessary, we agree in writing how it will be transferred, stored and deleted, consistent with your approvals and agreements. Never send sensitive data through the contact form.
Yes. We will be explicit about what the data can and cannot support, and frame the results as exploratory where that is the honest description.
That depends on the contribution and on the authorship criteria of the journal, which you should check. Many analysis projects warrant an acknowledgment rather than authorship. We will discuss it at the start, not at submission.
Yes. We assess each comment, run any additional analyses that are warranted, and draft responses that explain clearly what changed and why.